thesis - swinburne · mohammad ben salamah thesis submitted in fulfillment of the requirements for...
TRANSCRIPT
i
Thesis Design and Support of Systems for Operation and
Maintenance of a Cooling Petrochemical Pumping Station
By
Mohammad Ben Salamah
Thesis submitted in fulfillment of the requirements for the degree of Doctor of Philosophy
Swinburne University of Technology, Melbourne, Australia
October 2010
ii
SWINBURNE UNIVERSITY OF TECHNOLOGY CANDIDATE DECLARATION I certify that the thesis entitled: Design and Support of Systems for Operation and Maintenance of a Cooling Petrochemical Pumping Station for the degree of Doctor of Philosophy contains no material that has been accepted for the award of any other degree or diploma. To the best of my knowledge, this thesis contains no material previously published or written by another author, except where due reference is made in the text of the thesis. All work presented is primarily the result of my own research. Full Name: Mohammad J. Ben Salamah Signed:.......................................
iii
To my family, friends and colleagues.
iv
Abstract Refineries and petrochemical plants are a very important part of our modern world. Their products include petrol (car fuel), diesel (truck fuel), fuel oil for electrical power plants, benzene, kerosene, chemical fertilizers and the raw material for plastic. These products, and many others, are being used extensively in transportation, energy production, agriculture, manufacturing and many industries including the pharmaceutical industry. Refineries and petrochemical plants generate a lot of heat in the process of making their products. This heat is removed by heat exchangers. For these heat exchangers to work, they need large amounts of water. The water for the heat exchangers is provided by a pumping station. Pumping stations, for the petrochemical industry, pump large amounts of water to the consuming plants. These cooling pumping stations are made of a group of pumps connected in parallel. These pumps deliver cooling water to the consuming plants through a network of pipes. An interruption of cooling water would have severe consequences on the petrochemical industry. Consequently, the reliable delivery of the cooling water can not be over emphasized. To achieve this end, a reliability model for the cooling water delivery must be made. In this thesis, a reliability model for cooling water delivery from a cooling-pumping station to a group of petrochemical plants was made. The model took into account the amount of flow that a plant needs to remain operational. A feature of this model is that it is affected by the operational conditions of the lines and valves in the system. As a result, instead of having one reliability model for a plant, each plant would have several reliability models depending on the operational conditions of the lines and valves in the system. The model developed, also, gives a way to look at the reliability of a pumping station having several independent consumers. Reliability modeling is important. Practically speaking, however, it is only a first step. To ensure the reliability of the pumping station, pumps should receive timely maintenance. This timely maintenance is obstructed by the consumer demand of cooling water i.e. there is a conflict between the production function and the maintenance needs of pumps. In this thesis, an attempt was made to minimize this conflict. To minimize the conflict between operation and maintenance, scheduling was used. With scheduling, the operation of pumps around the year would be planned. It was noticed that the consumption of cooling water depended on two things: the weather and a plant’s production level or capacity utilization. Regression analysis was used to find the relationship between a plant’s water consumption and both the weather and production level. The elements of the weather that affected the cooling water consumption of a plant were the ambient air temperature, humidity and seawater temperature. Scheduling is an activity done for future planning. In the case of the pumping station it was for the future planning of pump operation around the year. The relationships developed with regression analysis required the previously mentioned weather factors
v
and a plant’s production level or capacity utilization. While the production levels or capacity utilization of a plant around the year can be obtained from the plant’s owners, the weather factors can not be known in advance. There were two ways to forecast the weather factors: either from a weather service or to develop methods for forecasting these factors. This thesis went with the second option and developed relationships for the local forecasting of the weather factors involved. In this thesis, the attempt to minimize the conflict between operation and maintenance was done by scheduling. The scheduling was done first by using regression analysis to find the relationship between water consumption and a plant’s production and the weather. Secondly, solving this scheduling problem required solving a forecasting problem for the weather factors involved. The results obtained were satisfactory. Increasing the reliability of a pumping station through reliability analysis and minimizing the conflict between operation and maintenance through the previously mentioned methods would help a pumping station in lowering its maintenance expenditure and in improving its service to its consumers. The fact remains, nevertheless, that a pumping station must also generate enough revenue for its owners. Revenue in a pumping station is achieved by selling cooling water to the consumers. The amount of cooling water sold is measured by a flow meter. Flow meters, just like all machines, are susceptible to failure. This failure directly affects the amount of water measured and, subsequently, the revenue. A dangerous type of flow-meter failure is the one that incrementally, but systematically and continuously, alters the readings of a flow meter. This failure is known as flow-meter drift. In this thesis two methods for detecting flow-meter drift were developed: One that used statistical process control (SPC) and the other used artificial neural networks. Both approaches were capable of working with the minimal existing data and were financially inexpensive in their development and application. The first approach, flow-meter-drift detection by using statistical process control, had to transform the widely oscillating data of water demand to a linear form. This was done by creating a virtual mean. The linear, transformed, data were then processed by the SPC method. The method was tested and found satisfactory. The second approach, flow-meter-drift detection by using artificial neural networks, used the same virtual mean developed in the first approach. The linear, transformed, data was further normalized to make the findings universal to all volumes of flow. The normalized output data were then processed by a three layer neural network. The input layer was made of seventeen numerical inputs and seven symbolic inputs. The output layer would show if the flow was normal or drifting upwards or drifting downwards.
vi
List of Publications AlSalamah, M., Shayan, E. and Savsar, M., “Reliability analysis of a cooling seawater
pumping station”, The International Journal of Quality and Reliability Management,
Vol.23, No. 6, 2006, pp 670-695, (27 pages).
Ben Salamah, M., Shayan, E. and Savsar, M., “Minimizing the Conflict between
Operation and Maintenance- a Case Study “, International Journal of Data Analysis and
Information Systems (IJDAIS), Vol. 2, No. 1, Jan-June 2010, pp 19-38.
Ben Salamah, M., Kapoor, A., Savsar, M., Ektesabi, M, Abdkhodaee, A., Shayan, E.,
“The detection of flow meter drift by using statistical process control”, International
Journal of Sustainable Development & Planning, Vol. 6, No. 1, Feb. 2011, pp 91-103.
Ben Salamah, M., Palaneeswaran, E. , Savsar, M., Ektesabi, M, “Detection of flow meter
drift by using artificial neural networks”, International Journal of Sustainable
Development & Planning, Vol. 6, No. 4, December 2011.
The evidence for publishing the above mentioned papers is shown in Appendix II.
vii
Contents Chapter 1
Introduction……………………………………………………………………………....1
1.1 Petroleum and Petroleum Refining…………………………………………1
1.2 The Cooling-Water System of a Refinery…………………………………..2
1.3 About Cooling Petrochemical Pumping Stations and their
Importance…………………………………..……………………...…………….3
1.4 The Cooling Pumping Station Understudy…………………………….…..4
1.5 The Structure of a Cooling Pumping Station………………………………7
1.6 Overview of the Research Problem……………………………….……….39
1.7 The Aims and Objectives of the Research………………………….……..40
1.8 The Significance of the Research and its Scope………………….……….40
Chapter 2 Reliability Analysis of a Cooling Pumping Station…………………….…42
2.1 Introduction to Cooling Pumping Station Reliability……………………42
2.2 Literature Review ………………………………………………………….43
2.3 Reliability Analysis of the System…………………………………………45
2.3.1. Introduction………………………………………………………45
2.3.2 Data Collection……………………………………………………52
2.3.3 Data Modification………………………………………………...57
2.3.3.1 Failure Rate for the Pipe Section……………………...58
2.3.3.2 Failure Rate for the Header Section…………………..58
2.3.3.3 Failure Rate for the Valve Section…………………….58
2.3.4 Reliability Calculations…………………………………………..60
2.3.5 The Reliability of Water Delivery to a Consumer
in All Cases……………………………………………………………..73
2.3.6 Considering the Reliability of All the Consumers
and the Entire Pumping Station………………………………………76
viii
Chapter 3 A Case Study of Minimizing the Conflict between Operation and
Maintenance…………………………………………………………..………81
3.1 Introduction……………………………………………………………...….81
3.2 Literature Review………………………………………………..…………83
3.3 Factors Influencing the Demand Variation……………………………….92
3.3.1 The Process Category…………………………………………….93
3.3.1.1 The Amount of Production or Capacity Utilization….93
3.3.2 The Weather Category……………………………………….…..93
3.3.2.1Seawater Temperature………………………….………94
3.3.2.2 Ambient Air Temperature…………………..…………95
3.3.2.3 Humidity………………………………………………..96
3.4 Reasons for Choosing Regression………………………………………….96
3. 5 Model Development………………………………………….…………….99
3.5.1 Data Collection and Analysis………………….………………..102
3.5.2 Regression models……………………………...………………..102
3.5.3 Exponential Smoothing……………………………….………...104
3.5.4 Relationship between Ta and Ts for the Specific Location of the
Pumping Station…………………………………………….…………104
3.5.5 Relationship between Ta, Ts and H for the Specific Location of
the Pumping Station……………………………………………..…....104
3.6 Prediction and Scheduling …………………………………………….…105
Chapter 4 A Method for Flow Meter Drift Detection………………..……………..108
4.1 Introduction………………………………………...………...……………108
4.2 Literature Review………………………………………………….……...113
4.2.1 Literature Review on the Phenomenon of Unaccounted
for Water………………………………………………………………114
4.2.2 Literature Review on Instrument (or Sensor) Drift Which is The
Root Cause of The Flow-Meter Problem………………...…………121
ix
4.2.3 Literature Review on Flow-Meter Drift Which is a Special Case
of The Above Problem and The Resulting Unaccounted-for-Fluid-
Loss Phenomenon…………………..………………………...………121
4.2.4 Literature Review on Hardware Solutions for The Problem of
Unaccounted-for- Fluid…………………..…………………………...122
4.2.5 Literature Review on Statistical Methods to Solve The Problem
of Unaccounted-for- Fluid……………..………………….…………..122
4.2.6 Literature Review on Statistical Process Control (SPC) Which is
The General Approach That Was Used………..…………….………122
4.2.7 Literature Review on CUSUM…………..…...………………...123
4.3 Factors Influencing Flow Meter Readings………………………….…..125
4.4 Reasons for Choosing Statistical Process Control (SPC), and Its
Underlying Assumptions and Limitations…………………………………...127
4.4.1. Reasons for Choosing Statistical Process Control (SPC)…….127
4.4.2. The Underlying Assumptions of SPC………………...………..127
4.4.3. Limitations of Statistical Process Control…………...………..128
4.5 The Research Method……………………………………………………..128
4.6 The Method of the CUSUM………………………………………………133
4.6.1. General…………………………………………………………..133
4.6.2 The Method of Tabular CUSUM………………………………134
4.7 Case Studies………………………………………………………………..135
4.7.1 Case Study #1……………………………………………………135
4.7.2 Case Study #2……………………………………………………137
2.7.3 Case Study #3……………………………………………………138
2.7.4 Case Study #4……………………………………………………138
2.7.5 Case Study #5……………………………………………………139
2.7.6 Case Study #6……………………………………………………139
4.8 Limitations of the Presented Method and Suggestions for
Further Study………………………………………………………………….140
x
Chapter 5 an Alternative Method for Flow Meter Drift Detection………………...142
5.1 Introduction……………………………………………………………..…142
5.2 Literature Review…………………………………………………………142
5.3. Production Processes and Their Quality Assurance…………………...144
5.3.1 The Nature of Production Processes…………………………..144
5.3.2 Introduction to the Shewhart Chart…………………………...145
5.4 Reasons for Choosing Artificial Neural Network Methods, Their
Assumptions and Limitations……………………………………………..… 146
5.4.1 Reasons for Choosing Artificial Neural Networks…………….146
5.4.2 Assumptions of ANN-Based Modeling…………………………147
5.4.3 Limitations of Artificial Neural Networks……………………..147
5.5 The Research Method………………………………………………..……147
5.5.1 The Inputs for the Artificial Neural Network…………………148
5.5.1.1 The Inputs for the Artificial Neural
Network- The Numerical Inputs…………………………..…148
5.5.1.2 The Inputs for the Artificial Neural
Network- The Symbolic Inputs……………………………….150
5.5.2 The Hidden and Output Layers………………………………..152
5.6 Results of the Simulation, Training,
Cross Validation & Testing………………………………………………...…154
5.6.1. The Simulation of Flow Data….……………………………….154
5.6.2. Training and Cross Validation of the
Artificial Neural Network……………………………………………155
5.6.2. Testing of the Artificial Neural Network……………………...156
5.7 A Possible Way of Improving the Results……………………………….157
5.8 Potential Applications……………………………………………………..158
Chapter 6 Conclusions…………………………………………...……………………159
6.1 Summary …………………………………………………………………..159
6.2 The specific contributions this study has
made to the existing body of knowledge and industry practice…………...166
xi
6.3 Conclusion…………………………………………………………………168
References……………………………………………………………...………………170
Appendix I.………………………………………………………………………...…..178
Appendix II…………………………………………………………………………….186
Evidence for publishing of
1. Reliability analysis of a cooling seawater pumping station ……….………..187
2. Data analysis technique to resolve the conflict between
operation and maintenance……………………………………………………188
3. The detection of flow meter drift by using statistical process
control………………………………………………………………………….189
4. The detection of flow meter drift by using artificial
neural networks………………..………………………….…………………..190
xii
List of Tables Table 1.1 Value of one day of produced goods by each consumer…………………...…..7
Table 2.1 Failure rates of pumps…………………………………………………………52
Table 2.2 Actual records of line failures………………………………………………....54
Table 2.3 Reciprocals of Mean Time between Failures (MTBF) for each line (failure
rates)……………………………………………………………………………………...54
Table 2.4 Failure rates for line and header sections calculated by equation (1)…………57
Table 2.5 Failure rates for line or header sections, actual
and a modified Equation (1)……………………………………………………………58
Table 2.6 Failure rates for failed system valves…………………………………………59
Table 2.7 Failure Rates for the system valves that did not fail………………………….60
Table 2.8 Description of every case for the consumers………………………………….66
Table 2.9 System reliability equations for different
consumers under different scenarios……………………………………………….…….68
Table 3.1 C12 equation coefficients………………………….……………………...…103
Table 3.2 statistical measures of equation 3.5……………..…………………..……….103
Table 3.3 C6 equation coefficients……………………………………………………..103
Table 3.4 Statistical measures of C6 equation…………………………….……………104
Table 4.1 C7 Consumption…………………………………………………………......135
xiii
Table 4.2 The statistical properties of the virtual mean and the parameters of the tabular
CUSUM for C7…………………………………………………………………………137
Table 4. 3 The statistical properties of the virtual mean and the parameters of the tabular
CUSUM for C7-Case #2………………………………………………………….……138
Table4. 4 The statistical properties of the virtual mean and the parameters of the tabular
CUSUM for C6…………………………………………………………………..…….138
Table 4.5 The statistical properties of the virtual mean and the parameters of the tabular
CUSUM for C12……………………………………………………………..…………139
Table4. 6 The statistical properties of the virtual mean and the parameters of the tabular
CUSUM for C3………………………………………………………………….……...139
Table 4.7 The statistical properties of the virtual mean and the parameters of the tabular
CUSUM for C5………………………………………………………………..………..140
Table5. 1 Training results for 1000 epochs…………….……………………..………..156
Table5. 2 The confusion matrix…………………………….…………………………..157
Table A. 1The tabular CUSUM method as applied to C7 Case #1…………………….178
Table A. 2 The CUSUM method as applied to C7 Case#2…………………………….179
Table A. 3 the recorded consumption of C6……………………………………………179
Table A. 4 the CUSUM method as applied to C6 (Case Study #3)…………………….180
Table A. 5 The recorded consumption of C12………………………………………….181
Table A. 6 the CUSUM method as applied to C12……………………………………..182
Table A. 7 the recorded consumption of consumer C3………………………………...183
Table A. 8 the CUSUM method as applied on C3……………………………………...183
Table A. 9 the recorded consumption for C5………………………………………..….184
Table A. 10 the CUSUM method as applied on C5…………………………………….184
xiv
List of Figures Figure 1.1 An aerial view of a part of the Shauiba industrial complex………...…………5
Figure 1.2 An aerial view of the Shauiba industrial complex near the pumping station.…6
Figure 1.3 An aerial view of the cooling pumping station understudy………………..…..8
Figure 1.4 an aerial view of the cooling pumping station understudy with numbers
showing the visible parts of it………………………………………………….………….9
Figure 1.5 part of the control room of the cooling water pumping station………………10
Figure 1.6 part of the control panel showing the controls for pump # 9…………………11
Figure 1.7 the left-most part of the control panel………………………………………..12
Figure 1.8 the right-most part of the control panel………………………………………14
Figure 1.9 the panel for the electrical system of the pumping station…………………...15
Figure 1.10 the panel for the emergency-electrical system……………………………...16
Figure 1.11 one of the diesel generators…………………………………………………17
Figure 1.12 the gantry crane of the pumping station…………………………………….18
Figure 1.13 the pump deck showing the motors over their stands……………………….19
Figure 1.14 a schematic of the pumping station…………………………………………20
Figure 1.15 chlorine cylinders…………………………………………………………...21
Figure 1.16 chlorination units……………………………………………………………22
Figure 1.17 the header basement housing the header……………………………………23
Figure 1.18 the header……………………………………………………………………24
Figure 1.19 the header resting on its concrete base……………………………………...25
Figure 1.20 the header with a line branching from it…………………………………….26
Figure 1.21 a gearbox for one of the valves in the header basement……………….……27
Figure 1.22 an apparatus over the header for venting it and for providing water to the
auxiliary system………………………………………………………………………….28
Figure 1.23 a flow meter…………………………………………………….…………...29
Figure 1.24 a schematic of the side view of the cooling pumping station…………….…30
Figure 1.25 the cover of the traveling band screen……………………………………....31
Figure 1.26 the traveling band screen after the cover has been removed …………….…32
Figure 1.27 the motor stand, the thrust bearing and the coupling………………………..33
Figure 1.28 the discharge valve connecting the pump to the header…………………….34
xv
Figure 1.29 the Valve Basement…………………………………………………………34
Figure 1.30 two butterfly valves…………………………………………………………35
Figure 1.31 the hydraulic unit over the pump deck…….………………………………..36
Figure 1.32 the hydraulic cylinder, the hydraulic shaft and the weight………………… 36
Figure 1.33 part of the bell bolted to a section of the casing……………………….……38
Figure 1.34 a more clear view of the bell………………………………………………..38
Figure 2.1 the grouping of pumps in the pumping station for the reliability study……...44
Figure 2.2 a reliability block diagram with parallel and series components…………….48
Figure2.3 a consumer with all of its components working……………………..……..…49
Figure2.4 the same consumer after closing the middle-header valve and a line valve…..51
Figure 2.4 the reliability of consumer C1 under different circumstances…………...…...52
Figure 2.6 the entire pumping station divided into its major components ………………61
Figure 2.7 a schematic diagram of pump group PIC…………………………………….62
Figure 2.8 a schematic diagram of pump group Equate…………………………………63
Figure 2.9 a schematic diagram of pump group MAR…………………………………..64
Figure 2.10 reliability block diagram of C1-caseI……………………………………….65
Figure 2.11.Block diagram of C6 -case IV………………………………………………71
Figure 2.12 Block diagram of the bypass system………………………………………..72
Figure 2.13 block diagram of C7-Case III……………………………………………….74
Figure 2.14 C1 reliability behaviors under different operational conditions…………….75
Figure 2.15 the reliability of each of the seven consumers………………………………78
Figure 2.16 the APCRS of the pumping station
for the cases shown in figure 2.15……………………………………………………….79
Figure 2.16 the reliability of each of the seven consumers with their respective case
numbers over a period of 48 thousand hours…………………………………………….80
Figure 2.17 the APCRS of the pumping station for the
cases shown in figure 2.16……………………………………………………………….80
Figure 3.1 Capacity utilization versus seawater consumption………………..………….97
Figure 3.2 Seawater temperature versus seawater consumption………………...………98
Figure 3.3 Ambient air temperature versus seawater consumption……………...………98
Figure 3.4 Humidity versus seawater consumption………………………...……………99
xvi
Figure 3.5 The model development process……………………………………………101
Figure 3.6 the pump curve……………………………………………………………...105
Figure 3.7 predicted and actual number of pumps operating…………………………...107
Figure 4.1 transforming the seasonal time series to a linear series.……………….……129
Figure4.2 the function given by equation (4.1)…………………………….………..…130
Figure 4.3 the sinusoidal function of equation (4.1)………………………..…………..130
Figure4.4 a seasonal time series………………………………………………...………131
Figure4.5 the virtual mean………………………………………………………….......133
Figure 4.6 C7 flow meter readings for the consumption……………………………….136
Figure 5.1 a process over time…………………………………………………….……145
Figure 5.2 the Shewhart chart………………………………………………...….……..146
Figure 5.3 a normal and a drifting process……………………………………....……..149
Figure 5.4 the structure of the artificial neural network…………………………..……153
Figure 5.5The learning curve for the ANN: MSE for training and cross
validation………………………………...…………………………………………...…156
- 1 -
Chapter 1 Introduction 1.1 Petroleum and Petroleum Refining
Petroleum is one of the most important resources in our modern world. Historically, it has
been known, and used, for a long time. Its importance, however, have soared only after
the industrial revolution.
“These surface deposits of crude oil have been known to human beings for thousands of
years. In the areas where they occurred, they were long used for such limited purposes as
caulking boats, waterproofing cloth, and fuelling torches. By the time of the Renaissance,
some surface deposits were being distilled to obtain lubricants and medicinal products,
but the real exploitation of crude oil did not begin until the 19th century. The Industrial
Revolution had by then brought about a search for new fuels, and the social changes it
effected had produced a need for good, cheap oil for lamps; people wished to be able to
work and read after dark.” (EncartaEncyclopedia 2004)
After the Industrial Revolution, the modern industrial societies were established. These
societies heavily depend on petroleum. “Modern industrial societies use it (petroleum)
primarily to achieve a degree of mobility—on land, at sea, and in the air—that was barely
imaginable less than a hundred years ago. In addition, petroleum and its derivatives are
used in the manufacture of medicines and fertilizers, foodstuffs, plastic ware, building
materials, paints, and cloth, and to generate electricity.” (EncartaEncyclopedia 2004)
Its export contributes to the national income of many countries while its derivatives
influence the lives of hundreds of millions of people around the globe. “In fact, modern
industrial civilization depends on petroleum and its products; the physical structure and
way of life of the suburban communities that surround the great cities are the result of an
ample and inexpensive supply of petroleum. In addition, the goals of developing
countries—to exploit their natural resources and to supply foodstuffs for the burgeoning
populations—are based on the assumption of petroleum availability.”
(EncartaEncyclopedia 2004)
- 2 -
Petroleum is rarely useful by itself and the many products that it contains can only be
obtained after petroleum goes through a refinery. A refinery processes the crude
petroleum, mainly through a distillation unit, to produce asphalt, greases, lubricants,
waxes, industrial fuels, diesel, kerosene, petrol, aircraft fuel…etc. (EncartaEncyclopedia
2004)
A refinery is a complex engineering system. It generates many products. It also needs a
lot of inputs “A typical refinery requires enough utilities to support a small city. All
refineries produce steam for use in process units. This requires water-treatment systems,
boilers, and extensive piping networks. Many refineries also produce electricity for
lighting, electric motor-driven pumps, and compressors and instrumentation systems. In
addition, clean, dry air must be provided for many process units, and large quantities of
cooling water are required for condensation of hydrocarbon vapours.”
(EncyclopediaBritanica 2010)
This thesis is about the cooling-water system of a refinery and/or petrochemical plant.
More specifically, it is about the pumping station part of it.
1.2 The Cooling-Water System of a Refinery
Parkash (Parkash 2003) lists the uses of water in a refinery,
“Water is used in an oil refinery for the following purposes:
Cooling.
Steam generation.
Domestic and sanitation purposes.
Washing products.
Flushing equipment, pipelines, and hydro tests.
Fire fighting.” (Parkash 2003)
Regarding the use of cooling water, Parkash writes that “Refining operations are
conducted at elevated temperatures. In a rough overall sense, a refinery must be in heat
balance. All heat added in the forms of fuel burned, steam consumed, or coke burned
- 3 -
must be removed by one of the various cooling systems. Water cooling is one such
system. The others are air cooling and heat exchange with other streams. Cooling
accounts for about 90% of the total refinery water requirements. Approximate cooling
water requirements of a refinery can be estimated as a function of refinery complexity.”
(Parkash 2003)
There are two types of refinery cooling systems: Once through and recirculating.
Parakash explains that “Water in a refinery's cooling system either travels through the
system once or is recirculated. In a once-through system, pumps suction water from a
source, such as a sea, river, or lake, and deliver it to process units or other water users
within the refinery. After passing through the cooling equipment, the hot cooling water is
conducted to a point of disposal through a pressure system of piping or through a gravity
flow system. In recirculated systems, pumps suction water from a cooling tower basin
and deliver it to cooling equipment. After passing through water user equipment, the hot
cooling water is discharged through a pressure return system to the top of the cooling
tower. The water cooling system includes heat exchangers, pumping equipment,
distribution piping, and water intake stations, and cooling towers.” (Parkash 2003) Most
of the cooling systems that are considered in this thesis are once-through cooling systems.
In this thesis, the subjects and problems related to the pumping station part of the refinery
cooling system are going to be studied.
1.3 About Cooling Petrochemical Pumping Stations and their Importance
A cooling petrochemical pumping station is a very important component in any refinery
or petrochemical plant. A refinery or a petrochemical plant usually produces large
amounts of heat. This heat is removed via heat exchangers. Heat exchangers usually use
water to remove the heat.
The process industry works around the clock. Consequently, the cooling pumping station
should pump continuously without any interruption. When the flow of cooling water to a
refinery or petrochemical plant is interrupted, the following scenarios may happen:
- 4 -
1. The sudden increase in temperature will cause damage to
equipment. This damage to equipment may in turn results in
hazards to both individuals and the environment.
2. The damaged equipment will incur equipment fixture or
replacement.
3. As a result of absence of cooling water, a petrochemical plant or
refinery will have to make an unscheduled shutdown, an event
that the petrochemical industry tries to avoid for several reasons,
including:
a) decrease in the generation of revenues: the unplanned
stoppage would cause loss of production and,
consequently, significant financial losses;
b) the unplanned shutdown would put a chemical plant or
refinery at predicament with its consumers, and monetary
penalties may apply in addition to the loss of reputation of
that particular plant or refinery with prolonged ill effects;
and
c) while the cause of an unplanned shutdown might only last
for seconds, minutes or hours, it would take several days to
restore the chemical plant or refinery to its previous steady
state production.
1.4 The Cooling Pumping Station Understudy
The work presented in this thesis was carried in a cooling petrochemical pumping station
in Kuwait. Kuwait does not have any rivers or lakes. This would make the sea the only
source for the large amounts of water required for the heat exchangers. The pumping
station understudy is located in a large petrochemical complex to the south of Kuwait and
serves several consumers (petrochemical plants). It is part of a once-through cooling
system. Figure 1.1 shows part of the petrochemical complex and the location of the
pumping station within it. Figure 1.2 shows a closer look at the pumping station.
- 5 -
To illuminate the importance of this pumping station, table1.1 below shows the cost of
loss of production per day for the consumers of the pumping station. These figures were
obtained from the consumers. The recorded losses are due to loss of production only and
do not include the costs of equipment damage, legal penalties for delays, extra man
power works…etc. It can be seen, therefore, the an interruption resulting in one day of
stoppage will cause the consumers an excess of $ 13 million , not to mention the human
and environmental risks associated with such an event.
Figure 1.1 An aerial view of a part of the Shauiba Industrial Complex. The arrow points
to the location of the cooling pumping station understudy (taken from Google earth).
- 6 -
Figure 1.2 An aerial view of the Shauiba Industrial Complex near the pumping station. In
the left of the picture parts of the petrochemical plants are shown. On the right of the
picture, the upper arrow points to the location of the cooling pumping station while the
lower arrow shows the return (hot) water from one of the plants (taken from Google
Earth).
- 7 -
Table 1.1 The cost of one-day loss of production
1.5 The Structure of a Cooling Pumping Station
The structure of the cooling pumping station under study is not largely different from
what Parksh (Parkash 2003) described of a typical seawater cooling pumping station. He
wrote that
“Because of the large volume of water intake for cooling, water is pumped from the sea
to an inlet sump through a battery of low lift pumps. Sea water next flows through a
system of bar screens, scrapper screens, and rotary screens to the suction of the high lift
pump manifold. The screen system prevents entry of marine life, seaweed, algae, and the
like into pump suction. To prevent growth of algae and fungi and suppress microbial
activity, chlorine is injected (0.5-1.0 ppm) into sea water before it enters the suction of
the high lift pump battery. As sea water is very corrosive, reinforced concrete pipes are
used for all piping greater than 30 in. diameter. Below 28 in. diameter, cement lined steel
pipes are used for sea water service. The battery of high lift pumps supplies sea water to a
sea water manifold running throughout the refinery, from which individual process units
and utilities tap their cooling water supply. To prevent any grit, debris, or scale from
entering the exchangers, strainers are placed at the entry of every process or utility unit.
Warm sea water coming out of an exchanger is segregated into two manifolds: a "clean"
sea cooling water and an "intermediate" sea cooling water. Intermediate, or oil-
contaminated, sea water, by definition, is that sea water used in service where the oil side
of exchanger is operating at 55 psig or greater. All contaminated sea water may be routed
Consumer Daily cost of loss of production ($)
C1 and C2 1,525,000 C3 581,480 C4 860,000 C5 50,000 C6 9,107,000 C7 1,000,000 Total 13,123,480
- 8 -
through the intermediate sea cooling water manifold to a battery of corrugated plate
interceptor separators before discharging it to sea. If the oil side of the exchanger is
operating at less than 55 psig, all such returning sea water is routed to the clean sea
cooling water manifold and discharged directly to the sea.”
Figure 1.3 shows an aerial view of the pumping station under study.
Figure 1.3 An aerial view of the cooling pumping station understudy (taken from Google
earth).
- 9 -
Figure 1.4 an aerial view of the cooling pumping station understudy with numbers
showing the visible parts of it (taken from Google earth).
The arrows and numbers in figure 1.4 point to the visible parts of the pumping station.
The numbers in figure 1.4 corresponds to the following
1. maintenance workshops
2. control room and administrative offices
3. gantry crane
4. pump deck showing the pump motors and the traveling band screen
5. chlorination building
6. fore bay
- 10 -
7. cleaning boat
8. emergency diesel generator building
9. store building
Figures 1.5 below shows a closer look at the control room (number 2 in the above list).
The control room is where the pumping system that includes the pumps, the traveling
band screen, the valves, the flow meters…etc. is controlled. The most prominent
component in the control room is the control panel. The control panel encircles the
control room and it has the parts responsible for the operation of pumps, auxiliary system,
electrical system, emergency-electrical system, and the industrial area’s fire-fighting
system. The part shown in figure 1.5 is the one for the pumping system
Figure 1.5 part of the control room of the cooling water pumping station.
- 11 -
Figure 1.6 below shows a closer view at a part of the control panel that is responsible for
the operation of pump # 9.
Figure 1.6 part of the control panel showing the controls for pump # 9.
Figure 1.6 above shows a cam switch with the word CCR lit beneath it. The word CCR
stands for Central Control Room. It can be seen that this word appears lit five times in the
figure. This means that the device is being controlled from the control room as apposed to
being controlled from the site (which is termed ‘LOCAL’ in the panel).
- 12 -
The upper part of figure 1.6 shows three columns of lamps. The left-most column shows
the status of the screen valve providing water to the traveling-band screen. The middle
column shows the status of the main discharge valve of the pump. The right-most column
shows the status of traveling-band screen itself. The lower part of the figure shows the
controls of two system valves.
Figure 1.7 below shows an operator standing at the left-hand side of the control panel.
Figure 1.7 an operator standing at the left-most part of the control panel. This part is
responsible for the operation of the auxiliary system and for displaying the alarms of
auxiliary system and the alarms of the system valves.
The left-most part of the control panel, shown in figure 1.7, is responsible for the
operation of the auxiliary system and for displaying the alarms of auxiliary system and
- 13 -
the alarms of the system valves. The auxiliary system is made up of four pumps called
the auxiliary pumps. These pumps are vertical pumps, just like the main pumps (that will
be discussed later). The auxiliary pumps, nevertheless, produce smaller amounts of water
in comparison with the main pumps. The auxiliary pumps deliver their water to the
auxiliary system.
The auxiliary system is responsible for delivering seawater to the following
a. the cooling and lubricating system of the main pumps (the cooling water pumps,
CWP’s).
b. the traveling-band screen system, and
c. the chlorination system.
A pump is a rotating machine that produces enough torque to increase the pressure of the
process medium (in this case sweater) to deliver it to a destination point. The vertical
pumps used as the main cooling-water pumps (CWP’s), just like most heavy-rotating
machinery, need lubricating oil to facilitate the process of rotation. The oil in the cooling
water pumps is located in a part of it called the thrust bearing (will be shown later). Both
of friction (that comes from rotation) and the torque generated produce heat. If this heat is
not dissipated, problems will result in the oil and, subsequently, the pump.
The auxiliary system provides cooling water that is delivered to the thrust bearing of each
main cooling-water pump. This water enters a serpentine pipe in the thrust bearing to
cool the oil. Afterwards, this water is discharged to the pump chamber.
In addition to the four auxiliary pumps, the auxiliary system takes water for thrust-
bearing cooling from pipes that branch out from several locations of the header. This was
done to increase the redundancy and, consequently, the reliability and availability of the
thrust-bearing-cooling process.
The traveling-band screen and chlorination system shall be discussed later.
Figure 1.8 below shows the right-most part of the control panel.
- 14 -
Figure 1.8 an operator standing at the right-most part of the control panel. Behind him on
right is the electrical system of the pump and on the left is the emergency-electrical
system. In front of him is a turned-off monitor of the station’s SCADA system.
The electrical system is made of four 132 KV/ 11.5 KV transformers. Three of these
transformers are always working while the fourth one is in standby duty or under
maintenance. There is a ring bus bar that the secondary outputs of these transformers are
connected two. The ring bus bar is divided into sections and feeds the 11 KV motors of
the main pumps. Two service transformers (11.5 KV/ 415 V) are connected to the ring
bus bar. Each service transformer is connected to its own bus bar and provides electricity
to low voltage applications such as the chlorination system, lighting, air conditioning and
several motor control centers (MCC’s). The motor control centers (MCC’s) house the
motors of the auxiliary pumps, the main valves, the system valves and the traveling band
screen valves. A closer look at the panel of the electrical system can be seen in figure 1.9
below.
- 15 -
Figure 1.9 the panel for the electrical system of the pumping station.
Figure 1.8 above shows the panel for emergency-electrical system on the left of the
picture. A closer look at this panel is shown in figure 1.10. The emergency-electrical
system is responsible for operating two main pumps for one of the consumers and
providing the electricity for the low-voltage system when an electrical blackout takes
place. The main components of this are two diesel generators. One of the diesel
generators is shown in figure 1.11. The two generators are housed in the building labeled
8 in figure 1.4 above.
- 16 -
Figure 1.10 the panel for the emergency-electrical system.
- 17 -
Figure 1.11 an operator standing at the side of one of the diesel generators.
Other parts of the control room not shown in the above figure are the programmable logic
controllers (PLC’s) and the supervisory control and data acquisition system (SCADA).
