fault detection and diagnosis of brahmanbaria gas
TRANSCRIPT
Fault Detection and Diagnosis of Brahmanbaria Gas
Processing Plant Using Artificial Neural Network Analysis
MASTER OF SCIENCE IN ENGINEERING (CHEMICAL)
Suman Ahmed
Department of Chemical Engineering
BANGLADESH UNIVERSITY OF ENGINEERING AND TECHNOLOGY, DHAKA
March, 2018
ii
Fault Detection and Diagnosis of Brahmanbaria Gas
Processing Plant Using Artificial Neural Network Analysis
by
Suman Ahmed
A thesis submitted to the Department of Chemical Engineering
in partial fulfillment of the requirements for the degree of
MASTER OF SCIENCE IN ENGINEERING (CHEMICAL)
Department of Chemical Engineering
BANGLADESH UNIVERSITY OF ENGINEERING AND
TECHNOLOGY, DHAKA
March, 2018
v
Abstract Natural gas (NG) plays an important role in different sectors such as power generation,
fertilizer, industrial and commercial sector in Bangladesh. There are many gas
producing fields in Bangladesh. Although process safety technology has been
gradually implemented over the years, several accidents have happened in gas field
industry in Bangladesh like BGFCL, Niko, Occidental, Tullow, and Chevron. Massive
blowout took place in Occidental operated field and the similar incident happened at
Tengratila field of Niko. Major gas leakage was found in Titas Gas Field (BGFCL).
Gas processing plants associated with the gas fields have encountered process
industries and outage due to lack of proper monitoring. Bangura gas plant was shut
down for lacking proper monitoring and the Bibiyana gas plant was shut down to
repair a gas leaking . Those incidents lead to emphasis on fault detection and diagnosis.
Advanced process control (such as supervisory control and data acquisition (SCADA)
and Distributed control) systems help to operate the plant more reliably. However,
operator is saturated by alarms due to disturbances in a chemical process, need tool to
rapidly identifying root cause of fault and to rapidly intense to mitigate consequences.
To reduce the frequency and consequences of accidents, several techniques of hazard
identification and fault diagnosis have been developed and implemented. Over the last
few years, several studies were carried out on detection and diagnosis of process plant
disturbances using NN based Fault Diagnosis Technique.
In this thesis, an attempt has been made to study the fault detection and diagnosis of gas
processing plant using NN based system. Firstly, the steady state model of the gas
processing plant was developed using HYSYS and be validated using Brahmanbaria gas
plant data. Secondly Dynamic model is developed within Aspen HYSYS to study the
transient behaviour and different states (normal and abnormal) of the plant. Thirdly,
Different states of the process plant were generated using dynamic model.
vi
Finally, a multi-layered feed forward NN based fault detection and diagnosis model has
been developed to identify the fault (disturbance) and no fault (normal) operation. The
developed NN based fault detection and diagnosis system has been trained using back
propagation algorithm. The NN based fault detection and diagnosis system has been
trained, validated and tested using the dynamic model data. Several neural networks
with different configurations and various learning strategies has employed in the training
process to obtain the optimum NN architecture for fault detection and diagnosis.
Preliminary results shows that NN based method successfully detect the faults of Gas
processing plant. It is expected that ANN based fault detection and diagnosis tool will be
popular in petrochemical process due to its simplicity to develop.
vii
Acknowledgements
Firstly, the author would like to express my sincere thanks to Dr. Md. Tanvir Sowgath
for proposing this research topic. I would like to thank him for his valuable suggestions,
excellent guidance, support, encouragement, and supervision that made this dissertation
possible.
The author would like to gratitude to Dr. Md. Ali Ahammad Shoukat Choudhury who
made my thesis work to bring in day light by providing proper guidance though I am
working on chevron Bangladesh a as national rotator shift duty (14/14).
The author would like to gratefully acknowledge Dr. Syeda Sultana Razia for her
kindness help as well as guidance enabled me to complete this thesis paper process. Her
guidance helps me to write up thesis work
The author would like to thanks to Dr. Syed Farid Ahmed for his proper guidance to
complete the thesis work.
We owe a debt of gratitude to the Professor I.M.Mujtaba, University of Bradford, UK in
neural networks whose unceasing research and his support has been a continual
challenge and inspiration to us.
The author would like to thank Mr. Shafiq of Brahmanbaria gas processing plant
(BGFCL) to provide gas process plant live data, are gratefully acknowledged.
Finally, the author would like to thank my family members, all my Chevron Bangladesh
colleagues and all my friends for their support and encouragement during the research
work.
viii
Table of contents
Abstract…………………………………………………………………………..............v Acknowledgements……………………………………………………………………..vii List of Figures……………………………………………………………………...........xi List of Tables…………………………………………………………………………...xiii Nomenclature………………………………………………………………………. ….xiv 1. INTRODUCTION
1.1 Background...............................................................................................................1
1.2. Problem Statement…….…………………………………………………………..3 1.3 Objective and Scope Research…………...………………………………………..4 1.4 Organization of the Thesis…………………………….……………………….…..5 2. LITERATURE REVIEW 2.1Introduction...............................................................................................................6
2.2 Principal of fault ......................................................................................................7 2.3 Quantitative fault detection method………………………………………….….....9 2.3.1 Process monitoring and Diagnosis……………………………………………10 2.4 Neural Network......................................................................................................11 2.4.1Neural Network Classification….………………………………………….....11 2.4.2 Neural Network Application in Chemical and Process Engineering……........12
2.4.3 Basics to artificial neural networks………………………..............................14
ix
3.0 Process modeling using neural network 3.1 Data preprocessing………………………………………………………………..22
3.2 Data Cleaning…………………………………………………………………….23
3.3 Normalization of Input and Output Data Sets……………....................................24
3.4 Coding for Data Pre-processing…..........................................................................26
3.5 ANN Structure………………………………………………….…………….…...27
3.5.1 Structure Selection…………………………….………...................................27
3.5.2 Sizing the Network Structure………………………………………………..29
3.5.3 Algorithm for Optimum Network Structure………………………………..30 3.6 Selection of Proper Transfer Function……………………………………..….....32
3.7 Initializing the Weight Factor Distribution……………………............................34
3.8 Selection of ANN Parameters………………………………................................34
3.9 Train the Network…………………………………………………………….….37 3.10 Validation & Testing………………………………..……………………….….37 4. PLANT SIMULATION 4.1 Introduction………………………………………………………………………38
4.2 Process description of Brahmanbaria gas processing plant (BGFCL), Brahmanbaria, Bangladesh…...........................................................................................39
4.3 Steady state simulation of Brahmanbaria Gas Processing Plant………….…..….40
4.4 Dynamic state simulation conversion from steady state simulation of Brahmanbaria Gas Processing Plant…………………………………………………….41 4.5 Dynamics Monitoring…………………………………………………………….42
x
4.6 Steady State Data and Dynamic Data…………………………………………….47
4.7 Simulation & Validation…………………………………………………………..48
4.8 Summary…………………………………………………………………………..52 5. ANN Fault detection and diagnosis modeling of Brahmanbaria gas
processing plant
5.1 Introduction………………………………………………………………………53
5.2 ANN Model Development of Brahmanbaria Gas Processing Plant……………..54
5.3 ANN Based Data Prediction for Brahmanbaria Gas Processing Plant (Normal Modeling)……………………………………………………………………………57
5.4 Neural Network Fault detection scheme………………………………………...59 5.5 Result……………………………………………………………………………66 6.0 CONCLUSION AND RECOMENDATION FOR FUTURE WORK
6.1 Conclusion………………………………………………………………………76 6.2 Recommendation for future work………………………………………………77 REFFERENCES…………………………………………………………………..78 Publication………………………………………………………………………….80 Appendix A…………………………………………………………………………81 Appendix B…………………………………………………………………………85 Appendix C…………………………………………………………………………87
xi
List of figures 2.1 Time-dependency of faults 7
2.2 Basic models of faults: (a) Additive fault (b) Multiplicative faults 8
2.3 Classification of Neural Network
11
2.4 Components of Biological neuron
16
2.5 Model of Artificial Neuron 16
2.6 A representation of a simple 3-layer feed-forward ANN 18
2.7 Several activation functions 19
3.1 Preprocessing and Post-processing within network object 26
3.2 The hyperbolic tangent function superimposed over the sigmoid function
33
4.1 Steady state HYSYS Simulation. 40
4.2 Temperature profile of fractionation column normal operation 42
4.3 Disturbance profile of fractionation column during feed valve full close
43
4.4 Disturbance profile of fractionation column during feed valve full open
44
4.5 Disturbance profile of fractionation column during separators valve full close
45
4.6 Dynamic state simulation of Brahmanbaria Gas Processing Plant (BGFCL)
46
5.1 Online NN based fault detection system
59
xii
5.2 Offline NN based fault detection
60
5.3 Neural Network Back propagation Training Scheme. 60
5.4 NN based fault detection system
63
5.5 Algorithm to extract weights and biases from optimized network 64
5.6 Normal Operation mode by HYSYS 69
5.7 Tower valve disturbance Operation mode by HYSYS 70
5.8 Normal operational trends
71
5.9 Training of NN Fault detection system 72
5.10 Statistical regression analysis of NN predicted data with fault 73
5.11 Statistical regression analysis of NN predicted data with fault
73
5.12 Statistical regression analysis of NN predicted test data
74
xiii
List of tables
3.1 Hierarchy of Artificial Neural Networks 28
4.1 Sales gas composition 48
4.2 Steady state and dynamic state data comparison 49
4.3 Steady state and dynamic state data comparison
49
4.4 Steady state and dynamic state data comparison
50
4.5 Steady state and dynamic state data comparison
50
4.6 Steady state and dynamic state data comparison
51
5.1 Input-output parameter 66
5.2 Types of operation mode/disturbances criteria for Neural Network analysis
67
5.3 NN architecture for different conditions
68
xiv
Nomenclature
ANN Artificial Neural Network
ANNs Artificial Neural Networks
NN Neural Network
AI Artificial Intelligence
FF Feedforward
FFANN Feed-Forward Artificial Neural Network
FLNs Functional Link Networks
RBFNs Radial Basis Function Networks
RBF Radial Basis Function
AF Activation Function
BA Backpropagation Algorithm
BP Backpropagation
FFBPN Feed forward Back propagation Network
LM Levenberg- Marquardt
RFTS Rapid Foundry Tooling System
FTA Fault Tree Analysis
BGPP Brahmanbaria Gas Processing Plant
HP & LTS High Pressure & Low Temperature Separator
FEMA Failure mode and effect analysis
xv
HAZOP Hazard and Operability
PHA Process Hazard Analysis
PSM Process Safety Management
1
Chapter 1
INTRODUCTION
1.1 Background
The chemical and petrochemical industry is becoming larger and more complex in
Bangladesh. Many of the gas fields are operated by government organization like
Bangladesh Gas Field Limited (BGFCL), rest of the gas fields are operated by
international oil company like Chevron, Kris energy, Santos. Process safety management
has not developed in comparison to the growth of chemical and petrochemical industry.
Consequences several accident history record available (Lee, 1996). The study of fault
detection and diagnosis will play vital rule in reducing the occurrence of sudden,
disruptive, or dangerous outages, equipment damage, and personal accidents, and to
assist in the operation with the maintenance program.
Research of quantitative fault detection system has earned more interest in recent times
not just only due to economical, yet more importantly; it functions as a safety
mechanism. The application of an advance controller in terms of fault detection will help
to reduce the probability of accident and loss as a result of human or mechanical error.
Real time fault detection enables supervisory control system to rapidly intervene and
prevent any incipient fault event into escalating into a process incident or accident
thereby preventing process outage and potential loss. The disaster in Bhopal and
Chernobyl is an excellent example why an advance controller can play a vital role in
preventing the incident from happen in the first place.
Fault detection is essentially a pattern recognition problem, in which a functional
mapping from the measurement space to a fault space is calculated A wide variety of
techniques have been proposed to detect and diagnose faults.
Generally, there are three different options available to approach a fault diagnosis
problem: state estimation methods, statistical process control methods, and knowledge-
based methods.
The emergence of artificial intelligence (AI) also plays a role in the development of
control system. The approach of AI is focusing on imitating the rational thinking of
2
human (Lee, 2006). AI system such as fuzzy logic, neural network and genetic
programming had been integrated with the conventional control system to produce an
intelligent controller system. A neural network, a type of knowledge-based system,
possesses many desirable and preferred properties for chemical process fault diagnosis.
These properties include its abilities to learn from example, extract salient features from
data, reason in the presence of novel, imprecise or incomplete information, tolerate noisy
and random data, and degrade gracefully in performance when encountering data beyond
its range of training (Venkatasubramanian, 2003).
Reviewing the development of neural network fault detection and diagnosis systems, the
general trend in research is to increase the robustness of the system to un-modelled
patterns, realize fast and reliable diagnosis in dynamic processes and dynamically filter
noisy data used for detection (Hamid, 2004).
Aspen HYSYS solves the critical engineering and operating problems that arise
throughout the lifecycle of a chemical process, such as designing a new process,
troubleshooting a process unit or optimizing operations of a full process like an Acrylic
Acid plant. The process simulation capabilities of Aspen HYSYS enables engineers to
predict the behavior of a process using basic engineering relationships such as mass and
energy balances, phase and chemical equilibrium, and reaction kinetics. With reliable
thermodynamic data, realistic operating conditions and the rigorous Aspen HYSYS
equipment models, they can simulate actual plant behavior. In this case, the intelligent
controller will help the operator to handle and deal with various abnormal conditions or
fault that happen with more reliable, efficient and faster. The implementations of Neural
Network for fault detection in Brahmanbaria Gas Processing plant are proposed. Steady
state simulation has made by using plant operations parameter of Brahmanbaria gas
processing plant (Sultana R, Syeda., Ahmed, Suman., Rahman, Md Bazlur., Mehfuz,
Omit., and Shamsuzzaman, Razib., 2008). The steady state simulations data will be
compared with the plant live data (Kamruzzaman, 1999). Dynamic state simulation
model will help to understand the real plant behavior (HYSYS 3.2). The dynamic state
simulation plant model data is compared with steady state simulation plant data.
