a fuzzy logic based diagnosis system for the on-line supervision of

13
A fuzzy logic based diagnosis system for the on-line supervision of an anaerobic digestor pilot-plant Antoine Genovesi, Je ´ro ˆme Harmand, Jean-Philippe Steyer * Laboratoire de Biotechnologie de l’Environement (LBE) – INRA, Avenue des Etangs – 11100, Narbonne, France Received 2 October 1998; accepted 17 February 1999 Abstract This paper deals with the development of a fuzzy logic based diagnosis system and its application as a fault detection and isolation (FDI) procedure in a wastewater treatment plant. Different fault detection methods are tested and their advantages and limitations are highlighted. An aggregate FDI strategy is implemented and tested. Results using the fuzzy residual generation module are presented and discussed based on experimental data from a 1 m 3 pilot-scale anaerobic digestion reactor for the treatment of raw industrial wine distillery vinasses. # 1999 Elsevier Science S.A. All rights reserved. Keywords: Waste-water treatment; Anaerobic processes; Fixed-bed bioreactors; Fuzzy logic; Fault detection and isolation; Diagnosis 1. Introduction 1.1. Problem statement Due to the increasing complexity and necessity of safety in industrial processes, efficient diagnosis systems are becoming more and more important. Indeed, even in normal operation conditions, several types of disturbances can be present and they can largely affect the operating conditions of a process. A clear need for advanced control and diag- nosis systems is thus expressed in order to keep the system performances as close as possible to the optimal conditions. This task can be divided into several sub-tasks (see Fig. 1). This is particularly true in biological processes (e.g., the biological wastewater treatment plants) where the state of the ‘‘living’’ part of the system has to be closely monitored. The fact that biological processes (or bioprocesses) depend on the activities of living beings leads to a high complexity and often very little is known about the phenomena that occur in a bioprocess. Unlike chemical reactions where information on stoichiometry, reaction rate, concentrations at the equilibrium, etc., are available, measurements obtained from on-line sensors do not yield a complete view of a bioprocess. Thus, the bioengineer concerned with control and diagnosis purposes must handle an ill-defined process. Anaerobic digestion is among the oldest biological waste- water treatment processes having first been studied more than a century ago [1]. It is a multi-step process in which organic matter is degraded into a gas mixture of methane and carbon dioxide. It thus reduces the chemical oxygen demand (COD) of the influent and produces valuable energy (methane). The biological scheme involves several multi- substrate multi-organism reactions that are performed both in series and in parallel (see, e.g., [2,3]). It has been experimentally demonstrated that anaerobic digestion is particularly adapted for concentrated wastes such as agricultural (e.g., plant residues, animal wastes) and food industry wastewaters. In addition, this process is able to operate under severe conditions: high-strength effluents and short hydraulic retention times. Anaerobic digestion is also often used as a sludge treatment for the stabilization of primary and secondary sludges. However, anaerobic digestion is intrinsically a very unstable process and variations of the input variables (hydraulic flow rate, influent organic load) may easily lead the process to a wash-out of the tank. This phenomenon takes place under the form of volatile fatty acids accumula- tion in the reactor and it is essential to implement carefully designed control strategies able to keep the process running in stable conditions. The control of anaerobic digestion is thus very complex and serious breakdowns, owing mainly to Biochemical Engineering Journal 3 (1999) 171–183 *Corresponding author. Tel.: +33-468-425-159; fax: +33-468-425-160; e-mail: [email protected] 1369-703X/99/$ – see front matter # 1999 Elsevier Science S.A. All rights reserved. PII:S1369-703X(99)00015-7

Upload: ngongoc

Post on 10-Feb-2017

217 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: A fuzzy logic based diagnosis system for the on-line supervision of

A fuzzy logic based diagnosis system for the on-line supervisionof an anaerobic digestor pilot-plant

Antoine Genovesi, JeÂroÃme Harmand, Jean-Philippe Steyer*

Laboratoire de Biotechnologie de l'Environement (LBE) ± INRA, Avenue des Etangs ± 11100, Narbonne, France

Received 2 October 1998; accepted 17 February 1999

Abstract

This paper deals with the development of a fuzzy logic based diagnosis system and its application as a fault detection and isolation (FDI)

procedure in a wastewater treatment plant. Different fault detection methods are tested and their advantages and limitations are highlighted.

An aggregate FDI strategy is implemented and tested. Results using the fuzzy residual generation module are presented and discussed based

on experimental data from a 1 m3 pilot-scale anaerobic digestion reactor for the treatment of raw industrial wine distillery vinasses. # 1999

Elsevier Science S.A. All rights reserved.

Keywords: Waste-water treatment; Anaerobic processes; Fixed-bed bioreactors; Fuzzy logic; Fault detection and isolation; Diagnosis

1. Introduction

1.1. Problem statement

Due to the increasing complexity and necessity of safety

in industrial processes, ef®cient diagnosis systems are

becoming more and more important. Indeed, even in normal

operation conditions, several types of disturbances can be

present and they can largely affect the operating conditions

of a process. A clear need for advanced control and diag-

nosis systems is thus expressed in order to keep the system

performances as close as possible to the optimal conditions.

This task can be divided into several sub-tasks (see Fig. 1).

This is particularly true in biological processes (e.g., the

biological wastewater treatment plants) where the state of

the `̀ living'' part of the system has to be closely monitored.

The fact that biological processes (or bioprocesses) depend

on the activities of living beings leads to a high complexity

and often very little is known about the phenomena that

occur in a bioprocess. Unlike chemical reactions where

information on stoichiometry, reaction rate, concentrations

at the equilibrium, etc., are available, measurements

obtained from on-line sensors do not yield a complete view

of a bioprocess. Thus, the bioengineer concerned with

control and diagnosis purposes must handle an ill-de®ned

process.