PLC’s have the control scheme for various equipment, most importantly pumps. The
SCADA system collects the data from the pumping plant for trending, monitoring,
alarming and special output generation such as the amount of water delivered or the
electricity consumed by using some algorithms.
Label 3 in figure 1.4 shows the gantry crane. The gantry crane is responsible for lifting
objects over the pump deck; specially, the main pumps and their motors. A look at the
gantry crane from a view different from the one taken in figure 1.4 is shown in figure
1.12.
- 18 -
Label 4 in figure 1.4 is the pump deck. The pump deck has over it the upper part of the
Figure 1.12 the gantry crane of the pumping station.
main pumps, the pump-motor stands, the pump motors, the upper part of the traveling-
band screens, various local control panels, the auxiliary pumps, the piping works for the
auxiliary system that include the lubricating, chlorination and traveling screen piping and
the gantry crane and its rail. Figure 1.13 shows part of the pump deck. The figure shows
part of the rail for the crane and the hydraulic panels for the main valves of the pumps
and the motors and upper part of the pumps.
Figure 1.4 shows a satellite image of the pumping station. Not all parts of the pumping
station are visible in that image. Figure 1.14 shows a schematic of the pumping station.
- 19 -
Figure 1.13 the pump deck showing the motors over their stands. Each group of pumps
has its own color.
This schematic shows the nine major components of which the cooling pumping station
under study is made of. The numbers in the list below corresponds to the numbers in
figure 1.14.
1. A sea intake which allows water to flow by gravity to,
2. A basin or fore bay, which collects the water.
3. A chlorinating system, to disinfect the seawater.
4. A traveling screen, which removes any particles greater than 1 cm in
diameter before water is sucked by the pumps.
5. A set of pumps to deliver the sea water to the consumer(s). The pumps are
separated into groups each having a specified discharge flow and pressure.
Each pump group delivers the seawater to its own designated consumer(s).
- 20 -
6. A designated header for each pump group, which collects the water from
all the pumps in the group.
7. A piping system, which is connected to a consumer.
8. A flow meter to measure the amount of seawater used by each consumer.
9. A group of butterfly and throttle valves to control the water flow to the
consumer
Figure 1.14 a schematic of the pumping station
Many parts in the schematic of figure 1.14 are shown in figure 1.4 and were explained
previously. Next, the parts that have not been shown and discussed will be mentioned.
- 21 -
The chlorination system which is labeled 3 in figure 1.14 is mainly located in the
chlorination building (labeled 5 in figure 1.4). This system adds chlorine to the seawater
to disinfect it. The disinfection largely minimizes the growth of marine life in the tubes of
the heat exchangers. The chlorine solution comes from a chlorine plant, located outside
the pumping station. When this solution enters the chlorine building, it is diluted by
seawater coming from the auxiliary system. For two hours every day, a ‘shock dose’ of
chlorine from the pumping stations own chlorination units is added to the chlorine
solution coming from the chlorine plant. The purpose of this shock dose is to kill any
biological organisms that may have developed immunity for the regular dose. Chlorine
cylinders (shown in figure 1.15) are used for this shock dose.
Figure 1.15 chlorine cylinders.
- 22 -
Figure 1.16 chlorination units
The chlorine in the cylinders is delivered by its own pressure and gravity to the
chlorination units shown in figure 1.16. These units, with the aid of seawater from the
auxiliary system, make a chlorine solution that is injected in the fore bay.
The header, labeled 6 in figure 1.14, collects water from several pumps. It is located in
the header basement below the pump deck. The header is divided to sections by several
valves. From the header, several lines branch out to the consuming plants. Both the
header and the lines are drained for maintenance purposes. When the header and the lines
are filled up again, the trapped air inside them could damage them if it is compressed by
the water. Therefore, venting apparatuses are installed over the header and the lines to
facilitate the escape of the trapped air.
- 23 -
Over the header, the venting apparatus is usually combined with another apparatus that
provides cooling water to the auxiliary system. This cooling water goes to a serpentine
tube inside the thrust bearing to cool the lubricating oil. Figures 1.17 to 1.22 show the
header and what is connected to it.
Labeled 8 in the schematic of figure 1.14, is the flow meter. A flow meter is a device that
measures the flow. There is a flow meter installed over every line for every consumer.
Flow meters are installed for billing purposes. The flow meters used in the pumping
station are simply pipe sections with sensors attached to them. The sensors used are either
ultrasonic or electromagnetic. Figure1.23 shows a flow meter that was removed from the
site.
Figure 1.17 the header basement housing the header. The header is not clearly shown in
the left of the picture. The four trays on the right of the picture are cable trays. The boxes
that are attached on the cable trays are local control panels for operating the header
valves.
- 24 -
Figure 1.18 the header clearly shown in this figure colored in red and stretching all the
way to the end of the basement. The checker plates that usually cover the header have
been removed. At the lower right of the picture shown, colored in blue, is the motor and
manual hand wheel that drives the gearbox of one of the valves that divides the header.
At the back of the picture is an apparatus for venting the header and for providing cooling
water for the auxiliary system. The auxiliary system delivers this cooling water to the
lubrication system of the pumps. A pipe is shown in the back leaving the apparatus and
going to the lubrication system.
- 25 -
Figure 1.19 the header resting on its concrete base. Also shown in this figure at the left is
a valve that divides the header. The protrusion on the left is for the bearing of the shaft
that rotates the butterfly valve.
- 26 -
Figure 1.20 the header on the right with a line branching from it to one of the consumers
on the left. At the top of the line, there is a venting apparatus at the upper left corner for
venting the line.
- 27 -
Figure 1.21 a gearbox, colored in blue, for one of the valves in the header basement. This
gearbox can be manually operated by the wheel on the left but is usually driven by a
motor. The lower-right corner shows cables leaving the cable try and going to the motor.
- 28 -
Figure 1.22 an apparatus over the header for venting it and for providing water to the
auxiliary system.
- 29 -
Figure 1.23 an operator standing besides a dismantled flow meter. This flow meter is
simply a pipe sections that connects with the rest of the pipeline going to a particular
consumer. However, it is provided with sensors that measure the amount of water
consumed. At the top of the flow meter is a metallic cylinder that houses one of the
sensors.
- 30 -
Even with the satellite image of figure 1.4 and the schematic in figure 1.14 not all parts of
the pumping station are shown. Figure 1.24 shows a schematic of the side view of the
pumping station.
Figure 1.24 a schematic of the side view of the cooling pumping station.
The letters in the following list corresponds to letters shown in figure 1.24
A- Traveling band screen cover.
B- Upper wheel of the traveling band screen.
C- Front wall of the pump chamber.
D- Lower wheel of the traveling band screen.
- 31 -
E- Traveling band screen.
F- Pump casing.
G- 11.5 K.V. electric motor
H- Motor stand.
I- Coupling.
J- Pump thrust bearing.
K- Back wall of the pump chamber.
L- Header.
M- Pump discharge valve.
N- Bell containing the impeller.
O- Motor shaft.
P- Pump shaft.
Labeled A in figure 1.24 is the traveling bands screen cover. This is also shown in figure
1.25. The traveling band screen, labeled E in figure 1.24, is mainly made of a wire mish
that has 10 mm X 10 mm openings. The function of this wire mish is to prevent any
object
Figure 1.25 an operator trying to remove the cover of the traveling band screen.
- 32 -
that has a diameter greater than 10 mm from going beyond the screen. Objects of
diameter greater than 10 mm would stick on the outer surface of the screen. Over time,
these objects would accumulate and obstruct the suction of the pump. This is a serious
condition that could damage the pump. To prevent this from happening, the traveling
band screen would automatically rotate every 8 hours (or when there is substantial
accumulation of material on its surface) to bring the lower parts of the screen upwards. At
the top, a strong jet of water from the auxiliary system would clean the wire mesh of the
screen. The traveling band screen is divided into baskets. Figure 1.26 shows a basket
with its wire mish.
Figure 1.26 the traveling band screen after the cover has been removed showing two
baskets of wire mish.
- 33 -
Labeled H in figure 1.24 is the motor stand. Its function is to support the motor. Labeled J
in figure 1.24 is the thrust bearing, the function of which has been previously explained.
Labeled I in the same figure is the coupling. The function of the coupling is to couple the
motor to the pump and by doing so transmits the motion of the motor to the pump which
eventually transforms it to torque. Figure 1.27 shows these parts in actuality.
Figure 1.27 the motor stand, the thrust bearing and the coupling. There is a protective
sheet of metal that surrounds the coupling.
Labeled M in figure 1.24 is the pump’s discharge valve. All discharge valves are located
in the Valve Basement. Figure 1.28 shows the valve at the site and Figure 1.29 shows the
basement where all the discharge valves are located. The pump discharge valve is of the
butterfly type. Butterfly valves are usually either fully open or fully closed. Figure 1.30
shows butterfly valves that have been dismantled from their pumps.
- 34 -
Figure 1.28 the discharge valve connecting the pump to the header. The protrusion in the
middle is for a bearing that houses the shaft that turns the valve disk.
Figure 1.29 the Valve Basement.
- 35 -
Figure 1.30 two butterfly valves that have been removed from their respective pumps.
The valves are resting on their side. The valve disk and valve body are shown.
The pump discharge valve is actuated by a hydraulic unit. Figures 1.31 shows the
hydraulic unit located above the valve in the pump deck. The pump discharge valve
would open when the shaft of a hydraulic cylinder, pushed by the hydraulic pressure
behind it, causes the valve disk to rotate upwardly. The pump discharge valve would
close when the hydraulic pressure pushing the hydraulic cylinder’s shaft is released and a
weight causes the valve disk to rotate downwardly.
- 36 -
Figure 1.31 the arrow points to the hydraulic unit located over the pump deck.
Figure 1.32 the arrow points to the hydraulic cylinder housing the hydraulic shaft. The
hydraulic shaft is not shown because it is retracted inside the cylinder due to the valve
closed position. The weight is at the back of the picture.
- 37 -
Labeled N in figure 1.24 is the bell which houses the pump impeller and labeled F in the
same figure is the pump casing. The impeller sucks seawater and considerably increases
its pressure. The pressurized seawater flows through the pump casing to the pump’s
discharge valve and, subsequently, to the header. Figure 1.33 shows a dismantled pump
casing connected to the bell.
- 38 -
Figure 1.33 part of the bell, at the right, is bolted to a section of the casing at the left.
Figure 1.34 a more clear view of the bell.
- 39 -
1.6 Overview of the Research Problem
The high cost of a sudden shutdown of the cooling pumping station made reliability and
availability the main management issues in the operation of the cooling petrochemical
pumping station. Accordingly, the purpose of this work is the following: Because of the
criticality of cooling sea water interruption to both the pumping station and its
consumers, the reliability of the pumping station is extremely important and can not be
over emphasized. Usually, the first step in improving the reliability of something is
making a reliability model of it. The problem encountered here was that classical
reliability analysis was not appropriate for making a reliability model for the pumping
station.
A reliability model of a system is the first step in the process of improving the reliability
of this system. It is certainly not the only step. Reliability implies making proper
maintenance to the working equipment. This maintenance, it was found, could not be
easily conducted in practice. The performance of maintenance was often antagonized by
the operational needs of the consumers. This resulted in a situation where production and
maintenance were in conflict. To solve this problem, a method had to be devised to
minimize this conflict. This minimization, it was hopped, would help in conducting
maintenance and, as a result, would make the pumping station more reliable.
The purpose of improving the reliability through proper maintenance, which does not
interfere with the pumping station’s operation, was to defend it against financial losses
and to increase its profit. The income of the pumping station is generated by charging its
consumers for the cooling seawater supplied. The charging is done by taking the readings
of flow meters installed on the pipelines for each consumer and including these reading in
monthly bills. The problem was that these flow meters, like all machines, failed
occasionally.
For a measuring device like a flow meter, one method of defining a failure status is when
the device produces an inaccurate reading that is inconsistent with the true phenomenon
being measured. In practice, this meant over charging or under charging a consumer.
- 40 -
When a flow meter under registers the consumption of a plant, the amount of cooling
water that was unregistered is called unaccounted-for-water. Understandably, the
pumping station owners would like to reduce the amount of unaccounted-for-water as
much as possible because it represents lost revenue.
1.7 The Aims and Objectives of the Research
The aims and objectives of this thesis are to
1. describe a model for the reliability of the pumping station,
2. minimizing the conflict between the operation of the pumping station and its need
for maintenance and
3. reduce the amount of unaccounted-for-water lost by detecting flow meter drift.
1.8 The Significance of the Research and its Scope.
The research done in this thesis was carried at a cooling petrochemical pumping station
called Pumping Station B. This station is one of two pumping stations owned by PAI of
Kuwait. As described earlier, the station is part of a once-through cooling system. As a
result, some of the issues mentioned in this research (problems, findings and solutions)
might be specific for it. Nevertheless, it is this author’s opinion that many more issues
might be of interest to other cooling petrochemical pumping station owners world wide.
The previous works on pumping stations include the book by Jones (Jones 2008) the
military manual for pumping station design (US Army Corps of Engineers 1994) . Both
of them concentrate on mechanical, electrical and civil aspects of pumping stations. As
for cooling water systems, Sharp and Sharp (Sharp and Sharp 1995) discussed the effect
of water hammer in a cooling water systems. Castro et al (Castro, Song et al. 2000)
discussed the minimization of operational costs in a cooling-water system. Their work,
however, concentrated more on cooling towers. Ponce-Ortega et al (Ponce-Ortega, Serna-
Gonzalez et al. 2009) presented a simultaneous optimization model for the synthesis and
detailed design of re-circulating cooling water systems. The objective was minimizing the
total annual cost for the cooling water system, which includes the capital costs for the
tower, coolers, and pumps, plus the operating costs for the fan of the tower, pumps and
- 41 -
water make up. Cortinovis et al (Cortinovis, Ribeiro et al. 2009) presented a paper with
the objective of developing and validating a procedure for the systemic performance
analysis and characterization of cooling water systems. Their approach combined
experimental design with mathematical modeling. Harish et al (Harish, Subhramanyan et
al. 2010) developed a theoretical model to establish viability of providing variable
frequency drives (VFD’s) for cooling water pumps (CWP’s) in power plants with
seawater based once through condenser cooling water system.
None of the previous works addressed the issues of pumping station reliability or the
problem of the conflict between the production function and the maintenance need for a
pumping station. Also, the problem of unaccounted-for-water loss in a cooling system
was not addressed. This thesis is an attempt at filling this gap. It is composed of four
works:
1. a look at how the reliability of a cooling pumping station can be described,
2. a method for minimizing the conflict between operation and maintenance ,
3. a method for detecting flow meter fault by using statistical process control, and
4. an alternative method for detecting flow meter fault by using artificial neural
networks.
In addition to the abstract and this introduction, this thesis has a chapter on the reliability
analysis of a cooling pumping station. Chapter 3 deals with minimizing the conflict
between the operation and maintenance of a cooling pumping station. Chapters 4 and 5
are about reducing the unaccounted-for-water loss. In both chapters, the method to
achieve this end is flow meter fault detection. In Chapter 4, a method for detecting flow-
meter fault by using statistical process control (SPC) is presented while in Chapter 5
flow-meter fault is detected by using artificial neural networks (ANN’s). Chapter 6 is
where the discussion and concluding remarks take place. There is also an appendix at the
end of this thesis.
- 42 -
Chapter 2 Reliability Analysis of a Cooling Pumping Station
2.1 Introduction to Cooling Pumping Station Reliability
The scenario under study is called Pumping Station B, located in Shuiba Industrial Area,
the main area for the petrochemical industry in Kuwait. It belongs to the Public Authority
for Industry (PAI), a branch of the Kuwaiti Government.
This study attempts to model the reliability and availability, from consumer point of
view, of the South Pumping Station (B), at Shuiba Industrial Area. First a functional
reliability block diagram of the system is constructed. Then, the operational reliability of
each component of the system is estimated by using the data from the past failures,
repairs, and maintenance of various components, including motors, pumps, valves, and
pipes in the system.
The steady state reliability of the system is then calculated based on the minimum
number of pumps needed for acceptable operation and in standby to meet the industrial
demand for cooling water, with the specified pressure and flow rate. The reliability figure
will help the management to determine the likelihood of water flow interruption and to
incorporate the necessary preventive measures into the system.
The South Pumping Station (B) includes a complex system of 16 huge vertical mixed
flow pumps, about 50 butterfly and throttle valves, seven header sections, more than 20
pipeline sections, several flow meters, and driving motors for the pumps and the valves.
The sea intake pipes and fore bay of the station are designed for a maximum pumping
capacity of 182,000 m3/h of sea water. The structure of the system has been explained in
the previous chapter and will be further elaborated on in the following paragraph and
figure.
- 43 -
The specific structure of the station and the arrangement of pumps, as in figure 2.1, show
that pumps 1-6, 7-12 and 13-16 are separated into three groups. The first group (pumps 1-
6) is called group PIC serving a set of specific consumers. The second group (pumps 7-
12) is called group Equate only available to a single consumer, namely Equate. The third
group (pumps 13-16) is called group MAR serving the same named refinery. The piping
system for each group and the connection to the consumers are also shown in figure 2.1.
2.2 Literature Review
The literature does not contain significant major research work directly related to
cooling-petrochemical-pumping stations. Even works on pumping station reliability in
general were difficult to obtain. Thomas (Thomas 1981) and Lydell (Lydell 2000) dealt
with vital components of a pumping station; namely, pipelines and pressure vessels. The
subject of their work, however, was nuclear power plants and not pumping stations. Other
related work can be seen in Billinton and Allan (Billinton and Allan 1983), Proctor et al.
(Proctor, Savsar et al. 1985) , Hoffmeister (Hoffmeister 1988), Nakamura (Nakamura
2001) , Cobb (Cobb 1998) , Shayan (Shayan 1986), Bevilacqua et al.(Bevilacqua,
Braglia et al. 2003), Baxter and Tortorella (Baxter and Tortorella 1994) and Nakamura et
al.(Nakamura, Nanayakkara et al. 1992).
Before embarking on a work on the reliability of a cooling-petrochemical-pumping
station, the term ‘reliability’ itself should be defined. Billinton and Allan (Billinton and
Allan 1983) have quoted Bagowsky (Bagowsky 1961) on the definition of reliability:
“Reliability is the probability of a device performing its purpose adequately for the period
of time intended under the operating conditions encountered”. (Billinton and Allan, p. 2)
There are many works on pump reliability. Nevertheless, the literature does not contain
significant major research works directly related to pumping station reliability, not to
mention the reliability of cooling petrochemical pumping stations.
- 44 -
PAI -SEA WATER HEADER PLANT (B)
P-7
P-8
P-9
P-10
P-11
DN 2400
P -12
P-6
P-5
P-4
P-3
P-2
P-1
DN1400
DN 2600
P -13
P -14
P -15
P -16
DN 2000
DN1200
V. speed V. speed
DN1400
DN2000
DN 2400 DN 2000
DN1400
DN1400
DN 800
DN 1200
DN 1400
DN 1400
DN 1000
DN 800DN 1200 DN 1000
DN 1400
DN 1400
DN 1400
DN 1400
DN 1400
DN 2000DN 2000
DN 20004G1 4G2
4M1 4M2
4M3 4M4
TO MINA ABD
ULLA
REFINERY
T P 1
T P 2
T P 3
TO EQUAIT
TO P IC
(B) H
P.
TO P IC (B
) L P
.
TO SALT & CHLO.
TO KNPC REFINERY
TO P IC (A)
TO BY-PASS SYSTEM
DN 1400DN 1000
TH .V.
TH .V.
TH .V.DN 1400
TH .V. DN 1000
DN 22004E1
4E3
4E2
4H1
4E4
4N1 4N2 4N34G 4G3 4L1
4B64B5
4B3
4B44H2
4D4
4D34J1
4H3
4I1
LIN
E A
(S
OU
TH
)
LIN
E B
(N
OR
TH
)
TP1A
TP1B
TP2A
TP2B
TP3A
TP3B
ORFICE PLATE
5F1 5F2 5E1 5E2
5H2
5C1
5B1
5H15D
1
FLOWMETER
Figure 2.1 the grouping of pumps in the pumping station for the reliability study.
- 45 -
Although the term ‘reliability’ can be clearly defined, the application of the definition can
be elusive i.e. the reliability of a system can be considered from many different views.
Billinton and Allan (Billinton and Allan 1983) have written that
” The criterion of ‘adequate performance’ is an engineering and managerial problem.
Failure of a system may be a catastrophe or a complete failure to operate, or it may be
caused by a violation of the required system function; for example, the power output of a
mechanical pump may fall below a minimum requirement although the pump may still be
operating. An assessment of adequate performance is a matter of engineering appraisal and
appreciation”. (Billinton and Allan, p.3).
As explained in the previous paragraph, it is important to define the term ‘failure’ in the
context of a cooling-petrochemical-pumping station. Certainly, when all the pumps in the
system are not working (as what would happen in an electrical blackout) this would be a
failure. Nevertheless, if the output of pumps is below the minimum requirement for a
consumer to operate, the operating pumps would be as useless as failing pumps.
Conversely, their might be enough operating pumps to satisfy more than the minimum
requirement for a plant to operate. Yet, the system of delivery might take a configuration
that would make operating these pumps degrading for the entire pumping system.
Subsequently, the term ‘failure’, in our case, would imply
1. not satisfying the minimum requirement of flow to a consumer to operate, and
2. not observing the operational constraints of the system.
The definition of reliability considered, therefore, in this study was cooling seawater
delivery to the consuming plants at the required pressure and flow rate while observing the
operational constraints on the system. This definition of reliability made it very close to a
definition that might be adopted by a consumer of cooling seawater.
2.3 Reliability Analysis of the System
2.3.1. Introduction
Just like any reliability study analysis, the functional reliability block diagrams of the
system were constructed. Then, the operational reliability of each component of the system
- 46 -
was estimated by using the data from the past failures, repairs, and maintenance of various
components, including motors, pumps, valves, and pipes in the system. The system
reliability was then calculated based on the minimum number of pumps needed for
acceptable operation and in standby to meet the variable industrial demand for cooling
water, with the specified pressure and flow rate. It was hoped that the reliability figure will
help the management to determine the likelihood of water flow interruption and to
incorporate the necessary preventive measures into the system. Still, achieving all the
previously mentioned tasks was only possible after solving some practical and theoretical
problems as shall be explained in the following paragraphs.
When the reliability of a system is thought of, it is usually based on a success/failure
criterion. Billinton and Allan (Billinton and Allan 1983) have written that
“In many applications, this criterion (the success/failure criterion) is the most appropriate
one to use. Examples of this are mechanical structures, aircraft flight control circuits, safety
or hazard monitors or detectors, and so on. However, in some engineering problems that
involve ‘flow’, a different criterion may be more appropriate. Examples of ‘flow’ occur in
electrical power systems, chemical process plants, cooling water circulators, manufacturing
industries, and so on. In these cases, a criterion based, not on numbers of components, but
on percentage throughput, flow or output is usually more desirable” (Billinton and Allan,
p.49)
The reliability of a cooling pumping station is, basically, the reliability of a cooling water
circulator. This mandated that the point of view of Billinton and Allan, that the reliability of
such a system should use “a criterion based, not on numbers of components, but on
percentage throughput, flow or output” should be considered. In this thesis, this point of
view was applied and further developed. Other issues related to flow, that this candidate
considered, were issues such as maximum amount of flow a pipe can withstand.
Reducing the amount of flow to a consumer below a limit might lead to a derated or even
failed state. On the other hand, exceeding the flow threshold in a pipe would lead to pipe
- 47 -
erosion and rupture which will result in failure. This is what was meant in the introductory
paragraphs of this chapter by writing that reliability in this study was thought of as the
reliability of cooling seawater arriving to the consumer at the required pressure and flow
rate and within the system’s constraints.
Pipelines presented a challenge in the analysis due to the two conflicting issues: the first
one is that each consumer had a demand that should be met while each line of the lines
delivering this demand had a delivering capacity that should not be exceeded. To express
these conflicting factors the concept of the conditional parameter was introduced.
Conditional parameters, thus, can take the forms:
1. Flow is less than X m3/h (considering the delivering pipe capacity when it is
relevant)
2. Flow is more than or equal to Y m3/h (considering the consuming plant demand
when it is relevant)
3. Y < Flow< X ( considering both pipe capacity and plant demand when both are
relevant)
Because reliability is basically a sort of a probability calculation, it is done by considering
the failure rate of each component of the system. The failure rate is how often the
component fails in a specified period of time. To make the conditional parameter fit into
reliability calculations, it too must be a probability. Actually, it is how often the conditional
parameter(s) is satisfied.
In a classical reliability analysis diagram, each component is represented as a square. The
squares are arranged according to their functional relations in the block diagram. When the
calculations are made, the failure rates (probabilities) are put into these squares.
Conditional parameters can be treated just like components both in reliability diagrams and
calculations. To distinguish them from components, conditional parameters were
represented by circles in the reliability diagrams as seen in figure 2.2 below.
- 48 -
A
B E
C
G F Y<Flow <X
Figure 2.2 a reliability block diagram with parallel and series components. A conditional
parameter is shown as a circle.
The classical analysis and representation were insufficient for some components, such as
valves and solenoids. These components can lead to system success even if they were in
some failed states (Proctor, Savsar et al. 1985). Whereas a typical component can be
thought of as being in either the operating or failed states, a valve, for instance, can be in
three states as:
(1) operating;
(2) failed in the open state (failed open); and
(3) failed in the closed state (failed closed).
Series of redundant conditions greatly enhances a system’s reliability when only open state
failures can occur, while system reliability improves for parallel redundancy of valves
which tend to fail only in the closed state (Proctor, Savsar et al. 1985). This means that a
valve in the failed open state would not cause system (or subsystem) failure if it was in
series with other components. Similarly, a valve in the failed closed state would not cause
system (or subsystem) failure if it was in parallel with other components.
The reliability of the pumping system can be dramatically affected by header section
isolation or a line valve closure. This means that for each consumer, reliability depends on
how many header sections and pipelines are there for the supply of cooling water. If one
section or a pipeline is closed for maintenance, a reduced supply of water would still be
- 49 -
available through other sections or pipe lines, with less reliability, due to the decrease in the
number of operating pumps and reduction in the amount of possible flow. These aspects of
the system have been taken into consideration in this research.
Consumer
Figure2.3 a consumer with all of its components working.
Normally, all the header valves in a pumping station are open to connect adjacent header
sections and all the line valves are open. Figure 2.3 above shows a consumer in this case.
Nevertheless, a header section can be isolated and a line valve might be closed for a
number of reasons including:
1. maintenance work in that section;
2. maintenance work in a pump valve in that section; and
- 50 -
3. draining a line connected to that header section (usually for maintenance purposes
either at the supplier end or the consumer ends).
Similarly, a line valve may be closed for a number of reasons that include:
1. maintenance work on the line; and
2. consumer requirement usually for making routine or emergency maintenance.
As there are many header sections and line valves connected to each group of pumps, the
reliability of the system can be dramatically affected by header section isolations or line
valve closures. This means that for each consumer, reliability depends on how many header
sections and pipelines are there for the supply of cooling water. If one section or a pipeline
is closed for maintenance, a reduced supply of water would still be available through other
sections or pipe lines, with less reliability, due to reduction in the flow of water supply. In
this study, these aspects of the system have been taken into consideration. However,
stoppages of water supply to the consumers due to their own maintenance and reliability
issues are not included, as they do not impact on the reliability of the system under study.
Figure 2.4 below shows the same consumer of figure 2.3 with a pipeline and a header
section that are isolated. A consequence of this is that the pumps attached to that header
section will not be operational. The black valves shown in figure 2.4 are closed valves. The
consumer in this case might still be operational but with a reduced reliability due to the
decreased number of pumps. This would illustrate the influence that header and/or pipeline
isolations would exert in the system.
The different combinations of section header and/or pipeline isolations that a consumer can
go through demanded that a reliability analysis must be made for each case. As each case
would have a different reliability than the other cases. For example, figure 2.5 below shows
the reliability of consumer C1 under different circumstances resulting from header and/or
pipeline isolations.
It is interesting to notice in C1-Case I that even at t = 0, its reliability R is 0.89. This is
unusual in reliability analysis (where at t = 0, usually R = 1). However, this occurs in the
- 51 -
case of the system studied here due to the effect of the conditional parameters. The
conditional parameter in this particular case is satisfied 89 % of the time, starting from the
initial time, and it is not satisfied in remaining 11 % of the time. Hence, the observed
behavior is obtained for the reliability at time 0. It can also be seen that the most reliable
case for operation is C1-Case III. In this case, all the header valves are open, all the header
sections are utilized and all the 6 pumps are available. This case is slightly better than C1-
Case II where only 5 of the 6 pumps are available.
Consumer
Figure2.4 the same consumer after closing the middle-header valve and a line valve.
Some consumers are supplied by more than one line (like the one in figures 2.3 and 2.4
above) while other consumers share a single pipeline that supplies all of them. In the later
- 52 -
case, section header and/or pipeline isolations will also have an influence in the final
reliability of the system.
Figure 2.5 the reliability of consumer C1 under different circumstances.
2.3.2 Data Collection
The operation and maintenance logbooks of the pumping station were examined between
the years 1997-2002. The aim was to develop a reliability model of the pumps based on the
time to failure (TTF) distributions of their respective components, such as the traveling
screen, the main electric motor, the mechanical pump and the discharge valve. However,
because of the scarcity of components failure, no useful conclusions could be drawn.
Therefore, each pump was considered as one component in the overall reliability model for
the respective station. The actual failure data for all pumps are shown in table 2.1.
Table 2.1 Failure rates of pumps.
Group of Pumps Pumps 1-6 Pumps 7-12 Pumps 13-16
Failure rate per hour 1.542 x 10-4 1.6092 x 10-4 3.164056 x 10-4
Based on preliminary studies of the related data and histogram plotting, it was concluded
that the equipment failures followed the exponential distribution. As for the theoretical
basis for choosing the exponential distribution, Billinton and Allan (Billinton and Allan
1983) wrote that
- 53 -
“The exponential, or strictly the negative exponential, distribution is probably the most
wildly known and used distribution in reliability evaluation of systems. The most important
factor for it to be applicable is that the hazard rate should be constant, in which case it is
defined as the failure rate λ…In practice, the negative exponential distribution has a much
wider degree of significance than just that of first failure and is extensively used in the
analysis of repairable systems in which the components cycle between operating or up
states and failure or down states (which is the case in the pumping station’s reliability
analysis)…It is frequently used in system reliability evaluation problems without
substantiating that the failure rate is constant or independent of time. There are usually
three justifications made for this:
1. First, the analytical techniques, particularly for large systems, are very
complex unless simplifications are made. In this case the assumption of constant
failure rates and the application of the exponential distribution considerably
simplify the problem.
2. Second, the data used in the evaluation exercise is often very limited and
insufficient to verify the correct underlying distribution. Consequently, it is
argued that it is unrealistic to use a technique more complicated than the data
justifies.
3. Third, it can be shown (Section 12.6 of the book) that if the concern is only
with limiting state values of system probability then the underlying distribution
loses its significance and the results are identical whatever distribution is used.”
(Billinton, R. and Allan, R. Reliability Evaluation of Engineering Systems: Concepts and
Techniques. Pages 149-150)
The failure data on headers and pipelines were very scarce, in the period from 1992 to
2002. Table2.2 shows the actual failures of the pipe lines, where question mark “?”
represents an unknown date of failure and operating hour. The reciprocals of time between
failures (TBF) or mean time between failures (MTBF) of pipelines are calculated as failure
rates in table 2.3. However, they are based on very few data points, sometimes a single
piece of data. For example, the only header failure recorded since 1992 was on header
section H7 and it is MTBF is 51,312 hours.
- 54 -
In addition, some lines never failed during this period, where “X” was put in tables 2.2 and
2.3 for their TBFs and reciprocals.
Table 2.2 Actual records of line failures (operating hours and TBF is from: 1/1/1992)
Table 2.3 Reciprocals of Mean Time Between Failures (MTBF) for each line (failure rates).
With many unexciting failure rates, the reliability calculation and analysis could not be
performed. A method had to be available for overcoming this obstacle. The literature was
first consulted. Several works on pipes were found. For instance, Al-Dakheel (Al-Dakheel
2004) explained the reasons for seawater pipe failure and suggested alternatives for the
classical “patch or replace” solutions. Wang et al. (Wang, Dong et al. 1993) and Khulief
and Emara-Shabiak (Khulief and Emara-Shabiak 2004) also presented experimental
methods for online leak detection of pipelines. All works, however, did not deal with
calculating the failure rate.
Failure Line
Date of Failure
OperatingHour
Date of Repair
OperatingHour TBF Notes
1 1a 24/01/1999 61,920 25/01/1999 61,944 61,920
A puncture was noticed in the line just after valve 4E2.
2 2a 10/08/2000 75,456 07/01/2002 87,816 75,456 A puncture was noticed in the line.
3 3 28/09/2000 76,632 01/10/2000 76,704 76,632 A puncture was noticed in the line.
4 3Ab ? ? 22/03/1992 1,944 1,944 TBF based on fixing date. 5 3Ab ? ? 18/12/1998 61,032 59,088 TBF based on fixing date. 6 3Ab ? ? 22/12/2001 87,432 26,400 TBF based on fixing date. 7 3Aa ? ? 18/12/1998 61,032 61,032 TBF based on fixing date. 8 3Aa ? ? 05/02/2000 70,968 9,936 TBF based on fixing date. 9 4 ? ? 23/01/2000 70,656 70,656 TBF based on fixing date.
10 4 ? ? 07/01/2001 79,056 8,400 TBF based on fixing date.
11 5 31/10/2000 77,424 01/11/2000 77,448 77,424 A puncture was noticed in the line.
12 5 ? ? 03/02/2001 79,704 2,256 TBF based on fixing date.
13 11a 30/06/1999 65,688 30/06/1999 65,688 65,688 A puncture was noticed in the line.
Line MTBF Reciprocal Line MTBF Reciprocal 1 61,920 1.61499E-05 7 x x 2 75,456 1.32528E-05 8 x x 3 39,172 2.55284E-05 9 x x 4 39,528 2.52985E-05 10 x x 5 39,840 2.51004E-05 11 65,688 1.5E-05 6 x x
- 55 -
Nakamura developed a method that can be used when there are few available data.
Nakamura’s method of dimensional reduction was used first with submarine electrical
cables (Nakamura, Nanayakkara et al. 1992) and later was used to determine the
maintenance scheduling for pump systems in thermal power stations (Nakamura 2001).
Still, it was not possible to draw a meaningful failure rate from the available data even by
using the method of dimensional reduction as applied by Nakamura et al. A method
developed by Thomas (Thomas 1981) was helpful in providing a path to the solution of the
problem determining the reliability of pipelines.
Thomas (Thomas 1981) paper dealt specifically with pipe and pressure vessel failures. The
Thomas method is an approximation strategy in order to estimate failure probability for
leakage and rupture of pipelines and pressure vessels. Leak is defined by Thomas as fluid
going through the wall of the pipe or vessel. Rupture is catastrophic leakage. Subsequently,
rupture is a subset of leakage, according to Thomas. Thomas mentions that it is estimated
that 5 per cent of leakages are ruptures. Both leakage and rupture would require stopping
seawater to the consumer in order to make the necessary repairs. Therefore, in this study,
we are only interested in the probability of leakage. Lydell (Lydell 2000) showed concerns about the Thomas approach, preferring a newly
developed SKI-PIPE database on piping failures in commercial nuclear power plants.