3
1.2 Problem Statement
As we are heading toward the future, the advance knowledge and technology are
contributing to the improvement on reliability, safety and efficiency of fault detection
and diagnosis system. This system is very important as it will prevent accident; failure
and disaster from happening and save many lives. Today, safety and health are becoming
a crucial agenda in developing and managing technical processes. As a result, the
development of Neural Network in various fields, especially in fault detecting has shown
great progress. Neural Network has the potential to be developed further to be applied in
chemical plant such as Brahmanbaria Gas Process Plant.
Furthermore, Matlab 7.0 used to model and stimulate the Neural Network in terms of
monitoring and supervising the Brahmanbaria Gas Process Plant. Matlab is a high-
performance language for technical computing software that is used widely in the
engineering field to calculate and solve many mathematical and technical problems.
Thus, this research will be focusing on fault detection on as Brahmanbaria Gas Process
Plant by using Neural Network. These researches will emphasis on how and how far
Neural Network can contribute to overcome as Brahmanbaria Gas Process Plant failure
and fault problems.
4
1.3 Objective and Scope Research
The main aim of this research is to develop a fault detection system using neural network
analysis. By using the Brahmanbaria gas process plant as the case study, the
implementation of neural network will help the controller to detect fault more efficient.
The work covered the following scope:
To develop a steady state model of the Brahmanbaria gas processing plant using
HYSYS and is validated using real plant data.
To develop a dynamic model to understand the real plant behavior and
disturbances.
To develop a NN based fault detection and diagnosis system.
The possible outcome of the thesis is a fault detection and diagnosis tool for
Brahmanbaria gas processing plant.
5
1.4 Research Methodology
Chapter 1 is the introduction to this thesis. Summary of the background, problem of
statement, objective and scope of research as well as research methodology are also
included in this chapter.
Chapter 2 provides an overview of quantitative fault detection system, artificial neural
network system, its uses in the chemical and petrochemical industries, feed forward and
backpropagation system are briefly discussed.
Chapter 3 depict model development process by using ANN, data pre processing, data
cleaning process, normalization input and output data sets, coding for data pre-
processing, ANN structure, training, validation and testing system are described.
Chapter 4 describes process description of Brahmanbaria gas processing plant and also
illustrate the steady state and dynamic state simulation process by using plant design data
and HYSYS simulator. Then it shows the validation process with steady state simulation
data and plant live data. Here it does also explain the conversation process from steady
state plant simulation process to dynamic state plant simulation process. It shows the
way to monitor of dynamic disturbances by using dynamic simulation. Finally, in this
chapter shows different fault construction process in HYSYS dynamic simulation and
data collection process for artificial neural network analysis.
Chapter 5 shows ANN fault detection and diagnosis model development process and
provides the results of the simulation and experimental studies after applying the
proposed methods. The observed results also discussed.
Chapter 6 states the conclusions drawn from the current work and suggests possible
direction for future research.
6
Chapter 2
LITERATURE REVIEW
2.1 Introduction In the area of plant-wide control at the supervisory level, the process fault detection and
system plays a key role. Fault detection, isolation, and recovery (FDIR) is a subfield of
control engineering which concerns itself with monitoring a system, identifying when a
fault has occurred, and pinpointing the type of fault and its location. Two approaches can
be distinguished: A direct pattern recognition of sensor readings that indicate a fault and
an analysis of the discrepancy between the sensor readings and expected values, derived
from some model. In the latter case, it is typical that a fault is said to be detected if the
discrepancy or residual goes above a certain threshold. Fault detection usually includes
the fault diagnosis and fault correction system. Fault diagnosis is the identification of the
root causes of process upset. Meanwhile, fault correction is the provision of
recommended corrective actions to restore the process to normal operating condition.
In this regard, real-time appropriate actions must be taken in present chemical and
petrochemical manufacturing plants. The technical personnel in most of these industries
is responsible for process monitoring status, detecting abnormal events, diagnosing the
source causes and administering proper intervention to bring the process to normal
operation.
A large variety of techniques for fault detection had been proposed in the literature in
recent times. Due to the broad scope of the process fault diagnosis problem and the
difficulties in its real time solution, various computer-aided approaches have been
developed over the years. (Himmelblau and Hussain, 1978) These cover a wide variety
of techniques such as early attempts using fault trees and diagraphs, analytical
approaches, and knowledge-based systems and neural networks in more recent studies.
From a modeling perspective, these methods require either accurate process models,
semi-quantitative models, or qualitative model.
7
Neural networks have been studied very intensively. New architectures and learning
algorithms are developed all the time. Even though the present neural network models
don‟t achieve human-like performance, they offer interesting means for pattern
recognition and classification. The traditional pattern recognition includes a large
collection of very different types of mathematical tools (preprocessing, extraction of
features, final recognition). In many cases it is difficult to say what kind of tool would
best fit to a particular problem. Neural network make it possible to combine these steps,
because they are able to extract the features autonomously. They are practical to use,
because they are nonparametric. It has also been reported that the accuracy of neural
classifiers is better than that of traditional ones.
8
2.2 Principle of Fault
A fault is defined as an unpermitted deviation of at least one characteristic property of a
variable from an acceptable behavior (Isermann, 1997). In the meantime, Himmelblau in
1978 defines a fault as a process abnormality or symptom, such as high temperature in a
reactor or low product quality. In general, fault is deviations from the normal operating
behavior in the plant that are not due to disturbance change or set point change in the
process, which may cause performance deteriorations, malfunctions or breakdowns in
the monitored plant or in its instrumentation. Therefore, the fault is a state that may lead
to a malfunction or failure of the system. The time dependency of faults can be
distinguished, as shown in Figure 2.1, as abrupt fault such as overheating and
overpressure, incipient fault such as continuing overflow, and intermittent fault such as
fault in gear or valve.
Fig.2.1: Time-dependency of faults: Abrupt (a), incipient (b), and Intermittent(c) by Isermann (1997).
9
According to Gertler in 1998, faults can be categorized into the following categories:-
i. Additive process faults: Unknown inputs acting on the plant, which are normally
zero. They cause a change in the plant outputs independent of the known input.
Such fault can be best described as plant leaks and load
ii. Multiplicative process faults: These are gradual or abrupt changes in some plant
parameters. They cause changes in the plant outputs, which also depend on the
magnitude of the known inputs. Such faults can be best described as the
deterioration of plant equipment, such as surface contamination, clogging, or the
partial or total loss of power.
iii. Sensor faults: These are difference between the measured and actual values of
individual plant variables. These faults are usually considered additive
(independent of the measured magnitude), though some sensor faults (such as
sticking or complete failure) may be better characterized as multiplicative.
iv. Actuator faults: These are difference between the input command of an actuator
and its actual output. Actuator faults are usually handled as additive though,
some kind (such as sticking or complete failure) may be described as
multiplicative.
10
2.3 Quantitative Fault Detection Method
There are three common methods of quantitative fault detection system. They are:
State Estimation Approaches Statistical Process Control Approach and Knowledge-
Based Approaches:
i. State Estimation Approaches: It can estimate immeasurable parameters, if they
are observable. However, they require an exact process model. It combines a
fundamental model of the process with on-line measurements to provide on-
line, recursive estimates of the underlying theoretical states of the process.
ii. Statistical Process Control Approach: Its capability to model any nonlinear
relationship is still limited by the assumed basis functions used in the
regression.
iii. Knowledge-Based Approaches: Knowledge-based approaches use expert
systems or artificial intelligence methods to process data. In rule-based expert
systems, the process model is represented by a set of qualitative and
quantitative governing rules. Other method is Neural Network (NN) based
system. NN is used to model complex relationships between inputs and
outputs or to find patterns in data. It serves as pattern recognition to identify
the process fault by reasoning based on generalizing a set of data. Firstly, the
different fault situations are trained, validated and tested. After the network
learning process, it can predict the fault detection.
One of the Knowledge based approaches is using Neural Network. Neural Network is
a non-linear statistical data modeling tools. They can be used to model complex
relationships between inputs and outputs or to find patterns in data. ANN is attractive
due to its information processing characteristic such as nonlinearity, high parallelism,
fault tolerance as well as capability to generalize and handle imprecise information
(Basheer and Hajmeer, 2000). These characteristics have made ANN suitable for
solving a variety of problems. In fault detection case, the neural network serves as
pattern recognition to identify the process fault by reasoning based on generalizing a
11
set of data. With parallel computation and ability to adapt to changes, neural network
is a best choice for fault detection system. NN analysis tool is one of the intelligent
tools that it uses widely in the world for fault detection and diagnosis analysis. In NN
analysis, lots of data can analysis but other techniques cannot handle lots of data
analysis.
2.3.1 Process Monitoring and Diagnosis
In process monitoring, neural network system has been successfully applied to reduce
cost of emission monitoring in modern chemical industries (Chementator, 2000).
AlphaMOS, France and Neural Computer Sciences, Southampton, UK, have termed
up to develop intelligent odor-sensing systems using neural networks, which makes
them suitable for online process monitoring, such as continuously monitoring
perfumes in soaps and cosmetic bases and aromas in the food industry.
The electronic nose, called FOX 2000, developed by AlphaMOS is being used for
monitoring odor from sulfur compound and natural gas (which is odorless to the
human nose). The next generation of electronic nose will be equipped with hybrid
arrays of different type of sensors. For example, sensors incorporating conduction
polymers, which are more discriminating than metal oxide silicon, will be
commercially available in mid- 2005. Using sensors for surface acoustic waves allows
monitoring of small molecules of toxic gases in the range of 1-10 parts per billion.
The devices can typically acquire, analyze and recognize a sample within seconds
compared to about 30-60 minutes required for conventional chromatography tests
(Chementator, 2000 & Blanchar, 1994).
Fujitsu and Nippon Steel Corporation, Japan, has implemented a neural network
system for monitoring and detecting process faults in the continuous casting of
steel. In such an operation, the molten steel enters a water-cooled mold. The outside
surface of the continuous slab of steel gradually solidifies, but must be kept
continuously moving and contained within the walls of the continuous caster. On
occasion, a processing defect called breakout occurs.
12
2.4 Neural Network:
The use of Neural Networks (NNs), in all aspects of process engineering activities,
such as modeling, design, optimization and control, has considerably increased in
recent years (Venkatasubramanian, 2003). Different NN based techniques
(architecture, training) have been adopted in different field of science to overcome the
difficulties of first principle based modeling. The non-linear relationship between
input and output of a system can be built up cost effectively by NNs.
2.4.1Neural Network Classification
Basically neural network can be generally being separated into two groups according
Lee in 2006:-
i. Supervised neural network- neural network operating with supervised learning
and training strategies, which is major of ANNs such as Hopfield Network,
FFBPN (Feed forward Back propagation Network), RBF (Radial Basis
Function), etc.
ii. Unsupervised neural network- neural network that do not need any supervised
learning and training strategies, including all kinds of self-organizing, self-
clustering, and learning networks such as SOM, ART (Adaptive Resonant
Theory), and so on.
Single-Layer ANNs
Multi-Layer ANNs
Recurrent ANNs
ANN system structure
Adaline Perceptron
Hopfield Network
LVQ
Madaline FFBPN
RBFN Neocognitron
BAM ART Hopfield
Network Boltzman
Machine
Fig. 2.3: Classification of Neural Network (Patterson 1996)
13
The Feed-forward network is when the data flow from input to output units. The data
processing can extend over multiple layers of units, but no feedback connections are
present. Moreover, for Recurrent networks, its contain feedback connections.
Different from the feed-forward networks, the dynamical properties of the network
are important. In certain times, the activation values of the units will undergo a
relaxation process such that the network will develop to a stable state in which these
activations do not change anymore. In other applications, the changes of the activation
values of the output neurons are significant, such that the dynamical behavior
constitutes the output of the network.
2.4.2 Neural Network Application in Chemical and Process
Engineering
Realistic process model is very complicated and time consuming because it consists
of a lot of non-linear relation. Even it is unachievable when the basic mechanism is
not understood. Neural networks can learn a non-linear relation from example (input
output) and solve the problem easily (Montague et al., 1994). NN has been widely
used extensively in chemical engineering. NN has been used over the years such as in
process modeling, adaptive control, model based control, hybrid process monitoring,
fault detection, dynamic modeling, and parameter estimation process flow sheet
simulations, on-line process optimization and visualization, parameter estimation,
fault diagnosis, error detection, data reconciliation, process analysis, oil and gas
exploration, manufacturing, process control, product design and analysis, visual
quality inspection system, machine analysis, project bidding, dynamics of chemical
process systems.
In chemistry, neural network determines the molecular structure by comparing the
data obtained by spectroscopic analysis. In process control, NN determine the
complex relationship between the controlled and manipulated variable comparing the
data obtained from the monitoring of the process and the fault detection. NN shows
great promise over the recent years to solve problems that have proven to be difficult
for the standard technique using digital computers. NN is inherently parallel in
14
structure like human brain and has capabilities for storing knowledge from analyzing
information. In the recent history NN has become very popular by different
researchers from a wide range of disciplines i.e. aerospace, automotive,
transportation; telecommunications, electronics, robotics, speech; financial, insurance,
securities, banking; manufacturing, oil and gas; medical and defense (Hagan et al.,
1996).