Anaerobic digestion is among the oldest biological waste-

water treatment processes having ®rst been studied more

than a century ago [1]. It is a multi-step process in which

organic matter is degraded into a gas mixture of methane and

carbon dioxide. It thus reduces the chemical oxygen demand

(COD) of the in¯uent and produces valuable energy

(methane). The biological scheme involves several multi-

substrate multi-organism reactions that are performed both

in series and in parallel (see, e.g., [2,3]).

It has been experimentally demonstrated that anaerobic

digestion is particularly adapted for concentrated wastes

such as agricultural (e.g., plant residues, animal wastes) and

food industry wastewaters. In addition, this process is able to

operate under severe conditions: high-strength ef¯uents and

short hydraulic retention times. Anaerobic digestion is also

often used as a sludge treatment for the stabilization of

primary and secondary sludges.

However, anaerobic digestion is intrinsically a very

unstable process and variations of the input variables

(hydraulic ¯ow rate, in¯uent organic load) may easily lead

the process to a wash-out of the tank. This phenomenon

takes place under the form of volatile fatty acids accumula-

tion in the reactor and it is essential to implement carefully

designed control strategies able to keep the process running

in stable conditions. The control of anaerobic digestion is

thus very complex and serious breakdowns, owing mainly to

Biochemical Engineering Journal 3 (1999) 171±183

*Corresponding author. Tel.: +33-468-425-159; fax: +33-468-425-160;

e-mail: [email protected]

1369-703X/99/$ ± see front matter # 1999 Elsevier Science S.A. All rights reserved.

PII: S 1 3 6 9 - 7 0 3 X ( 9 9 ) 0 0 0 1 5 - 7

Page 2: A fuzzy logic based diagnosis system for the on-line supervision of

the organic overload of various origins, have been noticed at

the industrial scale [4]. For these reasons, a fuzzy logic

based diagnosis system, including advanced fault detection

and isolation capabilities, has been developed. This system

is presented in the following.

1.2. Fault detection and isolation (FDI)

As de®ned by Isermann and Balle [5], a fault is an

unexpected change in a system such as a malfunction or

an unexpected variation in operational conditions that tends

to degrade the overall system performance. Such malfunc-

tions may occur in the sensors (instruments), in the actua-

tors, in the components of the process (e.g., inhibition of the

biomass in a biological reactor) or in the control system if

the process is running in closed loop. Corrupted data

transmissions between the sensors/actuators and the data

processor faults in the data processing procedure and errors

in the A/D±D/A converters are some common modes of

failure.

Fault modes are mainly of two kinds: abrupt (i.e., step-

like changes) and incipient (e.g., bias or drift). The ®rst

mode plays an important role in safety-relevant systems

where `̀ hard'' failures have to be detected early enough so

that catastrophic consequences can be avoided. The second

one is small and not easy to detect. It is related to main-

tenance problems where early detection is required [6]. In

this case, the correct estimation of the bias is also useful

when working with closed-loop systems.

Automatic fault detection and isolation (FDI) uses ana-

lytical or heuristic knowledge in order to detect as early as

possible deviations from the normal operation of a process.

To this end, FDI can be divided into several steps (see

Fig. 2); the ®rst one (i.e., the residual generation) consisting

in the use of an algorithm that generates a vector to carry

information about a particular fault.

The residual generation algorithm should work even if (i)

the time evolution of the fault is unknown, (ii) the math-

ematical model of the nominal system is uncertain (with

unknown tolerances), (iii) there are system and measure-

ment noises (with unknown characteristics) and (iv) con-

straints exist with respect to the time residual generation.

But the existence of such an algorithm depends on (i) the

knowledge of the `̀ normal'' or nominal behavior, (ii) a good

de®nition of the faulty behavior, (iii) the existence of

analytical redundancy relations and (iv) a satisfactory relia-

bility of the redundant information (i.e., invariance or at

least robustness with respect to the unknown inputs).

Fault detection can be achieved using a model-based

approach. Research in this ®eld started in the early 1970s

with some results on observer-based fault detection in linear

systems [7±9]. In the late 1970s, a ®rst book appeared on

model-based methods for the fault detection and diagnosis

in chemical processes [10]. At the same time, instrument

failure detection based on analytical redundancy of multiple

observers was shown [11] and the use of parameter-estima-

tion methods was demonstrated for technical systems [12±

15]. The development of process fault-detection methods

based on modeling, parameter and state estimation was then

summarized by some survey papers [6,16,17] and by some

books [18±20].

There has been rapidly growing interest for using tech-

niques such as arti®cial neural networks or fuzzy logic for

fault detection purposes [5,21]. A few demonstrations of the

advantages of fuzzy-based FDI systems can be found in the

literature (see, e.g., [22±25]). Other recent studies use this

technique for sensor validation [26] or combine it with

arti®cial neural networks for process diagnosis [27].

Over the last two decades, fault detection has thus gained

in methodology and robustness using powerful techniques

from different ®elds: control theory (mathematical model-

ing, state estimation and parameter identi®cation), arti®cial

intelligence (expert systems, arti®cial neural networks,

fuzzy logic, qualitative reasoning) and statistics (likelihood

ratio test, decision theory, principal component analysis).

Fig. 1. The different sub-tasks to be performed by a diagnosis system.

Fig. 2. The several steps involved in FDI systems.

172 A. Genovesi et al. / Biochemical Engineering Journal 3 (1999) 171±183

Page 3: A fuzzy logic based diagnosis system for the on-line supervision of

However, fault detection is not the only task when

developing an advanced diagnosis system. In a more general

way, advanced diagnosis can indeed be de®ned as the ability

of a system to support the process operator in choosing the

right information that provides the best overview of the

process over a span of time during which past, present and

future process outputs, or some combinations of these are

considered.