Lydell suggested using the Thomas approach with caution and only if actual, current failure
statistics, cannot be accessed. As will be shown later, the authors agrees with Lydell with
regard to the cautious use of the Thomas method. Neither Thomas approach nor the SKI-
PIPE database explicitly includes seawater as the process medium. For example, data in
Lydell’s paper, which is taken from the SKI-PIPE database, has the piping failure event
populations for seven process media not including seawater. It is not clear whether the
eighth medium called “Others” does include seawater or not. Owing to lack of access to
SKI-PIPE database, Thomas approach will be followed in this work with some
modifications, as shall be explained later.
- 56 -
Another concern was that both papers were not clear on how to deal with composite
thicknesses, i.e. thicknesses that are made of layers of different materials. For instance, a
header thickness usually consists of a 5mm rubber lining and 16mm steal casing. A line
thickness may include a cement mortar lining, a steal casing and a cement encasement. The
approach taken in this study is to consider the total thickness as the sum of all the
thicknesses despite the fact that different materials would have different strengths and
resistances to liquid penetration. An alternative for the Thomas approach might be a
method developed by Jarrett et al.(Jarrett, Hussain et al. 2002), (Jarrett, Hussain et al. 2003)
and (Jarret, Hussain et al. 201) from municipal water background. Equation (1) is taken
from Thomas (1981) to calculate the failure rate of pipelines; the notations are from Lydell
(2000). The calculation results for different headers and line sections are shown in table
2.4.
F Tot Base EQ FB (2.1)
Where λF-Tot is plant specific total leakage frequency; λBase= 1 x 10-8 failure/year =1.142 x
10-12 failure/h, a constant value; QE is change in reliability by piping size and shape
differences; F is plant age factor; B is the design learning curve, which is ignored in the
calculations.
50E P WQ Q Q (2.2)
2PDQ Lt
(2.3)
1.75WDQ Nt
(2.4)
Where L is the length of the pipe; D is the pipe diameter; t is the pipe wall thickness;
and N is the number of circumferential welds in the piping system
- 57 -
The calculated results, shown in table 2.4, seem to be much more optimistic than the truth,
assuming lower failure rates than what is actually happening. The reason for this
exaggeration might be due to the method adopted in calculating the reliability of the pipes.
Table 2.4 Failure rates for line and header sections calculated by equation (1)
Line or Header Section
Failure Rate (failure/ hr)
Line or Header Section
Failure Rate (failure/ hr)
Line or Header Section
Failure Rate (failure/ hr)
Line or Header Section
Failure Rate (failure/ hr)
H1 3.33 E-07 L3 4.45E-07 H6 1.45E-07 L 11a 1.00E-07 H2 1.60E-07 3Aa 3.30E-08 H7 2.40E-07 L11b 5.10E-08 H3 7.71E-08 3Ab 6.00E-09 L9 2.60E-07 H5 4.00E-07 L1a 4.00E-08 3Ba 2.00E-09 L10 2.60E-07 L4 4.85E-07 L2a 4.00E-08 3Bb 6.60E-08 L6 4.00E-09 L5 4.85E-07 L1b 8.70E-08 3Ca 4.75E-08 L7 4.00E-09 L2b 8.70E-08 H4 3.96E-07 L8 4.00E-09
2.3.3 Data Modification
The problems exhibited by some categories of the data in this research demanded their
modification. Data modification is, sometimes, practiced in reliability engineering.
Practicing it in some instances might be a sign of good engineering. Moon et al (Moon et
al. 1998) when describing an experienced reliability engineer has written the following
about him “Through system development and years of experience in using the system, he
has developed a ‘second sense’ about the data that is not readily apparent to others. In
particular, data selection, data modification and parameter adjustments depend upon his
judgment and experience”. Examples of failure rate modification include the work done by
Martorell et al (Martorell et al 2010). In that work, the failure rate of a safety device was
lowered ten times to represent a more reliable component. The American military standard
MIL-HDBK217F (DoD, 1992) uses the following equation to modify the failure rate of an
electronic component λp=λb.π. Where λp is the part failure rate, λb is the base failure rate
and π modifies the base failure rate for a variety of parameters including environmental
conditions.
The modification process for each category is explained in the following sub sections.
- 58 -
2.3.3.1 Failure Rate for the Pipe Section.
To overcome the shortcomings of equation (1), this candidate developed a modified
Thomas approach as explained here. For the pipe sections that failed several times, the
average MTBF was taken and its reciprocal was considered a constant failure rate for that
entire pipe section. Second, for pipe sections that failed only once, the reciprocal of TTF
was taken and considered a constant failure rate. Finally, for the pipe sections that never
failed the Thomas approach was used but with modified λBase. It was assumed that since the
Thomas paper came from nuclear reactor background, the pipes used in the study were of
extreme high reliability, which may not be required or used in other industries. Therefore,
two new λBase were calculated: one for the pipes and the other for the headers. λBase for the
pipes was calculated in reverse for each failing pipe. Next, the average of those λBase was
taken and considered in calculation of the failure rate for the pipes that never failed. The
new λBase for the pipes is 1.07 x 10-9 failure/h.
2.3.3.2 Failure Rate for the Header Section
The headers, due to their construction had a much lower failure rate than pipes. It was
deemed appropriate, hence, to calculate λBas for headers independent of the pipes. The
reciprocal for the TTF of header section 7 (H7) was considered as its failure rate and the
header’s λBas was calculated as 5.7 x 10-11 failure/h. The new results are shown in table 2.5.
Table 2.5 Failure rates for line or header sections, actual and a modified Equation (1).
Line Lambda Line Lambda Line Lambda
1 1.60E-05 8 3.80E-06 H4 1.84E-05 2 1.30E-05 9 2.43E-04 H5 1.87E--5 3 2.50E-05 10 2.43E-04 H6 7.30E-06 4 2.53E-05 11 1.50E-05 H7 1.95E-05 5 2.51E-05 H1 1.38E-05 Dark cells represent 6 3.80E-06 H2 9.00E-06 actual failures. 7 3.80E-06 H3 3.20E-06
2.3.3.3 Failure Rate for the Valve Section
The TBF of system valves that failed from 1 January 1992 until 31 December 2002 are
shown in table 2.6.
- 59 -
Table 2.6 Failure rates for failed system valves.
System Valve 4G4 4N2 4N3 4B5 4H1 4H2 4H3 TTF/MTBF(hr) 79,152 79,272 79,128 45,600 24,240 17,952 7,176
Failure Rate 1.26339E-05 1.261E-05 1.26E-05 2E-05 4E-05 6E-05 0.0001
All valves failed once except 4H1 which failed 3 times. Other valves did not fail at all from
1 January 1992 until 31 December 2002. There were no methods in the literature to
calculate the valve reliabilities in this situation. To overcome this difficulty, the following
approach was used:
For valve 4H1 which failed three times, the MTBF was taken. The reciprocal was
considered a constant failure rate as in table 2.6
For the valves that failed only once, the TTF was registered and its reciprocal was
considered a constant failure rate, this is shown in table 2.6.
For the remaining system valves that did not fail since 1 January 1992, it was
decided to consider the time of failure as the last day of the observation, i.e. 31
December 2002. This is a conservative way of estimating failure rates and thus the
reliability. The TTF (TTF = 87,600 h) was registered and its reciprocal was
considered a constant failure rate of 1.141 x 10-5 failure/h for the valves. This is
applied to all the valves except the ones with the prefix TP and 4B6 installed in
1997 with TTF of 52,560 h and failure rate of 1.9 x 10-5 failure/h. Table 2.7 shows
these failure rates.
There was no record of valves failing close or open. The records only mention that
the valves failed. It has been decided to consider the failure rate as satisfying for the
conditions including “Operating” “failed close” “failed open”.
- 60 -
Table 2.7 Failure Rates for the system valves that did not fail.
Valve Name
Failure Rate
(Failure/ hour)
Valve Name
Failure Rate
(Failure/ hour)
Valve Name
Failure Rate
(Failure/ hour)
Valve Name
Failure Rate
(Failure/ hour)
4G1 1.14E-05 4D3 1.14E-05 TP1B 1.90E-05 4M1 1.14E-054G2 1.14E-05 4J1 1.14E-05 TP2A 1.90E-05 4M2 1.14E-054E1 1.14E-05 4D4 1.14E-05 TP2B 1.90E-05 4M3 1.14E-054E2 1.14E-05 4D3 1.14E-05 TP3A 1.90E-05 4M4 1.14E-054E3 1.14E-05 4J1 1.14E-05 TP3B 1.90E-05 4B3 1.14E-054E4 1.14E-05 4I1 1.14E-05 4G3 1.14E-05 4B4 1.14E-054N1 1.14E-05 TP1A 1.90E-05 4L1 1.14E-05 4B6 1.90E-05
2.3.4 Reliability Calculations
A schematic diagram showing the whole system is shown in figure 2.6. Owing to space
limitations, only the names of the major components are shown. The system is divided into
three distinct groups:
(1) Group PIC. The schematic diagram of this group is shown in figure 2.7.
(2) Group EQUAT. The schematic diagram of this group is shown in figure 2.8.
(3) Group MAR. The schematic diagram of this group is shown in figure 2.9.
The water supply to each consumer can be in different operational cases, which are
mutually exclusive. In other words, the system can be in only one of the cases at a given
time.
- 61 -
Figure 2.6 the entire pumping station divided into its major components for the reliability analysis and calculations.
- 62 -
Figure 2.7 a schematic diagram of pump group PIC (pumps 1-6 and their associated headers, lines and valves)
- 63 -
Figure 2.8 a schematic diagram of pump group Equate (pumps 7-12 and their associated headers and lines)
- 64 -
Figure 2.9 a schematic diagram of pump group MAR (pumps 13-16 and their associated headers and lines)
- 65 -
In order to analyze supply reliability, one has to consider all possible cases as seen in table
2.8. Analysis of the reliability of each consumer is done for each different case. This in turn
was represented in diagrams and equations. The system equations for each consumer in
each case are also tabulated and presented in table 2.9. Duplicated cases and those leading
to zero reliability were not listed in either table, which explains why the case numbers are
not continuous. The following examples illustrate the method of analysis and calculation
for some selected cases.
Example 1: C1-Case I (if valve 4G1 fails to open).
The following seven requirements must be met for the supply to arrive at consumer C1:
(1) At least: P1 and V1; or P2 and V2; or P3 and V3 must be operational.
(2) Line 1b must be operational.
(3) And valve 4E3 (operating or failed open).
(4) And line 1a must be operational.
(5) And valve 4E1 (operating or failed open).
(6) And header H1 must be operational.
(7) And consumer C1 must utilize at least 10,500 m3/h of water (minimum production)
(figure 2.10).
Figure 2.10 reliability block diagram of C1-caseI.
- 66 -
Table 2.8 Description of every case for the consumers
Consumer/Case Case Description C1-Case I
4G1 is closed.
C1-Case II
4G1 is open but 4G2 is closed
C1-Case III
4G1 is open and valve 4G2 is also open
C2-Case I
4G1 and 4G2 are closed
C2-Case II
4G1is closed and 4G2 is open
C2-Case III
4G1is open and 4G2 is open
C3- Case I
4G1 open, 4G2 open, 4N1 is open
C3- Case II
4G1 closed, 4G2 open, 4N1 is open
C3- Case III-1 4G1 closed, 4G2 closed, 4N1 is open -C4 and C5 are operating.
C3-Case III-2
4G1 closed, 4G2 closed, 4N1 is open- C4 operating C5 not operating.
C3- Case III-3
4G1 closed, 4G2 closed, 4N1 is open -C4 not operating C5 operating
C3-Case III-4
4G1 closed, 4G2 closed, 4N1 is open- C4 not operating C5 not operating
C4- Case I
4G1 open, 4G2 open, 4N1 is open
C4-Case II 4G1 closed, 4G2 open, 4N1 is open C4-Case III-1
4G1 closed, 4G2 closed, 4N1 is open -C3 and C5 are operating.
C4-Case III-2
4G1 closed, 4G2 closed, 4N1 is open-C3 operating and C5 is not
C4-Case III-3
4G1 closed, 4G2 closed, 4N1 is open -C3 not operating and C5 operating
C4-Case III-4 4G1 closed, 4G2 closed, 4N1 is open-C3 and C5 are not operating
- 67 -
Consumer/Case Case Description C5-Case I
4G1 open, 4G2 open, 4N1 is open
C5-Case II
4G1 closed, 4G2 open, 4N1 is open
C5-Case III-1
4G1 closed, 4G2 closed, 4N1 is open C3 and C4are operating.
C5-Case III-2
4G1 closed, 4G2 closed, 4N1 is open-C3 operating and C4 is not
C5-Case III-3
4G1 closed, 4G2 closed, 4N1 is open-C3 not operating and C4 is operating.
C6-Case I
line L4 closed, valve 4G4 closed and line L5 closed
C6-Case II
line L4 open, line L5 closed and valve 4G4 closed
C6-Case III line L4 closed, line L5 opened and valve 4G4 closed C6-Case IV line L4 open, line L5 open and valve 4G4 closed
C6-Case VII line L4 closed, line L5 open and valve 4G4 open
C6-Case IX
line L4 closed, line L5 closed and valve 4G4 open (Water well not go to C6. The reliability is zero)
C6-Case X line L4 open, line L5 closed and valve 4G4 open
C6-Case XII
line L4 open, line L5 open and valve 4G4 open
C7-Case II
Valve 4L1 closed, line L9 closed, and line L10 open
C7-Case III
Valve 4L1 closed, line L9 open, and line L10 closed
C7-Case IV
Valve 4L1 closed, line L9 open, and line L10 open
C7-Case VII
Valve 4L1 open, line L9 open, and line L10 closed
C7-Case X
Valve 4L1 open, line L9 close, and line L10 open
C7-Case XII
Valve 4L1 open, line L9 open, and line L10 open
- 68 -
Table 2.9 System reliability equations for different consumers under different scenarios.
Consumer / Case
Subsystem Reliability Equation
C1-Case I R s = 0.89 e-6.86 x10^-5 t (1-(1-e-1.45x10^-4 t)3 ) C1-Case II R s = e-6.86 x10^-5 t (1- ( (1-e-1.45x10^-4 t)3 (1-e-1.54x10^-4 t)2 ) ) C1-Case III R s = e-6.86 x10^-5 t (1- ( (1-e-1.45x10^-4 t)3 (1-e-1.54x10^-4 t)2
(1-e-1.48x10^-4 t) ) ) C2-Case I R s = 0.77 e-5.8x10^-4 t (1-(1-e-1.45x10^-4 t)2 ) C2-Case II R s = 0.77 e-5.8x10^-4 t (1- ( (1-e-1.45x10^-4 t)2 (1-e-1.48x10^-4 t) ) ) C2-Case III R s = 0.77 e-5.8x10^-4 t (1- ( (1-e-1.45x10^-4 t)2 (1-e-1.6x10^-4 t)3
(1-e-1.48x10^-4 t) ) ) C3- Case I R s = e-6x10^-5 t (1-( (1-e-2.64x10^-4 t) (1-e-3x10^-4 t) ) ) (1- ( ( 1-
(e-1.38x10^-5 t (1-(1-e-1.45x10^-4 t)3) ) ) (1-e-9x10^-6 t (1- (1-e-1.45x10^-4 t)3 ) ) ) (1-e-1.45x10^-4 t) ) )
C3- Case II R s = e-6x10^-5 t (1- ( (1-e-9x10^-6 t (1- (1-e-1.45x10^-4 t)2 ) )
(1-e-1.45x10^-4 t) ) ) (1- ( (1-e-2.64x10^-4 t) (1- e-3x10^-4 t ) ) ) C3- Case III-1
R s = e-3.35x10^-4 t
C3-Case III-2 R s = e-2x10^-4 t (1- ( ( 1-e-1.6x10^-4 t) (1- e-1.73x10^-4 t ) ) ) C3- Case III-3
R s = e-2x10^-4 t (1- ( ( 1-e-1.6x10^-4 t) (1- e-1.73x10^-4 t ) ) )
C3-Case III-4 R s = e-2x10^-4 t (1- ( ( 1-e-1.6x10^-4 t) (1- e-1.73x10^-4 t ) ) ) C4- Case I R s = e-6x10^-5 t ( 1- ( ( 1-e-1.38x10^-5 t (1- (1-e-1.45x10^-4 t)3 ) )
(1- e-9x10^-6 t (1- (1-e-1.45x10^-4 t)2 ) ) (1-e-1.45x10^-4 t) ) ) (1- ( ( 1-e-7.3x10^-5 t) (1- e-1x10^-4 t) ) )
C4-Case II R s = e-6x10^-5 t ( 1- ( (1- e-9x10^-6 t (1- (1-e-1.45x10^-4 t)2 ))- (1-e-1.45x10^-4 t) ) ) (1- ( ( 1-e-7.3x10^-5 t) (1- e-1x10^-4 t) ))
C4-Case III-1
R s = e-2x10^-4 t (1- ( ( 1-e-7.3x10^-5 t) (1- e-1x10^-4 t) ) )
C4-Case III-2 R s = e-2x10^-4 t (1- ( ( 1-e-7.3x10^-5 t) (1- e-1x10^-4 t) ) ) C4-Case III-3 R s = e-2x10^-4 t (1- ( ( 1-e-7.3x10^-5 t) (1- e-1x10^-4 t) ) )
- 69 -
Consumer / Case
Subsystem Reliability Equation
C4-Case III-4 R s = e-2x10^-4 t (1- ( ( 1-e-7.3x10^-5 t) (1- e-1x10^-4 t) ) ) C5-Case I R s = e-6x10^-5 t ( 1- (1- e-9.6x10^-5 t)2) (1- ( (1-e-1.38x10^-5 t
(1-e-1.45x10^-4 t)3 ) (1-e-9x10^-6 t (1-e-1.45x10^-4 t)2 ) (1-e-1.45x10^-4 t) ) )
C5-Case II R s = e-6x10^-5 t ( 1- (1- e-9.6x10^-5 t)2) (1-( (1-e-9x10^-6 t
(1-e-1.45x10^-4 t)2 ) (1-e-1.45x10^-4 t) ) ) C5-Case III-1 R s = e-2.05x10^-4 t ( 1- (1- e-9.6x10^-5 t)2 ) C5-Case III-2 R s = e-2.05x10^-4 t ( 1- (1- e-9.6x10^-5 t)2 ) C5-Case III-3 R s = e-2.05x10^-4 t ( 1- (1- e-9.6x10^-5 t)2 ) C6-Case II R s = 0.1212 e-1.82x10^-4 t (1- (1-e-1.61x10^-4t)3 )
C6-Case III R s =0.1212 e-1.82x10^-4 t (1- (1-e-1.61x10^-4t)3 ) C6-Case IV R s =0.1212 e-1.14x10^-5 t ( 1- ( 1- ( e-1.13x10^-4 t
(1- (1-e-1.61x10^-4 t)3 ) ) ) (1-e-2x10^-4 t) (1- (1-e-1.61x10^-4t)3) ) ) )
C6-Case VII R s =0.1212 e-2.6x10^-4 t(1- (1-e-1.61x10^-4t)3 ) C6-Case X R s =0.1212 e-1.82x10^-4 t (1-( (1-e-1.87x10^-5 t (1- (1-e-1.61x10^-4t)3 ) )
(1-(1- (1-e-1.61x10^-4t)3 ) ) ) C6-Case XII R s = e-2.4x10^-5 t ( 1- (1- e-1.9x10^-5 t )2 )3 (1- ( ( 1- e-5.6x10^-5 t
(1- (1-e-1.61x10^-4t)3 ) ) ( 1 – e-1.45x10^-4 t ( 1- ( 1- e-1.61x10^-4 t )3 ) ) ) )
C7-Case II R s = 0.986 e-5.3x10^-4 t ( 1- ( 1- e-6.32x10^-4 t )3 ) C7-Case III R s = 0.986 e-5.3x10^-4 t ( 1- ( 1- e-6.32x10^-4 t )3 ) C7-Case IV
R s = e-1.95x10^-5 t ( 1- ( 1- e-6.32x10^-4 t )3 ) ( 1- ( 1- e-5.09x10^-4 t )2 )
C7-Case VII R s = 0.986 e-5.3x10^-4 t ( 1- ( 1 - ( 1- ( 1- e-6.32x10^-4 t )3 )
( 1 – ( 1- e-6.4x10^-4 t )3 ) ) ) C7-Case X
R s = 0.986 e-5.3x10^-4 t ( 1- ( 1 - ( 1- ( 1- e-6.32x10^-4 t )3 ) ( 1 – ( 1- e-6.4x10^-4 t )3 ) ) )
C7-Case XII R s = e-1.95x10^-5 t ( 1- ( 1 - ( 1- ( 1- e-6.32x10^-4 t )3 )
( 1 – ( 1- e-6.4x10^-4 t )3 ) ) ) ( 1- ( 1- e-5.09x10^-4 t )2 )
- 70 -
The equations of this subsystem are:
Ro=1-(1-RpRv)3 (2.5)
Rs= Ro RH1R4E1R1aR4E3R1B P( C1 consumption >10,500 m3/hr) (2.6)
P (C1 consumption > 10,500 m3/hr)= 0.887323944 (from past production data)
Note that because we consider the valve, motor, screen and pump as constituents of one
unit called the main pump, then RpRv= Rp. The final reliability equation of this case is as
follows:
Rs(t) = 0.89 Exp[-6.83x10-5t](1-(1-Exp[-1.54x10-4t])3) (2.7)
Example 2:C6-Case IV (line L4 open, line L5 open and valve 4G4 closed) For consumer C6 to operate in this case, it must have all of the following components
operating as a minimum (see fig. 2.11):
1. Line sections 6, 7, and 8
2. TP1A or TP1B
3. TP2A or TP2B
4. TP3A or TP3B
5. L4 and L5
6. H4 and H5
7. Bypass (4B6 and 11a and 11b and 4H1) as shown in figure 2.12
8. Flow passing through L4 and L5 each is not more than 41,000 m3
9. Consumption is not less than that required by the plant
There is no data on single line flow. Thus, point 8 can not be calculated. However, because
the terminal points are connected to different units within that plant, one can not assume
that each line is going to deliver the same amount of water. To be on the safe side, it is
prudent to assume that the total consumption of C6 should not exceed 41,000 m3/hr which
is the maximum amount of water a single line (either L4 or L5) can withstand.
- 71 -
Figure 2.11.Block diagram of C6 -case IV
- 72 -
Figure 2.12 Block diagram of the bypass system.
The equations for C6-case IV are:
Ro1 = 1-(1-Rp) 3 (2.8)
Rx= Ro1 R(H4) R(4N2) R(L4) R(TP1B) R(TP2B) R(TP3B) (2.9)
Ro2= 1-(1-Rp) 3 (2.10)
Ry= Ro2 R(H5) {R(4B6) R(11a) R(11b) R(4H1)} R(4N3) R(L5) R(TP1A) R(TP2A) R(TP3A) (2.11)
Rs= [1-((1-Rx) (1-Ry))] R (L6) R (L7) R (L8) P (C6 consumption < 41,000 m 3/hr) (2.12)
P (C6 consumption < 41,000 m 3/hr) = 0.121212 (From past production data)
The final reliability equation for this case is obtained as follows:
Rs(t) = 0.121212 exp[-1.14x10-5t] (1- (1-(1-exp[-1.13x10-4t] (1-(1-exp[-1.61x10-4t)3) )
(1- exp[-2x10-4t] (1-(1-Exp[-1.61x10-4t)3)))) (2.13)
Example 3 C7-Case III (Valve 4L1 closed, line L9 open, and line L10 closed)
For consumer C7 to operate in this case, it must have all of the following components
operating as a minimum:
1. 9b
2. 4M3(operating or failed open)
3. 9a
4. 4M1 (operating or failed open)
5. H7
Bypass System
4B6 11a 11b 4H1
- 73 -
6. (P14 and V15 and P15 and V15) or (P14 and V14 and P16 and V16) or (P15
and V15 and P16 and V16)
7. Consumption is less than 20,000 m3/hr.
The block diagram for this case is shown in Figure 2.13.
The equations for C7-case III are:
Ro1= Rp 2 (2.14)
Ro= 1-(1-Ro1) 3 (2.15)
Rs= Ro R (H7) R (4M1) R (4M3) R9 P (C7 consumption < 20,000 m3/hr) (2.16)
P (C7 consumption < 20,000) = 0.985915
Reliability equation is: Rs(t) = 0.985915 exp[-5.3x10-4t] (1-(1-exp[-6.32x10-4t])3) (2.17)
2.3.5 The Reliability of Water Delivery to a Consumer in All Cases
The probability that each consumer receives the water supply within a specified period of
time is the reliability of the subsystem related to that particular consumer for that period
and the specified operational condition. The operation manager can give an estimate of
conditional subsystem reliability for each consumer under each operational condition or
case. For example, for consumer 1 to receive supply under the subsystem condition of case
I, the supply reliability would be 0.867356 for a period of 1,000 h calculated using the
related equation for C1-Case I. Figure 2.14 shows the reliability of C1 under each case.
C1-Case I can be seen in figure 2.14 as the case with the least reliability. It is interesting to
notice that even at t = 0, its reliability R is 0.89. This is unusual in reliability analysis
(where at t =0, usually R =1). However, it occurs in the case of the system studied here due
to the effect of the conditional parameters. The conditional parameter in this particular case
is satisfied 89 per cent of the time, starting from the initial time, and it is not satisfied in
remaining 11 per cent of the time. Hence, the observed behavior is obtained for the
reliability at time = 0.
- 74 -
Figure 2.13 block diagram of C7-Case III
- 75 -
C1 Reliability Under Different Operational Conditions
0
0.2
0.4
0.6
0.8
1
0 5000 10000 15000 20000 25000 30000 35000 40000
t
Rs
C1-Case I C1-Case II C1-Case III
Figure 2.14 C1 reliability behaviors under different operational conditions
It can also be seen that the most reliable case for operation is C1-Case III. In this case, all
the header valves are open, all the header sections are utilized and all the 6 pumps are
available. This case is slightly better than C1-Case II where only 5 of the 6 pumps are
available.
Full reliability over a period of t hours for consumer 1 under any operational condition can
be calculated by the following approach provided that related probabilities of each
operational condition are available.
Rc1= Pcase1Rcase1+Pcase2Rcase2+Pcase3Rcase3 (2.18)
Assume that all the cases have an equal opportunity of occurring (which is an unrealistic
assumption but sufficient for illustration of calculations). We can calculate the reliability of
C1 as follows:
Rc1=(0.33)( 0.867356)+(0.33)( 0.976925)+(0.33)( 0.976963) = 0.931008
Similarly, reliability calculations are made for all seven consumers with seven subsystems.
- 76 -
2.3.6 Considering the Reliability of All the Consumers and the Entire Pumping
Station
The reliability for each of the seven subsystems has been considered, in this study, as the
reliability of cooling seawater arriving to the consumer at the required pressure and flow
rate while observing the operational constraints on the subsystem. At any one time,
therefore, there are seven reliability indices corresponding to the seven subsystems (plants).
The reliability of a system with many outputs has been discussed, for example in
Nakashima and Yamoto (Nakashima and Yamoto 1984), Ramamoorty and Gupta
(Ramamoorty and Gupta 1976) , Yin and Silio (Yin and Silio 1994) and Abulma’atti and
Qamber (Abulma’atti and Qamber 2001). In this thesis, it is suggested that the reliability of
the entire pumping station needs a different treatment from what is already available. The
reliability of the entire pumping station can be considered from three different points of
view :
1. A pumping station can be considered working only if all its consumers are
satisfied simultaneously. An implication of such a definition is that all outputs
should be thought of as connected in series.
2. A pumping station can be considered working only if n-out-of-7 consumers are
satisfied.
3. A pumping station can be considered working at all times (except for the case of
complete shutdown) and its reliability (or an indication of it) can be thought of
as the average percentage of reliability satisfaction of its individual consumers.
Proposition 1 does not apply because when a consumer is not satisfied (having a reliability
of zero as a result of seawater not reaching it with the desired pressure and flow rate) the
pumping station does not stop its production, i.e. it keeps “working”. If one or more of
consumers are having zero reliability, the reliability indices of the rest of the consumers
will not decrease at all. In fact the opposite might occur! Consider the following
hypothetical situation: there are four pumps working for group PIC, three pumps working
for group EQUATE and two pumps working for group MAR. Then, Line 3 (Figure 2.6) is
punctured, leaking significantly. To rectify the situation valve 4N1 must be closed
immediately. Consequently, consumers C3, C4 and C5 will not have seawater at all
- 77 -
(reliability for them is zero now). Because 3 consumers in group PIC are not taking
seawater, the pressure in the header and towards the remaining two consumers (C1 and C2)
will largely increase. To overcome this, at least one pump of the four pumps working for
group PIC should be turned off .Now, only three pumps are working for group PIC. As a
result, C1 and C2 would have three pumps as standby in comparison with two pumps in
standby before the accident. A consequence of having more standby pumps is the increase
of reliability for C1 and C2. Meanwhile, consumers C6 and C7 are not affected at all. Their
reliability would be the same as before and after the accident. Proposition 2 is difficult to administer as there is no practical way of setting a value for n.
A pre-requisite of n-out-of-k systems is that all the components of such a system must be
identical which is not the case here, where different consumers utilize different flow rates
of seawater at different pressures. Practically speaking also, each consumer (petrochemical
plant) has a different owner with contractual rights to get seawater.
Proposition 3 does not consider the reliability of the entire pumping station directly. Rather,
it defines the average percentage of consumer reliability satisfaction (APCRS) to give an
indication of the reliability of the pumping station. The (APCRS) can be defined as the
average reliability of the seven consumers multiplied by 100 (equation 2.19). This approach
provides a way of combining the reliabilities of seven independent subsystems that serve
seven independent consumers having different owners. Therefore, proposition 3 was
adopted for this study.
1
1 *100%n
Cii
APCRS Rn
(2.19)
where APCRS is the average percentage of consumer reliability satisfaction and Rci is
the reliability for each consumer from i =1,2,…,n.
In our case n = 7 because there seven consumers.
Example 4 suppose at time =t, the reliability for each consumer is the following C1=0.8, C2=0.97, C3=0.65, C4= 0.88, C5= 0.99, C6=0.91 and C7= 0.78
- 78 -
The average percentage of consumer reliability satisfaction (APCRS) would be APCRS= ((.8+.97+.65+.88+.99+.91+.78)/7) x100 = 85.4%
Although the above example demonstrates an “instantaneous” look at APCRS, it can be
calculated as a function of time over a specific time period for a given condition. The
following examples illustrate this.
Example 5 Consider the most common status of the system in which all the line valves are
open and all the header sections are connected. This would be C1-Case III, C2-Case III, C3-
Case I, C4-Case I, C5-Case I, C6-Case XII and C7-Case XII. The reliability of each
consumer in his respective case over a period of 48,000 hours is shown bellow in figure
2.15. APCRS for this case is shown in figure 2.16.
Figure 2.15 the reliability of each of the seven consumers with their respective case
numbers over a period of 48 thousand hours.
00.10.20.30.40.50.60.70.80.9
1
0 10000 20000 30000 40000 50000 60000
t
R(t
)
C1-Case IIIC2-Case IIIC3-Case IC4-Case IC5-Case IC6-Case XIIC7-Case XII
- 79 -
Figure 2.16 the APCRS of the pumping station for the cases shown in figure 2.15.
Example 6 Consider another status of the system in which all the line valves are open and
the entire header sections are connected with exception of pump group Equate. Consumer
C6 would be in case II which means that line L4 is open, line L5 is closed and valve 4G4 is
closed. This would be C1-Case III, C2-Case III, C3-Case I, C4-Case I, C5-Case I, C6-
Case II and C7-Case XII. The reliability of each consumer in its respective case over a
period of 48,000 hours is shown above in figure 2.16 .APCRS for this case is shown in
figure 2.17.
APCRS
0
20
40
60
80
100
0 10000 20000 30000 40000 50000 60000
t
APCR
S
APCRS
- 80 -
00.10.20.30.40.50.60.70.80.9
1
0 10000 20000 30000 40000 50000 60000
t
R(t)
C1-Case IIIC2-Case IIIC3-Case IC4-Case IC5-Case IC6-Case XIIC7-Case XII
Figure 2.16 the reliability of each of the seven consumers with their respective case
numbers over a period of 48 thousand hours.
APCRS
0
20
40
60
80
100
0 10000 20000 30000 40000 50000 60000
t
APCR
S
APCRS
Figure 2.17 the APCRS of the pumping station for the cases shown in figure 2.16
- 81 -
Chapter 3 A Case Study of Minimizing the Conflict between
Operation and Maintenance.
3.1 Introduction
The previous chapter emphasized the importance of pumping station reliability and
explained a method for modeling its reliability. In order to achieve the required reliability
of a pumping station, maintenance must be performed on all of its equipment at regular
times. This, however, is rarely achieved with ease. The reason for this difficulty has to do
with the relationship between operation and maintenance within the pumping station. The
following paragraphs shall explain this issue more.
A very important issue in almost all plants is the coordination between production and
maintenance activities. Both of them are necessary: the operation of machinery would
produce revenue for the owner(s) while maintenance will keep these machines running.
Nevertheless, one function is usually performed at the expense of the other. If a machine is
stopped for maintenance, it is stopped from producing revenue. Similarly, if a machine is
operated continually without proper maintenance, it will eventually fail. Lack of
coordination, hence, results in degradation for both operation and maintenance. The
pumping station understudy suffered from a conflict between its production (operation) and
maintenance functions. The result was unreliable operation due to the failure of the
unmaintained machines and inconvenient maintenance that interrupted production.
Something had to be done to solve (or at least minimize) this problem. What this study
presents is a practical method of data analysis for minimizing the conflict between
operation and maintenance activities in this industrial facility.
One way of minimizing the conflict, it was thought, was making an optimal schedule for
the operation and maintenance of the pumps supplying cooling seawater. The making of
this optimal schedule required knowing the seawater demand pattern first. Next, the
maintenance function could be done around this demand. It was expected that the
preplanned maintenance would enhance the reliability of seawater supply to the customers
and would reduce the operation and maintenance costs of the stations.
- 82 -
To analyze the cooling seawater demand, it was imperative to identify the major factors
that produced it. The demand for seawater at the pump station, it was thought, depended on
two things: a plant’s aggregate production level (or capacity utilization) and the weather
conditions. Plant capacity utilization (CU) or production (P) were thought to be involved in
determining a plant’s seawater consumption because of the observation of lower
consumption during a plant's partial or total shutdown. The reason the weather was thought
of as a determining factor in the consumption of cooling seawater was the observation of
the variability of this consumption during the seasons: the cooler the weather, the less the
cooling seawater that is required by the petrochemical plants and vice versa. A more
thorough discussion on the factors influencing the demand variation will take place at a
later section.
The function of the analysis, then, was to express the seawater demand as a function of
these meteorological factors (seawater temperature, Ts, air temperature, Ta, and humidity,
H) and the aggregate production levels (P) or plant capacity utilizations (CU) of the users
(petrochemical plants).
The method of data analysis used for modeling the plant consumptions in this research was
regression analysis. The reasons for choosing linear regression as the primary method of
analysis shall be explained in a separate section. In addition for the reasons mentioned in
that section, regression analysis is a long-established method and is familiar to engineers.
This familiarity, it is hoped, will make engineers less reluctant in applying the method
developed in this research to other cooling pumping stations in the future. Regression
analysis involves identifying the relationship between a dependent variable and one or more
independent variables.