Application of Neural Networks to bio-processing and chemical engineering have
increased significantly since 1988. One of the first applications of neural network is to
fault diagnosis of a chemical reactor system (Anderson, 1992). Since then, the
number of research publications on neural network application in bio-processing and
chemical engineering has risen significantly. Neural computing provides a good
overview of potential application of neural networks as listed below (Jacobsson,
2001):
i. Classification: Use input values to determine the classification e.g. is the input
the letter A, is the blob of the video data a plane and what kind of plane is it.
ii. Prediction: ANNs have been shown to be successful as predictive tools by
predicting that some event will or will not occur, predicting the time at which
an event will occur, or predicting the level of some event outcome. To predict
with an acceptable level of accuracy, an ANN must be trained with a sizable
set of examples of past pattern/ future outcome pairs. The ANN must then be
able to generalize and extrapolate from new patterns to predict associative
outcomes. Many process industries are now using ANNs in a big way
for production, stock selection, predicting time series temperature profile of
reactors and portfolio management (Bishop, 1996).
iii. Data association: Like classification but it also recognizes data that contains
errors e.g. not only identify the characters that were scanned but identify when
the scanner is not working properly.
iv. Data Conceptualization: Analyze the inputs so that grouping relationships can
be inferred e.g. extract from a database the names of those most likely to by a
particular product.
15
v. Data Filtering: Smooth an input signal e.g. takes the noise out of a telephone
signal (Hoskins, 1988).
vi. Optimization: ANNs have been used for a number of problems that require
finding an optimal or near optimal solution e.g. the scheduling of
manufacturing operation, finding the shortest of all possible ways to profit
maximization, minimization of some cost function under an asset of
constraints, and so on (Bishop, 1996).
Some integrated product formulation and optimization system such as CAD/Chem
consists of neural networks to estimate the properties of a given product formulation,
an expert system to run the formulation model repeatedly in essence asking many
what-if questions relating to product formulation, a set of user-defined design goals
that apply a fuzzy logic inspired method to specify ranking and preferences of product
properties and constraints in ingredients, processing and costs; and a product
optimizer to drive the repeated what-if trials in the direction required to meet these
goals (Patterson, 1995). Commercial applications of the CAD/Chem system have
included the optimal design of formulated products such as coatings, plastics,
rubber, specially polymers, building materials (Islam and Hossain, 2001).
2.4.3 Basics to artificial neural networks
Artificial Intelligence (AI)
The term Neural Network resulted from artificial intelligence (AI) research, which
attempts to understand and model brain behavior. According to Barr and Feigenbaum
(1981): “Artificial Intelligence is the part of computer science concerned with designing
intelligent computer system that exhibits characteristics we associate with intelligence
in human behavior”. This definition simply states that the goal of AI is to make computer “Think” to
make them solve problems requiring human intelligence. Focusing on the means of
achieving this goal, Buchanan and Short life (1983) offer another definition of AI:
16
“Artificial intelligence is the branch of computer science dealing with symbolic, non-algorithmic methods of problem solving.” This second definition of AI emphasizes two aspects of AI-based methods for
problem solving. First, AI is a formal procedure specifying a step-by-step execution
path that guarantees a correct or optimal solution at some point. Second, AI involves
Symbolic processing, a branch of computer science that deals with non-numerical
symbols and names. In contrast, the more classical numerical processing deals with
numerical calculation and process.
Artificial Neural Networks (ANNs)
Artificial Neural Networks (ANNs) are simplified models of the central nervous
system. They are networks of highly interconnected neural computing elements that
have the ability to respond to input stimuli and to learn to adapt to the environment.
Robert Hecht- Nielson defines neural networks as follows: “A Neural Network is a computing system made up of number of simple, highly interconnected nodes or processing elements, which process information by its dynamic state response to external inputs.” To achieve good performance, neural networks employ a massive interconnection of
simple computing cells referred to as “Neurons” or “Processing units.” the following
definition is given by Alekxander and Morton (Haykin, 1999). “A Neural network is a massively parallel distributed processor that has a natural propensity for storing experimental knowledge and making it available for use. It resembles the brain in two respects:
i. Knowledge is acquired by the network through a learning process.
ii. Inter-neuron connection strengths known as synaptic weights are used to store
the knowledge.”
The procedure used to perform the learning process is called a Learning
Algorithm, the function of which is to modify the Synaptic Weights of the
network in an orderly fashion so as to attain a desired design objective. Neural
17
networks are also referred to in the literature as neural computers, connectionist
networks, parallel distributed processors, etc. (Haykin, 1999).
Neuron
In the human brain, neuron within the nervous system interacts in a complex fashion.
Typical neuron collects signals from others through a host of fine structures called
Dendrites. The neuron sends out spikes of electrical activity through a long, thin stand
known as an Axon, which splits into thousands of branches. At the end of each
branch, a structure called a Synapse converts the activity from the axon into electrical
effects that inhibit or excite activity from the axon into electrical effects that inhibit or
excite activity in the connected neurons. When a neuron receives excitatory input that
is sufficiently large compared with its inhibitory input, it sends a spike of electrical
activity down its axon. Learning occurs by changing the effectiveness of the synapses
so that the influence of one neuron on another changes. In simulating biological
neurons, first the essential features of neurons and their interconnections are deduced.
Then typically program a computer to simulate these features. However, everyone
have incomplete knowledge about neurons and has limited computing power, so
models are necessarily gross idealizations of real networks of neurons. A Neuron is an
information-processing unit that is fundamental to the operation of a neural network.
It behaves as activation or mapping function and it produces an output when the
cumulative effect of the input stimuli exceeds a threshold value. Figure 2.5 shows the
model for a neuron. We may identify three basic elements of the neuron, as described
here (Haykin, 1999):
18
Fig.2.4: Components of Biological neuron
Fig.2.5: Model of Artificial Neuron
i. A set of Synapses or Connecting links, each of which is characterized by a
Weight or Strengths of its own. This weight is a sort of filter, which is a part of
linkage connecting the input to the neuron. It models the synaptic neural
connection in biological nets act to either increase (excitatory input) or
decrease (inhibitory input).
ii. An Adder for summing the input signals, weighted by the respective synapses
of the neuron; the operation described here a linear combiner.
19
An Activation function for limiting the amplitude of the output of a neuron. The
activation function is also referred to in the literature as a squashing function in that it
squashes (limits) the permissible amplitude range of the output signal to some finite
value.
A detailed mathematical model of a neuron is shown in Figure 2.6. Models may
include an externally applied threshold/bias that has the effect of lowering the net
input of the activation function. On the other hand, the net input of the activation
function may be increased by employing a bias term rather than a threshold. In
mathematical terms, we may describe a neuron m by writing the following pair of
equations:
rm= ∑Nwmn xn bm 2.1
ym= f (rm) 2.2 Where x1, x2,…, xN are the inputs; wm1, wm2, …, wmN are the synaptic weights of
neuron m; rm is the linear combiner output; bm is the bias term; f is the activation
function; and ym is the output signal of the neuron. The scalar input x is transmitted
through a connection that multiplies its strength by the scalar weight w, to form the
product wx, again a scalar. The neuron shown in equation 2.1 has a bias bm but bias
term may not be used depending on the situation. You may view the bias as simply
being added to the product wx as shown by the summing junction or as shifting the
function f to the left by an amount bm. The bias is much like a weight, except that it
has a constant input of 1. The transfer function net input rm, again a scalar, is the sum
of the weighted input wx and the bias bm. This sum is the argument of the transfer
function f which takes the argument rm and produces the output ym. Note that w and
b are both adjustable scalar parameters of the neuron.
20
Feed-Forward Artificial Neural Network (FANN)
The power of single neuron can be greatly amplified, using multiple neurons in a
network of layered connectionist architecture. Neurons layered in such a way is also
called Feed- Forward Artificial Neural Network and abbreviated to FANN. A feed
forward network has a layered structure. Feed-forward artificial neural networks
include MLPs, Functional Link Networks (FLNs) and Radial Basis Function
Networks (RBFNs) (Looney, 1996)
MLPs are perhaps the most popular network architecture in use for many problems.
This is the type of network shown in Figure 2.6. The network has a simple
interpretation as a form of input-output model, with the weights and thresholds
(biases) the free parameters of the model. Such networks can model functions of
almost arbitrary complexity, with the number of layers, and the number of units in
each layer, determining the function complexity. Important issues in MLP design
include specification of the number of hidden layers and the number of units in these
layers (Neural Networks, 1984-2003)
Fig. 2.6: A representation of a simple 3-layer feed-forward ANN
21
On the left is the layer of inputs, or branching nodes, which are not artificial neurons.
A feature vector x = (x1,…,xN) that represents a pattern enter the input layer on the
left with each component xn entering one and one input node. From each nth input
(branching) node, the nth component xn fans out to each of the M neurons in the
middle layer. Thus each mth hidden (middle) neurons has a fan-in of all N input
components. As each xn enters the mth neuron of the hidden layer, it is modified via
multiplication by the synaptic weight wmn for that connection line. All resulting
products wnmxn at the mth hidden neuron are summed over n to yield:
rm= ∑n Wnm Xn 2.3 and
ym= h(rm) 2.4
is the activation output. Doing the same for the output layer we have
sj= ∑m umj ym 2.5
and zj= g(sj) 2.6 Backpropagation Algorithm: The Levenberg-Marquardt Method
A single layer network has severe restrictions: the class of tasks that can be
accomplished is very limited. A two layer feedforward network can overcome many
restrictions but did not present a solution to the problem of how to adjust the weights
from input to hidden units. The central idea behind this solution is that the errors for
the units of the hidden layer are determined by backpropagating the errors of the units
of the output layer. For this reason the method is often called the backpropagation
learning rule.
Understanding the Backpropagation
The backpropagation supervised learning algorithm is used to find weights in
multilayer feedforward networks. The backpropagation algorithm is conceptually
simple. Following each input data vector, the network performance is evaluated on the
target values in the validation set. The errors resulting from the comparison of the
22
actual and target output values are propagated backward through the network, and
the weight values are adjusted to minimize error. With respect to neural networks,
the performance criterion is the minimization of squared error. Therefore, the total
system error is expressed as follows:
E= 1 ∑ ( ɛn )2 = ½ ||ɛ||2 2.7
Where, ɛn is the error of the nth pattern, and ε is a vector with elements ɛn.
The problem of learning in neural networks is formulated in terms of the minimization
of the error function E. This error is a function of the adaptive parameters (weights
and biases) in the network. Many learning laws are in common use. Most of these
laws are some sort of variation of the best known and oldest learning law, Hebb's
Rule. Research into different learning functions continues as new ideas routinely
show up in trade publications. Some researchers have the modeling of biological
learning as their main objective. Others are experimenting with adaptations of their
perceptions of how nature handles learning. Either way, man's understanding of how
neural processing actually works is very limited. Learning is certainly more complex
than the simplifications represented by the learning laws currently developed
(Anderson, 1992).
The backpropagation algorithm is the most practical and commonly used model for
neural networks. Some of the well-known types of backpropagation algorithms are
(Jacobsson, 2001):
i. Gradient descent with adaptive learning rate backpropagation: is a network
training function that updates weight and bias values according to gradient
descent with adaptive learning rate.
ii. Gradient descent with momentum & adaptive learning rate backpropagation: is
a network training function that updates weight and bias values according to
gradient descent momentum and an adaptive learning rate.
iii. Levenberg-Marquardt backpropagation: is a network training function that
updates weight and bias values according to Levenberg-Marquardt
optimization.
23
Chapter 3
MODEL DEVELOPMENT USING ANN
Artificial neural network is being used for model development. Four steps are
followed to develop the model. There four steps are laid down below in network
designing process:
i. Assemble and pre-process raw data to get training data
ii. Create the network object.
iii. Train the network
iv. Simulate the network response to new inputs
Neural network design requires a lot of hard work, collected raw data are pre-
processed before using them in network training, network object is created,
performances are monitored, parameters adjusted, connections added, rules modified,
and on and on until the network achieves the desired results.
3.1 Data Pre Processing
The objective of data pre-processing is to produce the training set of the NN,
which represents the relationship of network inputs and outputs. Preprocessing
means that the existing data is processed (in some way) before the network is trained
on it. By preprocessing the data, the problem may be much more suitable for the
network. The major tasks in data preprocessing are (Williams, 2000):
Data cleaning Normalization input and output data sets.
24
3.2 Data Cleaning Many data sets are imperfect due to the presence of missing values and noise in
the data. Incomplete data comes from
data value not available when collected
Different consideration between the time when the data was collected and
when it is analyzed.
human/hardware/software problems
Noisy data comes from the process of data
collection
entry and transmission Inconsistent data comes from
Different data sources
Functional dependency violation To handle data imperfection, data cleaning algorithms must be developed which can
Fill in missing values
Identify outliers and smooth out noisy data
Correct inconsistent data
Resolve redundancy caused by data integration The various methods that can be used to deal with data cleaning requirements are
thoroughly discussed in issues around data mining. In this section a program is
written to identify outliers and remove them before the data is used to train a network
3.3 Normalization of Input and Output Data Sets
This normalization is very critical, if the input and output variables are not of the
same order of magnitude, some variables may be given more significance than they
otherwise would have. The training algorithm is forced to compensate for order-of-
magnitude differences by adjusting network weights, which is not very effective for
25
many of the training algorithm i.e. backpropagation algorithm. For example, if input
variable 1 has a value of 10000 and input variable 2 has a value of 10, the assigned
weight for the second variable going into a node of hidden layer 1 must be greater
than that of the first variable to have any significance. In addition, the typical transfer
function, such as a sigmoid function, cannot distinguish between two different values
of input values when the latter are large because they yield identical threshold output
values of 1.0. For example, using the sigmoid function, when xi =5, we find f (xi)=
0.993.
This section mainly describes different normalization procedures. The first
normalizes each variable, xi, in the data set to between 0 to 1 by dividing its value by
the upper limit of that variable, xi, max to give a normalized variable, xi, norm.