This point calls for more explanation since advanced

diagnosis systems are not only devoted to collecting and

displaying process data in a useful way but additional tasks

are to be achieved. That is to analyze causes of alarms,

provide the primary cause without displaying other alarms

and assist the process operator by advising him on recovery

actions and by indicating the urgency of these actions.

A good FDI system for a bioprocess also requires inclu-

sion of speci®c biological knowledge. Arti®cial intelligence

methods such as knowledge-based systems or fuzzy logic

bring new insights to biological process control and intro-

duce a `̀ biological dimension'' [28±47]. All these studies

were designed to cope with uncertain and incomplete

knowledge and they allowed the coupling of quantitative

information with the qualitative or symbolic expressions so

as to reproduce the actions and decisions of an experienced

process operator. Representation of time-dependent knowl-

edge and model-based reasoning were also characteristics

introduced in these diagnosis systems.

For anaerobic digestion, an expert system to control the

dilution rate by managing four different control laws has

been presented in the literature [48]. This expert system

reproduced the decisions of a skilled human operator and

prevented organic overloads and digester imbalance. The

present paper details an approach capable of dealing with

organic overloads as well as with other disturbances like

technical problems (i.e., pipe clogging, foam forming,

sensor bias, etc.) and bad settings of local control loops.

It is organized as follows. First, the anaerobic digestion pilot

plant and its instrumentation are described. Then, the meth-

odology used to perform residual generation and the fault

detection and isolation strategy are explained. Finally, some

examples of problems are presented and discussed along

with the results of the on-line application of the FDI module.

2. Material and methods

The experiments were performed using an anaerobic up-

¯ow ®xed bed reactor (Fig. 3) at the `̀ Laboratoire de

Biotechnologie de l'Environnement'' of INRA in Narbonne,

France.

The feeding system of the reactor is made of three tanks

(27 m3 each) which are connected by a piping system of

0.1 m diameter and 60 m long. The experiments were

carried out with raw industrial wine distillery ef¯uents

Fig. 3. Schematic representation of the pilot plant.

A. Genovesi et al. / Biochemical Engineering Journal 3 (1999) 171±183 173

Page 4: A fuzzy logic based diagnosis system for the on-line supervision of

obtained from local wineries in the area of Narbonne. This

substrate, neither sterile nor homogeneous, has changing

characteristics in those tanks and in the pipes as depicted in

Table 1.

The reactor is a circular column of 3.5 m height, 0.6 m

diameter and 0.982 m3 total volume. The media used (i.e.,

Cloisonyl: 180 m2/m3 of speci®c surface) ®lls 0.0337 m3

leaving 0.948 m3 of effective volume. The support develops

135 m2 of surface.

The 0.2 m3 dilution tank has a ¯oat in order to control the

volume and a degassing system that removes the gas con-

tained in the liquid phase of the in¯uent. Connected to this

tank is a remotely controllable peristaltic pump which

ensures the desired in¯uent ¯ow rate fed into the reactor.

Fresh substrate is mixed with the recycled one just before

entering the heat exchanger (which regulates the tempera-

ture to 358C).

The pH measuring and regulating system (HANNA

instruments) is made of a pH sensor, a PID hardware

controller, an NaOH storage tank and a dosing pump.

The addition of the soda can be performed either in the

dilution system or in the recirculation loop just before the

heat exchanger. The heated liquid is then introduced at the

bottom of the reactor where it is homogenized by the mixing

pump.

In the liquid output of the reactor, there is another

degassing system and a gas evacuating system. The liquid

from the top of the reactor is collected by over¯ow in a

receiving vessel. Some of this liquid is recycled at 0.05 m3/h

and the rest is sent to the sewer.

The input ¯ow rate is measured and processed by an

electromagnetic sensor (KRHONE) with a normalized ana-

log output (4±20 mA). The gas analyzing loop (see Fig. 4) is

composed by a dryer which eliminates the humidity by

cooling the gas.

The ULTRAMAT 22P sensor (SIEMENS) measures the

CO2 and CH4 percentage composition of the analyzed gas.

This sensor works on the principle of the nondispersive

absorption of infrared light, that is, a one beam method with

a two-layer optopneumatic detector.

The gas ¯owmeter is located at the output of this loop. It

uses an electromagnetic ¯oater to measure the produced gas

continuously. Hydrogen concentration in the gas phase is

measured by an AMS 6400 H2 analyzer (Pekly Hermann-

Moritz). This sensor is very sensitive to H2S contained in the

gas, so a `̀ pura®l'' trap which changes its color on becoming

saturated was added. A peristaltic pump was also installed to

ensure a ®xed gas ¯ow rate throughout the cell (electro-

chemical cell). It uses capillary diffusion, which has the

advantage of a low temperature coef®cient, a direct indica-

tion of the concentration (in ppm) and consequently, the

in¯uence of pressure over the measure is low.

The following variables are thus measured on-line:

� the input liquid flow rate,

� the recycled liquid temperature,

� the heater temperature (from the temperature regulation

loop),

� the recycled liquid pH,

� the output gas flow rate,

� the biogas composition (CO2/CH4 percentage and H2

concentration).

These sensors are connected to an input/output device

that allows the acquisition, treatment and storage of data on

a PC using the Modular SPC1 software developed by the

French Company SERI Environnement. This software per-

forms advanced control law calculations as well as process

supervision.