Refineries and petrochemical plants, like any business, plan in advance. They make yearly
plans that include the intended production or plant capacity utilization for each month of
the incoming year. It was intended to take these figures and insert them in the consumption
model of each plant to solve the optimization and scheduling problem. An obstacle
appeared here: If the plants production or capacity utilization can be provided by the plant
owners, there is no way of knowing the seawater temperature, Ts, the air temperature, Ta,
- 83 -
and humidity, H in advance i.e. they must be forecasted. This meant that in order to solve
the optimization and scheduling problem, a forecast problem must be solved first. To
achieve this requirement, there were two options: to hire the services of a meteorological
company or to find a method to (locally) predict the three factors. The second option was
chosen.
The weather is a complex phenomenon and developing a weather forecast system can be a
sophisticated task. It was thought, however, that developing a forecast system for only
some aspects of the weather may be more achievable. Only the relevant aspects of the
weather (which were the air temperature, Ta, and its humidity, H, in addition to the sea
temperature, Ts) would be forecasted. Accordingly, another goal of the research was to
find a method to predict the three input (meteorological) variables. To achieve this goal,
regression analysis and exponential smoothing were used to analyze and, then, predict the
three weather variables.
It was intended to use the predicted weather variables in addition to plant capacity
utilization (or production) to establish the demand for cooling seawater. It was intended
also to use this predicted demand to develop an optimal operation and maintenance
schedule for the pumps over the planning horizon. Nevertheless, before all of this could be
done, it was necessary to test the assumptions of the research against actual data.
Data were collected over 14 years, from 1991 to 2004. It has been decided to use the data
up to the end of the year 2003 for analysis and use the year 2004 to test this analysis.
3.2 Literature Review
The occasional conflict between operation and maintenance has been studied before and
the need for a sort of an optimal solution has been recognized. Cassady and Kutanoglu
(Cassady and Kutanoglu 2005) have written the following about production scheduling and
preventive maintenance (PM) planning:
“In practice, these activities are typically performed independently despite the clear
relationship that exists between them. PM activities take time that could otherwise be used
for production, but delaying PM for production may increase the probability of machine
- 84 -
failure. Hence, there are trade-offs and conflicts between PM planning and production
scheduling.”
In an attempt to resolve this conflict, Cassady and Kutanoglu (Cassady and Kutanoglu
2005) developed a mathematical model which incorporates production scheduling and
preventive maintenance planning for a single machine. Tam et al (Tam, Chan et al. 2006)
extended the work of Cassady and Kutanoglu on maintenance scheduling to optimize both
reliability and cost. They considered a multi-component system. Coudert et al (Coudert,
Grabot et al. 2002) also dealt with the conflicting relationship between production and
maintenance. They have suggested that the multi-agent paradigm may provide an
implementation framework allowing the modeling of the negotiation process between the
maintenance and production functions. They showed that fuzzy logic provided facilities for
modeling the degrees of freedom of the negotiation. To consider both the production and
maintenance requirements Brandolese et al (Brandolese, Franci et al. 1996) developed an
expert system to schedule the operation of parallel machines. Optimization methods have
been, also, used in the process industry to solve the conflict between maintenance and
production. Ashareti et al (Ashareti, Teelen et al. 1996) presented a mixed-integer linear
programming model to simultaneously plan preventive maintenance and production in the
process industry.
Sometimes, as in this research, the optimization problem would mandate solving a forecast
problem. Forecasting demand has been an important issue in the 20th century and the
current century. Lapide (Lapide 1997) in his review of the book Against the Gods: the
Remarkable Story of Risk by Peter Bernstein had also presented a short history of the
development of forecasting. Forecasting has been used extensively across many businesses
that included engineering and non-engineering fields. Work has been done on both short-
term and long-term forecasting. Also, many methods have been applied for achieving the
purpose of forecasting.
- 85 -
Forecasting had been, and still is, a subject of interest to engineering. Electrical power
engineering, for example, studied it for a long time. In this branch of engineering the
object of forecasting is usually the electric load. The basics of electrical power engineering
can be found in Crow et al (Crow, Gross et al. 2003) .A historical review of short-term
load forecasting (defined as the prediction of the system load over an interval ranging from
one hour to one week) up to 1987 can be found in Gross (Gross and Galiana 1987). Ho et al
(Ho, Hsu et al. 1992) designed a multilayer neural network with an adaptive learning
algorithm for the Taiwanese short-term load forecasting. Only the peak and valley loads
were forecasted. To speed up the convergence rate of the learning process, an adaptive
learning algorithm in which the momentum is automatically adapted in the training process
was presented. This algorithm, it was found, converged much faster than the conventional
algorithm making it more convenient. A different approach for forecasting short term load
was used by Smith (Smith 1989). His approach employed time series to predict the short
term electric demand. The method used optimally combined two Box-Jenkins ARIMA
models and a spectral decomposition model for prediction.
Electrical long term load forecasting has also received some attention. Parlos et al (Parlos,
Oufi et al. 1996) worked on this subject. They stated that long term load forecasting is
made to help in making long term investment decisions regarding the electric industry.
According to them, long term forecasting is radically different from short term forecasting
both theoretically and practically making a different treatment unavoidable. In their paper,
the development and testing of a hybrid intelligent long-term load forecasting system was
presented. The system was made of several neural networks forecasting blocks, genetic
algorithms for network architecture selection and optimization and fuzzy rules for forecast
combination.
The work in this thesis attempted to optimize pump operation and maintenance by making a
pump scheduling system. Many of the research work done on pump optimization and
scheduling have been about water distribution pumping stations. In such stations, a pump or
a group of pumps would occasionally run for a short period (an hour or so) to distribute the
water or fill a tank. Accordingly, many of the research work done on pump scheduling
- 86 -
concentrated on operating the pumps on low tariff times. According to Lopez-Ibanez et al
(Lopez-Ibanez, Parsad et al. 2005) “The main goal of the Pump Scheduling Problem is to
schedule the operation of N pumps over a time period, typically 24 hours, in such a way
that system constraints and boundary conditions are satisfied, while the operational cost is
minimized. The most important costs associated with the operation of pumps are electrical
and maintenance costs” (Lopez-Ibanez, Parsad et al. 2005).
A cooling pumping station for the petrochemical industry runs continuously. Thus, the
thought of shifting the operation of pumps to low tariff times is invalid. With the
continuous operation, only the number of pumps in this station changes. This change
depends, largely, on weather and the consuming plants’ capacity utilization. Consequently,
the goal of this research was exactly what Lopez-Ibanez et al have stated, except that the
scheduling period is 12 months.
Lopez-Ibanez et al (Lopez-Ibanez, Parsad et al. 2005) used genetic algorithms for optimal
pump scheduling. Their objectives were to minimize the electrical costs and the
maintenance expenses. Another work having even more objectives to optimize was done by
Lucken et al (Lucken, Baran et al. 2004) which “… proposes the use of parallel
asynchronous evolutionary algorithms as a tool to aid in solving an optimal pump-
scheduling problem. In particular, this work considers a pump-scheduling problem having
four objectives to be minimized: electrical energy cost, maintenance cost, maximum power
peak and level variation in a reservoir” (Lucken, Baran et al. 2004).
Multi-source, multi-storage tank water supply systems use a lot of pumps. Beckwith and
Wong (Beckwith and Wong 1995) developed a method for scheduling electric pumps in
such a system using a genetic algorithm (GA). The objective of their scheduling problem
was to ensure that the volume of water required by the water distribution system is
adequately provided by the pumps in the system whilst minimizing the cost incurred in the
use of electrical power by the pump motors in the system. Their algorithm took into
account the characteristic curves, the efficiency curves and the flow limits of pumps in the
system and the system characteristic curves.
- 87 -
Another work on optimizing water distribution networks was done by (Biscos, Mulholland
et al. 2003). Their paper presented an approach for the optimization of potable water
distribution networks. The optimization objectives were the minimization of energy costs
and chlorine injection. The method used was mixed integer non-linear programming,
MINLP.
Finally, work on the optimal scheduling of pumps in the mining industry was done by
(Grobler and Heijer 2006). Their work aimed at reducing the energy cost used by mine
pumps by scheduling their operation. The purpose of this scheduling was shifting the
electrical load (consumed by the pumps) out of the critical peak period (the period with the
higher tariff) to other times of the day. The core of this work, however, was about selecting
a representative data period for the development of a baseline. Using data reaching too far
back in the past led to having energy trends that are no longer present. On the other hand,
using too recent data may not reflect of the true operation of the system. The authors wrote
"...the purpose of a baseline is to give a true representation of the operational characteristics
of a system...”.
In this thesis, regression analysis was used for modeling plant seawater consumption and
some weather factors. Regression analysis was first described by Francis Galton (Pearson
1930). The term regression was used for the first time by Galton (Galton 1886), who
thought that he found out a law that ''gives the numerical value of the regression towards
mediocrity in the case of human stature’’ (Galton 1886). Regression analysis has been used
before to model production and/or consumption of both organic and manufactured systems.
In organic systems, for example, Krakacier et al (Krakacier, Goktolga et al. 2006) reported
the results of a regression analysis of the relationship between energy use and agricultural
productivity. Time series data were used in the regression analysis. Double log. linear
regression analysis was used to express the index of agricultural productivity (API) as a
function of both energy consumption (EC) and gross addition of fixed assets (AFA). The
results showed that there was a positive relation between energy use and productivity.
Another example of the use of regression analysis with organic products is by Amacher et
- 88 -
al (Amacher, Hyde et al. 1999) were household data from Nepal's two major populated
regions were used to examine fuel wood consumption and production.
An example of the use of regression analysis in the study of production and/or consumption
in a manufactured system would be in the work of Al-Ghaniam (Al-Ghaniam 2003). His
paper used multiple regression models to represent relationships between energy
consumption and the related control variables. The regression models proved the existence
of valid relationships between electricity consumption and maintenance /production
management factors (failure rate and production rate). The regression models were further
used to formulate an economic treatment that demonstrated that good management
practices can result in significant savings in energy. Arize (Arize 2000), also, used
regression, among other tools, to investigate the long run relationship between U.S.
petroleum consumption and its determinants.
Some papers compared regression with other tools. For example, (Pao 2006) tried to model
and forecast the energy consumption in Taiwan using both regression analysis and
Artificial Neural Networks (ANN’s). In addition to temperature, his paper considered the
national income, population, gross domestic production and consumer price index.
According to Pao, these factors and the price are the most possible factors to affect
electricity consumption in the literature world wide. In a situation similar to seawater
consumption by the petrochemical plants in Kuwait, the electricity consumption in Taiwan
increases in summer and decreases in winter. Pao noticed that different researchers used
different models in different countries. He also stated that variables affecting demand and
energy consumption may vary from one region to another. Consequently, a model
developed for one region may not be appropriate for another region. In his study, it was
found that ANN’s have a higher forecasting capability than that of the regression model.
Another study used regression analysis and ANN's in manufacturing (Jaouadi, Msahli et al.
2006). The study tried to devise methods to accurately predict the amount of sewing thread
required to make up a garment. Three modeling methodologies were analyzed: theoretical
model, linear regression model and ANN model. The later two models were much better
than the first with the ANN model giving the best accurate prediction. This would be the
- 89 -
second study showing that ANN’s have a higher forecasting capability than that of the
regression model.
Finally, regression analysis was used to study the total consumption of both manufactured
and organic components. In his economical study, Ghatak (Ghatak 1998) investigated the
consumption behavior of India from 1919 until 1986.
The research presented in this thesis studied the influence of some weather factors in the
consumption of cooling seawater by petrochemical plants. The issue of the influence of
weather on production/ consumption of both organic and manufactured system has been
studied considerably.
An example of the influence of weather on a system can be seen on many works. For the
production of organic systems, Oury (Oury 1965) presented a method to measure the effect
of weather on crop production. According to him, the method presented was universal and
applicable to all crops. In the many mathematical models presented, crop yield (Y) is a
function of temperature, precipitation and the de Martonne and Angstrom indices. Another
study on the effect of weather on agricultural production was done by (Zagaitov 1982). His
paper studied the effects of weather on agricultural grain production on former Soviet
Russia. Amongst many findings, the author has discovered that unfavorable conditions for
grain crops have reoccurred discernibly once every six years. Zagaitov linked this
phenomenon to another one: the existence of 6-7 year cycles of fluctuation of precipitation
and temperature in Russia.
Dong et al (Dong, Lee et al. 2005) presented a holistic utility bill analysis method for base-
lining whole commercial building energy consumption in the tropical region. Six buildings
were used for case studies. Multiple linear regression analysis was performed .The results
showed that the variation of energy consumption in most of these buildings has to do with
temperature only. Another example is Bowers (Bowers 2001) who studied the effect of
weather on floating oil production systems. This study, according to Bowers, allowed the
full 'cost of weather ' to be assessed balancing the costs and benefits of investing in
contingent capacity against the possibilities of losing production time. Forecasting
- 90 -
consumer demand with relation to weather was done by Sivillo and Reilly (Sivillo and
Reilly 2004/2005) who concluded that the weather plays an important role in forecasting
consumer products demand.
The weather plays an important role in determining electrical demand (load). Thus, many
short term electric-load forecasting techniques depend on forecasted temperatures to predict
the electric load. Temperature forecasting needs the expertise of a meteorological service.
Not all electric utilities have access to (or are willing to pay for) such a service. Khotanzad
et al (Khotanzad, Davis et al. 1996) tried to close the gap in temperature forecasting for the
needs of electric utilities. In their research, a technique has been developed to forecast
hourly temperatures for up to seven days in advance. This technique utilized ANN's. Chen
et al (Chen, Yu et al. 1992) tried to forecast short-term electrical loads taking the weather
into account. Their paper presented an ANN model for forecasting weather sensitive loads.
The authors first stated that ANN's has an advantage over statistical methods such as time
series or regression analysis. This advantage, according to them, lies in ANN's ability to
model a multivariate problem without making complex dependency assumptions among
input variables and in extracting the implicit nonlinear relationship amongst them. The
ANN model developed by the authors is not fully connected to reduce the training time.
The model was made of one main ANN and three supporting ANN's. The main ANN was
used to provide the model's basic forecast reference while the three supporting ANN's were
used to increase the learning capacity of the proposed model. The results indicated that this
model can achieve greater forecasting accuracy than the traditional statistical model.
In this thesis, it is claimed that one of the influencing factors of seawater demand is plant
capacity utilization. There has been some works before that studied capacity utilization.
They include the works of Adam and Ebert (Adam and Ebert 1992) who discussed the
influence of capacity utilization on operations planning. Kirkely et al (Kirkley, Paul et al.
2002), also, defined a sequence of technological-economic definitions of capacity and
excess capacity for fishing industries, and provided empirical estimates of these measures.
The work by Anderson (Anderson 2001) supported the hypothesis that product mix acts
through capacity management decisions to reduce performance from the level implied by
- 91 -
direct effects alone. Delgado et al (Delgado, Jaumandreu et al. 1999) developed an
econometric model to simultaneously assess the degree of substitutability between labor
and materials and the impact of capacity utilization in the relative shares. They found
evidence of strong input share variations according to the degree of capacity utilization.
To solve its optimization and scheduling problem, the research presented in this thesis also
tried to forecast some weather factors such as ambient air temperature, seawater
temperature and humidity. As mentioned above, the prediction of weather factors usually
requires a meteorological service and not all companies are willing to use them. As a result,
many in-house methods have been developed to forecast the weather. The previously two
mentioned works of Khotanzad et al (Khotanzad, Davis et al. 1996) and Chen et al (Chen,
Yu et al. 1992) are examples of such methods.
In this study, seawater temperature was predicted by using air temperature and regression
analysis. Prediction of water temperature based on air temperature using regression analysis
was done before. Fore instance, Saila et al (Saila, Cheeseman et al. 2004) have noticed that
“the statistical relationship between air and water temperatures is traditionally established
by classical regression analysis or by using time series analysis procedures”. According to
them, the advantage of this latter approach is its simplicity and minimal data requirements
in contrast to the deterministic models which require much more data and are more
complex mathematically. The objective of their study was to develop and test a stochastic
model to accurately predict maximum daily water temperatures during the summer season
for small streams in the Wood-Pawcatuck Watershed using local air temperature and other
available meteorological data. The model was then utilized to more precisely define and
predict the extent of suitable habitat for brook trout in the study area under past drought or
other adverse environmental conditions. Five daily weather related inputs (maximum air
temperature, minimum air temperature, precipitation, evaporation and dry bulb
temperature) and one output variable (maximum daily stream temperature) were utilized to
train, calibrate and validate a neural network model designed to predict maximum summer
stream temperatures from the above-mentioned atmospheric input variables.
- 92 -
Kettle et al (Kettle, Thompson et al. 2004) also empirically modeled daily mean lake
surface temperatures (LST's) in summer for some lakes as a function of local air
temperature and theoretical clear-sky solar radiation.
Modeling water consumption was done in many works including those by Smaoi et al
(Smaoui, BuHamra et al. 2002) and BuHamra et al (BuHamra, Smaoui et al. 2003).These
two papers used ANN's in conjunction with Box-Jenkins approach to model the monthly
water consumption in Kuwait. First, the Box-Jenkins approach was used to predict the
missing values of the monthly water consumption due to the Iraqi invasion of Kuwait.
Second, the Box-Jenkins approach was used with the task of discovering the appropriate
lagged variables or input nodes in the input layer of the ANN's. It was found that when the
variables of the input layer in ANN's are chosen based on the Box-Jenkins approach rather
than on traditional methods, the average relative error for training and testing data sets was
reduced by 24%.
3.3 Factors Influencing the Demand Variation
The factors influencing the demand variation can be divided to two major categories:
process and weather. The process category has one factor only which is the amount of
production or the capacity utilization of the plant. The second category, the weather,
contains three factors: the ambient temperature, humidity and seawater temperature. The
factors of the two categories have been, empirically, noticed to influence the demand for
cooling seawater.
In the next paragraphs the reason why each factor influences the demand is explained. Both
categories have to do with the mechanisms of heat transfer in heat exchangers. Explaining
heat transfer in a thorough and detailed manner is beyond the scope of this thesis. For an
introduction to the subject of heat transfer, the interested reader may refer to the Heat and
Properties of Matter section on the work of Harrison (Harrison 1991). A more detailed
treatment of heat transfer and its applications can be found in Çengel (Çengel 1998).
- 93 -
3.3.1 The Process Category
3.3.1.1 The Amount of Production or Capacity Utilization
Heat exchangers are used to ascertain that the process fluid is kept in a certain range of
temperatures. All the mechanical processes and most of the chemical processes in a
petrochemical plant produce heat; if this heat is not removed from the process equipment,
these equipment are going to be damaged. Heat exchangers are used to remove this heat.
When a certain (mechanical or chemical) process produces heat, it is, almost always,
noticed that decreasing this process is going to produce less heat while increasing the same
process is going to produce more heat.
When the production of plant is increased (or, equivalently, when its capacity utilization is
increased) more heat is going to be produced. To keep the process equipment in the
specified range of temperature, more cooling seawater is required to remove this excess
heat. On the other hand, when the production of a plant is decreased (or, equivalently, when
its capacity utilization is decreased) less heat is going to be produced. To keep the process
equipment in the specified range of temperature, less cooling seawater is required to
remove this heat.
If the operators of the petrochemical plant do not order more cooling water when the
production is increased, their equipment are going to be damaged by the heat of the
process. On the opposite situation, if the operators of the petrochemical plant do order more
cooling water than needed when the production is decreased, their action is going to make
the plant owners incur a, needlessly, high cooling-water bill and would , eventually,
decrease the revenue of the plant and its profits. Therefore, plant operators tend to order
just enough quantity of cooling seawater that would put the operating equipment in their
specified temperatures.
3.3.2 The Weather Category
This category involves three factors which are seawater temperature, ambient air
temperature and the humidity of the ambient air.
- 94 -
3.3.2.1 Seawater Temperature
Çengel (Çengel 1998) mentioned that “Heat transfer in a heat exchanger usually involves
convection in each fluid and conduction through the wall separating the two fluids” (p.569).
Both conduction and convection depends on the temperature difference. The law that
governs conduction is
2 1( )Cond
AQ k
(3.1)
Where Q. cond is the rate of heat transfer in conduction,
A is the area subject to conduction,
‘l’ is the length of the area subject to conduction,
θ1 is the temperature of the fluid, and
θ2 is the temperature of the wall of the tube of the heat exchanger.
Çengel (Çengel 1998) defined convection as “the mode of energy transfer between a solid
surface and the adjacent liquid or gas that is in motion and it involves the combined effect
of conduction and fluid motion”. Convection heat transfer is expressed by Newton’s law of
cooling
( )Conv s fQ hA (3.2)
Where Q. conv is the convective rate of heat transfer
h is the convective heat transfer coefficient
A is the surface area of convection
θs is the surface temperature
θf is the temperature of the fluid to which the convective-heat transfer is directed to
provided that this fluid is provided in sufficiently large quantities.
As mentioned previously, the plant operators tend to order just enough quantity of cooling
seawater that would put the operating equipment in their specified temperatures. When the
seawater is cold (having lower temperatures) the temperature difference in both of the
conductive and convective heat transfer equations is going to be large. Subsequently, there
will be a high rate of heat transfer. On the other hand, when the seawater is warm or hot
- 95 -
(having higher temperatures) the temperature difference in both of the conductive and
convective heat transfer equations is going to be small. Subsequently, there will be a low
rate of heat transfer.
Regarding convective heat transfer, Çengel (Çengel 1998) mentioned that “The fluid
motion enhances heat transfer…In fact, the higher the fluid velocity, the higher the rate of
heat transfer” (p.350). Fluid velocity is related to the flow by the following equation
3 2( / ) ( ). ( / )Flow m s Area m Velocity m s (3.3)
When the temperature of the cooling seawater is relatively high, the temperature difference
between the cooling seawater and the fluid being cooled becomes small. Consequently,
there will be less heat transfer. To overcome this situation and enhance heat transfer, a
plant’s operators take advantage of the relationship between fluid velocity and heat transfer.
Practically speaking, this is done by increasing the flow. The end result is that the plant is
going to consume more water.
3.3.2.2 Ambient Air Temperature
At a heat exchanger, the final phase of heat transfer takes place between the outer surface
of it and the ambient air. The equation that governs this process is the following
( )Conv s aQ hA (3.4)
Where Q. conv is the convective rate of heat transfer from the surface of the heat exchanger
to the ambient air.
h is the convective heat transfer coefficient.
A is the surface area of convection.
θs is the surface temperature of the outer surface of the heat exchanger.
θa is the temperature of the ambient air to which the convective-heat transfer is directed to.
Because of the above relationship, it can be seen that when the ambient air temperature is
low, the temperature difference is going to be high and there will be a high rate of heat
transfer. When the ambient air temperature is high, the temperature difference is going to
- 96 -
be small and there will be a low rate of heat transfer. When low heat transfer is present,
plant operators will try to enhance it by increasing the velocity of cooling seawater which
will increase the flow.
3.3.2.3 Humidity
As mentioned in the previous subsection, the final stage of heat transfer in a heat exchanger
involves the transfer of heat from the outer surface of the heat exchanger to the ambient air.
This heat transfer is directly proportional to the temperature difference between the heat
exchanger surface and the ambient air. It was observed that the heat transfer is also directly
proportional to humidity in the air. This is consistent with the literature. For example, Still
et al (Still, Venzke et al. 1998) studied the convective heat transfer from a cylinder to a
humid air stream flowing normal to the cylinder. They mention that “The determination of
the rate of convective heat transfer to or from a circular cylinder in a cross-flow is
important in numerous applications in engineering, for example in heat exchangers and
tube banks.” (Still, Venzke et al. 1998).They found that “For molar fractions of water
vapour up to 0.27, the heat transfer increased with increasing humidity.” (Still, Venzke et
al. 1998)
3.4 Reasons for Choosing Regression.
Bowerman et al (Bowerman, O'Connell et al. 2005) wrote on when to use linear
regression,
“The simple linear regression assumes that the relationship between the dependent variable,
which is denoted y, and the independent variable, denoted x, can be approximated by a
straight line. We can tentatively decide whether there is an approximate straight-line
relationship between y and x by making a scatter diagram, or scatter plot, of y versus x.
First, data concerning the two variables are observed in pairs. To construct the scatter plot,
each value of y is plotted against its corresponding value of x. If the y value tend to
increase or decrease in a straight-line fashion as the x value increases, and if there is a
scattering of the (x,y) points around the straight line, then it is reasonable to describe the
relationship between y and x by using the simple linear regression model” (Bowerman,
O'Connell et al. 2005).
- 97 -
Scatter diagram of the relationship between
cooling water consumption and air temperature,
cooling water consumption and humidity,
cooling water consumption and seawater temperature,
cooling water consumption and plant capacity utilization (or production)
were made and it was found that the relationship was linear. Figures 3.1 to 3.4 show the
linear relationship between seawater consumption and every one of the aforementioned
variables. The existence of a linear relationship justified the use of regression analysis in
this work.
Capacity Utilization Vs. Consumption
0
10,000,000
20,000,000
30,000,000
40,000,000
50,000,000
60,000,000
0 10 20 30 40 50 60 70 80 90 100Capacity Utilization
Con
sum
ptio
n
Figure 3.1. Capacity utilization versus seawater consumption.
- 98 -
Seawater Temp Vs. Consumption
0
10,000,000
20,000,000
30,000,000
40,000,000
50,000,000
60,000,000
0 5 10 15 20 25 30 35 40
Seawater Temperature
Con
sum
ptio
n
Figure 3.2 Seawater temperature versus seawater consumption.
Ambient Air Temperature Vs. Consumption
0
10,000,000
20,000,000
30,000,000
40,000,000
50,000,000
60,000,000
0 5 10 15 20 25 30 35 40 45
Ambient Air Temperature
Cons
umpt
ion
Figure 3.3 Ambient air temperature versus seawater consumption.
- 99 -
Humidity Vs. Consumption
0
10,000,000
20,000,000
30,000,000
40,000,000
50,000,000
60,000,000
0 10 20 30 40 50 60 70 80 90
Humidity
Cons
umpt
ion
Figure 3.4 Humidity versus seawater consumption.
3. 5 Model Development
The model development process done in the work presented in this thesis can be divided
into five steps:
1. data collection,
2. analysis of data,
3. statistical modeling,
4. prediction and
5. scheduling
The data collected were plant data (seawater consumption (Con) and capacity utilization
(CU)) and weather data (ambient air temperature Ta, seawater temperature Ts and humidity
H). Then, an analysis of the data collected was done. First, regression analysis was
conducted to determine the consumption equations of each plant. Second, regression
analysis was done on seawater temperature (Ts) and humidity (H). It was found that air
temperature can be best modeled (for predictive purposes) by using exponential smoothing.
The analysis of the plant and the weather data resulted in many statistical models. These
models were used for prediction: First, ambient air temperature Ta was predicted using
- 100 -
exponential smoothing. Second, the predicted air temperature Ta was used to predict the
seawater temperature Ts. Third, the predicted seawater temperature Ts, in turn, was used to
predict humidity (H). Fourth, all the predicted weather factors and the plant capacity
utilization (assumed to be provided by the plants) were used to predict the seawater
consumption of each plant (Con.). Finally, this predicted seawater consumption was
translated to a predicted number of pumps. The last step, scheduling, was done by using the
predicted number of pumps to schedule their operation and maintenance. The model
development process can be seen in figure 3.5 below.
- 101 -
Figure 3.5 the model development process.
- 102 -
3.5.1 Data Collection and Analysis
The monthly consumption of each plant was obtained from the consumption bills of the
pumping station. The average air temperature (Ta) and humidity (H) were obtained from the
Meteorological Department of the Kuwait Civil Aviation Authority. Seawater temperature
(Ts) at receiving points was obtained from the petrochemical plants’ measure of input
water.
3.5.2 Regression models
The variables of concern were established for preliminary investigation of available data
and expert opinion. The independent variables were:
1) The average monthly temperature of the ambient air (Ta),
2) The average monthly temperature of the seawater supplied (Ts),
3) The average monthly humidity (H),
4) The percentage of capacity utilization of each petrochemical plant (CU),
5) The production level at the petrochemical plant (P)
And the output variable or the response was:
6) The seawater consumption (Con.)
The idea was to find a relationship between the output variable as a function of the input
variables. This relationship was investigated by using regression analysis. Then, for any
estimated set of input variables, the output level can be predicted as a basis for planning. To
predict the future ambient air temperature, exponential smoothing was used on the
historical air temperature data. Note that Kuwait weather does not experience significant
fluctuations over long periods of time.
An example of the equations that resulted from regression analysis, is the equation for the
consumption of consumer C12 shown below
12 0 1 2C s s aCon b b PT b T T (3.5)
- 103 -
The equation coefficients and other relevant statistical measures are shown in the tables
below.
Table 3.1 C12 Equation coefficients. Coefficients
P value Std Error -95% 95% t Stat VIF b0 13155055.9 1.05122E-08 1881745.429 9365026.108 16945085.69 6.991 b1 8.860 7.20452E-10 1.139 6.567 11.15 7.781 1.330 b2 -53223.9 0.02670 23231.7 -100015 -6432.8 -2.291 1.330
Table 3.2 statistical measures of equation 3.5
Summary |R| 0.844R2 0.712R2 adjusted 0.699Standard Error 2719617.472# Points 48PRESS 377684775351136.00R2 for Prediction 0.673Durbin-Watson d 0.821First Order Autocorrelation 0.564Collinearity 0.752Coefficient of Variation 13.167
Another equation of consumption is for consumer C6.
6 0 1 2C sCon b b T b CU (3.6)
The equation coefficients and other relevant statistical measures are shown in the tables
below.
Table 3.3.Equation coefficients for C6 Coefficients
P value Std Error -95% 95%t Stat VIF b0 4655410.798 0.03466 2171588.992 343066 8967756.059 2.144 b1 744414 9.0102E-26 51030.0 643078 845749 14.59 1.012b2 180121 1.64867E-15 18818.3 142752 217491 9.572 1.012
- 104 -
Table 3.4. Statistical measures of equation 3.6
Summary |R| 0.865R2 0.749R2 adjusted 0.744Standard Error 2934460.517# Points 96PRESS 892896232058345.00R2 for Prediction 0.720Durbin-Watson d 1.585First Order Autocorrelation 0.198Collinearity 0.988Coefficient of Variation 7.588Precision Index 36.189
3.5.3 Exponential Smoothing
As mentioned above, exponential smoothing was used for the prediction of temperature.
Thirteen years of monthly data were compiled. The best value of α was found to be 0.4 with
mean square error of 1.477.
3.5.4 Relationship between Ta and Ts for the Specific Location of the Pumping Station
It was found that for the specific location of the pumping station, seawater temperature is
related to ambient air temperature by the following relationship:
0 1s aT b bT (3.7)
Where b0= 9.574 and b1=0.584
3.5.5 Relationship between Ta, Ts and H for the Specific Location of the Pumping
Station
It was found that for the specific location of the pumping station, humidity is related to both
seawater temperature and ambient air temperature by the following relationship:
0 1 2a s aH b bT b T T (3.8)
- 105 -
Where b0=114.52, b1= - 4.606 and b2= 0.06616.
3.6 Prediction and Scheduling
The data analyzed and the models reached in the previous section were used for making
predictions. Year 2004 was used for testing the models. As previously explained, the
average ambient air temperature for every month, Ta was predicted first. Then, the
predicted ambient air temperature, Ta, was used for predicting the seawater Temperature,
Ts. Next, both of Ta and Ts were used to predict the humidity, H. Third, these predicted
parameters along with the plant capacity utilization and production figures available were
put in to the regression models for each plant to predict its seawater consumption. Finally,
the obtained seawater consumption was translated into the number of operating pumps.
Predicting the number of pumps based on the predicted seawater consumption may need
further explanation. A vertical centrifugal pump is usually characterized by its flow (Q) and
its total head (TH) with the latter being mainly a function of the discharge pressure. What
links the flow (Q) and the total head (TH) of a pump is the pump curve shown in figure 3.6.
Q
TH
Minimum Pump Flow Maximum Pump Flow
The Pump Curve
Figure 3.6 the pump curve.
In this curve, a pump normally operates between two extreme points, the minimum and the
maximum flow points. The minimum flow point is characterized by low flow and high
pressure while the maximum flow point is characterized by high flow and low pressure.
- 106 -
Predicting the number of pumps needed for a particular consumer was done by dividing the
(predicted) consumption (flow) of that consumer over the maximum flow a pump can
produce and then rounding that number up.
An example for the above procedure is the following. Assume that the predicted cooling
water consumption for a consumer in June is 40 million cubic meters. This consumer is
supplied by a group of pumps. Each pump has the maximum capacity of 16,000 m3/hr. The
predicted number of pumps is calculated as follow: Fist, the predicted consumption of the
month (40 million cubic meters) is divided by the number of days in June to get the daily
consumption.
66 3
Monthly Consumption =Daily ConsuptionNumber of Days40×10 =1.333×10 m /Day
30
Second, the daily consumption is divided by 24 hours to get the hourly consumption.
63 3
Daily Consuption Hourly Consumption24 hours
1.333 10 55.55 10 /24
m hr
Third, the hourly consumption is divided by the maximum capacity of a pump (in this case
16,000 m3/hr) to get the number of operating pumps.
3
3
Hourly Consumption No. of PumpsMax. Pump Capacity55.55 10 3.47 Pumps
16 10
The result shows that the required number of pumps for June is 3.47 pumps. Nevertheless,
this face value for the required number of pumps should not be taken literally. Mainly
because it does not make actual sense and it is not applicable. The problem with this value
is with the number beyond the decimal point, 0.47. It is not possible to operate a 0.47 of a
pump. In fact, it is not possible to operate any fracture of a pump. A pump can either be
turned on or off.
- 107 -
When an experienced operation engineer sees that required number of pumps is 3.47
he/she would interpret it as four pumps working at a capacity lower than the maximum
capacity. The fourth step, hence, is rounding up the number of operating pumps.
The above procedure was repeated for all the plants. The results were satisfactory. As an
example, the results obtained for C6 are shown in figures 3.7.
Actual vs.P red icted N um ber o f P um ps
0
1
2
3
4
5
6
1 2 3 4 5 6 7 8 9 10 11 12
M onth
Num
ber o
f Pum
ps
A ctual Num ber of P um ps P redic ted Num ber of P um ps
Figure 3.7 predicted and actual number of pumps operating.
- 108 -
Chapter 4 A Method for Flow Meter Drift Detection 4.1 Introduction The previous chapter presented a case study of minimizing the conflict between operation
and maintenance of a pumping station. The purpose of minimizing that conflict was to
allow for maintenance to be conducted without interrupting the operation of the pumping
station. Maintenance is carried out in order to increase the reliability of the pumping
station, the issue of which was discussed in chapter 2. Reliability, however, is not sought
for its own sake. It is only a mean to reduce maintenance costs, unplanned down times and
interruptions to cooling-seawater supply. The consequences of the later two are increased
operational costs and less revenue.
While the operational costs are measured by labor, the electricity consumed, the price of the
spare parts used and other consumables, the revenue is measured by the amount of seawater
supplied to the consumers. The measuring device is the flow meter. The pumping station
has a flow meter installed over every delivery pipe and each consumer is charged by the
amount recorded by its flow meters.
Flow meters are susceptible to errors and inaccurate readings. Therefore, calibration is
needed to ensure the accuracy of the readings. It involves the checking of a measuring
instrument against accurate standards to determine any deviation and correct for errors
(Encarta 2004).
The calibration process takes time and is usually done by the staff of the instrumentation
section in any plant. However, this staff is also involved in other major tasks and
downsizing has left fewer people to carry out the ever increasing tasks of calibration and
other responsibilities.