When xi ranges between 1500-4500, and normalization factor of xi, norm= 5000, is
assigned, a normal distribution results between 0.3 to 0.9. One limitation of this
method is that it does not utilize entire range of the transfer function and the most
important that data are oscillating with higher value to lower one. Uniform
distribution is not ensured. Equation 3.1 illustrates that only a small portion of the
transfer function corresponds to xi values of 0.3 to 0.9 and -0.3 to -0.9. The weight
factors can broaden and shift this range to include a larger region of the transfer
function.
However, as the number of variables and weight factors increase, these adjustments
become more difficult for training algorithm. As a result, this normalization method is
adequate for many simple networks, but problems can arise as the network
architectures become more complex. The second method expands the normalization
range so that the minimum value of the normalized variable, xi,norm is set at 0 and
the maximum value, xi,norm is set at one. Normalized variable xi,norm is defined by
using the minimum and maximum values of the original variable, xi,min and xi,max
respectively.
26
This method significantly improves on the first method by using entire range of the
transfer function, as equation 3.2 illustrates. Another benefit from this method is that
every input variable in the data set has a similar distribution range, which improves
training efficiency.
The third technique normalizes the data set between limits of -1 and +1, having the
average value set at 0. We call this technique the zero-mean normalization method
and represent the normalization variable, xi, norm by:
and Ri, max = maximum [ Xi, max- Xi, avg- Xi, min] 3.4
Where xi is an input or output variable, xi,avg is the average minimum value of
the variable, xi,max is the maximum value of the variable and Ri,max is the
maximum range between the average value and either the minimum or the maximum
value.
As in the second method, the zero mean method utilizes the entire range of the
transfer function and every input variable in the data set has similar distribution
range. Moreover, this method gives some meaning to the values of the normalized
variable: 0 represents normal state of the variable, -1 represents a very low level of
the variable and +1 represents a very high level of the variable.
In addition, by setting all of the normal states of the variables to zero, the network
will always have a standard structure that makes training more consistent from one
problem to the next and also more efficient. Specifically, all networks should
normally predict output responses of approximately 0 (nominal value) for a set of
input variables at their nominal values of 0. Therefore, the network is essentially only
training deviations in the output variable to various deviations in the input variables.
27
3.4 Coding for Data Pre-processing
The accuracy of the network entirely depends on the data that are used to train the
network. During pre-processing, it is important that the data cover the range of inputs
for which the network will be used. Multilayer networks can be trained to generalize
well within the range of inputs for which they have been trained. However, they do
not have the ability to accurately extrapolate beyond this range, so it is important that
the training data span cover the full range of the input space. Sigmoid transfer
functions are generally used in the hidden layers. These functions become essentially
saturated when the net input is greater than three. When this happens at the beginning
of the training process, the gradients will be small, and the network training will be
very slow. In the first layer of the network, the net input is a product of the input
times the weight plus bias. Functions get saturated when input is large or vice versa.
But normalization before applying is standard practice. Generally, the normalization
step is applied to both the input vectors and the target vectors in the data set. In this
way, the network output always falls into a normalized range. The network output
can then be reverse transformed back into the units of the original target data when
the network is put to use in the field. It is easiest to think of the neural network as
having a preprocessing block that appears between the input and the first layer of the
network and a post-processing block that appears between the last layer of the
network and the output, as shown in the following figure.
Fig. 3.1: Preprocessing and Post-processing within network object.
To the input and output of a network, processing function is assigned by: net.inputs{1}.processFcns net.outputs{2}.processFcns
where the index 1 and 2 refers to the first input vector and output return from a two-
28
layer network. According to data transformation and integration steps all data is
normalized to a range between 0.15-0.85. Afterwards all pre-processed input and
output data is arranged in array form by the following MATLAB code.
In this code, data set is divided to 50% data for training, 25% testing and 25% for validation. 3.5 ANN Structure
After getting the training data set, the neural network can be built. To do this the
network structure should be defined first. Defining the network structure includes:
Network structure selection
Sizing the network structure
Training the neural network 3.5.1 Structure Selection
There is no well-defined procedure or rule to be used in building a neural network.
Data preprocessing – structure selection – network sizing – network training steps
are interrelated and the designer need to establish methods to develop a set of
competing neural network models and a way of performance measure to select the
best one. Artificial neural networks are a set of several different models featuring a wide
variety of different architectures, learning strategies and applications. At present
several types of networks specialized in carrying out various tasks are distinguished.
Table 3.1 categorizes the different types of artificial neural networks.
Category 1 networks are the most powerful, versatile, and reliable nonlinear
classifier recognizers (Looney, 1997). The networks in Category 2 may be trained to
some extent to adjust the field of attraction for the different classes, but are not yet
sufficiently reliable or efficient. Category 3 networks are self-organizing and
perform linearly separable clustering of data. The nature of the problem we are trying
to solve determines which neural network will be employed.
29
MLPs are perhaps the most popular network architecture in use today; this network
has a simple interpretation as a form of input-output model, with the weights and
biases the free parameters of the model. Such networks can model functions of
almost arbitrary complexity. Largely most researchers who use artificial neural
networks usually use the feed forward type and MLPs in particular (Looney, 1997).
Basing network topology selection on this confirmed fact, this thesis applies MLP
type neural networks to different chemical engineering modeling problems. Network
sizing and network training, discussed in the following section are mainly concerned
with issues related to MLP.
Table 3.1: Hierarchy of Artificial Neural Networks
Category Type Name of network type
1 Feedforward
(FANNs)
MLPs (multiple-layered perceptrons)
FLNs (functional link networks)
RBFNs (radial basis function
networks)
LVQNs (learning vector quantization
networks) 2 Recurrent
(RNNs)
Hopfield networks (random serial)
3 Self-organizing Maps(SOMs)
Kohonen‟s SOFMs (self-organizing
feature maps)
Sprecht‟s probabilistic networks
Bezdek‟s fuzzy c-means networks
Hybrid learning vector quantization
Grossberg‟s ART networks (adaptive
resonance theory) SOLVQNs (self-
organizing LVQNs)
30
3.5.2 Sizing the network structure
Let N be the number of input branching nodes, M be the number of hidden neurons, J
the number of output neurons, Q be the number of exemplar vectors for training. The
network architecture is determined by the numbers N, M, and J. The main questions in
sizing a neural network are:
How many layers of neurons to use?
How many input nodes to use?
How many neurons in the hidden layers should we use?
How many neurons should we use in the output layer?
The number of layers to use is provided by the Hornik-Stinchcombe-White result
stated as a feed-forward artificial neural network with two layers of neurodes and non-
constant non decreasing activation function at each hidden neurode can approximate
any piecewise continuous function from a closed bounded subset of Euclidean N-
dimensional space to Euclidean J-dimensional space with any pre-specified accuracy,
provided that sufficiently many neurodes be used in the single hidden layer.
The data determines N, J and Q. The number N of input nodes must be the number N
of feature in the feature vectors, so that once a set of feature is chosen, the number N
is fixed. J will fix the number of output neurons. The remaining task in network
sizing is to set the number of hidden neurons- M. unfortunately, there is currently
no universal guideline for determining the optimal number of hidden neurons. The
selection of the number of hidden neurons is often the result of empirical
experimentation combined with trial and error.
31
Selection of Proper Transfer Function
Three main transfer functions used in network training are sigmoid, hyperbolic tangent and
purelin. The sigmoid and hyperbolic tangent transfer functions perform well for the
prediction and process-forecasting. The hyperbolic tangent transfer function generally
outperforms the sigmoid transfer function. As the biological and chemical processing systems
become more complex and nonlinear, the advantages of the hyperbolic tangent transfer
function become more apparent. When hyperbolic tangent transfer function is superimposed
over sigmoid function, Figure 3.2 shows, two features distinguish the hyperbolic tangent
function:
i. The slope of the hyperbolic tangent function is much greater than the slope of the
sigmoid function.
ii. The hyperbolic tangent function has a negative response for a negative input value and
a positive response for a positive input value, while the sigmoid function always has a
positive response.
The fact that the hyperbolic tangent function has the greater slope means that it shows a
greater response to a small deviation in the input variable. Therefore, it can better distinguish
between small deviations in the input variable and can generate a much more nonlinear
response.
The second main feature of the hyperbolic tangent transfer (an output response has the same
sign as the input value) is critical for network nodes. This feature of the hyperbolic tangent
transfer function gives some meaning to a nodes output value: 0 represents the normal state
(average) of a node, -1 represents a very low response level and +1 represents a very high
response level.
Both the zero-mean normalization method and the hyperbolic tangent transfer function should
normally predict output responses of approximately 0 (nominal value) for a set of input
variables at their nominal values of 0. Within this structure, when the input variables are
nominally 0, the inputs to the nodes of the first hidden layer are also 0. Consequently, the
outputs of those nodes are 0 when using a hyperbolic tangent transfer function.
32
Similarly, the inputs and outputs of the remaining hidden layer and the output layer also 0. In
short, before any training takes place, the network already correctly predicts the nominal case
and essentially only has to be trained for deviations from that case. In comparison, a 0 input to
a sigmoid transfer function produces an output response of 0.5, which means that the network
must also adjust the initial weights to train the nominal case.
Fig.3.2: The hyperbolic tangent function superimposed over the sigmoid function.
3.7 Initializing the weight Factor Distribution
Prior to training a neural network, one must first initialize the weight factors, wij, between
the nodes of the hidden layers. The weight factors are set randomly with either a uniform or
Gaussian distribution. Here Gaussian distribution have found effective for the case study.For
neural network that are relatively simple, the initial distribution of the weight factors is not
particularly critical. The initial distribution set by the NN Tool of MATLAB, for instance,
almost always performs adequately in network training. The initial weight-factor distribution
is normally set to a fairly narrow range and allowed to broaden using high learning rates and
high momentum coefficients in the early stages of the training process. Problems can occur,
33
however, when using very large data sets and/or complex network architectures.
For complex networks, the weight-factor distribution do not broaden much during network
training and therefore the initial values is set to coincide with our normalized input and
output variables. 3.8 Selection of ANN Parameters
The ANN topology corresponds to a feed-forward multilayer perceptrons. This approach is
the most common architecture of an ANN. Multilayer feed-forward ANNs have two different
phases: a training phase (sometimes also referred as the learning phase) and an execution
phase. In the training phase the ANN is trained to output a specific value when given a set of
inputs. This is done in a set of input/output pairs called the training set.
An ANN training algorithm uses a set of parameters that will determine the weights between
the layers of the ANN are adjusted. If these parameters are chosen in a proper way, the
training will adjust the weights successfully and the ANN will perform perfectly in the
training set. This is why it is extremely important to determine correctly these parameters.
This part proposes a way to detect which parameters affect the training the most. This can be
done by proving different sets of parameters and then measuring the performance in training
set and the performance in validation set. Once having these two measures, then one can use
a feature selection algorithm to find those variables (parameters) that are the most correlated
with the results obtained. The performances in the training and validation sets are easily
obtained by testing the trained ANN in both sets and then calculating the absolute percentage
of error.
Parameter Influence Determination picks an ANN with the same architecture, initial weights
on its layers and the same training algorithm but with a different set of parameters every time
and trains it for 1000 epochs. The number of epochs was chosen arbitrarily and is sufficiently
large to guarantee that if the training algorithm behaves strangely.
This particular set of parameters is then added as a row to an observation matrix that the
feature selection algorithm needs as an input. Similarly the performance in the training set
and the performance in validation set are added to an entry of two different vectors. At the
34
end, the variable selection algorithm is going to solve the systems given by equation 3.5 and
3.6, where A is the observation matrix and b is the performance in either the training or
validation set. Ax=bt 3.5
Ax=by 3.6
The variable selection algorithm is going to obtain identified x that fits the best the
performance vectors by bT or bY. Even more important that the coefficient x are the order
import by the feature selection algorithm to the columns of the matrix A. The columns of A
are the parameters of the learning algorithm so the column select first by the feature selection
algorithm is the parameter most correlated with the performance of the ANN.
This procedure is done in the training and validation set. One thing is that one has to keep in
mind that the procedure is used to vary the parameters. It is well known that certain
parameters have default values, such case the learning rate. The combination of parameters
will have to contain this default values but also others that in the best case lead us to better
results, but obviously one need to take into consideration the time needed to train the network
and so one cannot test an infinite number of different parameter values. Procedure is
performed in a time series as follows:
1. Creation of a set containing sets of different values for the parameters.
2. Training of an ANN with a set of values for the parameters and the time series.
3. Evaluation of the performance of the ANN in the training and validation set.
4. Adding this set of parameters as a row of the feature selection input matrix.
5. Addition of the two measurements obtained in 3 to the input vectors.
6. If there is still a set with values that has not been evaluated go to 2.
7. Run the feature selection algorithm two times, each time with the matrix formed in 4 and one vector created in 5.
35
3.9 Train the Network
Once a network has been structured for a particular application, that network is ready to be
trained. There are two approaches to training - supervised and unsupervised. Supervised
training involves a mechanism of providing the network with the desired output either by
manually "grading" the network's performance or by providing the desired outputs with the
inputs. Unsupervised training is where the network has to make sense of the inputs without
outside help.
The vast bulk of networks utilize supervised training. Un-supervised training is used to
perform some initial characterization on inputs. However, in the full blown sense of being
truly self-learning, it is still just a shining promise that is not fully understood, does not
completely work, and thus is relegated to the lab (Anderson, 1995).
3.10 Validation & Testing Artificial neural networks are increasingly used as non-linear, non-parametric prediction
models for many engineering tasks such as pattern classification, control and sensor
integration. Neural network models are data driven and therefore resist analytical or
theoretical validation. Neural network models are constructed by training using a data set, i.e.
the model alters from a random state to a “trained” state, and must be empirically validated.