3. The fault detection and isolation (FDI) strategy

As already presented, different methods have been used

in the literature to perform the residual generation. Some of

them, the process-model-based methods, use mathematical

observers (states and outputs), parity equations, identi®ca-

tion and parameter estimations. The models used can be

either ®xed or adaptive, parametric or nonparametric, deter-

ministic or stochastic. These methods have different diag-

nosis capabilities depending on the kind of fault to be treated

(i.e., additive or multiplicative, pulse or step like faults,

etc.). Other methods, the signal-model-based methods, only

Table 1

Characteristics of the industrial wine distillery wastewater

Component Range

Volatile fatty acids (g/l) [5.00±6.00]

Total organic carbon (g/l) [2.50±6.00]

Phenol (mg/l) [90.0±275]

Total COD (g/l) [9.00±17.4]

Soluble COD (g/l) [7.60±16.0]

Total suspended solids (g/l) [2.40±5.00]

Volatile suspended solids (g/l) [1.20±2.70]

Alkalinity (meq/l) [30.8±62.4]

pH [5.00±5.40]

Fig. 4. Schematic description of the output gas flow pattern (NP�normal

position).

174 A. Genovesi et al. / Biochemical Engineering Journal 3 (1999) 171±183

Page 5: A fuzzy logic based diagnosis system for the on-line supervision of

use the available on-line measurements. They can detect

changes from normal behavior by using some characteris-

tics of the signal (e.g., the mean values and variance). They

also employ some techniques from signal processing (e.g.,

segmentation methods).

Since many methods are available to perform FDI and

since their usefulness depends on the variables monitored,

an open structure was developed that allows the aggregation

of the residuals and the use of different methods to

perform fault ®ltering and fault isolation. Three kind of

fault signals, each of them being related to the available

on-line measurements, were de®ned: sensor faults (SFs),

sub-process faults (SPFs) and process faults (PFs).

Each fuzzy fault signal was generated using the fuzzy

logic theory and a Mamdani structure was used for all

these signals. The membership functions for the inputs as

well as the associated rule-bases were de®ned from the

knowledge of the human operators while the output mem-

bership functions were de®ned using two cases: `̀ problem''

and `̀ no problem''. The maximum and the centroid

defuzzi®cation algorithms were used to obtain a real-valued

output. This choice was made in order to mimic the

reactions of a skilled human operator when facing a faulty

situation.

In the following, the algorithms associated with these

three different fault signals are explained.

3.1. Sensor fault (SF)

This signal represents the degree of con®dence in the

information directly given by the sensors. For example, if a

measurement is out of the range of the sensor (e.g., bad

calibration of the sensor) or extremely constant (e.g., the

sensor can simply be disconnected from the input/output

device), the SF signal is set to a value greater than 95%. On

the other hand, if the data are too noisy, their fault-asso-

ciated values are close to zero depending on the magnitude

of the noise.

Fig. 5 shows the general structure of the fuzzy residual

generation of SF.

For the initialization of the membership functions of the

raw data, the information given in Table 2 is used. The

values A1 and B1 are the corresponding minima and maxima,

while the variables A2 and B2 are 10% different from A1 and

B1, respectively. This 10% threshold has been chosen after

analyzing the different signals available on the process and

it was shown that it allows early enough fault detection

without being too speci®c of the signals dynamics.

In the case of the initialization of the membership func-

tions for the standard deviation, the minimum normal

standard deviation (B3) is set on the basis of direct data

observation. Values A3 and C3 are 0.5 and 1.5 times the B3

value. For the membership function `̀ zero'', a ®xed very low

range [0±0.001] is used. For the normal case, even though

one would normally choose a symmetric form such as an

equilateral or isosceles triangle for the membership func-

tions, it was decided to use a right angle triangle-like shape

in order to put more weight on the low values of the standard

deviation. These values are indeed considered to be normal

as long as they differ from zero while not being too high.

Since two different cases were de®ned to represent the

problems on the sensors, two membership functions were

de®ned too. The `̀ Pb'' function (values higher than 95%)

Fig. 5. Schematic representation of the fuzzy generation of SF.

Table 2

Range validity for the on-line measurements

On-line measurement Range validity

Recycled liquid temperature (8C) 25±55

Output (heater) temperature (8C) 30±70

pH (UpH) 5±8.5

Recycled liquid flow rate (l/h) 0±100

Liquid flow rate (l/h) 0±50

Gas flow rate 0±400

CO2/CH4 percentage 0±100

H2 (ppm) 0±1000

A. Genovesi et al. / Biochemical Engineering Journal 3 (1999) 171±183 175

Page 6: A fuzzy logic based diagnosis system for the on-line supervision of

means that the sensor is out-of-order while the `̀ dyn_Pb''

function means that there exists a problem either on the

sensor or on the process associated with the measured

variable; the third is the `̀ no problem'' case. The rule base

is composed as shown in Table 3.

3.2. Sub-process fault (SPF)

In the same way than the SF signals are related to the

measurements directly received by the sensors, the SPF

signals are dedicated to problems with local control loops

(i.e., association between a sensor and an actuator).

As already explained (see Section 2) four local loops

were de®ned within the anaerobic digestion process: (i) the

temperature, (ii) the pH, (iii) the input liquid in¯ow rate and

(iv) the recycled liquid ¯ow rate. In all these cases, there is a

feedback (i.e., from an on-line measurement) of the control

action. For example, the desired input liquid ¯ow rate is

controlled (the control action) and measured (the liquid

input ¯owmeter), so a direct comparison can be made

between the desired values and the measured values allow-

ing very simple and robust detection schemes.

It is here to be pointed out that measurements of variables

such as temperature, pH and ¯ow rates, which are regulated

to local set-points, are not always suf®cient to predict

failures. For fault detection purposes, it is indeed very often

more useful to check the actuator signal rather than the

regulated variable. As an example, it can be demonstrated

that by using the frequency of sodium hydroxide addition

instead of the pH measurement alone, malfunctioning can

be anticipated and minimized, thus improving the monitor-

ing of the process (see Fig. 6). In the same way, if a fault

occurs in a liquid ¯ow rate that is regulated through a valve,

it is more ef®cient to analyze the valve opening rather than

only liquid ¯ow rate measurements (see Fig. 7). In this case,

pipe clogging can be indeed detected as early as the valve

opening increases (in a nonfaulty situation, it should be

constant) instead of waiting for the decrease of the liquid

¯ow rate (i.e., when the valve opening reaches 100%).