Flow meters, the revenue measuring devices, are susceptible to drift which can produce
readings that are inconsistent with the flow being measured. Flow meter drift is a serious
- 109 -
problem that flow meter owners face every now and then. The subsequent losses that result
from this problem can be huge especially if the flow meters are used for billing purposes.
Flow meters, as shall be discussed below, do not usually get much of attention and a drift
can go unnoticed for a long time.
Maintenance in general and calibration specially is done as a matter of priority. Firstly, on
only very critical devices related to cases that may impose danger to the life of the workers.
Secondly, maintenance and calibration are done on the pumping station devices that are
critical to production to ensure meeting contractual obligations to the consumers for
continuous and undisturbed supply of seawater. Any interruption of seawater supply to a
consumer would incur a heavy cost.
Consumption flow meters are neither life nor production critical. This would make them in
the category of devices not receiving much attention. Indeed, they are usually not thought
of except once every month at billing time.
There are many things that can go wrong in a flow meter: its chamber might be flooded by
rain or underground water causing damage or its cables might get electrically grounded
causing problems in the electrical grid of the station. What we are concerned with,
however, are the faults that have to do with the accuracy of its readings.
Therefore, in the research presented in this thesis, a flow meter fault is defined as giving
inaccurate readings. Accordingly, a fault may be salient such as what will happen when a
flow meter gives a reading that the plant operators know is far above or below what a
particular consumer will take. On the other hand, a flow meter fault might be hidden. This
will happen when a flow meter incrementally but systematically gives inaccurate readings
that add up and go unnoticed by the operation staff. This is known as flow meter drift. The
detection of the later type of flow meter fault (flow meter drift) is the subject of this
chapter.
- 110 -
Flow meter errors (including drift) are a major cause for a condition known as unaccounted
for water. Unaccounted for water in a municipal water distribution network is present when
the quantity of water billed is less than the quantity of water produced. In a municipal water
distribution network this might happen as a result of many things including
1. leakage,
2. unregistered municipal use of water in public utilities such as schools, parks and the
fire-fighting hydrant network,
3. unregistered water consumption by house holds,
4. illegal water taping (water theft),
5. administrative and accounting mistakes, and
6. flow meter faults including flow meter drift.
In an industrial water distribution network, the use of high quality pipes, the fact that the
flow to all consumers is monitored and measured and the absence of illegal water taping
would make unaccounted for water results from two reasons
1. administrative and accounting mistakes, and
2. flow meter faults including flow meter drift.
Usually, administrative and accounting mistakes that have to do with a plant’s water
consumption, if discovered, can be corrected and many flow meter faults can be recognized
and dealt with. Flow meter drifts, on the other hand, are extremely difficult to discover
and, if discovered, there is no way to rectify the damage that has been caused by them.
This is because an industrial consumer is not expected to accept to pay in retrospect for a
fault that has to do with the inaccuracy of the flow meter; flow meters are the responsibility
of the pumping station owners. Therefore, it is extremely important to detect flow meter
inaccuracy as fast as possible in order for the pumping station owners to calibrate it and
avoid any financial loss.
Flow meter designers and manufacturers gave some attention to the issue of flow meter
fault (where their definitions of it might not exactly coincide with the one used here,
although being fairly close). The way manufacturers usually tackle this problem is by
- 111 -
designing hardware devices in order to detect flow meter faults (according to their
definitions). End users, on the other hand, have already invested in their flow meters and
cannot afford to change them every few years when the technology improves. It would be
difficult for them, also, to make any modifications in the electronic circuitry of their flow
meters for many reasons including that there are many circuits depending on the flow meter
type and manufacturer. In addition, making a continuous online flow meter monitoring
system would be complicated, expensive and requiring a huge amount of data.
An alternative method must be simple, inexpensive and needing minimal data. In the
previous chapter, it was shown that the cooling seawater consumption by each plant
followed a certain pattern. The alternative method, therefore, could make use of the
observed pattern of consumption. Any deviation from the right pattern would be
considered as an indication of a possible inaccurate reading (fault) requiring investigation.
The alternative method would use the following as a mean to achieving its objective: the
cooling-seawater consumption by every plant has a pattern and any deviation from this
pattern could be an indication of a possible fault. The alternative method has the following
constraints: it must be simple, inexpensive and needing minimal data. With the given
means and constraints, the best choice for the alternative approach sought after would be a
statistical method to be devised.
The use of a statistical method for the detection of flow meter inaccuracy was thought of as
an attractive option for many reasons including
a. It did not involve tampering with flow meter circuitry and the risks
involved in such an option.
b. It would be a universal solution independent of the flow meter type
(electromagnetic, ultrasonic...etc.) or manufacturer.
c. It would be able to work with the data from the monthly bills i.e. it
will needed minimal data.
d. It would not require any knowledge about the mathematical relations
and models that governed the process.
- 112 -
e. It would be inexpensive.
Of the many statistical methods available, statistical process control (SPC) was thought of
as the most appropriate one. Also, of the many SPC methods existing, the tabular CUSUM
was thought of as the most suitable method for detecting flow meter drift. Nevertheless, a
serious obstacle had to be overcome. The consumption data takes the form of a seasonal
time series that widely oscillates (increasing in the summer and decreasing in the winter).
This time series is not suitable for use in SPC in general and more specially in a cusum
method. The cusum method is basically a method that detects mean shift. Therefore, a
method had to be thought of to transform the seasonal time series consumption data to a
medium suitable to be used in SPC-cusum. This was done by creating a virtual mean,
( )t vf X .
The theory of the virtual mean is that for every point in the time series (regardless of its
location) their exits a virtual point that represents the virtual mean corresponding to this
point. It is this virtual mean that would be processed by the cusum method. This method is
explained in the research method section.
In this project, when any deviation from the normal consumption pattern is identified, it is
considered as an indication of a possible flow meter fault that would require the
Instrumentation Section to investigate and calibrate the flow meter in question and check
for its accuracy.
Adapting the process can be done inexpensively. The equations developed here can be
turned to simple algorithms in any spreadsheet software. The method developed works on
monthly bills. This would require the operation engineer or the accountant to type the flow
meter reading of every consuming plant and see the feedback from the software. This
process would only take few minutes every month i.e. it is not a time-consuming process.
Comparing the huge financial losses resulting from flow-meter drift with the inexpensive
cost of adapting the proposed method and the fact that it only uses billing data (which most
- 113 -
companies keep for accounting purposes) makes it, in this author’s opinion, attractive for
practical use.
It was shown in the previous chapter that the consumption of cooling seawater by a
petrochemical plant largely depends on the weather and the plant’s capacity utilization or
production. The weather is a seasonal phenomenon. Subsequently, the consumption of
cooling seawater by a petrochemical plant follows a seasonal pattern: it increases in the
summer and decreases in the winter. A plant’s production has a lot to do with its units (sub
plants). A typical unit might cost tens or hundreds of millions of dollars. Thus, after a plant
is erected, it is rare to add new units or permanently shut working units. The addition of a
new unit would require consuming more water while permanently shutting a working unit
would result in consuming less water.
A petrochemical plant’s seasonal pattern of consumption is represented by a seasonal time
series. A typical seasonal time series representing the consumption of a plant can be seen in
figure 4-4. Since adding a new unit or shutting a working unit is a rare event, it is safe to
assume that the seasonal time series representing the plant’s pattern of consumption will
continue without change.
In summary, the purpose of this chapter is to present a suitable drift detection technique that
can capture the deviation from the normal behavior at any given period and determine if the
flow meter is in or out of tune. The remainder of this chapter has the literature review, the
factors that influence flow-meter readings, the reason for choosing SPC and its underlying
assumptions and limitations, the research method, the cusum method, case studies and the
limitations of the presented method including suggestions for further study.
4.2 Literature Review
In this section a literature review on previous works is done. It was thought that it might be
convenient to divide the literature review into several sections. Accordingly, the literature
review was divided to
i. Literature review on the phenomenon of unaccounted for water.
- 114 -
ii. Literature review on instrument (or sensor) drift which is the root
cause of the flow-meter drift problem.
iii. Literature review on flow meter drift (which is a special case of the
above problem) and literature review on unaccounted-for- fluid lost
as a result of the flow-meter mal functioning.
iv. Literature review on hardware solutions for the problem of
unaccounted-for- fluid.
v. Literature review on statistical methods to solve the problem of
unaccounted-for- fluid.
vi. Literature review on statistical process control (SPC) which is the
general approach that is used.
vii. Literature review on CUSUM which is a method of SPC that is
specifically used for solving the problem at hand.
4.2.1 Literature review on the phenomenon of unaccounted for water.
The pumping station under study is a part of a refinery and petrochemical cooling system.
This would make it an industrial water utility. An industrial water utility has similarities
and differences to a typical water utility. The many similarities include that both pump
from a source to a group of consumers. Another similarity is that both use the same set of
equipment: pumps, headers, valves, pipes, flow meters…etc. The major difference is that a
typical water utility (whether public or privet) usually has three types of consumers:
residential, commercial and industrial. An industrial water utility, on the other hand, only
has industrial consumers. Another difference is that because that these industrial consumers
are usually major plants, certain requirements are expected in an industrial water utility.
These requirements are not needed in a typical water utility. The phenomenon of
unaccounted for water does happen in both typical and industrial water utilities.
When a flow meter is drifting, it is producing inaccurate readings. If these readings are less
than the actual amount of water a consumer had, this would mean that there is an amount of
water that went unaccounted for. This unaccounted for water will cause the provider of the
water to charge for less than what has been provided and, subsequently, to lose money.
- 115 -
As mentioned above, unaccounted for water is a condition that is suffered by all water
utilities. There is often a difference between the amount of water produced and the amount
of water billed. A drifting flow meter is one cause of this, although (as shall be explained
below) it is not the only cause.
In this section, the literature on unaccounted for water is going to be reviewed. Most of
this literature is about typical water utilities and their distribution networks. Some of it,
nevertheless, is about agriculture since this section of the economy suffers form “lost
water”. The literature on unaccounted for water usually includes works on flow meters and
leakages.
Abushamsa (Abushamsa 2001) wrote a simple definition for unaccounted for water. He
wrote that “Unaccounted for water represents the difference between net production and
consumption of water. “ (Abushamsa 2001).
Stathis and Loganathan (Stathis and Loganathan 1999) referred to the categories of the
American Water Works Association (AWWA). They wrote that “The American Water
Works Association has identified three major categories of ‘losses’ in a water distribution
system. These categories are (AWWA 1987): (1) Accounted-for losses; (2) Real losses; (3)
Unaccounted-for losses.” (Stathis and Loganathan 1999).
For the first category, accounted-for losses, Stathis and Logan than (Stathis and
Loganathan 1999) wrote that “Accounted-for losses occur at metered locations within the
water distribution system”(Stathis and Loganathan 1999). These metered locations,
however, are for non-billable customers. Stathis and Loganathan (Stathis and Loganathan
1999) explain that “Non-billable customers include municipal users and the fire station”
(Stathis and Loganathan 1999).
For the second category, real losses, Stathis and Loganathan (Stathis and Loganathan
1999) wrote that “ a large percentage of water entering the water distribution system is
- 116 -
neither metered nor put to a beneficial use. Water that falls into this second category of
losses is called “real losses.” (Stathis and Loganathan 1999). They explained that “Real
losses cannot be tracked by a utility and include such losses as leakage or theft. Leakage is
the main culprit with regard to real losses in a water distribution system accounting for
approximately 14% of the total water supply (Smith 1986).” (Stathis and Loganathan 1999)
As for the third type of loss in a water distribution system, unaccounted-for losses, Stathis
and Loganathan (Stathis and Loganathan 1999) wrote that “ Unaccounted-for losses are
losses from the system that are put to beneficial use. However, these beneficial uses are
either not metered or are under registered due to meter errors.” (Stathis and Loganathan
1999). They explain that “Defective water meters may under-register actual water use.
Therefore, much of the water entering the system becomes unaccounted-for losses.”
(Stathis and Loganathan 1999)
To summarize the situation, Stathis and Loganathan (Stathis and Loganathan 1999) wrote
“The bulk of water produced (70%) does bring a return on
investment in the form of metered water sales. However, 30% of
the water produced does not bring in any revenue for the utility.
The largest non-revenue producing use is underground leakage in
water mains, which accounts for approximately 14% of the total
water produced. The second most troubling part of the water
supply system is the problem of inaccurate meter readings, which
account for 10% of the water produced (Smith, 1986). The
ultimate goal of any water utility should be to maximize the
quantity of revenue-producing water in the system.” (Stathis and
Loganathan 1999)
They also mentioned that “The two main rehabilitation techniques for maximizing revenue-
producing water involve repairing leaks in the pipe network and fixing meters.”(Stathis and
Loganathan 1999).
- 117 -
Abushamsa (Abushamsa 2001) wrote that “Although the definition is simple, determining
the true figures can pose some difficulties since UFW consists of two components: physical
and non-physical (administrative).” According to him, (Abushamsa 2001) , the physical
component contributing to UFW is usually leakage. Abushamsa wrote that “Water
consumed but not recorded by the consumer’s meter or otherwise not accounted for by
government or public use is refered to as non-physical loss (Administrative losses) and is
reflected in loss revenue”. He mentioned many reasons for the non-physical losses
including “Under registration of consumer meters (especially large diameter, industrial
meters)” (Abushamsa 2001).
While Stathis and Loganathan (Stathis and Loganathan 1999) has quoted Smith (Smith
1986) that the unaccounted for water in the USA is 30%, Abushamsa (Abushamsa 2001)
has mentioned that this figure was 56.3% in Jordan. These two figures show the size of the
problem in a typical water utility. The water loss in water utilities, subsequently, concerned
many researchers. For instance, Almandoz et al (Almandoz, Cabrera et al. 2005) presented
“a methodology for evaluation of water losses based on discrimination of the two
components of uncontrolled water in a water distribution network: physical losses in mains
and service connections, and the volume of water consumed but not measured by
meters.(Almandoz, Cabrera et al. 2005)” The methodology “presumes that real losses
(leakage) in certain physical states of a network are a function of pressure, while apparent
losses defined as non metered consumed water are a function of consumption patterns i.e.,
domestic, industrial, institutional, etc.”(Almandoz, Cabrera et al. 2005).
While Andersen and Powell (Andersen and Powell 2000) used state-estimation for district
metered areas (DMA) demand monitoring and leak detection, Clark et al (Clark,
Sivaganesan et al. 2002) gave equations that can be used for cost estimation for replacing
pipes in a water distribution network. The water demand was thought as a contributing
factor. Therefore, Arniella (Arniella 2007) used billing data for developing demand
allocation that was used for setting up a hydraulic and water quality model for the water
network system of a major metropolitan area in Georgia. Alcocer-Yamanaka et al (Alcocer-
- 118 -
Yamanaka, Tzatchkov et al. 2006) used many statistical methods including the Poisson
Rectangular Pulse (PRP) method to model residential water consumption in a water
network. Another work that studied the residential demand pattern was done by Arbue´s
and Barbera´n (Arbue´s, Barbera´n et al. 2004) .In their paper they formulated a model for
residential water demand for the city of Zaragoza (Spain). Their aim of this study was to
evaluate the potential of pricing policies as a mechanism for managing residential water.
Water utilities often need to study the water demand during a day. For this purpose, they
use the raw data from flow meters and tank levels. Walski et al (Walski, Lowry et al.
2000) notes that “there are frequently problems with the raw data and how those raw data
are used” (Walski, Lowry et al. 2000). Walski et al explain that “The problem usually lies
with small errors or inaccuracies in the raw data being magnified into large errors in
demand curves” (Walski, Lowry et al. 2000). For them, the solution lays in eliminating the
noise from the data. For doing this they suggest using smoothing.
Jankovic´-Nisˇic´ et al (Jankovic´-Nisˇic´, Maksimovic´ et al. 2004) made a paper to
“provide a more systematic approach in designing target oriented data acquisition systems
for the control of water distribution networks.” (Jankovic´-Nisˇic´, Maksimovic´ et al.
2004). The core of their work was a sampling design problem “which is that of defining the
location and the sampling time interval of the measurements to be taken.” (Jankovic´-
Nisˇic´, Maksimovic´ et al. 2004) In their work, Jankovic´-Nisˇic´ et al tried to find the best
location and sampling rate for flow meters in order to detect leakage.
Nagar and Powell (Nagar and Powell 2000) addressed the issue of water-network
observability. According to them, a water network is not highly observable i.e. it has a high
uncertainty; they have written that “The measurement uncertainty is largely due to the
predominance of pseudo-demand measurements necessary to make up for the lack of real
transducers” (Nagar and Powell 2000). Thy presented a method, which separates the
parametric uncertainty in the network parameters from the nominal system. Kang and
Lansey (Kang and Lansey 2010) made a study to optimally locate field measurement sites
and leads to more reliable state estimation of a water network.
- 119 -
Water utilities, usually, have to balance two competing demands. The unaccounted for
water that results from faulty residential meters leads to lost revenue which demands
calibrating the flow meters more often. The large number of residential flow meters could
make the cost of pulling, testing, repairing, and replacing them exceeds that of the revenue
lost from unaccounted for water. Noss et al (Noss, Newman et al. 1987) addressed the issue
of optimal residential meter testing.
In their work, aiming at making an efficient water management system for irrigation,
Hamdy et al (Hamdy, Ragab et al. 2003) mentioned that “water use efficiency in this sector
(agriculture) is very poor not exceeding 45% with more than 50% water losses”. Baum et al
(Baum, Dukes et al. 2003) discussed the selection and use of water meters for irrigation
water measurement while Roberts (Roberts) discussed the issue of inaccurate flow meter in
the dairy farms.
Flow meters in a water distribution network, received the attention of many. Tamarkin
(Tamarkin 1992) wrote about the history, methods and benefits of Automatic meter reading
(AMR). Hauber-Davidson and Idris (Hauber-Davidson and Idris 2006) discussed the use of
smart meters. They gave the following definition for one “A Smart Water Meter is a
normal water meter linked to a device that allows continuous electronic reading and display
of the water consumption. It negates the need to manually read the meter dial. Once this
information is available as an electronic signal, it can be captured, logged and processed
like any other signal.” (Hauber-Davidson and Idris 2006). They mentioned that “Many
water authorities experiment with smart water meters for their residential customers,
largely to better understand consumption patterns.” .The authors, however, dismissed the
idea that smart water meters will be used for all residential consumers on economical basis.
Nevertheless, Hauber-Davidson and Idris, wrote that” The situation is completely different
for large water users, though. As shown in this paper, for them smart metering their water
consumption will soon become the norm rather than the exception.” (Hauber-Davidson and
Idris 2006).
- 120 -
Arregui et al (Arregui, Enrique Cabrera et al. 2005) presented several factors that affect a
flow meter’s accuracy while Ferréol (Ferréol 2005) addressed water meters’ park
inefficiency. He gave the following definition for a water meter park, “A water meters'
park is constituted of residential meters and of commercial and industrial meters”. Ferrol
differentiates between accuracy and efficiency. According to him,
“Accuracy is what is obtained
when a meter is tested on a test bench.
According to the tested flow rates, the meter
gives its answer in term of which percentage of
the volume it can measure. To present the
efficiency, the pattern of consumption must be
first explained. As a meter has an accuracy
depending on flow rate, it is important to look at
the consumption of a user by representing it
according (to) ranges of flow rates and
indicating for each range the proportion of water
which is passing. By doing this, it gives a
weight for each flow rate interval. This chart is
called the pattern of consumption.”(Ferréol
2005)
He further elaborates on this by writing “The efficiency corresponds to what the meter can
measure when it sees a certain pattern of consumption. It is the "multiplication" of the
accuracy curve by the pattern of consumption.” (Ferréol 2005).
Fulford (Fulford 2002) presented a paper showing the results of an investigation into the
performance of four models of current meters used in hydrography. The models were tested
for accuracy and consistency. Finally, Van Vugt and Samuelson (van_Vugt and Samuelson
1999) demonstrated the positive effect of using flow meters in the case of a drought. They
compared the residential water consumption of two groups during a drought. One group
- 121 -
was aware of its water consumption while the other was not. Van Vugt and Samuelson
wrote “It was predicted that metering would be beneficial in promoting conservation, in
particular, when people experienced a shortage. Consistent with expectations, the results of
both studies revealed that conservation efforts were greater among metered (vs. unmetered)
participants when they perceived the water shortage as severe.” (van_Vugt and Samuelson
1999). They have also written that “Additional analyses suggested that the positive effect
of metering could be partially explained by a greater concern with the collective costs of
overconsumption during the drought.”(van_Vugt and Samuelson 1999).
4.2.2 Literature review on instrument (or sensor) drift which is the root cause of the
flow-meter problem.
Bolton (Bolton 2000)defined drift by writing that “An instrument is said to show drift if
there is a gradual change in output over a period of time which is unrelated to any change in
input”. Sydenham et al (Sydenham, Hancock et al. 1989) defined drift in a different way
“Drift is the rate of change of the signal output with time”. As for the nature of drift and its
predictability Sydenham et al have written that “Drift is a complex effect usually poorly
understood and seldom occurs at a predictable, fixed rate”. They have also written that drift
“… is often a key limiting factor on system discrimination and accuracy and cannot be
ignored”.
4.2.3 Literature review on flow meter drift which is a special case of the above
problem and the resulting unaccounted-for-fluid-loss phenomenon.
Nilsson (Nilsson 1998) has quoted Meshkati and Groot (Meshkati and Groot 1993) that
inaccurate measurement is responsible for more than %80 of the unaccounted-for loss in
the gas supply industry. A similar phenomenon, unaccounted-for water loss, is known in
the potable water industry as was shown in section 4.2.1 and by Johnson (Johnson 1996)
and Hopkins (Hopkins, Savage et al. 1995). The issue of flow meter inaccuracy is a
problem wherever flow meters are used regardless of the fluid being measured. For
example, a flow meter error that resulted in unaccounted-for loss in petroleum caused a
political crisis in Kuwait as was written by AlHermi (AlHermi 2007) AlJasem (AlJasem
2007) and AlJasem (AlJasem 2007).
- 122 -
4.2.4 Literature review on hardware solutions for the problem of unaccounted-for-
fluid.
Not surprisingly, the issue of flow meter accuracy was the subject of many research papers
and patents. However, the majority of these have been about improving the sensing devices
and electronic circuitry like the patents done by Herwig et al (Herwig, Keese et al. 1994)
and Keech (Keech 2005). Therefore, the immediate benefits of such works are for flow
meter manufacturers.
4.2.5 Literature review on statistical methods to solve the problem of unaccounted-
for- fluid.
Very few research or practical works have considered using statistics for detecting flow
meter faults. These include the paper done by Nilsson (Nilsson 1998) who presented a
method for finding inaccurate gas flow meters using billing data. The method used billing
readings to find inaccurate meters by assigning an individual load index (LI), which was
primarily affected by the climate rather than the customer’s individual behavior. The
individual (residential) LI is compared with an average LI, and meters that differ from the
average LI were examined. The method excluded gas flow meters that are used in industrial
establishments because the consumption is not affected by the climate.
As mentioned above, what affects gas consumption in a residential area is largely the
weather. Hence, investigating outliers is a good method for finding faulty flow meters.
Unfortunately, this method cannot be used in an industrial environment because industrial
consumers vary greatly in their consumption. For instance, the quantity of seawater
consumed in a month by one plant is more than that consumed by another plant in a year.
4.2.6 Literature review on statistical process control (SPC) which is the general
approach that was used.
The seminal work on statistical process control is often attributed to Walter Shewhart and
his book Economic Control of Quality of Manufactured Product, which was published in
1931(ASQ 2009) .Several introductory works on statistical process control (SPC) include
- 123 -
the books of Montgomery (Montgomery 2005), Wheeler and Chambers (Wheeler and
Chambers 1992), Norton (Norton 2005), Amsden et al (Amsden, Butler et al. 1998), Abbott
(Abbott 1999) and Wheeler (Wheeler 2004).
Statistical process control (SPC) is a method for monitoring the mean of a process and is a
very popular method for doing so. Nevertheless, Han and Tsung (Han and Tsung 2006)
ascertain that the current methods of SPC focus mostly on monitoring and detection of
constant shifts in the mean. Other conditions where the mean shift is dynamic have not
been thoroughly studied according to them. To fill this gape, Han and Tsung (Han and
Tsung 2006) designed a reference-free Cuscore (RFCuscore) chart for tracing and detecting
dynamic mean changes.
4.2.7 Literature review on CUSUM.
If the reader is unfamiliar with the cusum method he/she is advised to read Chapter 8 of
Montgomery’s book (Montgomry 2005). The seminal work on the Cumulative Sum
(CUSUM) was done by Page (Page 1954). In his paper, Page explained that in the industry,
with its continuous production process, the quality of the output is approximately constant.
It only worsens as a result of a fault at some point in the process. Page wrote that “In
general, it will be possible to assign a quality number, θ, to the output which may be taken
as a parameter of the distribution. We are interested in the changes in θ” (Page 1954).
Detecting changes in the parameter θ is done by, what was called, process inspection
schemes. These schemes detect deterioration in the quality of the output from a continuous
production process. A widely used scheme consists of examining samples of a fixed size at
regular intervals of time. In that paper, rules were developed that "use all the observations
since action was last taken and that are suitable for the detection of any magnitude of the
change in the parameter" (Page 1954). Unlike the Shewhart chart, “With this rule the
decision whether or not to take action is made after each sample and all the previous
samples are used in making the decision” (Page 1954). Page was also interested in
estimating the point at which the change took place.
Later, Page (Page 1961) presented a more thorough explanation of the cumulative sum
charts and methods. Advancements since the publication of the first paper in 1954 up to
- 124 -
1961 were also presented. In 1962, Page (Page 1962) presented a research paper about
using gauging as an input for the CUSUM rather than measured observations. The scheme
developed was for controlling the mean and standard deviation of a normal distribution.
The developed scheme using gauging was better than the Shewhart chart for detecting
small changes but slightly less than the CUSUM schemes based on observations.
The cumulative sum is made in two ways: tabular and graphic. The graphic method utilizes
a V-mask. Some works on the construction of the V-mask include that of Johnson (Johnson
1961). That work was about how the theory of sequential probability ratio tests can help in
constructing the V-mask.
Ewan (Ewan 1963), showed the conclusions of applying the cusum method since its
introduction up to the time of publishing his work. Ewan outlined the various types of
continuous graphical control schemes which were available (at that time) and the type of
process for which cusum charts were most appropriate.
Ewan has noticed that “...the cu-sum chart is effectively a running mean chart and that it is
more effective in detecting sustained changes... than the standard control chart. On the
other hand, the control chart is more effective in detecting larger, shorter term changes and
is extremely simple to use.”(Ewan 1963).
Just like any monitoring scheme, the purpose of the cusum method is to stop production
when an alarm is given. This alarm could either be true or false. The economic cost of
stopping production in both cases and the cost of running the cusum method itself was first
considered by Taylor (Taylor 1968). In his paper an approximate formula for the long run
average cost per unit time as a function of the parameters of the cumulative sum chart was
developed. The purpose of doing this was to enable these parameters to be chosen
optimally under the average cost criterion.
Chiu (Chiu 1974) had some criticisms on Taylor’s method. Chiu presented a study of the
economic design of cusum charts which, according to Chiu, overcame the above mentioned
criticisms. Chiu also provided a simple scheme which determined a control plan that was
close to optimum.
- 125 -
More recent works about the cusum method can be found in Zhao et al (Zhao, Tsung et al.
2005) who produced a dual cusum (DCUSUM) by combining two cusum charts to detect a
range of mean shifts. Han et al (Han, Tsung et al. 2007) stated that although the cusum
method is simple, its performance deteriorates when the actual mean shift is unknown. To
overcome this condition, an alternative approach called the CUSUM chart with local signal
amplification (LSA-CUSUM) was presented. This scheme worked by amplifying or
weakening local signals to improve the power of the traditional CUSUM chart in detecting
an unknown mean shift over a range. Han et al claimed that measurable weakening and
amplifications of local signals can improve the ability of the CUSUM chart in detecting the
local mean shift. Han et al (Han, Tsung et al. 2007) also created a multi chart consisting of
several CUSUM or EWMA charts with different reference values that were used
simultaneously to detect anticipated process changes. Their work showed that the multi-
chart has the merits of quick detection of a range of mean shifts, was easy and had a
flexible design for various situations and great reduction in computational complexity. Han
and Tsung (Han and Tsung 2007) also made a cusum multi-chart scheme consisting of
multiple cusum control charts for detecting and diagnosing unknown abrupt changes in a
process. They showed that this scheme performed better than the single cusum or EWMA,
cusum-multi chart, EWMA-multi chart and GLR (Generalized Likelihood Ratio) charts.
4.3 Factors Influencing Flow Meter Readings
A flow meter is a device that measures the amount of flow. The reading of a flow meter
must accurately reflect the amount of fluid transferred between two points within a small
margin of error. There are many types of flow meters. Each type of flow meter is designed
and manufactured by many companies. If flow meter failure is defined as producing
inaccurate readings, the different designs and manufacturing methods will produce different
modes of failure and different reliability indices for each flow meter brand and model.
The investigation of the factors influencing the readings of each type, brand and model of
flow meters is something beyond the scope of this thesis. Still, the literature review in this
chapter showed some works that compared the performance of a single type of flow meter
that was designed and manufactured by different companies.
- 126 -
In the pumping station understudy, two types of flow meters were used: ultrasonic and
magnetic. A paper by Daniel Measurement and Control (Daniel Measurement And Control
2010) investigated the factors influencing the readings of one of their ultrasonic flow
meters. The paper warns that “This paper is based upon the Daniel USM design and the
information presented here may or may not be applicable to other manufacturers”. It also
cautions that “It is important to understand that the meters being analyzed in this paper are
of the chordal design, and therefore some of the analysis would not apply to other designs”.
The paper mentioned several reasons for producing inaccurate readings.
The first reason for producing inaccurate readings was transducer deterioration.
“Transducers typically generate the same level of ultrasonic signal time after time. Any
increase in gain on any path indicates a weaker signal at the receiving transducer. This can
be caused by a variety of problems such as transducer deterioration, fouling of the
transducer ports, or liquids in the line.” The paper also mentions that “other factors that
affect signal strength include metering pressure and flow velocity”
A fourth factor that influences flow meter reading is transducer performance “All ultrasonic
meter designs send multiple pulses across the meter to the opposing transducer in the pair,
before updating the output. Ideally all the pulses sent would be received and used.
However, in the real world, sometimes the signal is distorted, too weak, or otherwise the
received pulse does not meet certain criteria established by the manufacturer. When this
happens the electronics rejects the pulse rather than use something that might distort the
results” The paper continue on when this might happen “Unless there are other influencing
factors, the meter will normally operate at 100% transducer performance until it reaches the
upper limit of the velocity rating. Here the transducer signal becomes more distorted and
some of the waveforms will ultimately be eliminated since they don’t fit the pulse detection
criteria within the specified tolerance. At this point the meter’s performance will drop from
100% to something less.”
- 127 -
Another factor influencing flow meter readings is noise. “Each transducer is capable of
receiving noise information from extraneous sources (rather than its opposite transducer).
In the interval between receiving pulses, meters monitor this noise to provide an indication
of the “background” noise. This noise can be in the same ultrasonic frequency spectrum as
that transmitted from the transducer itself.” The paper states that “The measure of signal
strength to the level of “background” noise is called the Signal to Noise Ratio, or SNR for
short…SNR is generally not an issue unless there is a control valve or other noise
generating piping component present. When that occurs, the SNR values will drop. The
magnitude of the SNR is a function of the manufacturer’s methodology of expressing the
value”. Other reasons that influence flow meter readings that were mentioned in Daniel
Measurement And Control (Daniel Measurement And Control 2010) include dirty flow
meters and blockage.
4.4 Reasons for Choosing Statistical Process Control (SPC), and Its Underlying
Assumptions and Limitations
4.4.1. Reasons for Choosing Statistical Process Control (SPC)
Norton (Norton 2005) has defined Statistical Process Control (SPC) as “The use of
statistical methods to control/improve a process” he continued “the general goal of SPC is
to produce better goods and services” Norton also wrote that” When data can be collected
to measure the quality of a manufactured product or of a service, the potential is there to
use the data to tell if quality is slipping or holding steady, and whether efforts to improve
quality are working. The whole philosophy of SPC is to use data to continually improve
quality”. On SPC’s capability in detecting a variation in a process, Norton wrote that
“Control charts and summary statistical measures …such as range and standard deviation
can be used to detect when there are unwanted sources of variation in a process”.
4.4.2. The Underlying Assumptions of SPC
Amsden et al (Amsden, Butler et al. 1998) have written about the six assumptions or
principles for relaying on SPC. They are
1. No two things are exactly alike,
- 128 -
2. Variation in a product or process can be measured,
3. Things vary according to a definite pattern,
4. Whenever things of the same kind are measured, a large group of the measurements
will tend to cluster around the middle,
5. It is possible to determine the shape of the distribution curve for parts produced by
any process, and
6. Variations due to assignable causes tend to distort the normal distribution curve.
4.4.3. Limitations of Statistical Process Control
There are many SPC methods. We are mainly concerned with the Shewhart chart and the
cusum method. Koutras et al (Koutras, Bersimis et al. 2007) have written that “Each of the
aforementioned categories of control charts has specific advantages and disadvantages. A
Shewhart chart uses the information contained in the most recently inspected sample; as a
consequence, it is not very efficient in detecting gradual or small shifts in a process
characteristic. In contrast, this type of control chart may instantly detect a large shift in the
process level and for this reason it has been used for well over the last 70 years. On the
contrary, CUSUM and EWMA control charts are more sensitive in detecting small shifts in
a process since they use information from a long sequence of samples.”
4.5 The Research Method.
As mentioned previously, seasonal time series are not the best medium for SPC in general
and the tabular CUSUM specially. All SPC methods require that the process data to be
closely gathered around a mean. This is not the case of a seasonal time series were data are
usually very far from the mean.
One of the research problems, therefore, became of finding a suitable method of
transforming the seasonal time series to a linear time series. This problem is depicted in
figure 4.1 below.
- 129 -
Figure 4.1 transforming the seasonal time series to a linear series. The transformation was achieved by exploiting two features of a sinusoidal wave and their
corresponding features of seasonal time series. The explanations follow.
Let us examine a sinusoidal wave which is given by the function
( ) ( )Y aSin (4.1)
Where a >0 and >0 ; lest us assume also that this function is sampled at equal time
periods as shown in figure 4.2. In every wave length (before the wave repeats itself) we
assume that it has been sampled “d” times i.e. “d” is the number of samples in every 2π (or
360o). Let “α” be defined as α:= 2 π n, were n=1,2,3… Consequently, when the wave
repeats itself again and again as θ increases, every two points that are d samples apart
would be equal because
( ) ( )Sin Sin (4.2)
This would mean that
( ) ( )Y Y (4.3)
Manipulating equation 4.3 makes it
( ) ( ) 0Y Y (4.4)
This is shown in figure 4.3.
- 130 -
+ a
- a
0
a
a
Y
Figure 4.2 the function given by equation (4.1).
Figure 4.3 the sinusoidal function of equation (4.1) sampled ‘d’ times in every period and having the same values between every θ and θ+α.