The evaluation and validation of an artificial neural network prediction model are based upon
one or more selected error metrics. Generally, neural network models which perform a
function approximation task will use a continuous error metric such as mean absolute error
(MAE), mean squared error (MSE) or root mean squared error (RMSE). The errors will be
summed over the validation set of inputs and outputs, and then normalized by the size of the
validation set. Some practitioners will also normalize to the cardinality of the output vector if
there is more than one output decision, so the resulting error is the mean per input vector and
per output decision.
36
Chapter 4
PLANT SIMULATION
4.1 Introduction
In the recent years, the study on the plant design, control, and optimization had been done on
plant simulation to generate better control system and optimize the process. Gas processing
plant was simulated both steady state & Dynamic state with the HYSYS simulator. The
Brahmanbaria gas processing plant contains several standard unit operations that are typical of
many chemical plants. The plant has specially one distillation unit operation. This chapter will
focus on Brahmanbaria gas processing plant simulation steady state & Dynamic state, its fault
detection & diagnosis by using Neural Network.
37
4.2 PROCESS DESCRIPTION OF BRAHMANBARIA GAS
PROCESSING PLANT (BGFCL), BRAHMANBARIA, BANGLADESH
Fig.4.1: Process Block Diagram for process description.
38
Wet gas comes from two wells through pipe lines. The two wells have different capacity and
their total capacity is 40 MMscfd. As gas contains different types of contaminants, it does
require refining. Then, gas goes to air cooler (AC-101) which helps to reduce gas temperature
and pressure. Gas contains heavier hydrocarbon, water. It allows for flowing through to two
phase vertical separator (V-100 & 101). Here, liquid separates from gas. The overhead product
gas goes to gas/gas exchanger (E-201) shell side where gas loses its temperature .After the
gas/gas exchanger (E-201) gas goes to another two phase vertical separator (V-101). Here also
gas loses its pressure and temperature and overhead product allow for passing gas/gas
exchanger (E-201) tube side. From this two phase‟s separator, bottom product liquid stores
liquid storage tank (V-102). From the gas/gas exchanger gas goes to sales line through the
meter skid.
The heavier liquid is collecting from two phase vertical separators in liquid storage tank (V-
102). The heavier hydrocarbon is flowing to distillation column after pre-heating by steam
heater (E-100). The distillation column has ten trays. The distillation column (D-501) has
condenser and re-boiler unit. The feed liquid is fractionated according to temperature
difference and produce different types of product. The overhead product and bottom product
ratio depends on feed quality, feed capacity, number of trays and reflux ratio. In this plant,
motor spirit (MS) produce 15-20% of total feed and High speed diesel (HSD) produce 85-80%
of total feed. The overhead product and bottom product can be maximize, minimize by
changing some parameters like reflux ratio, temperature as numbers of tray and feed is fixed.
39
4.3 Steady state simulation of Brahmanbaria Gas Processing Plant
The plant simulation has completed by using Brahmanbaria gas processing plant well-1 and
well-7 gas composition, flow rate, temperature, pressure, different unit operations
specification like valve specification, cooler specification, vessel specification, heat exchanger
specification, pump specification, distillation column speciation. The whole plant steady state
simulation of the plant was configured in HYSYS simulator.
Fig.4.2: Steady state HYSYS Simulation.
40
4.4 Dynamic state simulation conversion from steady state simulation of
Brahmanbaria Gas Processing Plant
Dynamic simulation HYSYS model has been developed from steady state HYSYS mode by
adding some unit operations like PID control valves, column sizing, using dynamic assistant,
and controller operations. Valve operations have been added between separator, mixer, and
column operations, a heater operation has been also added between the mixer and column
operation for dynamic simulation purposes as well as installed a heater allows varying the
temperature of the feed entering the column.
Before run the simulation case in dynamic mode, the degrees of freedom for the flow sheet
has been reduced to zero by setting the pressure-flow specifications. It is also necessary to size
the existing valves, vessels, coolers, and heat exchangers in the main flow sheet and the
column sub- flow sheet. The sizing parameters like valve Cv value, vessel volume and cooler/
heat exchanger K- value must be specified for these unit operations. In addition, it
automatically sets the sizing parameters of the equipment in the simulation flow sheet.
Later on process parameters has been monitored in case of normal operations and abnormal
operations in dynamics simulation state.
41
Process simulation diagram for Dynamics state condition
Fig.4.3: Dynamic state simulation of Brahmanbaria Gas Processing Plant (BGFCL)
42
4.5 Dynamics Monitoring
Different parameters observe in dynamic simulation for normal and abnormal condition. It has
been created a strip chart to monitor the general trend of key variables. From the data book, it
has been added all of the variables that would like to manipulate or model. In dynamic
simulation, parameter can observe by running strip.
For neural network fault detection and diagnosis analysis, data has been created by
considering for different fault scenario like Tower valve full open (D-501), Tower valve full
close (D-501), High pressure separator (V-101) liquid outlet valve and low temperature
separator (V-201) liquid outlet valve full open and full close condition.
The disturbance was created in the dynamic simulation utility of HYSYS to simulate process
condition in the fractionation column feed valve full close and monitored trend analysis of it.
From the trend analysis, it is observed that the temperature profile of feed line has
disturbances and as well as molar feed profile disturbances in figure.4.4. In operating plant
operations team face difficulties during fractionation column feed line valve problem. As
consequences, top production and bottom production become low as well as re-boiler
temperature become abnormal and also tripped at high high temperature. As long term,
production become hampered and need to count lose production opportunity (LPO).
43
Fig. 4.4: Disturbance profile of fractionation column during feed valve full close.
44
The fault has created for considering high pressure separator liquid valve and low temperature
separator liquid valve full close to observe tower feed trend analysis. By trending analysis, it
is very clear to us that tower molar feed, temperature and pressure profile become disturbances
in figure 4.5 the tower top product and bottom product become abnormal range where in
normal conditions productions within design range. For this reason, top production and bottom
production become low as well as re-boiler temperature become abnormal and it also tripped
at high high temperature. As long term, production become hampered and need to count lose
production opportunity (LPO).
45
Fig. 4.5: Disturbance profile of fractionation column during separators valve full close.
46
4 .6 Steady State Data and Dynamic Data
The steady state simulation has stimulated by using brahmanbaria gas processing plant unit
operations design data, well gas composition, well flow, pressure, temperature etc. The steady
state simulation data has compared with the plant live data. Dynamic simulation has been
enthused from steady state simulation in HYSYS simulator. In dynamic simulation, different
process parameters monitored by using strip chart as well as sample data has collected by
considering normal and abnormal operations for NN analysis. From the comparison, it is
found that steady state and dynamic state simulation data‟s are almost similar.
47
4.7 Simulation & Validation
The Brahmanbaria gas process plant simulation data had to be validated with the actual
process to ensure its reliability, exactness and relevant. The Brahmanbaria gas process plant
simulation is compare with several variables on the actual plant data with steady state and
dynamic state to make the comparison as in Table 4.1, 4.2, 4.3, 4.4, 4.5 & 4.6 The result is
very overwhelming and the Brahmanbaria gas process plant simulation is proven to be a
reliable and relevant plant simulation.
Table 4.1: Sales gas composition Sales gas compositions comparisons
Component Plant data
Steady state
data
Dynamic state
data
Methane 0.95649 0.95649 0.948688234 Ethane 2.59E-02 2.59E-02 2.63E-02 Propane 7.16E-03 7.16E-03 7.65E-03 i-Butane 1.82E-03 1.82E-03 2.12E-03 n-Butane 1.46E-03 1.46E-03 1.80E-03
i-Pentane 1.16E-03 1.16E-03 1.76E-03 n-Pentane 7.19E-04 7.19E-04 1.23E-03 n-Hexane 4.87E-04 4.87E-04 1.44E-03 n-Heptane 4.42E-04 4.42E-04 2.73E-03 n-Octane 8.94E-05 8.94E-05 1.29E-03
H2O 2.94E-05 2.94E-05 8.46E-04 CO2 1.66E-03 1.66E-03 1.66E-03 N2 2.56E-03 2.56E-03 2.53E-03
48
Table 4.2: Steady state and dynamic state data comparison Sales gas Steady state data Plant data Dynamic state data
Flow (MMscfd) 39.17 39.17 39.7 Tower Feed
Pressure (psia) 200 200 199.61 Top Product
Pressure (psia) 200 200 198.35 Temperature (°F) 338 338 304 Bottom product
Pressure (psia) 200 200 199.24 Temperature (°F) 480 482 482.98
Table 4.3: Steady state and dynamic state data comparison Sales gas Steady state Dynamic state
Sales gas Vapor
Phase Sales gas
Vapor
Phase
Vapor/Phase fraction 1 1 0.99 0.99 Temperature (F) 39.38 39.38 90.31 90.31 Pressure (psia) 1030 1030 1090 1090 Molar flow (MMscfd)
39.17
39.17
39.7
39.7
Mass flow (Ib/hr) 73162.94 73162.94 76271.71 76254.1 Std Ideal Liq Volm (bbl/d)
16182.04 16182.04 16553.65 16551.71
Molar Entropy (Btu/Ibmole-F)
34.17 34.17 35.16 35.16
49
Table 4.4: Steady state and dynamic state data comparison Tower Feed Steady state Dynamic state
Tower
Feed
Vapor
Phase
Liquid
Phase
Tower
Feed
Vapor
Phase
Liquid
Phase
Vapor/Phase fraction 0.0013 0.00136 0.785 0.21 0.21 0.48 Temperature (F) 247 247 24.73 106.25 106.25 106.25 Pressure (psia) 200 200 200 199.61 199.61 199.61 Molar flow (MMscfd) 0.57 0.0078 0.45 0.27 0.006 0.13 Mass flow (Ib/hr) 4350.11 156 4106.47 1703.8 129.73 1411.1 Std Ideal Liq Volm (bbl/d) 452.51 3.34 433.52 181.2 26.51 143.5 Molar Entropy (Btu/ Ibmole-F) 17.16 38.37 18.4 24.08 39.96 23.21 Table 4.5: Steady state and dynamic state data comparison Top product Steady state Dynamic state
Ovhd
Vapor
Phase Ovhd-1
Vapor
Phase
Liquid
Phase
Vapor/Phase fraction 1 1 1 1 0 Temperature (F) 338.07 338.07 304.31 304.31 304.31 Pressure (psia) 200 200 198.35 198.35 198.35 Molar flow (MMscfd) 0.49 0.49 0.23 0.23 0 Mass flow (Ib/hr) 3445.95 3445.95 1181.92 1181.92 0 Std Ideal Liq Volm (bbl/d) 364.48 364.48 130.43 130.43 0 Molar Entropy (Btu/Ibmole-F) 51.2 51.2 47.67 47.67 41.84
50
Table 4.6: Steady state and dynamic state data comparison Bottom Product Steady state Dynamic state
Liquid
Prod
Vapor
Phase
Liquid
Phase
Liquid
Prod-1
Vapor
Phase
Liquid
Phase
Vapor/Phase fraction 0.0033 0.0033 0.96 0.2 0.2 0.79 Temperature (F) 480.42 480.42 480.42 482.98 482.98 482.98 Pressure (psia) 200 200 200 199.24 199.24 199.24 Molar flow (MMscfd) 0.0073 0.00244 0.0707 0.0042 0.00867 0.0033 Mass flow (Ib/hr) 904.16 30.06 874.1 521.86 107.39 414.46 Std Ideal Liq Volm (bbl/d) 88.03 2.92 85.1 50.75 10.45 40.3 Molar Entropy (Btu/Ibmole-F) 52.94 62.08 52.62 54.21 61.62 52.28 Liq Volm flow @ std cond (bbl/d) 87.75 2.92 84.83 50.59 10.41 40.17
51
4.8 Summary
Steady state and Dynamic state simulation has stimulated by using Aspen HYSYS Simulator
guide line. Steady state simulation feed well 1&2 data has collected from Brahmanbaria gas
processing plant data. Also vessel, mixer, cooler, pump, heat exchanger, distillation column
data collected from plant data specification. Further information was found Sultana R, Syeda.,
Ahmed, Suman., Rahman, Md Bazlur., Mehfuz, Omit., and Shamsuzzaman, Razib., 2008
B.Sc. Design.
52
Chapter 5
ANN Fault detection and diagnosis modeling
5.1 Introduction
The use of Neural Networks (NNs), in all aspects of process engineering activities, such as
modeling, design, optimization and control, has considerably increased in recent years
(Mujtaba and Hussain, 2001). Different NN based techniques (architecture, training) have
been adopted in different field of science to overcome the difficulties of first principle based
modeling. The non-linear relationship between input and output of a system can be built up
cost effectively by NNs.
In this chapter, a wide range of non-linear data sets from HYSYS simulated Brahmanbaria gas
processing plant have been presented for training, testing and validation
53
5.2 NN architecture NN provides a non-linear mapping between input and output variables and is useful in
providing cross-correlation among these variables without modeling and simulating the
system. The mapping is performed by the use of processing elements and connection weights
(Aldrich and Slater, 2001). The architecture of NN consists of a number of layers, a number of
neurons; transfer functions and weights and biases and how layers are connected among
themselves. With the increase of number of layers and neurons, the NN‟s capabilities of
approximating complex functions increases (provided data are not over fitted). In process
engineering feed forward network whose signals flow in the forward direction from the input
units to the output units and incorporates feedback in its operation are widely used because of
its simplicity and available mathematical algorithms to perform its function. A typical NN
architecture is shown in Figure 3.1.
Figure 5.1: A Typical NN Architecture
54
5.3 ANN Model Developments
The neural network model comparison is mainly used to choose the optimum number of
neurons in the hidden layer. This program let the user to do an experiment on the effect of
changing different network parameters like:
Number of neurons in the hidden layer.