The SPF algorithms use the information available from

the sensors (mainly the raw data, standard deviation, deri-

vatives and the mean values) as well as the information

given by the SFs in order to evaluate the presence of a fault

in one of the components of the loop (i.e., the actuator, the

controller or the sub-process itself).

In order to generate the residuals related to these signals,

several methods are used. For example, in the case of the

temperature loop (see Fig. 8), a fuzzy inference system is

applied that uses the standard deviation of both the recycled

liquid temperature (i.e., the sensor) and the heater tempera-

ture (i.e., the actuator).

The set values of the constants Ai, Bi and Ci were obtained

by direct observation of the evolutions of the target vari-

ables. In the case of A1 and A2, values of 0.2 and 0.4,

respectively, were used. B1 and B2 were set to 1. For the

outputs, values of 0.05, 0.2, 0.6 and 10 were used for the A3,

B3, C3 and D3, respectively. The inference rules for the

temperature sub-system are shown in Table 4.

Another method used to generate SPF residuals which is

complementary to the previous one was the direct compar-

ison of the raw data with the calculated mean (based on a

Table 3

Inference rules for the SF residual generation

Standard deviation

Zero Normal High

Raw data Low Pb Pb Pb

Normal Pb No_Pb Dyn_Pb

High Pb Pb Pb

Fig. 6. Improvements in increasing the `̀ relevance'' of a measurement by analyzing the actuator evolution.

Table 4

Inference rules for the SPF residual generation on the temperature sub-

system

Recycled temperature STD

Normal High

Heater temperature STD Normal No_Pb Pb

High Pb Pb

176 A. Genovesi et al. / Biochemical Engineering Journal 3 (1999) 171±183

Page 7: A fuzzy logic based diagnosis system for the on-line supervision of

moving window). The residual is expressed in percentage

error:

r � mÿ Di

m� 100% with m �

Pnj�k Dj

nÿ k;

where m is the mean value, nÿk the window width, i the

current time, Di the most recent data read, and r is the

residual.

This method allows us to detect slow varying phenomena,

the time window width being directly related to the variable

dynamics, whereas the previous one (i.e., the fuzzy infer-

ence system managing the standard deviations ± see Fig. 8)

is more dedicated to faults occurring without any change of

the mean value of the variable.

3.3. Process fault (PF)

This is the most important fault. The PF signals are indeed

®rst an indication of the importance of the previous faults

(SF and SPF) on the overall anaerobic digestion process: if

an SF or SPF signal is different from zero for too long, the

related PF signal increases (i.e., the longer the duration of

the problem, the greater the PF signal). This indicates that a

problem had occurred which avoids the process to be

operated in normal conditions.

On the other hand, the PF signals also quantify how

largely the biological process is affected by a change in the

organic loading rate. The problems detected in this case

concern either a failure in the input ¯ow rate or in the

organic load coming into the reactor. In other words, when

the PF related residual signals diverge from zero, the

problem that is detected is intense enough to affect the

overall biological process. The algorithm used in this last

case is the result of the following design procedure.

A ®rst order model (only valid around a known operating

point) was ®rst identi®ed, i.e.,

Qgas�s� � K

�s� 1Qin�s� with K � 14:14 and � � 2:05:

The comparison between the simulation of the model and

the output gas ¯ow rate measurements is shown in Fig. 9. As

it can be seen, this simple model represents very well the

relationship between the input (i.e., the liquid ¯ow rate) and

the output (i.e., the gas ¯ow rate) of the process. The

assumption is then made that every change in the liquid

input (i.e., normal or faulty COD changes) has an in¯uence

on the output gas ¯ow rate.

The process being operated in closed loop, it is expected

that the theoretical value of the control (i.e., the input ¯ow

rate) to be applied at time t is known and actual on-line

values are available as speci®ed before.

Fig. 7. Evolution of the liquid flow rate and of the valve opening during a pipe clogging.

Fig. 8. Schematic representation of the fuzzy generation of SPF concerning the temperature loop.

A. Genovesi et al. / Biochemical Engineering Journal 3 (1999) 171±183 177

Page 8: A fuzzy logic based diagnosis system for the on-line supervision of

Let us now call r1(t) the difference between the theore-

tical value of the input liquid ¯ow rate (i.e., computed within

a control algorithm framework) and its on-line measure-

ment. Similarly, let r2(t) be the difference between the

theoretical value of the output gas ¯ow rate (i.e., computed

using the simple previously presented model with the

measured input liquid ¯ow rate as the input of the model)

and its on-line measurement.

When computing the residuals, the signals were ®ltered

using a simple Kalman ®lter based upon the ®rst order

model previously presented. Residuals were then compared

to some small thresholds, �1 and �2, and introduced, at each

sample time, into the algorithm described in Fig. 10.

The reasoning scheme described in Fig. 10 is used to test

the consistency of the different variables in order to detect,

and with further assumptions, to isolate the problem:

� First, the consistency of residual r1(t) is tested ± it is

indeed related to the input of the process and it overrides

the output signals, i.e., r2(t). If the measured control input

is different from the theoretical one, then r1(t) points to a

problem in the actuator (e.g., in the input liquid flow-

meter in the connection between the pump and the data

processor, failure of the pump, etc.). At this step, the

operator has to be informed of the possible causes

together with their degree of urgency. Among all statis-

tical knowledge available on the device, the magnitude of

the residual is the main indication about the urgency of

the problem: the higher the value of the residual, the more

important the problem is.