A discrete seasonal time series (with no trend), as the one shown in figure 4.4, looks almost
like the sinusoidal function shown above in figure 4.3. It oscillates around a mean (or more
precisely an estimate of the mean, X ) and repeats itself once every year. Therefore, if a
seasonal time series is given by
- 131 -
1 2( ) , ,..., nX t X X X (4.5)
And if this series is sampled d times (d could be 12 month or 4 quarters…etc.), then
equation 4.4 becomes
1( ) ( ) 0X t X t d (4.6) Or similarly 1( ) ( ) 0X t X t d (4.7)
Where ε1 is i.i.d. (independent and identically distributed).
Seasonal Time Series
X
Xt
t
Figure 4.4 a seasonal time series.
Another property of sinusoidal waves is that when every two points that are d/2 time –
periods apart are averaged, the resulting average is the mean
( ) ( )
22
dY Y
(4.8)
Or, similarly,
- 132 -
( ) ( )
22
dY Y
(4.9)
In a seasonal time series, equation 4.8 becomes
( )
222
t dtX X
X
(4.10)
Or similarly
( )
222
t dtX X
X
(4.11)
Where X is the estimate of the mean and ε2 is is i.i.d. In the case of a seasonal time series, adding equations 4.7 and 4.11 would result in the
virtual or running mean for every point. This function is denoted by ( )t vf X and it
represents the virtual mean for every point.
( )2
( )( )2
t dt
t v t t d
X Xf X X X
(4.12)
Where
1 2 (4.13)
And ε is i.i.d.
- 133 -
The graph of the function for the virtual mean against the time series shown in figure 4.4 is
shown in figure 4.5.
Virtual Mean
X
t
Xt
Figure 4.5 the virtual mean.
It can be shown that the running or virtual mean given in equation 4.12 equals the seasonal-
time series’ estimated mean plus some residual as is shown in equation 4.14 below.
( )t vf X X (4.14)
The form given in equation 4.12 (or 4.14) is exactly what is needed for a process to be
monitored by SPC methods and, more specifically, the tabular CUSUM.
4.6 The Method of the CUSUM.
4.6.1. General
According to Montgomery (Montgomery 2005), the cumulative sum (CUSUM) is done by
plotting the cumulative sums of the deviations of the sample values from a target value, in
our case the virtual mean.
(4.15)
1
( )i
i jj
C x x
- 134 -
Or, put in another way,
1( )i i iC x x C (4.16) Where Ci is the cumulative sum up to and including the ith sample; the starting value C0=0.
There are two ways to represent cusums:
1. Tabular (or algorithmic) cusum.
2. V-mask form of the cusum.
Only the first method is going to be used.
4.6.2 The Method of Tabular CUSUM. According to Montgomery (Montgomery 2005), the tabular cusum works by accumulating
deviations from the target that are above it with one statistic C+ and accumulating
deviations from the target that are below it with another statistic C-.
If the mean is the target, then
1max[0, ( ) ]i i iC x x k C (4.17)
1max[0,( ) ]i i iC x k x C (4.18)
Where C+0=C-
0=0, and k is a reference value; k is one-half the magnitude of the shift we are
interested in detecting. In the work presented in this thesis, the magnitude of the shift we
are interested in detecting is one standard deviation .Accordingly,
2
k (4.19)
Note that C+i and C-
i accumulate deviations from the target value (in our case the virtual
mean) that are grater than k with both quantities being rest to zero on becoming negative.
If either C+i or C-
i exceeds a decision interval, H, the process is considered to be out of
control. Montgomery suggests that a reasonable value for H is five times the process
standard deviation.
5H (4.20)
His suggestion was followed.
- 135 -
4.7 Case Studies
In this section, we are going to give some case studies of the flow of some plants.
4.7.1 Case Study #1
This case study is used because it is a documented case of flow-meter drift. The flow of a
plant having the code name C7 is given in table 4.1 and the flow is shown in figure 4.10.
Table 4.1 C7 Consumption Month Consumption Month Consumption Month Consumption Mar-99 12,046,300 Jul-01 13,538,100 Nov-03 8,901,200 Apr-99 12,151,500 Aug-01 13,324,800 Dec-03 8,552,600 May-99 12,783,100 Sep-01 12,537,700 Jan-04 8,142,600 Jun-99 12,506,400 Oct-01 14,092,700 Feb-04 8,026,100 Jul-99 13,446,300 Nov-01 13,293,200 Mar-04 8,728,800 Aug-99 14,095,300 Dec-01 9,631,400 Apr-04 8,448,400 Sep-99 12,695,200 Jan-02 13,243,200 May-04 9,717,900 Oct-99 13,348,600 Feb-02 11,900,000 Jun-04 9,949,700 Nov-99 12,361,700 Mar-02 11,753,096 Jul-04 10,647,200 Dec-99 10,362,000 Apr-02 10,546,200 Aug-04 12,583,400 Jan-00 11,768,700 May-02 12,292,500 Sep-04 8,969,000 Feb-00 10,615,000 Jun-02 11,910,500 Oct-04 9,434,700 Mar-00 12,069,800 Jul-02 12,135,500 Nov-04 8,235,300 Apr-00 12,238,400 Aug-02 12,408,800 Dec-04 7,366,200 May-00 12,847,700 Sep-02 12,442,100 Jan-05 6,730,900 Jun-00 13,376,000 Oct-02 12,446,000 Feb-05 6,151,100 Jul-00 13,897,700 Nov-02 11,711,000 Mar-05 5,924,900 Aug-00 15,004,100 Dec-02 11,091,500 Apr-05 5,912,600 Sep-00 14,279,400 Jan-03 10,430,300 May-05 11,402,400 Oct-00 13,940,400 Feb-03 8,813,200 Nov-00 12,888,000 Mar-03 9,436,600 Dec-00 12,825,600 Apr-03 10,867,500 Jan-01 12,233,800 May-03 11,562,800 Feb-01 11,519,300 Jun-03 11,219,100 Mar-01 13,684,700 Jul-03 10,707,800 Apr-01 13,133,600 Aug-03 9,648,700 May-01 13,816,100 Sep-03 9,648,700 Jun-01 13,490,000 Oct-03 9,669,500
- 136 -
C7 Consumption
4,000,000
6,000,000
8,000,000
10,000,000
12,000,000
14,000,000
16,000,000
Jul-98 Dec-99 Apr-01 Sep-02 Jan-04 May-05 Oct-06
Month
Con
sum
ptio
n
Figure 4.6 C7 flow meter readings for the consumption.
It is obvious that the flow meter readings for the consumption are decreasing over the years.
At that time it was thought that the decrease in consumption was due to C7's economy in
the use of seawater and not to a flow meter fault. What led to this belief was that the owners
of C7 expressed that they were taking measures to decrease the use of cooling seawater at
that time. Simultaneously, a decline in the flow meter readings was noticed.
Suspicions in the abnormality of the plant consumption only started when it was,
coincidently, noticed that the pressure produced by the operating pumps (for C7) should
give more flow than what was recorded by the flow meter. This meant one of three
possibilities
1. The pressure gauges were wrong, or
2. The flow meter was wrong, or
3. There was a major leak in the pipeline.
The pressure gauges were calibrated and were found to be OK and there was no leak in the
pipeline. This made the flow meter the only suspect. When the flow meter in question was
inspected and calibrated, it became clear that it was inaccurate. This meant that C7 did not
- 137 -
use less water during that period and its efforts to save water were a failure. The proof of
this is the "jump" in flow meter readings after calibration (the flow meter drift was
discovered in the first week of May 2005 which is the last point in the graph).
Because this case was a known case of flow-meter drift, it was an ideal situation for
checking the validity of the tabular CUSUM method in detecting flow-meter drift. If the
method was capable of detecting the drift, the time of the drift’s onset would also be of
interest to estimate the revenue lost.
The statistical properties of the virtual mean and the parameters of the tabular CUSUM for
this specific consumer are shown in table 4.2 while the application of the method is shown
in table A1 in the appendix.
Table 4.2 The statistical properties of the virtual mean and the parameters of the tabular
CUSUM for C7 Mean (X bar) 13,566,054
Standard Deviation 915,876 K 457,938
Xbar+K 14,023,992 Xbar-K 13,108,116
H 4,579,380 It is obvious from table A1 the tabular CUSUM method was able to detect the drift on
December 2001. The method also indicated that the start of the drift happened some time at
July 2001. What is interesting about this particular case is that the records show that a
calibration took place on October 2004. It may be concluded that this calibration was not
perfect.
4.7.2 Case Study #2
This case study is about the same consumer, C7. Another study was conducted on the
period from Jun 2005 to November 2008. The statistical properties of the virtual mean and
the parameters of the tabular CUSUM for this specific consumer on this period are shown
in table 4.3 while the application of the method is shown in table A2.
- 138 -
Table 4. 3 The statistical properties of the virtual mean and the parameters of the tabular CUSUM for C7-Case #2
Mean (X bar) 12,164,264 Standard Deviation 695,409
K 347,705 Xbar+K 12,511,968Xbar-K 11,816,559
H 3,477,046
Table A2 shows that on August 2007, the alarm persisted for three consecutive months.
This may be an indication of drift. It also shows that the onset of this drift was on May
2007.
2.7.3 Case Study #3
This case study is about a plant with the code name C6. Its recorded flow-meter readings
for consumption are shown in table A3. The statistical properties of the virtual mean and
the parameters of the tabular CUSUM for this specific consumer are shown in table 4.4.
The CUSUM method as applied to C6 is shown in table A4.
Table 4. 4 The statistical properties of the virtual mean and the parameters of the tabular
CUSUM for C6 Table A4 shows that on May 2008 a large enough drift was detected and this drift started
on Mars 2007.
2.7.4 Case Study #4
This case study is about a plant with the code name C12. Its recorded flow meter readings
for consumption are shown in table A5. The statistical properties of the virtual mean and
the parameters of the tabular CUSUM for this specific consumer are shown in table 4.5.
The CUSUM method as applied to C12 is shown in table A6.
Mean (X bar) 31,696,783 Standard Deviation 10,993,078
K 5,496,539 Xbar+K 37,193,322 Xbar-K 26,200,244
H 54,965,389
- 139 -
Table 4.5 The statistical properties of the virtual mean and the parameters of the tabular CUSUM for C12
Table A6 shows that there has been a consistent alarm about a drift starting from September
2008. This supposed drift started on January 2008.
2.7.5 Case Study #5
This case study is about a plant with the code name C3. Its recorded flow meter readings
for consumption are shown in table A7. The statistical properties of the virtual mean and
the parameters of the tabular CUSUM for this specific consumer are shown in table 4.6.
The CUSUM method as applied to C12 is shown in table A8.
Table4. 6 The statistical properties of the virtual mean and the parameters of the tabular
CUSUM for C3 It can be seen from table A8 that there has been an alarm for drift since April 2006. This
drift started on July 2005.
2.7.6 Case Study #6 This case study is about a plant with the code name C5. Its recorded flow meter readings
for consumption are shown in table A9. The statistical properties of the virtual mean and
the parameters of the tabular CUSUM for this specific consumer are shown in table4.7. The
CUSUM method as applied to C12 is shown in table A11. It can be seen from Table A10
that the flow meter of this consumer did not experience drift during the period under study.
Mean (X bar) 20,367,209Standard Deviation 6,737,318
K 3,368,659 Xbar+K 23,735,868 Xbar-K 16,998,550
H 33,686,588
Mean (X bar) 3,886,171 Standard Deviation 616,733
K 308,366 Xbar+K 4,194,537 Xbar-K 3,577,804
H 3,083,665
- 140 -
Table 4.7 The statistical properties of the virtual mean and the parameters of the tabular
CUSUM for C5
Mean (X bar) 2,074,663 Standard Deviation 630,887
K 315,444 Xbar+K 2,390,106 Xbar-K 1,759,219
H 3,154,435 4.8 Limitations of the Presented Method and Suggestions for Further Study As mentioned previously, the method presented has many advantages, including
a. It doesn't involve tampering with flow meter circuitry and the risks
involved in such an option.
b. It is a universal solution that is independent of the flow meter type
(electromagnetic, ultrasonic...etc.) or manufacturer.
c. It can work with the data from the monthly bills i.e. it needs minimal
data.
d. It does not need any knowledge about the mathematical relations and
models that govern the process.
e. It is inexpensive.
However, this method has limitations and drawbacks. The first limitation is that (in time
series terms) it needs at least two d’s of data (in our case two years) before it starts
functioning. Obviously, this would not make it applicable to new plants. The second
limitation is that it works with seasonal time series only. Other consumption patterns such
as linear and mixed seasonal and linear time series can not have their drift detected by this
method. A potable (municipal) water network, for example, has a consumption pattern that
is usually mixed linear and seasonal. Finally, the time between the onset of drift and its
detection is in the order of months. In many situations, this may not be acceptable.
Expanding the limitations and eliminating the drawbacks of the presented method might be
the subject of further studies. A method might be introduced that does not involve the long
waiting period before application of the method previously presented. Other improvements
- 141 -
might be done on making the method applicable on other time series such as linear and
mixed linear and seasonal time series. Certainly, the issue of the time between the onset
of drift and its detection can be the subject of a study aiming to shorten it.
- 142 -
Chapter 5 An Alternative Method for Flow Meter Drift Detection
5.1 Introduction
In the previous chapter, Chapter 4, a method for the detection of flow meter drift was
introduced. That method depended on developing a virtual mean and, then, using the
tabular cusum. The tabular cusum is one of many methods of statistical process control. In
this chapter, the virtual mean developed in the previous chapter will be used once again.
Flow-meter drift, nevertheless, will be detected by using an alternative method, artificial
neural networks, ANN’s. What follows is a literature review on the subject. Then, a
discussion about production processes and their quality is made. Next, the reasons for
choosing artificial neural network methods are explained. The assumptions for using
ANN’s and their limitations are also presented. The structure of the artificial neural
network used is presented in the research method section. Afterwards, the results of the
ANN work are shown in tables and figures. Finally, suggestions for potential applications
of the method developed are made.
5.2 Literature Review
For an introduction to the subject of artificial neural networks, the reader is referred to the
work of Basheer and Hajmeer (Basheer and Hajmeer 2000) . More details on the subject
can be found in Mehrotra et al (Mehrotra, Mohan et al. 1997) and Principe et al (Principe,
Euliano et al. 2000) .A very good introduction can also be found at a web page by the
makers of the software NeuroSolutions (NeuroSolutions 2009).
Artificial neural networks have many applications. For example, Palaneeswaran et al
(Palaneeswaran, Love et al. 2008) used artificial neural networks to map rework cause and
effect in 112 Hong Kong construction projects. In this thesis, however, artificial neural
networks (ANN’s) were used for detecting change in a process. One of the early papers that
applied ANN’s to detect changes in a process mean was that of Cheng (Cheng 1995). In
that paper, Cheng was trying to provide an alternative to statistical process control (SPC)
methods. The SPC methods, namely the Shewhart and CUSUM control schemes, were
replaced by artificial neural networks. The neural network architecture that Cheng used was
- 143 -
the three-layer fully connected feed-forward network with back propagation. For the inputs
for his network, Cheng used both numerical and symbolic inputs. For the numerical inputs,
Cheng used a string of 16 past data. This string of data, called window, did not use the
original values of the process. Rather, it used transformed values; the transformed values
were obtained by a coding scheme. For the symbolic input, Cheng used one run rule (an
explanation for run rules will follow). The run rule that he used was the one regarding
exceeding ± 3σ (where σ is the process’s standard deviation). Cheng reported that ANN’s
were 20-40% faster in detecting small process changes than the traditional Shewhart and
CUSUM control schemes. Later, Cheng (Cheng 1997) used two types of pattern
recognizers based on different neural network architectures: a multilayer perceptron trained
by back-propagation and a modular neural network. Cheng (Cheng 1997) noticed that the
modular neural network provided better recognition accuracy than back-propagation when
high strong interference effects existed.
One important application of artificial neural networks is in pattern recognition. For a
general introduction to pattern recognition that includes using ANN’s for it, the reader is
referred to the book by Friedman and Kandel (Friedman and Kandel 1999); for the more
specialized subject of pattern recognition by ANN’s only, the reader is referred to the book
by Bishop (Bishop 2005). Consequently, there have been many papers on the subject of
using artificial neural networks (ANN’s) for pattern recognition of statistical process
control (SPC) methods. For example, Anagun (Anagun 1998) used a multi-layered neural
network trained with a back propagation algorithm for pattern recognition of control charts.
A method called histogram representation was employed. Hassan et al (Hassan, Baksh et al.
2003) used ANN's for pattern recognition of SPC charts and compared two methods of data
input to the ANN's: raw data and statistical features of data. The ANN with the statistical
features performed better than the one with raw data.
Pacella et al (Pacella, Semeraro et al. 2004) applied the adaptive resonance theory (ART)
neural networks in their work. They have presented a fuzzy ART neural system for quality
control. The purpose of the system was the detection of abnormal process behavior. Pecella
et al mentioned that the advantage of their system over other neural techniques was that it
- 144 -
did not require previous knowledge about the abnormal patterns or their mathematical
models or probability distribution functions.
In practice, the Shewhart control chart is used with what is called “Supplementary Run
Rules”. These rules indicate when to investigate a process when the points plotted on the
Shewhart chart exhibit certain behaviors. The run rules can be thought of as primitive
pattern recognition methods. Koutras et al (Koutras, Bersimis et al. 2007) presented the
subject of Shewhart control charts that are supplemented with additional rules. Yasui et al
(Yasui, Ojima et al. 2006) introduced two additional run rules. According to their work, a
process might be considered out of control if
1. Two of three successive observations exceed ± 2.0698 sigma control limits.
2. Two successive observations exceed ±1.9322 sigma control limits.
5.3. Production Processes and Their Quality Assurance
In this section the nature of production processes is going to be described. These processes
have the tendency to deteriorate with time. This deterioration will decrease the quality of
products. Consequently, there is a need for quality-assurance methods to prevent this. The
most common method is the Shewhart Chart.
5.3.1 The Nature of Production Processes. Any process aiming to produce a consistent and constant product, such as a reinforcing bar
of a certain diameter or a brick of certain dimensions or a chemical product with specific
properties …etc. would usually produce this product according to the standard required but
with some deviation. This deviation is usually expected of any process and the allowance
made for deviation is called the tolerance of the process or product. If the quality we are
interested in is called X and the desired value of X is X , the plot of X against time, t,
would look like what is shown if figure 5.1. The vertical axis shows the magnified area of
the desired value, X , and its tolerances.
- 145 -
Figure 5.1 a process over time. X at a certain time, is designated by Xt. For example, we would have X1, X2, X3
corresponding to values of X at times t=1, t=2, t=3, respectively. The values of Xt oscillate
closely around the desired value, X .It can be proved that the desired value, X , is actually
the mean of the quality or process we are interested in.
5.3.2 Introduction to the Shewhart Chart
The Shewhart chart is one of the most widely used methods for ensuring that a product is
produced according to what is desired. It is a quality assurance method achieved by
monitoring the behavior of a process. More precisely, it is about observing the extent to
which a process stays close to its mean (or strays away from it). To observe this, the chart
needs to use the mean of the process or quality, X , and its standard deviation, σ . The
Shewhart chart is made up of seven lines: One line at the mean of the process and a line at
the following values: ±σ, ± 2 σ, ± 3 σ. Figure 5.2 shows the Shewhart chart made against
the process of figure 5.1.
Several run rules have been developed to utilize the Shewhart chart. The purpose of these
rules is to prevent the process from deviating beyond what is permitted. An example of a
run rule is the following.
Rule 1
Take action if one point lays outside the ± 3 σ lines.
- 146 -
Figure 5.2 the Shewhart chart.
In general, run rules take advantage of the previously mentioned lines of the Shewhart chart
and the number of points that have passed them. The run rules would signal whenever they
are satisfied and would not signal otherwise.
5.4 Reasons for Choosing Artificial Neural Network Methods, Their Assumptions and
Limitations.
5.4.1 Reasons for Choosing Artificial Neural Networks
Basheer and Hajmeer (Basheer and Hajmeer 2000) have cited Jain et al (Jain, Mao et al.
1996) that “The attractiveness of ANNs comes from the remarkable information processing
characteristics of the biological system such as nonlinearity, high parallelism, robustness,
fault and failure tolerance, learning, ability to handle imprecise and fuzzy information, and
their capability to generalize” Basheer and Hajmeer (Basheer and Hajmeer 2000) has
added that “ Artificial models possessing such characteristics are desirable because (i)
nonlinearity allows better fit to the data, (ii) noise-insensitivity provides accurate prediction
in the presence of uncertain data and measurement errors, (iii) high parallelism implies fast
processing and hardware failure-tolerance, (iv) learning and adaptivity allow the system to
update (modify) its internal structure in response to changing environment, and (v)
generalization enables application of the model to unlearned data.”
- 147 -
5.4.2 Assumptions of ANN-Based Modeling
Rudolf and Kröplin (Rudolph and Kröplin 1997) have written that “the two necessary and
sufficient conditions for the correct generalization in neural networks can now be
established in form of the two following consecutive steps. The formerly unresolved
generalization capability of non-linear multi-layered feed-forward neural networks can be
now proven to be
pointwise correct, if and only if a training pattern pcan be learned and recalled
error-free by the new similarity neural network topology, F. “ and
totally correct, if and only if the neural network approximates after the training the
correct similarity function… The correct similarity function Fis approximated if and
only if the correct point-wise generalization property is fulfilled for each point in
the whole domain of definition of F”
5.4.3 Limitations of Artificial Neural Networks
Basheer and Hajmeer (Basheer and Hajmeer 2000) mentioned that ”ANNs also have
limitations that should not be overlooked. These include (i) ANNs’ success depends on
both the quality and quantity of the data, (ii) a lack of clear rules or fixed guidelines for
optimal ANN architecture design, (iii) a lack of physical concepts and relations, and (iv) the
inability to explain in a comprehensible form the process through which a given decision
(answer) was made by the ANN (i.e., ANNs criticized for being black boxes).” They assert
that “ANNs are not a panacea to all real-world problems; for that, other traditional (non-
neural) techniques are powerful in their own ways.”
5.5 The Research Method
A dataset of monthly water consumption for an industrial consumer was simulated for this
research. The main purpose of the research was to develop rational investigations of drifts
so as to have optimal billing with minimal cost and efforts. Also, the modeling frameworks
from this research would mainly provide systematic ‘alarm’ mechanisms to the operation
and maintenance staff, especially whenever a flow meter drift might occur in the future. In
the previous chapter, a seasonal time series consumption mapping was developed (i.e. for a
consumption typical of an industrial consumer) with linear patterns and virtual means.
- 148 -
Subsequently, the virtual mean mapping was augmented through systematic statistical
process control (SPC) with CUSUM to rationally detect the drift errors in industrial water
metering. In this chapter, further extension with ANN modeling of flow meter drift issues is
considered. The virtual mean developed in the last chapter was used again. Its outputs were
considered as the raw inputs to the input layer of the ANN.
The neural network models used normalized numerical inputs relating to past values of the
flow and run rules from the Shewhart chart for detecting the status of the flow meter.
5.5.1 The Inputs for the Artificial Neural Network
The artificial neural network used in this research is the three layer back propagation
network. It has an input layer, a hidden layer and an output layer. The raw inputs for the
input layer are the outputs of the virtual mean. The input layer is made up of two parts: a
numerical part and a symbolic part. This is similar to what Cheng (Cheng 1995) has done.
The purpose of including a numerical part is to study the behavior of the process
quantitatively while the purpose of including a symbolic part is to study the behavior of the
process qualitatively.
5.5.1.1 The Inputs for the Artificial Neural Network- The Numerical Inputs
When a process is drifting, the process or quality of interest might take a shape similar to
what is shown in figure 5.3. The figure shows a process or quality with an upward drift.
Similarly, a downward drift may exist.
In figure 5.3, after the upward drift has started, each point or sample Xt is going to be grater
than the previous point(s). The opposite would exist if there was a downward drift.
Consequently, examining the behavior of Xt in relation to its past values might give a clue
about the existence of a drift. Hence, an index bi was thought of where
i t t ib X X (5.1)
- 149 -
Figure 5.3 a normal and a drifting process.
This index would numerically compare the current value of the process Xt to one of its past
values Xt-i . Comparing the current value of the process to one past value would not be
enough, however. To have a good understanding of the behavior of the process, the current
value must be compared to many, n, past values. Therefore, for every value of Xt, n values
of bi are going to be produced. To further explain, for every Xt we would have b1, b2,
b3,….,bi,…,bn.
The n values of the index bi would examine the behavior of the current value of a process in
relation to its past values. Nevertheless, the values of bi obtained would be unique to its
particular process. The aim of this study is to detect flow meter drift and the pumping
station under study has many flow meters. The flow that is measured by one flow meter
differs greatly from that measured by other flow meters; sometimes, by several magnitudes.
This would mean that each flow meter would require its own simulation which would be
exhaustive and computationally expensive. To only make one simulation that is capable of
being generalized to many flows (processes), normalization was used. The normalized
value is computed as follows
- 150 -
tt
X XZ
(5.2)
Where X is the mean and σ is the standard deviation of the quality or process.
To study the behavior of Zt in relation to its past values an index zbi is defined as z
i t t i
z t t ii
b Z Z
X X X Xb
z t t ii
X Xb
(5.3)
Then, for every value of Xt, n values of zbi are going to be produced i.e. for every Xt we
would have zb1, zb2, zb3,….,zbi,…,zbn. The number of past values to consider, n, was decided
to be 17. Cheng (Cheng 1995) considered n to be 16. He also used a coding scheme as
mentioned.
Next, the mean and standard deviation for the produced ( )t vf X was calculated. The virtual
mean for the data was calculated using
( )2
( )( )2
t dt
t v t t d
X Xf X X X
(5.4)
The standard deviation of the virtual mean would be used in calculating the values of zbi in
equation (5.3): For very point of the virtual mean ( )t vf X the values of zbi were computed
by equation 5.3. Finally, the values of zbi made for the numerical inputs to the artificial
neural network.
5.5.1.2 The Inputs for the Artificial Neural Network- The Symbolic Inputs
To study the behavior of any process, the numerical characteristics of this behavior would
certainly be helpful. The previous sub-section dealt with that; it dealt with the quantitative
aspect of the process. A process, nevertheless, can be examined by using a different point
of view: The qualitative point of view.
The qualitative point of view would see if the behavior of the process is having certain
qualities or not. Because these qualities either exist or not exist or exist in a certain
- 151 -
condition, numerical values can not be assigned to them. Rather, logical or symbolic values
would be used for representing them.
One good method for examining a process qualitatively is to use the previously mentioned
run rules. Cheng (Cheng 1995) used one run rule as his symbolic input. Namely, the rule
when the process exceeds 3X . In this thesis, seven run rules were used as the symbolic
inputs to the neural network. The first five run rules from Koutras et al (Koutras, Bersimis
et al. 2007). The run rules demand that the process be investigated if
1. One point is outside ± 3 σ lines.
2. Two out of three consecutive points are beyond ±2 σ lines.
3. Four out of five consecutive points are ± 1 σ or beyond from the mean.
4. Eight consecutive points are on one side of the mean.
5. Six points in a raw steadily increasing or decreasing.
The remaining two rules are from Yasui et al (Yasui, Ojima et al. 2006)
6. Two of three successive observations exceed 2.0698X .
7. Two successive observations exceed 1.9322X .
Each rule, R, would give one of three responses.
R1,2,…,7 = +1 if the rule is satisfied and the points are greater than the mean. -1 if the rule is satisfied and the points are less than the mean. 0 if the rule is not satisfied. All the above mentioned run rules examined the qualitative behavior of the virtual mean,
( )t vf X . The mean X and standard deviation σ of ( )t vf X were calculated to establish the
values by which the run rules would trigger. Every value of the virtual mean, ( )t vf X , was
tested against the seven run-rules. If any rule was triggered, it would give a value of +1
or -1 depending on the position of the points with respect to the mean. On the other hand, if
a rule was not triggered, it would give a value of 0. The seven run-rules, R1,2,…,7 , made for
the symbolic inputs of the artificial neural network.
- 152 -
5.5.2 The Hidden and Output Layers.
As in any artificial neural network, there are hidden and output layers. The output layer
gives one of three statuses for the flow meter: normal or upward drifting or downward
drifting. These outputs are qualitative or symbolic. Hence, they were represented by three
symbols in the simulation. The normal case had the symbol ‘0’, the upward drifting case
had the symbol ‘1’ and the downward drifting case had the symbol ’-1’. The structure of
the artificial neural network is seen in figure 5.4 below.
- 153 -
Figure 5.4 the structure of the artificial neural network.
- 154 -
5.6 Results of the Simulation, Training, Cross Validation & Testing
5.6.1. The Simulation of Flow Data
The simulation made for the cooling water consumption had 7,515 points of data. For every
datum after the 16th datum, the values of zbi (making for the numerical inputs for the ANN)
were calculated. Also, each datum after the 16th datum was tested against the seven above
mentioned run-rules (making for the symbolic inputs for the ANN). If a rule was satisfied,
the corresponding input for it would be 1or -1. Else, it would be zero.
The data included both normal and drifting flows. The drifting flows were of the two types:
drifting upwards (producing higher flow-meter readings than the actual consumption) and
drifting downwards (producing lower flow-meter readings than the actual consumption).
The data representing normal flow numbered 2,499. The remaining 5016 data had the two
types of drift in addition to the normal ‘recovery’ flow.
In the remaining 5016 data, an external influence was deliberately inserted in the original
data to produce the effect of an upward or downward drift and, subsequently, faulty data.
The amount of the external influence representing the upward or downward drift varied
each time to give the artificial neural network a chance to be exposed to different
magnitudes of drift. After the end of each drift, 20 consecutive data representing a recovery
period of normal flow were added. The reason for adding these 20 data was to ‘clear’ the
numerical inputs of the input layer of the ANN. The numerical inputs for the input layer, as
mentioned previously, were the indices zbi. These indices compared past values of flow to
the current value of flow. The furthest in the past this comparison is made is with the past
17th value of flow. After the end of a drift, and after 20 consecutive data of normal flow
were inserted, the influence of this drifting past was supposed to be cleared on all zbi’s .
Before the above mentioned data were introduced to the input layer of the ANN, their order
was randomized. NeuroDimension (NeuroDimension, Inc.) recommended that for
classification problems (as is the case here), the order of the data be randomized before
presenting them to the network. “Neural networks train better if the presentation of the data
is not ordered” (NeuroDimension, Inc.).
- 155 -
5.6.2. Training and Cross Validation of the Artificial Neural Network.
The software used for building the artificial neural networks was NeuroSolutions by
NeuroDimension, Inc. The software divides the data into three sets: training, cross
validation and testing. On the purpose of this division, the authors of the software
(NeuroDimension, Inc.) wrote that “One of the primary goals in training neural networks is
to ensure that the network performs well on data that has not been trained on (called
“generalization”). The standard method of ensuring good generalization is to divide your
training data into multiple data sets. The most common data sets are the training, cross
validation and testing data sets”.
The purpose and mechanism of use of the cross validation data set is described by the
authors of the software. “The cross validation data set is used by the network during
training. Periodically, while training on the training data set, the network is tested for
performance on the cross validation set. During this testing, the weights are not trained, but
the performance of the network on the cross validation set is saved and compared to past
values. If the network is starting to over train on the training data, the cross validation
performance will begin to degrade. Thus, the cross validation data set is used to determine
when the network has been trained as well as possible without over training (i.e. maximum
generalization).” (NeuroDimension, Inc.).
Of the 7,515 points of data available, 5,645 points were used for training and cross
validation. One thousand training epochs were performed. A measure of the efficiency of a
neural network is its learning curve. The learning curve of a neural network is displayed as
the mean squared error (MSE) for each of the training and cross-validation data sets versus
the training epochs. The learning curve for the training and cross validation data sets of the
artificial neural network (ANN) constructed is shown in figure 5.5. The numerical values
for the MSE of the training and cross validation data sets are shown in table 5.1.
- 156 -
MSE versus Epoch
00.050.1
0.150.2
0.250.3
0.350.4
0.450.5
1 100 199 298 397 496 595 694 793 892 991
Epoch
MSE
Training MSE
Cross Validation MSE
Figure 5.5 the learning curve for the ANN: MSE for training and cross validation versus
the number of epochs.
Table5. 1: Training results for 1000 epochs
Best Networks Training Cross Validation Epoch # 997 968 Minimum MSE 0.01574738 0.016462574 Final MSE 0.01576385 0.017043336
On identifying an unsuccessful training, the authors of the software (NeuroDimensions,
Inc.) wrote “It may happen that the network does not learn the problem. This is best
evidenced by a learning curve that does not approach zero”. On what might lead to this, the
authors of the software (NeuroDimensions, Inc.) mentioned three factors
1. The network is capable of learning the problem but has not been trained long
enough.
2. The network is capable of learning the problem but is stuck in local minima.
3. The network is not powerful enough to learn the problem.
The learning curve of figure 5.5 did approach zero. This is a good indication that the
training was successful.
5.6.2. Testing of the Artificial Neural Network.
NeuroDimension (NeuroDimensions, Inc.) wrote that “Although the mean square error is a
good overall measure of whether a training run was successful, sometimes it can be
- 157 -
misleading. This is particularly true for classification problems”. They also mention that
“When ‘classification’ is selected as the problem type, the NeuralExpert (a part of the
software) stamps a pair of confusion matrix probes-one for the training set and one for the
cross validation set…The confusion matrix tallies the results of all exemplars of the last
epoch and computes the classification percentages for every output vs. desired
combination”.
A number of data, 1870, was saved for testing. The confusion matrix for the test is shown
below in table 5.2.
Table5. 2: The confusion matrix
Desired
Output Status(-1) Status(1) Status(0) Status(-1) 266 9 6 Status(1) 21 296 13 Status(0) 59 88 1112
For the data spared for testing, the confusion matrix shows the performance of the artificial
neural network. The confusion matrix compares the actual output with the desired output.
It shows how many times the neural network made the right decision and how many times
it was “confused”. For example, there have been a total of 393 cases with an actual upward
drift (Status (1)). The artificial neural network rightly identified 296 of them as they were,
having an upward drift. The network confused 9 cases as having downward drift and 88
cases as being normal while actually they were cases of upward drift.
5.7 A Possible Way of Improving the Results.
In this author’s opinion, the results were satisfactory. There could be, however, a room for
improvement. It was previously mentioned that the input for the ANN was made of 7,515
peaces of data. Of these data, 2,499 represented normal flow while 5016 represented
drifting and normal ‘recovery’ flows. The necessary calculations were made on all of them
and they were randomized before being introduced to the numerical part of the input layer
of the ANN.
- 158 -
The 20 consecutive data after each drift, representing the normal ‘recovery’ flow were
necessary to clear the values of zbi’s. Nevertheless, these data, and their zbi’s and run rules,
may have hindered the learning of the ANN and, consequently, lowered the efficiency of
the model constructed. This may have happened because although the values of these data
were normal, the values of their zbi’s and the rules associated with them were not. The
values of their zbi’s and the rules associated with them bore the effect of the past drift. In
hindsight, if the 20 consecutive data after each drift, representing the normal ‘recovery’
flow were only used to clear the zbi’s and the rules and not as inputs to the ANN, the results
may have been better.
5.8 Potential Applications
The findings of this research have the potential to be practically applied in any facility
involved in the consumption of industrial water for cooling purposes. The inputs for the
method developed are the monthly bills of industrial water consumption which are easily
obtained in any industrial facility. The method’s dependence on mathematical algorithms
and already kept records does give it an economical advantage over the more expensive
industrial hardware for solving the same problem. The method’s independence of the
volume of flow (it uses ‘normalized’ flow) makes it applicable to a wide range of
consuming plants.