Data transformation.
Transfer function in the hidden layer.
Transfer function in the output layer.
Training algorithm.
Parameters of the given training algorithm.
The performance of a neural network during training is measured based on the mean squared
error. After completing a training step the overall performance of a model is measured. Using
only numeric values of error functions are usually deceptive while judging whether the model
is efficient or not. So, in the neural network modeling approach of this thesis, graphical
outputs of error distributions were also used to help the model selection step.
55
Data Normalization/Preprocessing Group the data into Training, Validation and Testing Set Construct of set of Neural Network with a varying number of neurons in hidden layer and select baseline neuron as 1.
Select the Network Architecture
Select training algorithm and stopping criteria
Train and Validate the network
Post-process the data Calculate % of abs error & avg. abs error for network model
No
Satisfactory
Yes Showing graphical error evaluation for different network model in terms of number of neuron within hidden layer Select the Optimum network architecture (Number of Neurons within Hidden Layer) Figure 5.5: Algorithm to extract weights and biases from optimized network
56
Neural Network model development for Brahmanbaria gas processing plant includes the
following steps:
Step 1: Data for the stated problem is collected from simulated HYSYS model. Data
integration requirements are done manually before using the MATLAB programs. The
collected and integrated data is then stored in a separate data file.
Step 2: Data transformation is done before starting the network training. The preprocessed
data is divided in to three different sets, training set (50%), testing set (25%) and validation set
(25%).
Step 3: 1st section of MATLAB program, do data transformation, network construction,
network training, and selecting the best model.
Step 4: 2nd section MATLAB program, further analyze the performance of the
model selected in step 3, then the neural network description is extracted and saved in a
separate file.
The computer program which is stated in Step 3, construct a set of competing neural network
models. After training them simultaneously it compare the performance of each model.
Graphical outputs from this computer program are used to select the best neural network
topology.
57
5.4 Neural Network Fault detection scheme
Any non-linear relationship between input and output of a system can be captured effectively
using NNs. NNs are consist of large number of primitive computational elements called
neuron. NN based fault detection are based on classification of historic process knowledge.
The development of the NN based system involves selecting suitable architecture to
differentiate between the faults of the process with the normal condition. The steps of NN
based fault detection system in this work is shown in Fig 5.1 and Fig 5.2 .NN is trained,
validated and tested are similar to state as is shown in Fig.5.3 and Fig.5.4 .Historic data of 10
key variables which affects the different mode of operations (types of faults and normal) is fed
to the input layer and mode of operations fed to the output layer of the NN.NN is trained,
validated and tested using dynamic data. Several multi layered feedforward neural networks
with varying configurations and Levenberg Marquardt back propagation algorithm are also
employed in the training and testing process of the to obtain optimum neural network
architecture.
Aspen HYSYS Dynamic
Model
Key Parameters to monitor by NN
Input Signal to Input layer of NN
Output Signal from Output layer of NN
Input parameters
and Disturbances
Identification of Normal operation and Types of Faults
Online Neural Network (Fixed Architecture)
Fig. 5.2: Online NN based fault detection system.
58
Aspen HYSYS Dynamic
Model
Key Parameters to monitor by NN
Input Signal to Input layer of NN
Output Signal from Output layer of NN
Input parameters
and Disturbances
Identification of Normal operation and Types of Faults
Offline Neural Network (Training, Validating and Testing )
Key Parameters of 1000 secs Simulation of Different States (Normal and Abnormal) Stored by NN
Fig. 5.3: Offline NN based fault detection After getting option, network architecture the offline process is completed. Fixed architecture base NN has been used for on line testing. In the online mode data is directly feed from the HSYS simulator. Fixed NN base architecture online NN base architecture identified normal operation types of fault.
Fig. 5.4: Neural Network Back propagation Training Scheme.
59
The development of neural network model follows the standard procedure of system
identification model. Generally, standard procedure of system identification model involves
several procedures in order to make sure that the model is properly developed:
Selection of input and output variable
Training data generation
Selection of network structure
Selection of training and validation
Selection of input and output variables
For the application machine learning approaches, it is important to properly select the input
variables, as ANN‟s are supposed to learn the relationship between input and output variables
on the basis of input-output pairs provided during training. In neural network based fault
detection model, the input variables represent the operating state of the pneumatic actuator,
and the output is the condition of normal or abnormal which may cause in turn the faults. Then
these normal and abnormal conditions are taken as the output of the ANN model.
Training Data Generation
The generation of training data is an important step in the development of ANN models. To
achieve a good performance of the neural network, the training data should represent the
complete the range of operating conditions of the pneumatic actuator which contains all
possible fault occurrences.
Selection of Network Structure
To make a neural network to perform some specific task, one must choose how the units are
connected to one another. This includes the selection of the number of hidden nodes and type
of the transfer function used. The number of hidden- units is directly related to the capabilities
of the network. For the best network performance, an optimal number of hidden- units must be
properly determined using the trial and error procedure.
The ANN model used here has two hidden layer of logarithmic sigmoidal neurons, which
receives the inputs, then broadcast their output to an output layer of linear neurons, which
60
compute the corresponding values. The back propagation training algorithm, which propagates
the error from the output layer to the hidden layer to update weight matrix, is most commonly
used for feed forward neural networks.
The generated training data are normalized and applied to the neural network with
corresponding output, to learn the input-output relationship. The neural network model was
trained using Matlab program using the neural network toolbox. Based on the developed
Matlab program, the feed forward neural network model is trained using back propagation
method. At the end of the training process, the model obtained consists of the optimal weight
and the bias vector. After training the generalization performance of the network is evaluated
with the help of the test data and it shows that the trained ANN is able to produce the correct
output even for the new input.
A multi layered Feed forward networks (Fig.5.2) has been employed along with back
propagation algorithm. NN based fault detection system is trained, validated and tested using
data generated using dynamic model. Key process parameters and fault are to NN. Data sets
are sorted to avoid uneven distribution. All data are scaled to the symmetrical range of -1 to
+l. The data sets are divided into training, validation and test subsets. Several neural networks
with varying configurations and various learning strategies are also employed in the training
process of the neural networks. The following steps have been taken when developing the
model for fault detection (Fig.5.3).
61
Fig. 5.5 NN based fault detection system
Selection of training and validation
After the output variable had been selected to be test, the neural network had to be train and
validate before ready to be implemented on the Brahmanbaria gas processing plant. Neural
network can be train by two different styles of training. In incremental training the weights
and biases of the network are updated each time an input is presented to the network. In batch
training the weights and biases are only updated after all the inputs are presented. Training is
important to achieve and train the weights and biases that can estimation that similar to the
actual plant. Meanwhile, validation is a process to verify the Neural Network using unseen or
other data to test the reliability and robustness of the created neural network.
62
5.5 Result
This section presents the details of the development and testing of ANN model for fault
detection on control valve in Brahmanbaria gas processing plant. The Back propagation
algorithm was developed using MATLAB 7 Neural Network Toolbox. Plant test data
generation was designed and conducted in HYSYS. Brahmanbaria Gas processing plant has
captured the dynamic behavior. Table 5.1 shows the design of the plant model (Fig.4.3)
within Aspen HYSYS. 1700 sample data for key parameters from the normal situation to each
type of faults are simulated using Aspen HYSYS simulator. The data has divided into three
sets, a training set, a validation set and a testing set. The key data (Table 5.2 ) of different
normal operation mode using Aspen HYSYS simulator are shown in chart forms from Fig 5.6
and sample fault (Table 5.2) in the process plant generated using dynamic model are shown in
chart forms from Fig. 5.7. However, visually distinguished such data pattern is quite difficult.
Table 5.1: Input-output parameter
Stream name S-101 S-302 Heater- 501 inlet
Temperature (°F) 150 98 62 Pressure (Psia) 3000 1024 19 Mass flow rate ( Ib/hr) 57112 39789 11314 Stream name Tower feed S 601 S 611
Temperature (°F) 62 14 350 Pressure (Psia) 19 14 16 Mass flow rate ( Ib/hr) 11314 10113 1201
63
Table 5.2: Types of operation mode/disturbances criteria for Neural Network analysis.
Ope
ratio
ns
Tow
er fe
ed -
Tem
pera
ture
S 60
1-1
- Pr
essu
re
Tow
er fe
ed -
Pres
sure
S 60
1-1
- Te
mpe
ratu
re
S 61
1-1
- Te
mpe
ratu
re
S 61
1-1
- Pr
essu
re
S 30
2 -
Mol
ar F
low
S 30
2 -
Pres
sure
S 40
1 -
Tem
pera
ture
S 40
1 -
Pres
sure
Nor
mal
ope
ratio
n
[F] [psia] [psia] [F] [F] [psia] [MMscfd] [psia] [F] [psia]
121 10 19 214 225 20 24 75 101 10
121 10 19 214 225 20 24 75 104 10
119 10 17 188 225 20 25 75 104 10
121 10 19 214 225 20 24 75 104 10
119 10 17 188 225 20 25 75 104 10
Tow
er v
alve
full
open
79 15 192 10 225 20 23 75 10 66
79 15 192 10 225 20 24 75 10 67
79 15 192 10 225 20 24 75 10 65
79 15 192 10 225 20 24 75 10 67
79 15 192 10 225 20 24 75 10 65
Tow
er v
alve
full
clos
e 181 7 -76 10 225 20 24 75 10 68
182 7 -76 10 225 20 24 75 10 68
180 7 -76 10 225 20 24 75 10 68
182 7 -76 10 225 20 24 75 10 68
180 7 -76 10 225 20 24 75 10 68
HP
& L
TS v
alve
fu
ll op
en
373 25 81 10 260 20 24 75 69 374
373 25 81 10 260 20 24 75 69 374
373 25 81 10 260 20 24 75 69 374
373 25 81 10 260 20 24 75 69 374
HP
& L
TS v
alve
full
clos
e
373 25 81 10 260 20 24 75 69 374
153 18 219 10 225 20 24 75 10 166
153 18 219 10 225 20 24 75 10 166
152 18 218 10 225 20 24 75 10 166
152 18 218 10 225 20 24 75 10 166
64
Table 5.3: NN architecture for different conditions.
Operations Output neuron value Output neuron value by NN
Normal operation
1 0 0 0 0 1 0 0 0 0
1 0 0 0 0 1 0 0 0 0
1 0 0 0 0 1 0 0 0 0
1 0 0 0 0 1 0 0 0 0
1 0 0 0 0 1 0 0 0 0
Tower valve full open fault
0 1 0 0 0 0 1 0 0 0
0 1 0 0 0 0 1 0 0 0
0 1 0 0 0 0 1 0 0 0
0 1 0 0 0 0 1 0 0 0
0 1 0 0 0 0 1 0 0 0
Tower valve full close fault
0 0 1 0 0 0 0 1 0 0
0 0 1 0 0 0 0 1 0 0
0 0 1 0 0 0 0 1 0 0
0 0 1 0 0 0 0 1 0 0
0 0 1 0 0 0 0 1 0 0
HP & LTS valve full open fault
0 0 0 1 0 0 0 0 1 0
0 0 0 1 0 0 0 0 1 0
0 0 0 1 0 0 0 0 1 0
0 0 0 1 0 0 0 0 1 0
0 0 0 1 0 0 0 0 1 0
HP & LTS valve full close fault
0 0 0 0 1 0 0 0 0 1
0 0 0 0 1 0 0 0 0 1
0 0 0 0 1 0 0 0 0 1
0 0 0 0 1 0 0 0 0 1
0 0 0 0 1 0 0 0 0 1
65
Those data for each key parameter as mentioned by NN based fault detection system and fed
to neural network input layer table (5.3). Types of operation mode/disturbances criteria (5.3)
are integrated with output layer neuron of NN based system. Neuron of output layer for each
states value equal to 1 and for absent of the state‟s value equal to 0 (Table 5.3). Data sets are
sorted to avoid uneven distribution. All data are scaled to the symmetrical range of -1 to +1.
Fig.5.6: Normal Operation mode by HYSYS
66
Fig.5.7: Tower valve disturbance Operation mode by HYSYS
Training of NN Fault detection system is shown in Fig.5.8. The statistical regression plots of
(Fig.5.9 to Fig.5.10) between predicted and target data of different operation mode are plotted.
Network architecture is updated until the regression value is almost close to 1. Optimum
network is found for this work with 9 neurons in hidden layer.
67
However, if the set point of normal operation is changed whose response is not similar to the
training pattern work, NN fault detection might detect as normal operation. Hyperbolic
tangent functions are used in the input and hidden nodes. Linear functions are used in the
hidden layer and output nodes.
NN based fault detection system (Fig.5.4) is trained, validated and tested using data generated
using the dynamic model. Different faults in valve are identified before sending to NN. Neural
Network relates measurements to the faults and identify between normal and abnormal states.
The output neurons of the NN are set between the values of 0 and 1 depending on types of
fault (Fig.5.4).
Design considerations of the neural network fault detection involved are stated as is shown in
Fig.5.5. Data sets are sorted to avoid uneven distribution. All data are scaled to the
symmetrical range of -1 to +l. The data sets are divided into training, validation and test
subsets. Several neural networks with varying configurations and various learning strategies
are also employed in the training process of the neural networks.
In this work, a three layered Feed forward network has been employed along with back
propagation algorithm. The input layer contains 4, the 1st hidden layer 4, 2nd hidden layer 1
and the output layer 4 nodes. Hyperbolic tangent functions are used in the input and 1st hidden
nodes. Linear functions are used in the 2nd hidden layer and output nodes. Input and output
architecture is shown in Table 5.2 & 5.3.