� If r1(t) is nearly zero, the residual r2(t) has to be eval-

uated. If r2(t)�0, there is no problem. However, if it

differs significantly from zero, then a problem has

occurred. Obviously, this problem can have several

causes. It can simply come from a change in the operat-

ing point indicating that either the input flow rate (which

is easy to check since the flow rate measurements are

available) or the input substrate COD concentration has

Fig. 9. Measurements and first order model prediction of the output gas flow rate Qgas using the input liquid flow rate Qliq.

Fig. 10. Simple model based FDI algorithm for a process fault.

178 A. Genovesi et al. / Biochemical Engineering Journal 3 (1999) 171±183

Page 9: A fuzzy logic based diagnosis system for the on-line supervision of

changed. Another cause can be a failure in the output gas

flow rate measurement. However, at this step and without

any further information, it is not possible to isolate the

fault using only this on-line information.

4. Results and discussion

To demonstrate the ef®cacy of the presented fuzzy

logic based FDI approach, it was ®rst decided to compare

the results obtained with those of a simple threshold

comparison. To this end, the fuzzy membership functions

of the SF detection sub-system were changed to square

shapes (Fig. 11) and they were applied to results obtained

when foaming occurs in the reactor. In this case, the

biogas ¯ow rate is largely disturbed since foam enters

the output pipes, and if this problem is not solved rapidly,

the foam can reach the ¯owmeter, hence damaging the

sensor.

In Fig. 12 (left), both detectors are optimized for this data

set and they show similar performances, the fuzzy system

detecting the problem a little bit earlier though. On the other

hand, in Fig. 12 (right), the same detectors with the same

parameters were used to detect foaming in other operating

conditions. Here, the threshold comparison clearly shows a

delay in the detection, while the fuzzy logic based FDI

approach gives a rapid and gradual increasing signal of the

fault.

Fig. 11. Membership functions used in (a) a threshold comparison and (b) a fuzzy logic based FDI approach.

Fig. 12. Comparison between fault (foam forming) detection (upper) using a threshold comparison (middle) and a fuzzy logic based approach (lower).

A. Genovesi et al. / Biochemical Engineering Journal 3 (1999) 171±183 179

Page 10: A fuzzy logic based diagnosis system for the on-line supervision of

These examples clearly demonstrate that the fuzzy logic

based FDI approach is more robust and more effective than a

threshold comparison algorithm. It can detect the faults

earlier, and more importantly, it provides the human opera-

tor with the degree of urgency of the problem. While the

threshold comparison approach generates a fault detection

signal of either 0% or 100%, the fuzzy FDI approach

provides a more accurate index between these limits: the

larger the signal, the more signi®cant is the fault.

In another experiment, the sensor measuring the CO2

percentage in the gas phase was disconnected on purpose

from the reactor and the sensor signal dropped down to 0

(see Fig. 13 from 1112 to 1128 h). As it can be seen, the SF

signal generated by the fuzzy fault detection (see Sec-

tion 3.1) is automatically and rapidly set to its maximum

showing the presence of the fault on this sensor. Further-

more, the fault signal is sensitive to small changes of this

measured variable, as for example at 1155 h when the

organic loading rate was manually decreased (data not

shown in Fig. 13). This information was however main-

tained since it is useful in the determination of the `̀ normal''

state of the process by taking into account these discrete

events (e.g., human intervention).

Another example of the SF detection ability is presented

in Fig. 14. In this case, the peristaltic pump bringing the

input wastewater into the reactor broke down because it was

too old; the measured input liquid ¯ow rate was thus

unstable and very noisy (see Fig. 14 starting at about

368 h). Again, the residual generator connected to this

variable rapidly indicated the presence of the problem by

increasing the SF signal.

Note: from 350 to 357 h, the SF signal is also high but in

this case, it was due to another problem connected to a

manual stop of the input liquid ¯ow rate.

When the overall operating conditions change, the local

regulation loops (i.e., pH, temperature, recycling ¯ow rate)

may not work properly. This is the case in the following

experiment where a decrease in the temperature of the liquid

in¯uent saturated the control signal of the heater to its

maximum capacity (see the upper curve in Fig. 15). Also,

the controlled temperature (i.e., the recycled liquid tem-

perature) had important oscillating deviations from the

desired optimum set point (i.e., at 358C). In fact, the

parameters of the local hardware PID temperature controller

were not correctly set and the controller could not com-

pensate the disturbance. This is clearly illustrated by the

SPF signal shown in the lowest part of Fig. 15.

Finally, an example concerning a PF detection and iso-

lation is described in Fig. 16. This experiment represents a

problem that has to be avoided in real plant operation: an

increase in the input liquid ¯ow rate (the input COD

concentration was maintained at 10 g/l) leading to an

organic overload of the process. This is clearly visible in

the upper left ®gure starting at 1610 h. The PF signal

immediately shoots up from zero. During this period, the

measured input ¯ow rate was consistent with the theoreti-

cally required rate (i.e., measurements were equal to the

control signal sent to the pump). Then, it can be concluded

that the detected problem is not an actuator failure. More-

over, as explained in the previous section, this problem

cannot be isolated using only the available information.

However, the fact that the input ¯ow rate is signi®cantly

different from the nominal operating point (i.e., the operat-

ing point de®ned when t<1610 h) allows the diagnosis

system to advise the operator that there is an organic over-

load and that the process has to be quickly secured. This factFig. 13. Disconnection of the sensor (upper) and SF residual generated

(lower).

Fig. 14. Problem with the input liquid flow rate (upper) and SF residual

generated (lower): (1) manual stop of the input liquid flow rate; (2) normal

operating conditions; (3) breaking down of the input liquid peristaltic

pump.

180 A. Genovesi et al. / Biochemical Engineering Journal 3 (1999) 171±183

Page 11: A fuzzy logic based diagnosis system for the on-line supervision of

is con®rmed by the response of the output gas ¯ow rate that

is observed in the upper right ®gure. Indeed, it can be

noticed that the system begins to show response to the load

increase before its trajectory decreases again, while the

input ¯ow rate remains constant at 40 l/h (i.e., 20 l/h over

the operating point). This evolution being very different

from the results provided by the mathematical model, the

decision given to the human operator is then to secure the

process by decreasing the input liquid ¯ow rate, which may

be observed in the upper left ®gure at t�1636 h, when the

¯ow rate is stopped.