- 159 -
Chapter 6 Conclusions 6.1 Summary
Petroleum is one of the most important resources in our modern world. It is, however,
rarely useful by itself and the many products that it contains can only be obtained after
petroleum goes through a refinery. A refinery processes the crude petroleum, mainly
through a distillation unit, to produce asphalt, greases, lubricants, waxes, industrial fuels,
diesel, kerosene, petrol, aircraft fuel…etc. A refinery is a complex engineering system. It
generates many products. It also needs many inputs. One important input for a refinery is
large quantities of cooling water for condensation of hydrocarbon vapors. The cooling
water for a refinery or a chemical plant is provided by its cooling-water system. The most
important component of this cooling system is the cooling pumping station.
The reliable operation of refineries and petrochemical plants can not be over emphasized.
The unreliable operation of such plants would have severe consequences on human life,
cause injury to workers and damage the environment. The other consequence of the
unreliable operation of refineries and petrochemical plants is financial. When these plants
fail, this failure will cause the price of their outputs to increase which will disturb other
dependent industries and will make the owners of these plants suffer from financial losses
resulting from lost production, damaged equipment and penalties.
One vital requirement for the reliable operation of refineries and petrochemical plants is the
reliable operation of their cooling system including the cooling-water pumping station
supplying this system. This station should operate reliably, continuously and generate
enough revenue for its owners to make profit and justify the investment that had been put
into it.
The aim of this thesis is to support the operation and maintenance of a cooling
petrochemical pumping station. This was done by applying some tools and techniques. The
tools and techniques applied on the pumping station and presented in this thesis were
reliability analysis to increase the reliability of the pumping station, regression analysis to
- 160 -
minimize the conflict between operation and maintenance, statistical process control for
flow meter drift detection and artificial neural networks for the same purpose.
Chapter 1 was an introduction that discussed the importance of petroleum refining and the
petrochemical industry in general in our modern societies. The importance of the cooling
system for that industry was emphasized. At the core of the cooling system is the cooling-
pumping station. A detailed description of the location and structure of the pumping station
understudy was presented.
The research problem was also laid down on Chapter 1. The criticality of cooling sea water
interruption made the reliability of the pumping station an extremely important issue.
Usually, the first step in improving the reliability of a system is making a reliability model
of it. The problem encountered here was that classical reliability analysis was not sufficient
for making a reliability model for the pumping station. This was the first problem.
A reliability model of a system is the first step in the process of improving the reliability of
this system. It is certainly not the only step. Reliability implies making proper maintenance
to the working equipment. In the case of the cooling pumping station, this maintenance, it
was found, could not be easily conducted in practice. The performance of maintenance was
often antagonized by the operational needs of the consumers. This was the second problem.
The income of the pumping station is generated by charging its consumers for the cooling
seawater supplied. The charging is done by taking the readings of flow meters installed on
the pipelines for each consumer and including these readings in monthly bills. These flow
meters, like all machines, failed occasionally. This was the third problem
The aim of this thesis was to solve the above mentioned problems. Consequently, there
were three objectives of the thesis. These objectives were
1. to describe a model for the reliability of the pumping station,
2. to minimize the conflict between the operation of the pumping station and its need
for maintenance and
- 161 -
3. to reduce the amount of unaccounted-for-water lost by detecting flow meter drift,
a. by using SPC, and
b. by using ANN’s.
Chapter 2 examined the reliability of a cooling seawater pumping station. One incentive for
making this study was the consideration of the huge losses that might result as a
consequence of unreliable operation. The reliability of cooling seawater arriving to the
consumer at the required pressure and flow rate while observing the operational constraints
on the system was of interest.
The literature showed many works on pump reliability but no works on pumping station
reliability. In this chapter a method for modeling the reliability of a cooling pumping
station for the petrochemical industry was introduced. In this chapter, the mater of a
criterion for an ‘adequate performance’ of a system was shown to be an engineering and
managerial problem. It was explained in Chapter 2 that it was important to define the term
‘failure’ in the context of a cooling-petrochemical-pumping station. Certainly, when all the
pumps in the system are not working (as what would happen in an electrical blackout) this
would be a failure. Nevertheless, if the output of pumps is below the minimum requirement
for a consumer to operate, the operating pumps would be as useless as failing pumps.
Conversely, their might be enough operating pumps to satisfy more than the minimum
requirement for a plant to operate. Yet, the system of delivery might take a configuration
that would make operating these pumps degrading for the entire pumping system.
Therefore, reliability was considered as seawater arriving to the consumer at the required
flow and pressure. This unique definition influenced the modeling process where flow, not
an equipment or part of an equipment, was included in the reliability block diagram where
only equipment or parts of equipment are usually included. Hence, the reliability model
included flow, several components and their physical relationships, and a set of operational
constraints, called conditional parameters. All of these were used in the reliability block
diagrams and equations.
- 162 -
Another unique requirement for the reliability modeling of a pumping station is the
consideration of the many states in which cooling seawater is supplied to a consuming
plant. It was found that the reliability of the system was dramatically affected by header
section isolation or a line valve closure. In contrast to the typical reliability modeling which
usually requires one reliability block diagram, the many states in which water could be
supplied required as many reliability block diagrams. Therefore, the reliability of each
consumer in each case of header-section isolation and line valve closure was considered in
addition to the normal case of operation. A reliability model was developed and applied to
the actual data from the pumping station. To overcome the problems exhibited by some
categories of data in this research, some methods given in the literature were extended to
suit this need. The modeling process gave the reliability indices for all the consumers.
The reliability of the pumping station as a whole was also considered. Three propositions
for the reliability of the entire pumping station were discussed leading to a new measure of
reliability of the entire pumping station that was called APCRS.
As stressed in the previous paragraphs, the reliability of the cooling-petrochemical-
pumping station can not be overemphasized. A consequence of this is that the reliable
pump operation for providing the cooling seawater is extremely important. To ensure this
reliability, pumps should receive timely maintenance.
A very important issue in almost all industries is the coordination between maintenance and
production activities. Both of them are necessary: the operation of machinery would
produce revenue for the owner(s) while maintenance will keep these machines running.
Nevertheless, one function is usually performed at the expense of the other. If a machine is
stopped for maintenance, it is stopped from producing revenue. Similarly, if a machine is
operated continually without proper maintenance, it will eventually fail. Lack of
coordination, hence, results in degradation for both operation and maintenance. The
pumping station understudy suffered from a conflict between its production (operation) and
maintenance functions. The result was unreliable operation due to the failure of the
unmaintained machines and inconvenient maintenance that interrupted production. Chapter
- 163 -
3 presented a practical method of data analysis for minimizing the conflict between
operation and maintenance activities for the pumping station understudy.
It was thought that making an optimal schedule for the operation and maintenance of the
pumps supplying cooling seawater would minimize this conflict. The making of this
optimal schedule required knowing the seawater demand pattern first. Next, the
maintenance function could be done around this demand.
To analyze the cooling seawater demand, it was imperative to identify the major factors
that produced it. The demand for seawater at the pumping station depended on two things:
a plant’s aggregate production level (or capacity utilization) and the weather. Regression
analysis was used for relating the weather, production and seawater consumption. In
Chapter 3, the seawater consumption of several chemical plants was modeled using
regression analysis in relation to ambient air temperature Ta, seawater temperature Ts,
humidity H and capacity utilization (CU) or production (P). The purpose of modeling was
the prediction of future seawater consumption of the plants and to schedule the pump
operation and maintenance. Two examples were shown in detail for the use of this
modeling procedure. The results were tested and were found quite satisfactory.
For a pumping station to continue operating, it must make profit for its owners. Any factor
that decreases the profit increases the risk of jeopardizing the operation of the pumping
station and this risk is passed on to the petrochemical industry.
A pumping station makes its revenue by selling water to the consuming plants. The amount
of water sold to a consumer is measured by a flow meter. A flow meter is, like any
machine, susceptible to failure. The flow meter failure that jeopardizes the revenue of a
pumping station is called flow meter drift.
Chapter 4 was about the problem of flow meter drift and how it may be detected. Flow
meter drift is a problem that flow meter owners face every now and then. The physical
problem will have financial consequences if it happened to flow meters that are used in
- 164 -
billing. Flow-meter drift is a part of a wider phenomenon which is instrument drift. This
phenomenon is poorly understood and can happen at any time.
The purpose of Chapter 4 was to find a method to detect flow meter drift. The method
developed, SPC-CUSUM, had the following advantages
1. It did not involve tampering with flow meter circuitry and the risks involved in such
an option.
2. It would be a universal solution independent of the flow meter type (electromagnetic,
ultrasonic...etc.) or manufacturer.
3. It would be able to work with the data from the monthly bills i.e. it will needed
minimal data.
4. It would not require any knowledge about the mathematical relations and models
that governed the process.
5. It would be inexpensive.
The method presented (SPC-CUSUM) only needed the monthly billing data. From this data
only, the flow meters were investigated for drift. The nature of the data, nevertheless, was
an obstacle for applying the SPC-CUSUM. Therefore, before sending these data to be
processed by the SPC-CUSUM, they were transformed first to what was termed as a virtual
mean.
The theory of the virtual mean is that for every point in the seasonal time series (regardless
of its location) their exits a virtual point that represents the virtual mean corresponding to
this point. It is this virtual mean that would be processed by the cusum method.
In this project, when any deviation from normal consumption pattern is identified, it is
considered as an indication of a possible flow meter fault that would require the
Instrumentation Section to investigate and calibrate the flow meter in question to check for
its accuracy.
- 165 -
Adapting the process can be done inexpensively. The equations developed here can be
turned to simple algorithms in any spreadsheet software. The method developed works on
monthly bills. This would require the operation engineer or the accountant to type the flow
meter reading of every consuming plant and see the feedback from the software. This
process would only take few minutes every month i.e. it is not a time-consuming process.
Comparing the huge financial loss resulting from flow-meter drift with the inexpensive cost
of adapting the proposed method and the fact that it only uses billing data (which most
companies keep for accounting purposes) makes it, in this author’s opinion, attractive for
practical use.
The method was tested against a case that was known as a case of flow meter drift and it
succeeded in both detecting the drift and determining the time of its onset. Other case
studies were also presented. The limitations of the method were also mentioned.
In Chapter 5, an alternative method for detecting industrial water flow meter drift was
introduced. The method introduced was artificial neural networks. The reasons for
choosing artificial neural networks were cited from the literature and they have to do with
ANN’s characteristics such as
(i) nonlinearity which allows better fit to the data,
(ii) noise-insensitivity which provides accurate prediction in the presence of uncertain
data and measurement errors,
(iii) high parallelism which implies fast processing and hardware failure-tolerance,
(iv) learning and adaptivity which allow the system to update (modify) its internal structure
in response to changing environment, and
(v) generalization which enables the application of the model to unlearned data.
The artificial neural network presented in this chapter was the typical three layer neural
network. The input layer was made of 24 inputs. Seven of them were symbolic inputs while
the remaining 17 were numerical inputs. The numerical inputs represented the quantitative
aspect of consumption. There are many flow meters in the pumping station. Making a
simulation for every one of them would be exhaustive and computationally expensive. To
- 166 -
only make one simulation that is capable of being generalized to many flows (processes),
normalization was used. The symbolic inputs represented the qualitative aspect of
consumption. When the consumption validates any of the seven rules represented by the
seven symbolic inputs, the neuron for that particular rule is turned on (triggered). The
output layer had three outcomes: normal, drifting upwards and drifting downwards. The
network was trained and tested and the results were satisfactory.
6.2 The specific contributions this study has made to the existing body of knowledge
and industry practice.
In this section, the specific contributions this study has made to the existing body of
knowledge and industry practice are going to be mentioned. The significance of this
research lies in its focus on the pumping station of a refinery or a petrochemical complex
cooling system. Cooling water in a refinery or a petrochemical complex has received some
attention before. Nevertheless, the pumping station part of it, to the best of this candidate’s
knowledge, did not receive much of attention. In addition to shedding light on the role of
the pumping station of a petrochemical cooling system, other contributions of this study to
the existing body of knowledge and industry practice are mentioned below.
First, a method to model the reliability analysis of a refinery (or a petrochemical complex)
cooling pumping station was presented. It was found that this method needed to included
flow and the system constraints. It was also discovered, in the reliability analysis done, that
it several block diagrams were needed to model the reliability of a cooling pumping station
in contrast to one only in classical analysis. Another contribution this study made was that
it presented a measure, the APCRS, to consider the reliability of the entire pumping station.
These findings can be used in future studies on the reliability analysis of cooling water
pumping stations. They, also, can be extended to other problems that involve ‘flow’. The
flow here does not, necessarily, have to be of cooling water. It can be the flow of municipal
water, electricity, material…etc.
Second, this study explicitly mentioned the existence of a conflict between operation and
maintenance in a cooling petrochemical pumping station. This study, also, presented a
- 167 -
method to minimize this conflict. In doing so, the study exposed the factors that influence
the cooling water consumption of a plant. It was found that, in addition to a plant’s
production level, the weather conditions greatly influenced the amount of cooling water
consumed. A method to locally and cheaply predict the weather conditions was also
presented. All of these findings can be used in other situations where
1. there is a conflict between operational function and the maintenance needs in a
facility,
2. the cooling water consumption of a facility is analyzed,
3. the impact of the weather conditions on an engineering process needs to be
measured, and
4. a method is needed to, quickly and cheaply, predict some weather conditions.
Third, this study presented a method to detect flow meter drift without adding or using any
hardware on the flow meter. Rather, the method used the monthly bills’ data only. This
would make the presented method both universal (independent of flow meter type or
manufacturer) and cheap. In developing a method to detect flow meter drift, several other
contributions to the existing body of knowledge were made. These include
1. the use of SPC methods in the problems of flow
2. the development of the concept of the virtual mean of a seasonal time series. This
concept of virtual mean can be used in any statistical study involving a seasonal
time series. It can be used to detect its deviation from the normal pattern.
Fourth, this study used the concept of the virtual mean and an artificial neural network with
numerical and symbolic inputs to detect a flow meter drift.
In summary, the contributions this study has made to the existing body of knowledge and
industry practice are listed below
1. A method to model the reliability analysis of a refinery (or a petrochemical
complex) cooling system was presented.
2. Clearly mentioned the existence of a conflict between operation and maintenance in
a cooling petrochemical pumping station.
- 168 -
3. A method to minimize the conflict between operation and maintenance in a cooling
petrochemical pumping station was presented.
4. The factors influencing the cooling water consumption of a petrochemical plant
were exposed.
5. Explicitly demonstrated the effect of weather on an industrial process.
6. Measured the above mentioned weather effect.
7. A method to locally and cheaply predict the weather for an industrial facility was
presented.
8. Presented a method to detect flow meter drift without adding or using any hardware
on the flow meter. Rather, the method used the monthly bills’ data only.
9. The method developed, used the virtual mean. The virtual mean developed can be
used in any statistical study involving a seasonal time series to detect its deviation
from the normal pattern.
10. Used an artificial neural network with numerical and symbolic inputs to detect a
flow meter drift.
6.3 Conclusion
The common theme between chapters 2, 3, 4 and 5 is that every chapter presented a
statistical tool to solve a different problem of the cooling-seawater-pumping station. All of
these chapters supported the operation of the pumping station by the use of a statistical
method. Another minor common theme can be identified also. This theme is flow.
Flow was involved in all the four chapters. In Chapter 2, the reliability analysis of the
pumping station was done with the consideration of the minimum flow a plant needs to
operate and/or the maximum flow a pipe can withstand. In Chapter 3, minimizing the
conflict between operation and maintenance needed the prediction of the consumption
(which is directly related to flow) for each plant. Another discovery made in this chapter is
that this flow has a seasonal pattern. In Chapter 4, the seasonal pattern of flow (described in
Chapter 3) was used to detect drift in the flow meter. In Chapter 5 the seasonal pattern of
flow was exploited once again to detect flow meter drift. In this chapter, however, the tool
of detection differed: it was ANN’s while it was SPC in Chapter 4.
- 169 -
In this thesis, statistical tools were used to support the operation and maintenance of a
pumping station. Certainly, other tools and methods could be used for the same purpose.
The tools used and the methods developed, could be used in other applications.
- 170 -
References:
Abbott, J. C. (1999). SPC Practical Understanding of Capability by Implementing Statistical Process Control. Easley, S.C., Robert Houston Smith Publishers.
Abulma’atti, M. and I. Qamber (2001). "Reliability analysis of a redundant system with non-repairable units and Common-cause failures using the SPICE circuit simulation program." Proc. Natl. Sci. Counc. ROC (A) 25(5): 322-328.
Abushamsa, I. (2001). Unaccounted for Water—Using the Untapped Resource to Improve Financial and Technical Performance in Jordan. Water International. 26: 602-604.
Adam, E. E. and R. J. Ebert (1992). Production and Operations Management: Concepts, Models, and Behavior, Prentice Hall.
Al-Dakheel, D. (2004). Internal pipe rehabilitation. Proceedings of IMEC2004, Kuwait. Al-Ghaniam, A. (2003). "A ststistical approach linking energy management to maintenance
and production factors." Journal of Quality in Maintenance Engineering 9(1): 25-37. Alcocer-Yamanaka, V. H., V. Tzatchkov, et al. (2006). Instantaneous water demand
parameter estimation from coarse meter readings. 8th Annual Water Distribution Systems Analysis Symposium, Cincinnati, Ohio, USA.
AlHermi, K. A. (2007). The Lost Barrels (in Arabic) AlQabas. Kuwait. AlJasem, M. A. (2007). Good! The Pipe is not Punctured but the Flow meter is Fualty! (in
Arabic). Kuwait. 2009. AlJasem, M. A. (2007). The Petrolum Flow meter is Faulty! 2009. Almandoz, J., E. Cabrera, et al. (2005). "Leakage Assessment through Water Distribution
Network Simulation." Journal of Water Resources Planning and Management 131(6): 458-466.
Amacher, G. S., W. F. Hyde, et al. (1999). "Nepali Fuelwood Production and Consumption: Regional and Household Distinctions, Substitution and Successful Intervention." The Journal of Development Studies 35(4): 138-163.
Amsden, R. T., H. E. Butler, et al. (1998). SPC Simplified Practical Steps to Quality. New York, Productivity Press.
Anagun, A. S. (1998). "A neural network applied to pattern recognition in statistical process control." Computers & Industrial Engineering 35(1-2): 185-188.
Andersen, J. and R. Powell (2000). DMA structured state-estimation for demand monitoring. Water Resources 2000.
Anderson, S. W. (2001). "Direct and indirect effects of product mix charateristics on capacity management decisions and operating performance." The International Journal of Flexible Manufacturing Systems 13(3): 241-265.
Arbue´s, F., R. n. Barbera´n, et al. (2004). "Price impact on urban residential water demand: A dynamic panel data approach." Water Resources Research 40(W11402).
Arize, A. C. (2000). "U. S. petrolum consumption behaviore and oil price uncertainty: tests of cointegration and parameter instability." Atlantic Economic Journal 28(4): 463-477.
Arniella, E. F. (2007). Demand Allocation for the Gwinnett County, GA, Hydraulic and Water Quality
Model Setup. World Environmental and Water Resources Congress 2007: Restoring Our Natural Habitat.
Arregui, F., J. Enrique Cabrera, et al. (2005). Key Factors Affecting Water Meter Accuracy. Leakage 2005.
- 171 -
Ashareti, J., A. Teelen, et al. (1996). "A production and maintenance planning model for the process industry." International Journal of Production Research 34(12): 3311-3326.
ASQ (2009). Walter A. Shewhart Father of statistical quality control. AWWA (1987). Leaks in Water Distribution Systems: A Technical/Economic Overview.
Denver, Colorado, USA, American Water Works Association. Bagowsky, I. (1961). Reliability Theory and Practice, Prentice-Hall. Basheer, I. A. and M. Hajmeer (2000). "Artificial neural networks: fundamentals,
computing, design, and application." Journal of Microbiological Methods 43(1): 3-31.
Baum, M. C., M. D. Dukes, et al. (2003). Selection and Use of Water Meters for Irrigation Water Measurement. F. C. E. S. The Agricultural and Biological Engineering Department, Institute of Food and Agricultural Sciences, University of Florida. 2010.
Baxter, L. and M. Tortorella (1994). Dealing with real field reliability data: circumventing incompleteness by modeling & iteration. 1994 Proceedings Annual Reliability and Maintainability Symposium.
Beckwith, S. F. and K. P. Wong (1995). A genetic algorithm approach for electric pump scheduling in water supply system. IEEE International Conference on Evolutionary Computation, Perth, Australia.
Bevilacqua, M., M. Braglia, et al. (2003). "The classification and regression tree approach to pump failure rate analysis." Reliability Engineering & System Safety 79(1): 59-67.
Billinton, R. and R. Allan (1983). Reliability Evaluation of Engineering System: Concepts and Techniques, Pitman Advanced Publishing Program.
Biscos, C., M. Mulholland, et al. (2003). "Optimal operation of water distribution networks by predictive control using MINLP." Water SA 29(4): 393-404.
Bishop, C. M. (2005). Neural Networks for Pattern Recognition. Oxford, Oxford University Press.
Bolton, W. (2000). Newnes Instrumentation & Measurement Pocket Book. Oxford, Newnes.
Bowerman, B. L., R. T. O'Connell, et al. (2005). Forecasting, Time Series and Regression An Applied Approach, Thomson Broks/Cole.
Bowers, J. (2001). "The cost of weather in a floating oil production system." Journal of the Operational Research Society 52: 135-142.
Brandolese, M., M. Franci, et al. (1996). "Production and maintenance integrated planning." International Journal of Production Research 34(7): 2059-2075.
BuHamra, S., N. Smaoui, et al. (2003). "The Box-Jenkins analysis and neural networks: prediction and time series modeling." Applied Mathematical Modelling 27(10): 805-815.
Cassady, C. R. and E. Kutanoglu (2005). "Integrating preventive maintenance planning and production scheduling for a single machine." IEEE Transactions on Reliability 54(2): 304-309.
Castro, M. M., T. W. Song, et al. (2000). "Minimization of operational costs in cooling water systems." Chemical Engineering Research and Design 78(2): 192-201.
Çengel, Y. A. (1998). Heat Transfer A Practical Approch. Boston, McGraw-Hill.
- 172 -
Chen, S.-T., D. C. Yu, et al. (1992). "Weather sensitive shor-term load forecasting using nonfully connected artificial neural network." Transactions on Power Systems 7(3): 1098-1105.
Cheng, C.-S. (1995). "A multi-layer neural network model for detecting changes in the process mean." Computers & Industrial Engineering 28(1): 51-61.
Cheng, C.-S. (1997). "A neural network approach for the analysis of control chart patterns." International Journal of Production Research 35(3): 667-697.
Chiu, W. K. (1974). "The Econmic Design of Cusum Charts for Controlling Normal Means." Applied Statistics 23(3): 420-433.
Clark, R. M., M. Sivaganesan, et al. (2002). "Cost Models for Water Supply Distribution Systems." Journal of Water Resources Planning and Management 128(5): 312-321.
Cobb, L. (1998). A scale for measuring very rare events. Cortinovis, G. F., M. T. Ribeiro, et al. (2009). "Integrated analysis of cooling water
systems: Modeling and experimental validation." Applied Thermal Engineering 29(14-15): 3124-3131.
Coudert, T., B. Grabot, et al. (2002). "Production / maintenance cooperative scheduling using multi-agents and fuzzy logic." International Journal of Production Research 40(18): 4611-4632.
Crow, M., G. Gross, et al. (2003). "Power system basics for business professionals in our industry." IEEE Power & Energy Magazine 16-20.
Daniel_Measurement_And_Control (2010). How Today’s Ultrasonic Meter Diagnostics Solve Metering Problems Technical White Paper, Daniel_Measurement_And_Control: 16.
Delgado, M. A., J. Jaumandreu, et al. (1999). "Input cost, capacity utilization and substitution in the short-run." Spanish Economic Review 1(3).
DoD. (1991), Military Handbook Reliability Prediction of Electronic Equipment, MIL-HDBK-217F.
Dong, B., S. E. Lee, et al. (2005). "A holistic utility bill analysis method for baselining whole commercial building energy consumption in Singapore." Energy and Buildings(37): 167-174.
Encarta (2004). Calibration, Microsoft Corporation. EncartaEncyclopedia (2004). Petroleum. EncyclopediaBritanica (2010). petroleum refining. Ewan, W. D. (1963). "When and How to use Cu-Sum Charts." Technometrics 5(1): 1-22. Ferréol, E. (2005). How to Measure and Reduce the Water Meter Park Inefficiency?
Leakage 2005. Friedman, M. and A. Kandel (1999). Introduction to Pattern Recognition: Statistical,
Structural, Neural and Fuzzy Logic Approaches. London, World Scientific. Fulford, J. M. (2002). Model Performance of Water-Current Meters. Hydraulic
Measurements and Experimental Methods 2002, Estes Park, Colorado, USA. Galton, F. (1886). "Regression towards mediocrity in hereditary stature." Journal of
Anthropological Institute XV: 246-263. Ghatak, A. (1998). "Aggregate consumption functions for India: A cointegration analysis
under structural changes, 1919-86." Journal of Applied Statistics 25(4): 475-488. Grobler, L. J. and W. L. d. Heijer (2006). "The effect of data period selection for the
development of measurement and verification baselines for mine pumping projects." Energy Engineering 103(3): 41-55.
- 173 -
Gross, G. and F. D. Galiana (1987). "Short-term load forecasting." Proceedings of IEEE 75(12): 1558-1573.
Hamdy, A., R. Ragab, et al. (2003). "Coping with water scarcity: water saving and increasing water productivity." Irrigation and Drainage 52(1): 3-20.
Han, D. and F. Tsung (2006). "A Reference-Free Cuscore Chart of Dynamic Mean Change Detection and a Unified Framework for Charting Performance Comparison." Journal of the American Statistical Association 101(473): 368-386.
Han, D. and F. Tsung (2007). "Detection and Diagnosis of Unknown Abrupt Changes Using CUSUM Multi-Chart Schemes." Sequential Analysis 26(3): 225-249.
Han, D., F. Tsung, et al. (2007). "Cusum and EWMA Multi-Charts for Detecting a Range of Mean Shifts." Statistica Sinica 17(3): 1139-1164.
Han, D., F. Tsung, et al. (2007). "A Cusum Chart with Local Signal Amplification for Detecting a Range of Unknown Shifts." International Journal of Reliability, Quality and Safety Engineering 14(2): 81-97.
Harish, R., E. E. Subhramanyan, et al. (2010). "Theoretical model for evaluation of variable frequency drive for cooling water pumps in sea water based once through condenser cooling water systems." Applied Thermal Engineering 30(14-15): 2051-2057.
Harrison, P. (1991). Illustrated Dictionary of Physics, Librarie du Liban Publishers. Hassan, A., M. S. N. Baksh, et al. (2003). "Improved SPC chart pattern recognition using
statistical features." International Journal of Production Research 41(7): 1587-1603. Hauber-Davidson, G. and E. Idris (2006). Smart Water Metering. Water: 38-41. Herwig, J., D. Keese, et al. (1994). Magnetic flowmeter with fault detection U. S. P. Office.
U.S.A, Fischer & Porter Company. Ho, K., L., Y. Hsu, Y., et al. (1992). "Short term load forecasting using a multilayer neural
network with an adaptive learning algorithm." IEEE Transactions on Power Systems 7(1): 141-149.
Hoffmeister, J. A. (1988). "Reliability analysis of pumps for uranium solutions." IEEE Transactions on Reliability 37(2): 144-148.
Hopkins, D., P. F. Savage, et al. (1995). "Problems encountered during research into flow rate, pattern of water consumption and unaccounted-for water losses in urban areas." Flow Measurement and Instrumentation 6(3): 173-179.
Jain, A. K., J. Mao, et al. (1996). "Artificial neural networks: a tutorial." Computer 29(3): 31-44.
Jankovic´-Nisˇic´, B., C. e. Maksimovic´, et al. (2004). "Use of Flow Meters for Managing Water Supply Networks." Journal of Water Resources Planning and Management 130(2): 171-179.
Jaouadi, M., S. Msahli, et al. (2006). "Analysis of the modeling methodologies for predicting the sewing thread consumption." Innternational Journal of Clothing Science and Technology 18(1): 7-18.
Jarret, R. G., O. Hussain, et al. (201). Predictive models for pipeline failure and identification of pipes for replacement. Proceedings of Annual Seminar on Pipeline Asset Planning and Operations.
Jarrett, R. G., O. Hussain, et al. (2002). Prediction of failures in water mains and prioritisation for replacements. Proceedings CISRO/AWA 2nd Annual Seminar on Asset Management of Pipelines
Jarrett, R. G., O. Hussain, et al. (2003). Reliability assessment of water pipelines using limited data. Proceedings of Oz Water 2003-19th AWA Convention, Perth.
- 174 -
Johnson, N. L. (1961). "A Simple Theoretical Approach to Comulative Sum Control Charts." Journal of the American Statistical Association 56(296): 835-840.
Johnson, P. V. (1996). "Unaccounted-For Water Puzzle: More Than Just Leakage." FLORIDA WATER RESOURCES JOURNAL.
Jones, G. M. (2008). Pumping Station Design, Butterworth-Heinemann. Kang, D. and K. Lansey (2010). "Optimal Meter Placement for Water Distribution System
State Estimation." Journal of Water Resources Planning and Management 136(3): 337-347.
Keech, R. (2005). Flowmeter Fault Detection. U. S. P. Office, ABB Limited. 6,843,137 Kettle, H., R. Thompson, et al. (2004). "Empirical modeling of summer lake surface
temperatures in southwest Greenland." Limnology and Oceangraphy 49(1): 271-282.
Khotanzad, A., M. H. Davis, et al. (1996). "An artificial neural network hourly temperature forecaster with applications in load forecasting." IEEE Transactions on Power Systems 11(2): 870-876.
Khulief, Y. and H. Emara-Shabiak (2004). Model-based detection and isolation of leaks in pipelines with experimental verification. Proceedings of IMEC2004, Kuwait.
Kirkley, J., C. J. M. Paul, et al. (2002). "Capacity and capacity utilization in common-pool resource industries
" Enviromental and Resource Economics 22(1-2): 71-97. Koutras, M. V., S. Bersimis, et al. (2007). "Statistical process control using Shewhart
control charts with supplementary runs rule." Methodology and Computing in Applied Probability 9(2): 207-224.
Krakacier, O., Z. G. Goktolga, et al. (2006). "A regression analysis of the effect of energy use in agriculture." Energy Ploicy 34(18): 3796-3800
Lapide, L. (1997). "Development in demand forecasting from ancient Greeks to present." The Journal of Business Forecasting Methods & System 16(3): 32.
Lopez-Ibanez, M., T. D. Parsad, et al. (2005). Optimal pump scheduling: Representation and multiple objectives. Proceedings of the Eigth International Conference on Computing and Control for the Water Industry (CCWI2005), University of Exeter, UK.
Lucken, C. V., Baran, et al. (2004). "Pump scheduling optimization using asynchronus parallel evolutionary algorithm." CLEI Electronic Journal 7(2).
Lydell, B. O. Y. (2000). "Pipe failure probability-the Thomas paper revisited." Reliability Engineering & System Safety 68(3): 207-217.
Martorell, S., I. Marton, et al. (2010). Reliability of a safety system and probabilistic risk assessment, 4th International Conference on Safety & Environment in Process Industry, Florence, Italy.
Mehrotra, K., C. K. Mohan, et al. (1997). Elements of Artificial Neural Network. Cambridge, Massachusetts, The MIT Press.
Meshkati, S. and J. Groot (1993). A study of the 1991 unaccounted-for gas volume at the Southern California Gas Company, vol. III Measurement.
Montgomery, D. C. (2005). Introduction to Statistical Quality Control, John Wiley &Sons. Inc.
Montgomry, D. C. (2005). Introduction to Statistical Quality Control, John Wiley &Sons. Inc.
- 175 -
Moon, Y. B., C.K. Divers, et al. (1998), “AEWS: an integrated, knowledge-based system with neural networks for reliability prediction”, Computers in Industry, 35(2):101-108.
Nagar, A. K. and R. S. Powell (2000). Observability Analysis of Water Distribution Systems under Parametric and Measurement Uncertainty. Building Partnerships - 2000 Joint Conference on Water Resource Engineering and Water Resources Planning & Management.
Nakamura, M. (2001). "Decision for maintenance-intervals of equipment in thermal power stations, based on few data." IEEE Transactions on Reliability 50(4): 360-364.
Nakamura, M., N. Nanayakkara, et al. (1992). "Reliability analysis of submarine power cables and determination of external mechanical protections." IEEE Transactions on Power Delivery 7(2): 895-902.
Nakashima, K. and K. Yamoto (1984). Reliability improvement of multi-output systems by redundancy technique. 8th Advances in Reliability Technology Symposium.
NeuroDimension, Inc. NeuroSolutionsTM The Neural Network Simulation Enviroment. NeuroSolutions (2009). What is a Neural Network? . Nilsson, U. R. C. (1998). "A new method for finding inaccurate gas flow meters using
billing data: Finding faulty meters using billing data." Flow Measurement and Instrumentation 9(?): 237-242.
Norton, M. (2005). A Quick Course In Statistical Process Control. Upper Saddle River, New Jersy, Pearson Prentice Hall.
Noss, R. R., G. J. Newman, et al. (1987). "Optimal testing frequency for domestic water meters." Journal of. Water Resources Planning and Management 113(1): 1-14.
Oury, B. (1965). "Allowing for weather in crop production model building." Journal of Farm Economics 47(2): 270-283.
Pacella, M., Q. Semeraro, et al. (2004). "Manufacturing quality control by means of a Fuzzy ART network trained on natural process data." Engineering Applications of Artificial Intelligence 17(1): 83-96.
Page, E. S. (1954). "Continuous Inspection Schemes." Biometrika 41(1/2): 100-115. Page, E. S. (1961). "Cumulative Sum Charts." Technometrics 3(1): 1-9. Page, E. S. (1962). "Cumulative Sum Schemes Using Gauging." Technometrics 4(1): 97-
109. Palaneeswaran, E., P. E. D. Love, et al. (2008). "Mapping rework causes and effects using
artificial neural networks." Building Research & Information 36(5): 450-465. Pao, H.-T. (2006). "Modeling and Forecasting the Energy Consumption in Taiwan Using
Artificial Neural Networks." Journal of American Academy of Business 8(1): 113-119.
Parkash, S. (2003). Refinery Water Systems. Refining Processes Handbook, Elsevier 242-269
Parlos, A. G., E. Oufi, et al. (1996). Development of an intelligent long-term electric load forecasting system. International Conference on Intelligent Systems Applications to Power Systems, 1996. Proceedings, ISAP '96, Orlando, FL
Pearson, K. (1930). The Life, Letters and Labours of Francis Galton. Cambridge. Ponce-Ortega, J. M., M. Serna-Gonzalez, et al. (2009). "A disjunctive programming model
for the optimal design of cooling water systems " Computer Aided Chemical Engineering 26: 1257-1262.
- 176 -
Principe, J. C., N. R. Euliano, et al. (2000). Neural and Adaptive Systems: Fundamentals Through Simulations. New York, John Wiley & Sons.
Proctor, C. L., M. Savsar, et al. (1985). Fault tree application to a system with redundunt three-state devices. Current Advances in Mechanical Design & Production Third Cairo University MDP Conference, Cairo.
Ramamoorty, M. and R. Gupta (1976). "Piecewise method for reliability evaluation of large power systems." J. Inst. Eng (India) Electr. Eng Div. 57(El 1): 35-37.