68
Fig.5.8: Training of NN Fault detection system
69
Fig.5.9: Statistical regression analysis of NN predicted data with fault
Fig.5.10: Statistical regression analysis of NN predicted data with fault
Predictions by optimum NN within the training range follow the expected trends and it is
within the engineering accuracy (Fig. 5.11 and Table 5.2 & 5.3). This probe that optimum
network able to predict types of fault (here liquid control valve failure for product quality loss)
even when the network is with new inputs (test data).
70
Training of NN Fault detection system is shown in Fig. 5.8 The statistical regression plot (Fig.
5.11) between predicted and target data‟s is plotted to ensure that results generated have
satisfied. Predictions by different NN within the training range follow the expected trends and
it is within the engineering accuracy.
Fig. 5.11: Statistical regression analysis of NN predicted test data
71
Chapter 6
6.0 CONCLUSION AND RECOMMENDATION
Overview
Fault detection system is one of the main elements in safety measurement in the chemical
plant. It is very ironic to think such small system would bring such big difference and impact
on the safety, reliability and cost effective of the process. Neural Network has the ability to
process information characteristic such as nonlinearity, high parallelism, fault tolerance as
well as capability to generalize and handle imprecise information.
72
6.1 Conclusion
The development of neural network in various fields especially in fault detecting has shown
great progress. The priority of fault detection and diagnosis system in chemical and
petrochemical industry is increased. The implementation of Neural Network had provided a
reliable prediction as fault detection on Brahmanbaria gas processing plant was successfully
developed.
In this research, preliminary results shows that NN based method successfully detect the faults
of Brahmanbaria Gas processing plant. Brahmanbaria Gas Plant behavior and disturbances are
studied using HYSYS dynamic model. Feedforward NN based fault detection was developed
to identify the fault (disturbance) and no fault (normal) in plant operation. The NN based fault
detection system was being trained validated and tested using the dynamic model data.
Preliminary results show that NN based fault detection able to identify realistic fault output.
We believe that, NN based fault detection will help to avoid accident events and productivity
losses in Gas processing industry in Bangladesh and help operators to identify and monitor
multiple faults in real-time.
In future, more faults and multiple faults need to be accommodated to visualize the real plant
situation.
73
6.2 Recommendation for future work
Although the development of Neural Network can be considered as successful, there are still
areas and aspect that can be improved in the future work:
In this thesis, preliminary results shows that NN based method successfully detect the faults of
Brahmanbaria Gas processing plant. In future, more faults and multiple faults needs to be
accommodated to visualize the real plant situation.
Plant HYSYS dynamic simulation data used to perform neural network analysis of
Brahmanbaria gas processing plant, it would be better to use plant live data for best fault
detection and diagnosis analysis.
To perform comparison fault detection and diagnosis method with expert system like fuzzy
logic.
74
REFERENCES
Less, Frank. P., 1996. Loss Prevention in the Process Industries, Butterworth-Heinemann, 2nd edn, Chap.9, pp.364-398, Reed Educational and Professional Publishing., London. V. Venkatasubramanian, R. Rengaswamy, S. N. Kavuri, and K. Yin, “A Review of process fault detection and diagnosis Part III: Process history based methods,” Computers and Chemical Engineering, vol. 27, pp. 327-346, 2003. Mohd. Kamaruddin Bin Abd. Hamid., 2004.Multiple faults detection using artificial neural network. Master‟s thesis. Universiti Teknologi Malaysia, Malaysia. Sultana R, Syeda., Ahmed, Suman., Rahman, Md Bazlur., Mehfuz, Omit., and Shamsuzzaman, Razib., 2008, „Simulation of Brahmanbaria gas processing plant,‟ B.Sc. Design, Department of Chemical Engineering, Bangladesh University of Engineering & Technology, Dhaka, Bangladesh. Kamruzzaman S. 1999, „Simulation of Kailashtilla II Gas Processing Plant,‟ M.Sc. thesis, Bangladesh University of Engineering & Technology, Dhaka. HYSYS, HYSYS 3.2 user guide, Hyprotech Ltd, http://www.hyprotech.com/ Himmelblau, D.M and Hussain, M. A., 1978. Fault Detection and Diagnosis in Chemical and Petrochemical Processes, Vol-8 Berlin, Amsterdam: Elsevier Scientific Publishing Company. D. M. Himmelblau., 1978. Fault Detection and Diagnosis in Chemical and Petrochemical Processes, Elsevier Scientific Publishing Company, Amsterdam, Oxford., New York. Isermann, R., 1997. Supervision, Fault-detection and Fault-diagnosis Methods - an introduction. Control Eng. Practice. Vol. 5, No. 5, pp. 639-652. Gertler, J. J., (1998), Fault detection and diagnosis in engineering systems, New York: Marcel Dekker. Basheer , I.A. and Hajmeer M. (2000), Artificial Neural Networks: Fundamentals, Computing, Design, and Application, Journal of Microbiological Methods, 43: 3–31. Chementator., 2000. „Neural Networks Optimize Chemical Production,‟Chemical Engineering, vol.97, issue no.8, pp.29. Blanchar, D., 1994. „Applied AI News,‟ AI Magazine, vol.15, issue no.4, pp.79. Lee, R. S. T., Fuzzy-Neuro Approach to Agent Applications, Spinger-Verlag Berlin Heidelberg (2006). Patterson D. W. (1996), Artificial neural Networks: Theory and Application, Prentice
75
Hall. Anderson, D. & Mcneil, G., 1992. Artificial Neural Networks Technology, Kaman Sciences Corporation, Newyork. Jacobsson, H., Bergfeldt, N. & Lundell, S., 2001. Matlab and Neural Network Toolbox Tutorial, Oxford University Press Inc, London Bishop, C. M., 1996. Neural Networks for Pattern Recognition, Oxford University Press Inc, London. Hoskins, J.C. & Himmelblau, D.M., 1988. Artificial neural network models of knowledge representation in chemical engineering, Computer & Chemical Engineering, vol.12, issue.9-10, pp.881-890. Patterson, D. W. 1995. Artificial Neural Networks, 1st edn, pp.1-20, Prentice Hall, London. Islam, A. and Hossain, T. Z., 2001. Modeling of a chemical process using artificial neural network, B.Sc. thesis, Department of Chemical Engineering, Bangladesh University of Engineering & Technology (BUET), Dhaka, Bangladesh. Haykin. S., 1999. Neural Networks: A comprehensive Foundation, 2nd edition, Prentice Hall, London. Looney, Carl G., 1997. Pattern Recognition Using Neural Networks, Theories and Algorithms for Engineers and Scientists, Oxford University Press, London. Neural Networks 1984-2003, <http://www.statsoftinc.com/textbook/stneunet>, accessed on 18 January 2013. Sarle W. 2001, Why use activation functions?, <http://www.faqs.org/faqs/ai- faq/neuralnets/part2/section-10.html >, accessed on 17 October 2012. Williams, C., 2000., Data Preprocessing, School of Informatics, University of Edinburgh. Looney, C. G., 1997. Pattern Recognition Using Neural Networks, Theories and Algorithms for Engineers and Scientists, Oxford University Press, London. Mujtaba, I. M and Hussain, M. A., 2004. Neural Networks and Other Learning Technologies in Process Engineering, Vol-3, Imperial College Press, London. Anderson, D., and Mcneil, G., 1997. Artificial Neural Networks Technology, Data &Analysis Center for Software, Daedalian.
76
PUBLICATIONS
Sowgath, M.S., and Ahmed, Suman. 2014, „„Fault Detection of Brahmanbaria Gas
Plant using Neural Network,‟‟ on 8th International Conference, IEEE, Bangladesh.
Sowgath, M.S., and Ahmed, Suman. 2012, „„Study of Fault Detection and Diagnosis of Brahmanbaria Gas Processing Plant, Bangladesh using Neural Network,‟‟ in the Proceedings of 62nd Canadian Chemical Engineering Conference, October 14-17, 2012, Vancouver, Canada.
77
Appendix A.1
Steady state material stream.
Stream Name Well 1 Well 7 Well 1 D/S Well 7 D/S MIX-100 D/S
Vapor Fraction 0.99 0.99 0.99 0.99 0.99754 Temperature (°F) 140.00 145.00 112.23 111.79 111.9894 Pressure (psia) 1890.00 2100.00 1130.00 1130.00 1130.0000 Molar flow (MMSCFD)
18.00 22.00 18.00 22.00 40.0000
Mass flow (lb/hr) 3.492e+4 4.309e+4 3.492e+4 4.308e+4 7.803e+4 Liq vol flow (bbl/day) 7520 9220 7520 9220 1674 Heat flow (Btu/hr) -6.648e+7 -8.160e+7 -6.648e+7 -8.160e+7 -1.4808e+8 Molar Enthalpy(Btu/Ibmole)
-3.364e+4 -3.378e+4 -3.364e+4 -3.378e+4 -3.372e+4
Stream Name AC-101 D/S V-100 G/O V-100 L/O E-201 G/O Sales gas
Vapor Fraction 0.99 1.00 0.00 0.99 1.00 Temperature (°F) 91.62 90.38 90.38103 49.38 39.38 Pressure (psia) 1120.00 1090.00 1090.00 1080.00 1030.00 Molar flow (MMSCFD)
40.00 39.71 0.28 39.71 39.17
Mass flow (lb/hr) 7.8003e+4 7.628e+4 1721 7.628e+4 7.312e+4 Liq vol flow (bbl/day) 16740.76 16558.02 182.75 16558.02 16182.04 Heat flow (Btu/hr) -1.493e+8 -1.466e+8 -2.725e+6 -1.488e+8 -1.453e+8 Molar Enthalpy(Btu/Ibmole)
-3.398e+4 -3.360e+4 -8.856e+4 -3.412e+4 -3.378e+4
Stream Name Chiller
D/S
V-101 G/O V-101 L/O LV-100
D/S
LV-101 L/O
Vapor Fraction 0.98 1.00 0.00 0.15025 0.25350 Temperature (°F) 0.00 -1.89 -1.89 85.67 -17.60711 Pressure (psia) 1070.00 1040.00 1040.00 440.00 440.00 Molar flow (MMSCFD) 39.71 39.17 0.54406 0.28012 0.54406 Mass flow (lb/hr) 7.628e+4 7.316e+4 3.118e+4 1.721e+4 3.118e+4 Liq vol flow (bbl/day) 16558.02 16182.05 375.97 182.75 375.97 Heat flow (Btu/hr) -1.520e+8 -1.475e+8 -4.068e+6 -2.725e+6 -4.068e+6 Molar Enthalpy(Btu/Ibmole)
-3.476e+4 -3.430e+4 -6.809e+4 -8.859e+4 -6.809e+4
78
Appendix A.2
Table steady state material stream.
Stream Name Mix-101
L/O
V-102 inlet Tower
valve in
Vapor HC E-100 U/S
Vapor Fraction 0.236 0.294 0.00 1.00 4.00e-003 Temperature (°F) 19.89 14.39 13.28 13.28 13.078 Pressure (psia) 440.00 250.00 220.00 220.00 210.00 Molar flow (MMSCFD) 0.82 0.82 0.57 0.25 0.57318 Mass flow (lb/hr) 4.84e+4 4.84e+4 4.35e+4 4.89e+3 4.35e+4 Liq vol flow (bbl/day) 558.72 558.72 452.51 106.20 452.51 Heat flow (Btu/hr) -6.79e+6 -6.79e+6 -5.85e+6 -9.37e+5 -5.85e+6 Molar Enthalpy(Btu/Ibmole)
-7.50e+4 -7.50e+4 -9.30e+4 -3.39e+4 -9.30e+4
Stream Name Tower Feed Ovhd Liquid Prod C3Duty H-Q
Vapor Fraction 1.36e-002 1.00 3.33e-002 Temperature (°F) 24.73 338.07 480.42 Pressure (psia) 200.00 200.00 200.00 Molar flow (MMSCFD) 0.57 0.49 7.31e-002 Mass flow (lb/hr) 4.35e+4 3.44e+4 904.16 Liq vol flow (bbl/day) 452.52 364.48 88.03 Heat flow (Btu/hr) -5.82e+6 -3.73e+6 -6.07e+5 2.84e+7 2.73e+7 Molar Enthalpy(Btu/Ibmole)
-9.26e+4 -6.81e+4 -7.56e+4
79
Appendix B.1
Table dynamic state material stream
Stream Name Well 1 Well 7 Well 1 D/S Well 7 D/S MIX-100 D/S
Vapor Fraction 0.99948 0.999428 0.99871 0.99684 0.99769 Temperature (°F) 140.00 145.00 112.56 112.39 112.46 Pressure (psia) 1890.00 2090.00 1138.49 1138.49 1138.49 Molar flow (MMSCFD)
17.98 21.99999 17.98 21.99 39.98
Mass flow (lb/hr) 34888.77 43086.79 34888.77 43086.79 77975.56 Liq vol flow (bbl/day)
7514.28 9220.57 7514.28 9220.57 16734.85
Heat flow (Btu/hr) -70088849.40 -86082905.57 -70088849.40 -86082905.0 -156171754.97 Molar Enthalpy(Btu/Ibmole)
-78239.64 -78560.30 -78239.64 -78560.30 0.99769
Stream Name AC-101 D/S V-100 G/O V-100 L/O E-201 G/O Sales gas
Vapor Fraction 0.9934 1.00 0.00 0.99973 0.99994 Temperature (°F) 91.9295 91.9295 91.9295 90.6975 90.31277 Pressure (psia) 1128.4991 1128.4991 1128.4991 1118.547 1090.00 Molar flow (MMSCFD)
39.985850 39.72327 0.262578 39.7232 39.7060
Mass flow (lb/hr) 77975.55 76398.93 1576.61 76398.93 76271.71 Liq vol flow (bbl/day)
16734.85 16567.56 167.28 16567.56 16553.65
Heat flow (Btu/hr) -157417362.67 -154708707.51 -2708655.16 -154757102.95 -1.545e+8
Molar Enthalpy(Btu/Ibmole)
-79041.50 -78194.93 -207111.44 -78219.39 -78149.89
80
Appendix B.2
Table dynamic state material stream
Stream Name Chiller D/S V-101 G/O V-101
L/O
LV-100
D/S
LV-101
L/O
Vapor Fraction 0.99 1.00 0.00 0.10 0.13 Temperature (°F) 90.09 90.09 90.09 89.07475 86.70 Pressure (psia) 1104.51 1104.52 1104.51 651.85 651.85 Molar flow (MMSCFD)
39.72 39.70 1.72e-00 0.26 1.72e-002
Mass flow (lb/hr) 76398.93 76271.71 127.22 1576.61 127.22 Liq vol flow (bbl/day)
16567.57 16553.65 13.92 167.28 13.91
Heat flow (Btu/hr) -154757102.95 -154600754.51 -156348.44 -2708655.16 -156348.44 Molar Enthalpy(Btu/Ibmole)
-78219.39 -78174.36 -181755.16 -207111.44 -181755.16
Stream Name Mix-101
L/O
V-102 inlet Tower valve
in
Vapor HC E-100 U/S
Vapor Fraction 0.11 0.17 0.17 1.00 0.20 Temperature (°F) 88.90 85.50 85.50 85.50 83.91 Pressure (psia) 651.85 332.22 332.22 332.22 238.46 Molar flow (MMSCFD)
0.27 0.27 0.27 0.00 0.27
Mass flow (lb/hr) 1703.83 1703.84 1703.83 0.00 1703.83 Liq vol flow (bbl/day)
181.20 181.20 181.20 0.00 181.20
Heat flow (Btu/hr) -2865003.60 -2865003.60 -2865003.72 0.00 -2865003.72 Molar Enthalpy(Btu/Ibmole)
-205546.57 -205546.58 -205546.58 -78276.46 -205546.58
Stream Name Tower Feed Ovhd LiquidProd C3Duty H-Q
Vapor Fraction 0.21 1.00 0.20 Temperature (°F) 106.25 304.40 483.50 Pressure (psia) 199.61 199.28 200.14 Molar flow (MMSCFD) 0.27 0.23 4.20e-002 Mass flow (lb/hr) 1703.84 1181.95 521.87 Liq vol flow (bbl/day) 181.20 130.44 50.75 Heat flow (Btu/hr) -2839120.05 -1903315.98 -361543.40 0000 25886.33 Molar Enthalpy(Btu/Ibmole) -203689.34 -160686.52 -172709.31
81
Appendix C.1
Neural Network Data Analysis.