Between t�1636 h and t�1645 h, the signal for PF is

greater than zero because the system is not operating any-

more around its operating point (since organic volatile fatty

acids have accumulated in the reactor). This differs from the

previous situation in that the diagnosis system senses an

input ¯ow rate of zero. As a consequence, the system is

diagnosed to be under control and no alarm is sent to the

operator even if the PF signal is non-zero.

This defective diagnosis scheme can be improved if

further information is available. Let us consider that the

pH in the reactor is available on-line (see Fig. 16, lower left

part). A decrease in the pH indicates an accumulation of

volatile fatty acids. In this case, the load has to be decreased

immediately in order to avoid an overload. Examining the

pH measurement allows the system to con®rm the over-

loading. Obviously, the diagnosis has been improved since a

failure in the output gas ¯ow rate measurement has been

excluded.

There is a ®nal important point. After the step changes in

the input ¯ow rate described in Fig. 16, normal operation

conditions are restored at time t�1643 h. However, it can be

observed that the biological process has been disturbed by

this experiment (because of volatile fatty acids accumula-

tion) since a constant bias is observed in the lower right

®gure from t�1646 h to the end. It is very important to take

this into account since the a priori de®nition of `̀ normal

Fig. 15. Inappropriate setting of the temperature regulator and SPF

residual generated.

Fig. 16. Organic overload and PF residual generated (lower right).

A. Genovesi et al. / Biochemical Engineering Journal 3 (1999) 171±183 181

Page 12: A fuzzy logic based diagnosis system for the on-line supervision of

operation conditions'' can be affected by such dynamic

situations. This point is currently under study.

5. Conclusions

An aggregative fault detection and isolation (FDI) strat-

egy was developed and tested. The use of different methods

for residual generation allows the early detection of pro-

blems. The results obtained demonstrate the ability of the

fuzzy logic FDI scheme to handle human knowledge in a

systematic way and represents a potentially powerful tool

for FDI in industrial environments.

Different kinds of faults were detected by the proposed

strategy: sensor disconnection, actuator breakdown, local

loop problems and biological process problems. Results

were compared to the wide-applied threshold methods

and better performances were obtained. In addition, the

fuzzy logic based method was able to perform satisfactory

fault detection even in varying operating conditions. How-

ever, although the detected fault signals are useful to

identify the primary cause of a fault, further work has to

be done for its isolation. In particular, the generated signals

have to be integrated and ®ltered in a more general scheme

in order to be useful in an industrial context. This work is

currently under study.

Acknowledgements

The study has been carried out with the ®nancial support

from the Commission of the European Communities, Agri-

culture and Fisheries (FAIR) speci®c RTD program, for the

project CT96-1198, `̀ Advanced Monitoring and Control of

the operation of wastewater treatment for the wood industry

in order to improve the process ef®ciency (AMOCO)''. It

does not necessarily re¯ect its views and in no way antici-

pates the Commission's future policy in this area.

References

[1] P.L. Mc Carthy, in: D.E. Hughes et al. (Eds.), One Hundred Years of

Anaerobic Treatment, Anaerobic Digestion 1981, Elsevier Biome-

dical Press, Amsterdam, 1981, pp. 3±22.

[2] M. Denac, A. Miguel, I.J. Dunn, Modeling dynamic experiments on

the anaerobic degradation of molasses wastewater, Biotech. Bioeng.

31 (1988) 1±10.

[3] E. Heinzle, I.J. Dunn, G.B. Ryhiner, Modeling and control of

anaerobic wastewater treatment, Adv. Biochem. Eng. Biotech. 48

(1993) 79±114.

[4] C. Feuillette, Inventaire des meÂthaniseurs industriels en France (in

french), JourneÂes Industrielles sur la Digestion AnaeÂrobie, JIDA 96,

Narbonne, France, 17±19 June 1996.

[5] R. Isermann, P. BalleÂ, Trends in the application of model-based fault

detection and diagnosis of technical processes, Contr. Eng. Pract.

5(5) (1997) 709±719.

[6] P.M. Frank, Fault diagnosis in dynamic systems using analytical and

knowledge-based redundancy ± A survey and some new results,

Automatica 26(3) (1990) 459±474.

[7] R.V. Beard, Failure accomodation in linear systems through self-

reorganization, Rept. MVT-71-1, Man Vehicle Lab., Cambridge,

MA, 1971.

[8] H.L. Jones, Failure detection in linear systems, Ph.D. Thesis,

Department of Aeronautics, MIT, Cambridge, MA, 1973.

[9] A.S. Willsky, A survey of design methods for failure detection

systems, Automatica 12(6) (1976) 601±611.

[10] D.M. Himmelblau, Fault Detection and Diagnosis in Chemical and

Petrochemical Processes, Elsevier, Amsterdam, 1978.

[11] R.N. Clark, Instrument fault detection, IEEE Trans. on Aerosp.

Elect. Syst. 14 (1978) 456±465.

[12] H. Hohmann, Automatic monitoring and failure diagnosis for

machine tools (in German), Dissertation, Darmstadt, Germany,

1977.

[13] C. Bakiotis, J. Raymond, A. Rault, Parameter identification and

discriminant analysis for jet engine mechanical state diagnosis, IEEE

Conference on Decision and Control, Fort Lauderdale, FL, 1979.

[14] G. Geiger, Monitoring of an electrical driven pump using

continuous-time parameter estimation models, Sixth IFAC Sympo-

sium on Identification and Parameter Estimation, Washington,

Pergamon Press, Oxford, 1982.