Roberts, J. Getting a Water Meter to Work for You. Rudolph, S. and B.-H. Kr¨oplin (1997). On Modeling Assumptions and Artificial Neural
Networks. Workshop Proceedings Selbstorganisation von adaptivem Verhalten (SOAVE’97), University of Ilmenau, Germany.
Saila, S., M. Cheeseman, et al. (2004). Maximum stream temperature estimation from air temperature data and its relationship to Brook Trout habitat requirements in Rhode Island.
Sharp, B. B. and D. B. Sharp (1995). Cooling water systems. Water Hammer Practicl Solutions, Elsevier.
Shayan, E. (1986). "A probabilistic model of a coal-burning power plant." IEEE Transactions on Reliability R-35(5): 488-493.
Sivillo, J. K. and D. P. Reilly (2004/2005). "Forecasting consumer product demand with weather information: a case study." The Journal of Business Forecasting Methods & Systems 23(4): 22-29.
Smaoui, N., S. BuHamra, et al. (2002). A combination of Box-Jenkins analysis and neural networks to model and predict water consumption in Kuwait. Proceedings of the 2002 International Joint Conference on Neural Networks.
Smith, D. G. C. (1989). "Combination of forecasts in electricity demand prediction." Journal of Forecasting 8(3): 349-356.
Smith, J. B. (1986). Proceedings: AWWa Seminar on Leak Detection and Water Loss Reduction.
Stathis, J. A. and G. V. Loganathan (1999). Analysis of Pressure-Dependent Leakage in Water Distribution Systems. Proceedings of WRPMD.
Still, M., H. Venzke, et al. (1998). "Influence of humidity on the convective heat transfer from small cylinders." Experiments in Fluids 24(2): 141-150.
Sydenham, P. H., N. H. Hancock, et al. (1989). Introduction to Measurement Science and Engineering. New York, John Wiley & Sons.
Tam, A. S. B., W. M. Chan, et al. (2006). "Maintenance scheduling to support the operation of manufacturing and production assets " The International Journal of Advanced Manufacturing Technology 34(3-4): 399-405.
Tamarkin, T. D. (1992). Automatic Meter Reading. Public Power Magazine. 50. Taylor, H. M. (1968). "The Economic Design of Cumulative Sum Control Charts."
Technometrics 10(3): 479-488. Thomas, H. M. (1981). "Pipe and vessel failure probability." Reliability Engineering 2: 83-
124. US_Army_Corps_of_Engineers (1994). Mechanical and Electrical Design of Pumping
Stations. D. o. t. Army, The United States Goverment: 171. van_Vugt, M. and C. D. Samuelson (1999). "The Impact of Personal Metering in the
Management of a Natural Resource Crisis:A Social Dilemma Analysis." Personality and Social Psychology Bulletin 25(6): 735-750.
- 177 -
Walski, T. M., S. Lowry, et al. (2000). Pitfalls in Calibrating an EPS Model. Proceedings of Joint Conference on Water Resource Engineering and Water Resources Planning and Management 2000
Wang, G., D. Dong, et al. (1993). "Leak detection for transport pipelines based on autoregressive modeling." IEEE Transactions on Instrumentation and Measurement 42(1): 68-71.
Wheeler, D. J. (2004). Advanced Topics in Statistical Process Control the Power of Shewharts Charts. Knoxville, Tennessee, SPC Press.
Wheeler, D. J. and D. S. Chambers (1992). Understanding Statistical Process Control. Knoxville, Tennessee, SPC Press.
Yasui, S., Y. Ojima, et al. (2006). Generalization of the Run Rules for the Shewhart Control Charts Frontiers in Statistical Quality Control 8. H.-J. Lenz and P.-T. Wilrich. New York, Physica-Verlag: 207-219.
Yin, J. and C. Silio (1994). "K-terminal reliability in ring networks." IEEE Transactions on Reliability 43(3): 389-401.
Zagaitov, I. (1982). "Fluctuations in weather conditions and the planing of agricultural production." Problems of Economics 25(8): 55-73.
Zhao, Y., F. Tsung, et al. (2005). "Dual CUSUM control schemes for detecting a range of mean shifts." IIE Transactions 37(11): 1047-1057.
178
Appendix I
Table A. 1The tabular CUSUM method as applied to C7 Case #1.
Month Consumption Xt-(X(t-12)) Xt+X(t-6)
2 F(Xbar) Ci+ N+ Ci- N- StatusMar-99 12,046,300 Apr-99 12,151,500 May-99 12,783,100 Jun-99 12,506,400 Jul-99 13,446,300 Aug-99 14,095,300 Sep-99 12,695,200 Oct-99 13,348,600 Nov-99 12,361,700 Dec-99 10,362,000 Jan-00 11,768,700 Feb-00 10,615,000 Mar-00 12,069,800 23,500 12,382,500 12,406,000 0 0 0 0 OK Apr-00 12,238,400 86,900 12,793,500 12,880,400 0 0 227,716 1 OK May-00 12,847,700 64,600 12,604,700 12,669,300 0 0 666,532 2 OK Jun-00 13,376,000 869,600 11,869,000 12,738,600 0 0 1,036,048 3 OK Jul-00 13,897,700 451,400 12,833,200 13,284,600 0 0 859,565 4 OK Aug-00 15,004,100 908,800 12,809,550 13,718,350 0 0 249,331 5 OK Sep-00 14,279,400 1,584,200 13,174,600 14,758,800 734,808 1 0 0 OK Oct-00 13,940,400 591,800 13,089,400 13,681,200 392,016 2 0 0 OK Nov-00 12,888,000 526,300 12,867,850 13,394,150 0 0 0 0 OK Dec-00 12,825,600 2,463,600 13,100,800 15,564,400 1,540,408 1 0 0 OK Jan-01 12,233,800 465,100 13,065,750 13,530,850 1,047,266 2 0 0 OK Feb-01 11,519,300 904,300 13,261,700 14,166,000 1,189,273 3 0 0 OK Mar-01 13,684,700 1,614,900 13,982,050 15,596,950 2,762,231 4 0 0 OK Apr-01 13,133,600 895,200 13,537,000 14,432,200 3,170,439 5 0 0 OK May-01 13,816,100 968,400 13,352,050 14,320,450 3,466,897 6 0 0 OK Jun-01 13,490,000 114,000 13,157,800 13,271,800 2,714,705 7 0 0 OK Jul-01 13,538,100 -359,600 12,885,950 12,526,350 1,217,062 8 581,766 1 OK Aug-01 13,324,800 -1,679,300 12,422,050 10,742,750 0 0 2,947,132 2 OK Sep-01 12,537,700 -1,741,700 13,111,200 11,369,500 0 0 4,685,748 3 AlarmOct-01 14,092,700 152,300 13,613,150 13,765,450 0 0 4,028,415 4 OK Nov-01 13,293,200 405,200 13,554,650 13,959,850 0 0 3,176,681 5 OK Dec-01 9,631,400 -3,194,200 11,560,700 8,366,500 0 0 7,918,297 6 AlarmJan-02 13,243,200 1,009,400 13,390,650 14,400,050 376,058 1 6,626,363 7 AlarmFeb-02 11,900,000 380,700 12,612,400 12,993,100 0 0 6,741,379 8 Alarm
179
Table A. 2 The CUSUM method as applied to C7 Case#2
Month ConsumptionXt-(X(t-
12)) Xt+X(t-6)
2 F(Xbar) Ci+ N+ Ci- N- Status Jun-05 11,402,400 Jul-05 14,059,300
Aug-05 14,001,200 Sep-05 10,741,200 Oct-05 12,481,200 Nov-05 11,488,700 Dec-05 10,784,400 Jan-06 10,374,400 Feb-06 8,807,800 Mar-06 11,201,000 Apr-06 10,889,300 May-06 12,068,300 Jun-06 11,245,300 -157,100 10,809,850 10,652,750 0 0 1,163,809 1 OK Jul-06 13,390,600 -668,700 11,099,200 10,430,500 0 0 2,549,868 2 OK
Aug-06 14,045,700 44,500 12,623,350 12,667,850 155,882 1 1,698,577 3 OK Sep-06 13,362,200 2,621,000 12,125,750 14,746,750 2,390,664 2 0 0 OK Oct-06 13,536,200 1,055,000 12,802,250 13,857,250 3,735,945 3 0 0 Alarm Nov-06 4,300,100 -7,188,600 7,772,700 584,100 0 0 0 0 OK Dec-06 10,205,300 -579,100 11,797,950 11,218,850 0 0 597,709 1 OK Jan-07 11,741,000 1,366,600 12,893,350 14,259,950 1,747,982 1 0 0 OK Feb-07 10,502,700 1,694,900 11,932,450 13,627,350 2,863,364 2 0 0 OK Mar-07 11,849,200 648,200 12,692,700 13,340,900 3,692,295 3 0 0 Alarm Apr-07 11,026,600 137,300 7,663,350 7,800,650 0 0 4,015,909 1 Alarm May-07 14,746,400 2,678,100 12,475,850 15,153,950 2,641,982 1 678,518 2 OK Jun-07 13,367,400 2,122,100 12,554,200 14,676,300 4,806,314 2 0 0 Alarm Jul-07 15,026,000 1,635,400 12,764,350 14,399,750 6,694,095 3 0 0 Alarm
Aug-07 15,641,100 1,595,400 13,745,150 15,340,550 9,522,677 4 0 0 Alarm Sep-07 15,812,700 2,450,500 13,419,650 15,870,150 12,880,859 5 0 0 Alarm
Table A. 3 the recorded consumption of C6.
Month Consumption Month Consumption Month Consumption Month ConsumptionJan-05 35,129,000 Jan-06 39,061,800 Jan-07 31,358,000 Jan-08 31,181,800 Feb-05 31,903,200 Feb-06 30,716,600 Feb-07 24,931,700 Feb-08 32,188,900 Mar-05 34,606,900 Mar-06 11,355,000 Mar-07 32,876,800 Mar-08 34,653,100 Apr-05 35,887,700 Apr-06 25,660,400 Apr-07 32,867,400 Apr-08 33,987,200 May-05 43,108,300 May-06 42,155,600 May-07 36,913,500 May-08 42,439,300 Jun-05 43,108,300 Jun-06 41,178,000 Jun-07 42,308,100 Jun-08 41,927,400 Jul-05 40,000,000 Jul-06 41,517,400 Jul-07 43,696,100 Jul-08 43,864,500
Aug-05 44,893,200 Aug-06 41,776,100 Aug-07 43,485,400 Aug-08 43,272,500 Sep-05 41,564,000 Sep-06 40,785,300 Sep-07 41,621,800 Sep-08 38,750,100 Oct-05 42,644,400 Oct-06 40,686,400 Oct-07 42,496,400 Oct-08 36,687,300 Nov-05 40,791,700 Nov-06 34,954,200 Nov-07 39,358,400 Nov-08 33,510,300 Dec-05 41,854,300 Dec-06 33,429,800 Dec-07 33,920,200
180
Table A. 4 the CUSUM method as applied to C6 (Case Study #3).
Month Consumption Xt-(X(t-12)) Xt+X(t-6)
2 F(Xbar) Ci+ N+ Ci- N- Status
Jan-05 35,129,000
Feb-05 31,903,200
Mar-05 34,606,900
Apr-05 35,887,700
May-05 43,108,300
Jun-05 43,108,300
Jul-05 40,000,000
Aug-05 44,893,200
Sep-05 41,564,000
Oct-05 42,644,400
Nov-05 40,791,700
Dec-05 41,854,300
Jan-06 39,061,800 3,932,800 39,530,900 43,463,700 6,270,378 1 0 0 OK
Feb-06 30,716,600 -1,186,600 37,804,900 36,618,300 5,695,356 2 0 0 OK
Mar-06 11,355,000 -23,251,900 26,459,500 3,207,600 0 0 22,992,644 1 OK
Apr-06 25,660,400 -10,227,300 34,152,400 23,925,100 0 0 25,267,789 2 OK
May-06 42,155,600 -952,700 41,473,650 40,520,950 3,327,628 1 10,947,083 3 OK
Jun-06 41,178,000 -1,930,300 41,516,150 39,585,850 5,720,156 2 0 0 OK
Jul-06 41,517,400 1,517,400 40,289,600 41,807,000 10,333,833 3 0 0 OK
Aug-06 41,776,100 -3,117,100 36,246,350 33,129,250 6,269,761 4 0 0 OK
Sep-06 40,785,300 -778,700 26,070,150 25,291,450 0 0 908,794 1 OK
Oct-06 40,686,400 -1,958,000 33,173,400 31,215,400 0 0 0 0 OK
Nov-06 34,954,200 -5,837,500 38,554,900 32,717,400 0 0 0 0 OK
Dec-06 33,429,800 -8,424,500 37,303,900 28,879,400 0 0 0 0 OK
Jan-07 31,358,000 -7,703,800 36,437,700 28,733,900 0 0 0 0 OK
Feb-07 24,931,700 -5,784,900 33,353,900 27,569,000 0 0 0 0 OK
Mar-07 32,876,800 21,521,800 36,831,050 58,352,850 21,159,528 1 0 0 OK
Apr-07 32,867,400 7,207,000 36,776,900 43,983,900 27,950,106 2 0 0 OK
May-07 36,913,500 -5,242,100 35,933,850 30,691,750 21,448,533 3 0 0 OK
Jun-07 42,308,100 1,130,100 37,868,950 38,999,050 23,254,261 4 0 0 OK
Jul-07 43,696,100 2,178,700 37,527,050 39,705,750 25,766,689 5 0 0 OK
Aug-07 43,485,400 1,709,300 34,208,550 35,917,850 24,491,217 6 0 0 OK
Sep-07 41,621,800 836,500 37,249,300 38,085,800 25,383,694 7 0 0 OK
Oct-07 42,496,400 1,810,000 37,681,900 39,491,900 27,682,272 8 0 0 OK
Nov-07 39,358,400 4,404,200 38,135,950 42,540,150 33,029,100 9 0 0 OK
Dec-07 33,920,200 490,400 38,114,150 38,604,550 34,440,328 10 0 0 OK
Jan-08 31,181,800 -176,200 37,438,950 37,262,750 34,509,756 11 0 0 OK
Feb-08 32,188,900 7,257,200 37,837,150 45,094,350 42,410,783 12 0 0 OK
Mar-08 34,653,100 1,776,300 38,137,450 39,913,750 45,131,211 13 0 0 OK
Apr-08 33,987,200 1,119,800 38,241,800 39,361,600 47,299,489 14 0 0 OK
May-08 42,439,300 5,525,800 40,898,850 46,424,650 56,530,817 15 0 0 Alarm
Jun-08 41,927,400 -380,700 37,923,800 37,543,100 56,880,595 16 0 0 Alarm
Jul-08 43,864,500 168,400 37,523,150 37,691,550 57,378,822 17 0 0 Alarm
Aug-08 43,272,500 -212,900 37,730,700 37,517,800 57,703,300 18 0 0 Alarm
181
Sep-08 38,750,100 -2,871,700 36,701,600 33,829,900 54,339,878 19 0 0 OK
Oct-08 36,687,300 -5,809,100 35,337,250 29,528,150 46,674,706 20 0 0 OK
Nov-08 33,510,300 -5,848,100 37,974,800 32,126,700 41,608,083 21 0 0 OK
Table A. 5 The recorded consumption of C12. Month Consumption Month Consumption Month Consumption Month Consumption Month Consumption Jan-04 16,817,508 Jan-05 15,778,552 Jan-06 18,210,020 Jan-07 22,741,112 Jan-08 18,586,944 Feb-04 15,325,728 Feb-05 13,535,592 Feb-06 14,106,084 Feb-07 18,936,452 Feb-08 12,863,256 Mar-04 17,341,908 Mar-05 14,451,176 Mar-06 15,524,356 Mar-07 21,477,400 Mar-08 11,822,276 Apr-04 18,793,576 Apr-05 15,783,244 Apr-06 22,679,840 Apr-07 19,044,828 Apr-08 13,835,420 May-04 21,909,984 May-05 20,213,504 May-06 24,256,536 May-07 22,028,480 May-08 11,237,800 Jun-04 22,397,308 Jun-05 23,904,912 Jun-06 21,389,724 Jun-07 15,910,020 Jun-08 18,531,468 Jul-04 24,078,792 Jul-05 22,475,729 Jul-06 21,854,508 Jul-07 15,135,196 Jul-08 15,591,056 Aug-04 29,638,720 Aug-05 26,174,883 Aug-06 22,000,972 Aug-07 17,155,424 Aug-08 16,773,624 Sep-04 22,360,784 Sep-05 23,780,068 Sep-06 19,734,000 Sep-07 25,650,888 Sep-08 18,184,536 Oct-04 24,223,876 Oct-05 22,971,020 Oct-06 19,605,752 Oct-07 27,569,916 Oct-08 18,719,516 Nov-04 14,529,284 Nov-05 21,738,772 Nov-06 7,103,964 Nov-07 24,044,292 Nov-08 15,121,948 Dec-04 8,826,940 Dec-05 23,851,368 Dec-06 9,196,044 Dec-07 19,876,232
.
182
Table A. 6 the CUSUM method as applied to C12
Month Consumption Xt-(X(t-12)) Xt+X(t-6) 2 F(Xbar) Ci+ N+ Ci- N- Status
Jan-04 16,817,508Feb-04 15,325,728Mar-04 17,341,908Apr-04 18,793,576
May-04 21,909,984Jun-04 22,397,308Jul-04 24,078,792
Aug-04 29,638,720Sep-04 22,360,784Oct-04 24,223,876Nov-04 14,529,284Dec-04 8,826,940Jan-05 15,778,552 -1,038,956 19,928,672 18,889,716 0 0 0 0 OKFeb-05 13,535,592 -1,790,136 21,587,156 19,797,020 0 0 0 0 OKMar-05 14,451,176 -2,890,732 18,405,980 15,515,248 0 0 1,483,302 1 OKApr-05 15,783,244 -3,010,332 20,003,560 16,993,228 0 0 1,488,625 2 OK
May-05 20,213,504 -1,696,480 17,371,394 15,674,914 0 0 2,812,261 3 OKJun-05 23,904,912 1,507,604 16,365,926 17,873,530 0 0 1,937,281 4 OKJul-05 22,475,729 -1,603,063 19,127,141 17,524,078 0 0 1,411,754 5 OK
Aug-05 26,174,883 -3,463,837 19,855,238 16,391,401 0 0 2,018,904 6 OKSep-05 23,780,068 1,419,284 19,115,622 20,534,906 0 0 0 0 OKOct-05 22,971,020 -1,252,856 19,377,132 18,124,276 0 0 0 0 OKNov-05 21,738,772 7,209,488 20,976,138 28,185,626 4,449,758 1 0 0 OKDec-05 23,851,368 15,024,428 23,878,140 38,902,568 19,616,458 2 0 0 OKJan-06 18,210,020 2,431,468 20,342,875 22,774,343 18,654,933 3 0 0 OKFeb-06 14,106,084 570,492 20,140,484 20,710,976 15,630,040 4 0 0 OKMar-06 15,524,356 1,073,180 19,652,212 20,725,392 12,619,564 5 0 0 OKApr-06 22,679,840 6,896,596 22,825,430 29,722,026 18,605,722 6 0 0 OK
May-06 24,256,536 4,043,032 22,997,654 27,040,686 21,910,540 7 0 0 OKJun-06 21,389,724 -2,515,188 22,620,546 20,105,358 18,280,030 8 0 0 OKJul-06 21,854,508 -621,221 20,032,264 19,411,043 13,955,205 9 0 0 OK
Aug-06 22,000,972 -4,173,911 18,053,528 13,879,617 4,098,954 10 3,118,933 1 OKSep-06 19,734,000 -4,046,068 17,629,178 13,583,110 0 0 6,534,374 2 OKOct-06 19,605,752 -3,365,268 21,142,796 17,777,528 0 0 5,755,396 3 OKNov-06 7,103,964 -14,634,808 15,680,250 1,045,442 0 0 21,708,504 4 OKDec-06 9,196,044 -14,655,324 15,292,884 637,560 0 0 38,069,495 5 AlarmJan-07 22,741,112 4,531,092 22,297,810 26,828,902 3,093,034 1 28,239,143 6 OKFeb-07 18,936,452 4,830,368 20,468,712 25,299,080 4,656,246 2 19,938,613 7 OKMar-07 21,477,400 5,953,044 20,605,700 26,558,744 7,479,122 3 10,378,420 8 OKApr-07 19,044,828 -3,635,012 19,325,290 15,690,278 0 0 11,686,692 9 OK
May-07 22,028,480 -2,228,056 14,566,222 12,338,166 0 0 16,347,076 10 OKJun-07 15,910,020 -5,479,704 12,553,032 7,073,328 0 0 26,272,299 11 OKJul-07 15,135,196 -6,719,312 18,938,154 12,218,842 0 0 31,052,007 12 OK
Aug-07 17,155,424 -4,845,548 18,045,938 13,200,390 0 0 34,850,167 13 AlarmSep-07 25,650,888 5,916,888 23,564,144 29,481,032 5,745,164 1 22,367,686 14 OKOct-07 27,569,916 7,964,164 23,307,372 31,271,536 13,280,832 2 8,094,700 15 OKNov-07 24,044,292 16,940,328 23,036,386 39,976,714 29,521,678 3 0 0 OKDec-07 19,876,232 10,680,188 17,893,126 28,573,314 34,359,124 4 0 0 AlarmJan-08 18,586,944 -4,154,168 16,861,070 12,706,902 23,330,158 5 4,291,648 1 OKFeb-08 12,863,256 -6,073,196 15,009,340 8,936,144 8,530,434 6 12,354,055 2 OKMar-08 11,822,276 -9,655,124 18,736,582 9,081,458 0 0 20,271,147 3 OKApr-08 13,835,420 -5,209,408 20,702,668 15,493,260 0 0 21,776,437 4 OK
May-08 11,237,800 -10,790,680 17,641,046 6,850,366 0 0 31,924,622 5 OKJun-08 18,531,468 2,621,448 19,203,850 21,825,298 0 0 27,097,874 6 OKJul-08 15,591,056 455,860 17,089,000 17,544,860 0 0 26,551,564 7 OK
Aug-08 16,773,624 -381,800 14,818,440 14,436,640 0 0 29,113,475 8 OKSep-08 18,184,536 -7,466,352 15,003,406 7,537,054 0 0 38,574,971 9 AlarmOct-08 18,719,516 -8,850,400 16,277,468 7,427,068 0 0 48,146,453 10 AlarmNov-08 15,121,948 -8,922,344 13,179,874 4,257,530 0 0 60,887,474 11 Alarm
183
Table A. 7 the recorded consumption of consumer C3. Month Consumption Month Consumption Month Consumption Month Consumption Month Consumption
Jan-04 2,915,800 Jan-05 4,051,000 Jan-06 3,220,200 Jan-07 3,144,300 Jan-08 4,214,100
Feb-04 3,563,600 Feb-05 3,687,600 Feb-06 2,960,100 Feb-07 2,480,000 Feb-08 3,331,900
Mar-04 2,659,600 Mar-05 3,626,000 Mar-06 3,660,600 Mar-07 3,265,900 Mar-08 3,324,600
Apr-04 2,896,000 Apr-05 3,691,900 Apr-06 1,846,700 Apr-07 3,172,400 Apr-08 3,640,100
May-04 3,716,500 May-05 3,760,800 May-06 2,139,300 May-07 3,197,300 May-08 4,073,500
Jun-04 3,530,900 Jun-05 3,760,800 Jun-06 3,707,100 Jun-07 2,888,800 Jun-08 3,669,600
Jul-04 3,657,300 Jul-05 3,454,600 Jul-06 3,986,500 Jul-07 3,125,500 Jul-08 3,329,400
Aug-04 3,940,700 Aug-05 3,717,000 Aug-06 4,225,900 Aug-07 2,666,500 Aug-08 2,754,100
Sep-04 3,811,700 Sep-05 3,653,700 Sep-06 3,914,600 Sep-07 3,095,500 Sep-08 2,034,900
Oct-04 3,753,900 Oct-05 3,168,200 Oct-06 3,775,200 Oct-07 3,740,600 Oct-08 3,091,400
Nov-04 3,562,000 Nov-05 3,857,100 Nov-06 1,993,300 Nov-07 3,705,900 Nov-08 2,937,700
Dec-04 3,785,300 Dec-05 3,493,200 Dec-06 3,217,300 Dec-07 3,907,300
Table A. 8 the CUSUM method as applied on C3.
Month Consumption Xt-(X(t-12)) Xt+X(t-6) 2
F(Xbar) Ci+ N+ Ci- N- Status
Jan-04 2,915,800 Feb-04 3,563,600 Mar-04 2,659,600 Apr-04 2,896,000 May-04 3,716,500 Jun-04 3,530,900 Jul-04 3,657,300
Aug-04 3,940,700 Sep-04 3,811,700 Oct-04 3,753,900 Nov-04 3,562,000 Dec-04 3,785,300 Jan-05 4,051,000 1,135,200 3,854,150 4,989,350 794,813 1 0 0 OK
Feb-05 3,687,600 124,000 3,814,150 3,938,150 538,425 2 0 0 OK Mar-05 3,626,000 966,400 3,718,850 4,685,250 1,029,138 3 0 0 OK
Apr-05 3,691,900 795,900 3,722,900 4,518,800 1,353,401 4 0 0 OK May-05 3,760,800 44,300 3,661,400 3,705,700 864,563 5 0 0 OK Jun-05 3,760,800 229,900 3,773,050 4,002,950 672,976 6 0 0 OK
Jul-05 3,454,600 -202,700 3,752,800 3,550,100 28,539 7 27,704 1 OK Aug-05 3,717,000 -223,700 3,702,300 3,478,600 0 0 126,909 2 OK Sep-05 3,653,700 -158,000 3,639,850 3,481,850 0 0 222,863 3 OK
Oct-05 3,168,200 -585,700 3,430,050 2,844,350 0 0 956,317 4 OK Nov-05 3,857,100 295,100 3,808,950 4,104,050 0 0 430,072 5 OK Dec-05 3,493,200 -292,100 3,627,000 3,334,900 0 0 672,976 6 OK
Jan-06 3,220,200 -830,800 3,337,400 2,506,600 0 0 1,744,181 7 OK Feb-06 2,960,100 -727,500 3,338,550 2,611,050 0 0 2,710,935 8 OK Mar-06 3,660,600 34,600 3,657,150 3,691,750 0 0 2,596,989 9 OK
Apr-06 1,846,700 -1,845,200 2,507,450 662,250 0 0 5,512,544 10 Alarm
184
May-06 2,139,300 -1,621,500 2,998,200 1,376,700 0 0 7,713,648 11 Alarm Jun-06 3,707,100 -53,700 3,600,150 3,546,450 0 0 7,745,002 12 Alarm
Jul-06 3,986,500 531,900 3,603,350 4,135,250 0 0 7,187,557 13 Alarm Aug-06 4,225,900 508,900 3,593,000 4,101,900 0 0 6,663,461 14 Alarm
Table A. 9 the recorded consumption for C5. Month Consumption Month Consumption Month Consumption Month Consumption Month Consumption Jan-04 1,487,800 Jan-05 1,431,400 Jan-06 1,542,100 Jan-07 1,479,900 Jan-08 1,521,300
Feb-04 1,332,100 Feb-05 452,600 Feb-06 1,501,700 Feb-07 1,336,200 Feb-08 1,542,400 Mar-04 1,501,700 Mar-05 1,385,800 Mar-06 1,744,800 Mar-07 1,371,300 Mar-08 1,643,100
Apr-04 1,615,100 Apr-05 1,706,900 Apr-06 1,771,600 Apr-07 1,519,600 Apr-08 1,683,300 May-04 2,403,100 May-05 2,477,100 May-06 2,571,000 May-07 2,113,500 May-08 2,336,100 Jun-04 2,428,200 Jun-05 2,477,100 Jun-06 2,636,800 Jun-07 2,503,800 Jun-08 2,470,800
Jul-04 3,120,400 Jul-05 3,006,500 Jul-06 2,876,200 Jul-07 2,687,200 Jul-08 2,881,100 Aug-04 2,239,900 Aug-05 3,196,600 Aug-06 2,968,300 Aug-07 2,857,000 Aug-08 2,911,600 Sep-04 3,994,200 Sep-05 3,028,100 Sep-06 2,926,900 Sep-07 2,791,200 Sep-08 2,733,600
Oct-04 2,603,200 Oct-05 2,733,500 Oct-06 2,745,200 Oct-07 2,607,200 Oct-08 2,629,000 Nov-04 2,012,000 Nov-05 1,879,400 Nov-06 1,015,800 Nov-07 2,033,500 Nov-08 1,689,200 Dec-04 1,567,400 Dec-05 1,869,700 Dec-06 1,518,300 Dec-07 1,631,100
Table A. 10 the CUSUM method as applied on C5. Month Consumption Xt-(X(t-12)) Xt+X(t-6) 2 F(Xbar) Ci+ N+ Ci- N- Status
Jan-04 1,487,800 Feb-04 1,332,100 Mar-04 1,501,700 Apr-04 1,615,100 May-04 2,403,100 Jun-04 2,428,200 Jul-04 3,120,400
Aug-04 2,239,900 Sep-04 3,994,200 Oct-04 2,603,200 Nov-04 2,012,000 Dec-04 1,567,400 Jan-05 1,431,400 -56,400 2,275,900 2,219,500 0 0 0 0 OK
Feb-05 452,600 -879,500 1,346,250 466,750 0 0 1,292,469 1 OK
Mar-05 1,385,800 -115,900 2,690,000 2,574,100 183,994 1 477,588 2 OK Apr-05 1,706,900 91,800 2,155,050 2,246,850 40,738 2 0 0 OK May-05 2,477,100 74,000 2,244,550 2,318,550 0 0 0 0 OK
Jun-05 2,477,100 48,900 2,022,250 2,071,150 0 0 0 0 OK Jul-05 3,006,500 -113,900 2,218,950 2,105,050 0 0 0 0 OK
Aug-05 3,196,600 956,700 1,824,600 2,781,300 391,194 1 0 0 OK
Sep-05 3,028,100 -966,100 2,206,950 1,240,850 0 0 518,369 1 OK Oct-05 2,733,500 130,300 2,220,200 2,350,500 0 0 0 0 OK Nov-05 1,879,400 -132,600 2,178,250 2,045,650 0 0 0 0 OK
Dec-05 1,869,700 302,300 2,173,400 2,475,700 85,594 1 0 0 OK
185
Jan-06 1,542,100 110,700 2,274,300 2,385,000 80,488 2 0 0 OK Feb-06 1,501,700 1,049,100 2,349,150 3,398,250 1,088,632 3 0 0 OK
Mar-06 1,744,800 359,000 2,386,450 2,745,450 1,443,976 4 0 0 OK Apr-06 1,771,600 64,700 2,252,550 2,317,250 1,371,120 5 0 0 OK May-06 2,571,000 93,900 2,225,200 2,319,100 1,300,114 6 0 0 OK
Jun-06 2,636,800 159,700 2,253,250 2,412,950 1,322,958 7 0 0 OK Jul-06 2,876,200 -130,300 2,209,150 2,078,850 1,011,702 8 0 0 OK
Aug-06 2,968,300 -228,300 2,235,000 2,006,700 628,296 9 0 0 OK
Sep-06 2,926,900 -101,200 2,335,850 2,234,650 472,840 10 0 0 OK Oct-06 2,745,200 11,700 2,258,400 2,270,100 352,834 11 0 0 OK Nov-06 1,015,800 -863,600 1,793,400 929,800 0 0 829,419 1 OK
Dec-06 1,518,300 -351,400 2,077,550 1,726,150 0 0 862,488 2 OK Jan-07 1,479,900 -62,200 2,178,050 2,115,850 0 0 505,857 3 OK Feb-07 1,336,200 -165,500 2,152,250 1,986,750 0 0 278,326 4 OK
Mar-07 1,371,300 -373,500 2,149,100 1,775,600 0 0 261,945 5 OK Apr-07 1,519,600 -252,000 2,132,400 1,880,400 0 0 140,764 6 OK May-07 2,113,500 -457,500 1,564,650 1,107,150 0 0 792,833 7 OK
Jun-07 2,503,800 -133,000 2,011,050 1,878,050 0 0 674,002 8 OK Jul-07 2,687,200 -189,000 2,083,550 1,894,550 0 0 538,671 9 OK
Aug-07 2,857,000 -111,300 2,096,600 1,985,300 0 0 312,590 10 OK
Sep-07 2,791,200 -135,700 2,081,250 1,945,550 0 0 126,259 11 OK Oct-07 2,607,200 -138,000 2,063,400 1,925,400 0 0 0 0 OK Nov-07 2,033,500 1,017,700 2,073,500 3,091,200 701,094 1 0 0 OK
Dec-07 1,631,100 112,800 2,067,450 2,180,250 491,238 2 0 0 OK Jan-08 1,521,300 41,400 2,104,250 2,145,650 246,782 3 0 0 OK Feb-08 1,542,400 206,200 2,199,700 2,405,900 262,576 4 0 0 OK
Mar-08 1,643,100 271,800 2,217,150 2,488,950 361,420 5 0 0 OK Apr-08 1,683,300 163,700 2,145,250 2,308,950 280,264 6 0 0 OK May-08 2,336,100 222,600 2,184,800 2,407,400 297,558 7 0 0 OK
Jun-08 2,470,800 -33,000 2,050,950 2,017,950 0 0 0 0 OK Jul-08 2,881,100 193,900 2,201,200 2,395,100 4,994 1 0 0 OK
Aug-08 2,911,600 54,600 2,227,000 2,281,600 0 0 0 0 OK
Sep-08 2,733,600 -57,600 2,188,350 2,130,750 0 0 0 0 OK Oct-08 2,629,000 21,800 2,156,150 2,177,950 0 0 0 0 OK Nov-08 1,689,200 -344,300 2,012,650 1,668,350 0 0 90,869 1 OK
186
Appendix II Evidence for Publication
Evidence for publishing of
1. Reliability analysis of a cooling seawater pumping station …………..P 187.
2. Data analysis technique to resolve the conflict between
operation and maintenance……………………………………………..P 188.
3. The detection of flow meter drift by using statistical process
control…………………………………………………………………P 189.
4. The detection of flow meter drift by using artificial
neural networks………………..………………………….…………..P 190.
PLEASE NOTE
Appendix II is unable to be reproduced online.
Please consult print copy held in the Swinburne Library or click on the links below.
Alsalamah, MJ, Shayan, E, Savsar, M (2006) Reliability analysis of a cooling seawater pumping station. International Journal of Quality
and Reliability Management 23(6): 670-695 DOI: 10.1108/02656710610672489
Salamah, MB, Shayan, E, Savsar, M (2010) Minimizing the conflict between operation and maintenance: a case study. International
Journal of Data Analysis and Information Systems 2(1): 19-38 Publisher URL: http://www.serialsjournals.com/journal-
detail.php?journals_id=68
Salamah, MB, Kapoor, A, Savsar, M, Ektesabi, M, Abdekhodaee, A, Shayan, E (2011) The detection of flow meter drift by using statistical process control. International Journal of Sustainable Development and
Planning 6(1): 91-103 DOI: 10.2495/SDP-V6-N1-91-103
Salamah, MB, Palaneeswaran, E, Savsar, M, Ektesabi, M (2011) Detecting flow meter drift by using artificial neural networks.
International Journal of Sustainable Development and Planning 6(4): 512-521
DOI: 10.2495/SDP-V6-N4-512-521