Table Normal operation parameters of BGPP
Tower feed - Temp
S 601-1
- Press
Tower feed - Press
S 601-1
- Temp
S 611-1
- Temp
S 611-1
- Press
S 302 - Molar Flow
S 302 -
Press
S 401 -
Temp
S 401 -
Press
[F] [Psia] [Psia] [F] [F] [Psia] [MMscfd] [Psia] [F] [Psia] 122 10 19 215 225 20 32 75 100 10 122 10 19 215 225 20 24 75 104 10 122 10 19 215 225 20 24 75 105 10 122 10 19 215 225 20 24 75 101 10 122 10 19 215 225 20 24 75 99 10 122 10 19 215 225 20 24 75 99 10 122 10 19 215 225 20 24 75 101 10 122 10 19 215 225 20 24 75 106 10 122 10 19 215 225 20 24 75 105 10 122 10 19 215 225 20 24 75 104 10 122 10 19 215 225 20 24 75 103 10 122 10 19 215 225 20 23 75 104 10 122 10 19 215 225 20 24 75 102 10 122 10 19 215 225 20 24 75 104 10 122 10 19 215 225 20 24 75 102 10 122 10 19 215 225 20 25 75 101 10 122 10 19 215 225 20 24 75 106 10 122 10 19 215 225 20 24 75 102 10 122 10 19 215 225 20 24 75 105 10 122 10 19 215 225 20 24 75 100 10 122 10 19 215 225 20 24 75 105 10 122 10 19 215 225 20 24 75 101 10 122 10 19 215 225 20 24 75 99 10 122 10 19 215 225 20 24 75 99 10 122 10 19 215 225 20 24 75 101 10 122 10 19 215 225 20 24 75 106 10 122 10 19 215 225 20 24 75 105 10 122 10 19 215 225 20 24 75 104 10 122 10 19 215 225 20 24 75 103 10 122 10 19 215 225 20 23 75 104 10 122 10 19 215 225 20 24 75 102 10 122 10 19 215 225 20 24 75 104 10
82
Appendix C.2
Table Tower valve full open operation parameters of BGPP
Tower feed - Temp
S 601-1
- Press
Tower feed - Press
S 601-1
- Temp
S 611-1
- Temp
S 611-1
- Press
S 302 - Molar Flow
S 302 -
Press
S 401 -
Temp
S 401 -
Press
[F] [Psia] [Psia] [F] [F] [Psia] [MMscfd] [Psia] [F] [Psia]
79 15 192 10 225 20 23 75 10 66
79 15 192 10 225 20 24 75 10 67
79 15 192 10 225 20 24 75 10 65
79 14 192 10 225 20 24 75 10 67
79 14 192 10 225 20 24 75 10 65
79 14 192 10 225 20 24 75 10 68
79 14 192 10 225 20 24 75 10 69
79 14 192 10 225 20 24 75 10 67
79 14 192 10 225 20 24 75 10 68
79 14 191 10 225 20 24 75 10 68
79 14 191 10 225 20 25 75 10 67
79 14 191 10 225 20 24 75 10 68
79 14 191 10 225 20 23 75 10 68
79 14 191 10 225 20 25 75 10 68
79 14 191 10 225 20 24 75 10 68
79 14 191 10 225 20 24 75 10 68
79 14 191 10 225 20 26 75 10 68
79 14 191 10 225 20 24 75 10 67
79 14 191 10 225 20 24 75 10 68
79 14 191 10 225 20 24 75 10 67
79 14 192 10 225 20 24 75 10 67
79 14 192 10 225 20 24 75 10 65
79 14 192 10 225 20 24 75 10 68
79 14 192 10 225 20 24 75 10 69
79 14 192 10 225 20 24 75 10 67
79 14 192 10 225 20 24 75 10 68
79 14 191 10 225 20 24 75 10 68
79 14 191 10 225 20 25 75 10 67
79 14 191 10 225 20 24 75 10 68
79 14 191 10 225 20 23 75 10 68
79 14 191 10 225 20 25 75 10 68
79 14 191 10 225 20 24 75 10 68
83
Appendix C.3
Table Tower valve full close operation parameters of BGPP
Tower feed - Temp
S 601-1
- Press
Tower feed - Press
S 601-1
- Temp
S 611-1
- Temp
S 611-1
- Press
S 302 - Molar Flow
S 302 -
Press
S 401 -
Temp
S 401 -
Press
[F] [Psia] [Psia] [F] [F] [Psia] [MMscfd] [Psia] [F] [Psia] 181 7 -76 10 225 20 24 75 10 68 182 7 -76 10 225 20 24 75 10 68 180 7 -76 10 225 20 24 75 10 68 179 7 -76 10 225 20 24 75 10 68 179 7 -76 10 225 20 24 75 10 68 180 7 -76 10 225 20 24 75 10 68 180 7 -76 10 225 20 23 75 10 68 180 7 -76 10 225 20 24 75 10 68 180 7 -76 10 225 20 24 75 10 68 180 7 -76 10 225 20 24 75 10 68 186 7 -76 10 225 20 25 75 10 68 181 7 -76 10 225 20 24 75 10 68 181 7 -76 10 225 20 24 75 10 68 181 7 -76 10 225 20 24 75 10 68 181 7 -76 10 225 20 25 75 10 68 181 7 -76 10 225 20 24 75 10 68 181 7 -76 10 225 20 25 75 10 68 182 7 -76 10 225 20 23 75 10 68 180 7 -76 10 225 20 24 75 10 68 180 7 -76 10 225 20 24 75 10 68 181 7 -76 10 225 20 24 75 10 68 182 7 -76 10 225 20 24 75 10 68 180 7 -76 10 225 20 24 75 10 68 179 7 -76 10 225 20 24 75 10 68 179 7 -76 10 225 20 24 75 10 68 180 7 -76 10 225 20 24 75 10 68 180 7 -76 10 225 20 23 75 10 68 180 7 -76 10 225 20 24 75 10 68 180 7 -76 10 225 20 24 75 10 68 180 7 -76 10 225 20 24 75 10 68 186 7 -76 10 225 20 25 75 10 68 181 7 -76 10 225 20 24 75 10 68
84
Appendix C.4
Table HP Separator and Low temperature separator valve full open operation
parameters of BGPPL
Tower feed - Temp
S 601-1
- Press
Tower feed - Press
S 601-1
- Temp
S 611-1
- Temp
S 611-1
- Press
S 302 - Molar Flow
S 302 -
Press
S 401 -
Temp
S 401 -
Press
[F] [Psia] [Psia] [F] [F] [Psia] [MMscfd] [Psia] [F] [Psia] 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 374 25 81 10 260 20 24 75 69 375 373 25 81 10 260 20 24 75 69 375 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374 374 25 81 10 260 20 24 75 69 375 373 25 81 10 260 20 24 75 69 375 373 25 81 10 260 20 24 75 69 374 373 25 81 10 260 20 24 75 69 374
85
Appendix C.5
Table HP Separator and Low temperature separator valve full open operation
parameters of Brahmanbaria gas processing plant
Tower
feed -
Temp
S
601-1
-
Press
Tower
feed -
Press
S 601-
1 -
Temp
S 611-
1 -
Temp
S
611-1
-
Press
S 302 -
Molar
Flow
S 302
-
Press
S 401
-
Temp
S 401
-
Press
[F] [Psia] [Psia] [F] [F] [Psia] [MMscfd] [Psia] [F] [Psia]
153 18 219 10 225 20 24 75 10 166 153 18 219 10 225 20 24 75 10 166 152 18 218 10 225 20 24 75 10 166 152 18 218 10 225 20 24 75 10 166 152 18 218 10 225 20 24 75 10 166 152 18 218 10 225 20 24 75 10 166 152 18 218 10 225 20 24 75 10 166 152 18 217 10 225 20 24 75 10 166 151 18 217 10 225 20 24 75 10 166 151 18 217 10 225 20 24 75 10 166 151 18 217 10 225 20 23 75 10 166 151 18 217 10 225 20 24 75 10 166 151 18 217 10 225 20 24 75 10 166 151 18 216 10 225 20 24 75 10 166 151 18 216 10 225 20 24 75 10 166 150 18 216 10 225 20 24 75 10 166 150 18 216 10 225 20 24 75 10 166 150 18 216 10 225 20 23 75 10 166 150 18 216 10 225 20 26 75 10 166 150 18 216 10 225 20 25 75 10 166 152 18 218 10 225 20 24 75 10 166 152 18 217 10 225 20 24 75 10 166 151 18 217 10 225 20 24 75 10 166 151 18 217 10 225 20 24 75 10 166 151 18 217 10 225 20 23 75 10 166 151 18 217 10 225 20 24 75 10 166 151 18 217 10 225 20 24 75 10 166 151 18 216 10 225 20 24 75 10 166 151 18 216 10 225 20 24 75 10 166 150 18 216 10 225 20 24 75 10 166 150 18 216 10 225 20 24 75 10 166 150 18 216 10 225 20 23 75 10 166
86
Appendix D.1
Algorithm for Optimum Network Structure A Matlab code first initiated to determine optimum network structure in terms of
hidden layer neuron number. Preprocessed data are grouped into training,
validation and testing set. A loop is launched that calculate the percentage of error
for cumulative number of hidden layer neuron.
% Construct set of neural networks with number of hidden neurons varying between "hni" and "hnf" %at "del" step hni=1; hnf=20;
del=1;
for hn=hni:del:hnf;
net = network;
net.numInputs = 1;
net.inputs{1}.size = 2;
net.numLayers = 2;
net.layers{1}.size =hn;
net.layers{2}.size = 2;
net.inputConnect(1) = 1;
net.layerConnect(2,1) = 1;
net.outputConnect(2) = 1;
net.targetConnect(2) = 1;
net.layers{1}.transferFcn = 'tansig';
net.layers{2}.transferFcn = 'purelin';
net.biasConnect = [1; 1];
% The network performance measuring function and the training algorithms are set here
net.performFcn = 'mse';
net.trainFcn = 'trainlm';
87
Appendix D.2
Terms used in above code are described below:
hni =1, initial number of neuron within hidden layer is 1.
hnf =20, possible maximum number of neuron within hidden layer is 20.
del=1, step.
net =network, create custom network
net.numInputs, number of input source is 1, not the number of
elements in input vector.
net.inputs{1}.size , number of input elements.
net.numLayers, indicates number of layer in the network.
net.layers{1}.size, number of neuron within in first layer. Here 1 indicates as 1st layer.
net.layers{2}.size, number of neuron within second layer.
net.inputConnect (i, j), input weight connection going to the ith layer from the jth
input.
net.layerConnect, layer weight connection going to the ith layer from the jth
layer.
net.outputConnect (2), output connection to external world from two layers.
net.layers{1}.transferFcn, activation function used in 1st layer. Here
hyperbolic tangent sigmoid function (tansig) is used within 1st layer.
net.layers{2}.transferFcn, activation function used in 2nd layer. Here
linear transfer function (purelin) is used within 2nd layer.
net.biasConnect =[1;1], bias connect to 1st layer is 1 & 2nd layer is 1
net.performFcn, performance function used for training feedforward neural networks. Here mean square error (mse) function is used to measure network performance.
net.trainFcn, network training function. For levenberg-Marquardt method trainlm
function is used.