[15] D. Filbert, K. Metzger, Quality test of systems by parameter

estimation, Ninth IMEKO Congress, Berlin, Germany, 1982.

[16] R. Isermann, Process fault detection based on modelling and

estimation methods ± A survey, Automatica 20(4) (1984) 387±404.

[17] J. Gertler, Survey of model-based failure detection and isolation in

complex plants, IEEE Cont. Syst. Magazine 8(6) (1988) 3±11.

[18] L.F. Pau (Eds.), Failure Diagnosis and Performance Monitoring,

Marcel Dekker, New York, 1981.

[19] R.J. Patton, P.M. Frank, J. Clark, Fault Diagnosis in Dynamic

Systems ± Theory and Application, Prentice-Hall, Hertfordshire,

1989.

[20] R. Isermann, UÈ berwachung und Fehler-Diagnose (Supervision and

Fault Diagnosis), VDI-Verlag, DuÈsseldorf, Germany, 1994.

[21] P.M. Frank, Application of fuzzy logic to process supervision and

fault diagnosis, IFAC Symposium on Fault Detection Supervision

and Safety for Technical Processes, SAFEPROCESS'94, vol. 2,

Espoo, Finland, 13±16 June 1994, pp. 531±538.

[22] M. Ulieru, R. Isermann, Design of a fuzzy-logic based diagnostic

model for technical processes, Fuzzy Sets and Systems 58 (1993)

249±271.

[23] N. Kiupel, P.M. Frank, Fuzzy supervision for an anaerobic waste-

water plant, Proceedings CESA'96 1 (1996) 362±367.

[24] J. Montmain, S. Gentil, Decision-making in fault detection: a fuzzy

approach, TOOLDIAG'93, Toulouse, France, 1993.

[25] D. Giraud, C. Aubrun, A fuzzy fault diagnosis method applied to a

steam circuit, IEEE International Conference on Fuzzy Systems

FUZZ-IEEE'96, vol. 3, New Orleans, LA, 1996, pp. 1944±1950.

[26] A.N. Boudaoud, M.H. Masson, An adaptive fuzzy diagnosis system

for on-line sensor data validation, Third IFAC Workshop on On-line

Fault Detection and Supervision in the Chemical Process Industries,

Lyon, France, 1998.

[27] J.-P. Steyer, D. Rolland, J.-C. Bouvier, R. Moletta, Hybrid fuzzy

neural network for diagnosis: application to the anaerobic treatment

of wine distillery wastewater in fluidized bed reactor, Water Sci.

Technol. 36 (1997) 209±217.

[28] M.N. Karim, A. Halme, Reconciliation of measurement data in

fermentation using on-line expert system, Fourth International

Congress on Computer Application in Fermentation Technology:

Modeling and Control of Biotechnological Processes, Cambridge,

UK, 25±29 September 1988.

[29] C.L. Cooney, C.L. O'Connor, F. Sanchez-Riera, An Expert System

for Intelligent Supervisory Control of Fermentation Processes,

Eighth International Biotechnology Symposium, Paris, France,

1988, pp. 563±575.

[30] C. Qi, W. Shu-Qing, W. Ji-Cheng, Application of Expert System to

the Operation and Control of Industrial Antibiotic Fermentation

182 A. Genovesi et al. / Biochemical Engineering Journal 3 (1999) 171±183

Page 13: A fuzzy logic based diagnosis system for the on-line supervision of

Process, Fourth International Congress on Computer Application in

Fermentation Technology: Modeling and Control of Biotechnologi-

cal Processes, Cambridge, UK, 25±29 September 1988.

[31] K. Koskinen, Expert systems as a top level controller of an activated

sludge process, Water Sci. Technol. 21 (1989) 1809±1814.

[32] R.J. Aarts, A. Suviranta, P. Rauman-Aalto, P. Linko, An expert

system in enzyme production control, Food Biotechnol. 4 (1990)

301±315.

[33] M. Kishimoto, M. Moo-Young, P. Allsop, A fuzzy expert system for

the optimization of glutamic acid production, Bioproc. Eng. 6 (1991)

163±172.

[34] M. Aynsley, A.J. Hofland, A.J. Morris, G.A. Montague, C. Di

Massimo, in: A. Fiechter (Eds.), Artificial Intelligence and the

Supervision of Bioprocesses (Real-time Knowledge-based Systems

and Neural Networks), Adv. Biochem. Eng./Biotech., vol. 48,

Springer, Berlin, 1993, pp. 1±27.

[35] M. Pokkinen, Z.R. Flores Bustamante, H. Asama, I. Endo, R. Aarts,

P. Linko, A knowledge based system for diagnosing microbial

activities during a fermentation process, Bioproc. Eng. 7 (1992) 331±

334.

[36] T. Siimes, M. Nakajima, H. Yada, H. Asama, T. Nagamune, P. Linko,

I. Endo, Knowledge-based diagnosis of inoculum properties and

sterilization time in lactic acid fermentation, Biotech. Technol. 6

(1992) 385±390.

[37] K.B. Konstantinov, T. Yoshida, The way to adequate control of

microbial processes passes via real-time knowledge-based super-

vision, J. Biot. 24 (1992) 33±51.

[38] K.B. Konstandinov, R.M. Matanguihan, T. Yoshida, Physiological

state control of recombinant amino acid production using a micro

expert system with modular, embedded architecture, IFAC Model-

ling and Control of Biotechnical Processes, Colorado, USA, 1992,

pp. 411±414.

[39] T. Kashihara, M. Mawarati, T. Inoue, J. Prior, C.L. Cooney, A pH

profile of beer fermentation using a knowledge-based system, J.

Ferment. Bioeng. 76 (1993) 157±159.

[40] J.-Ph. Steyer, I. Queinnec, D. Simoes, Biotech: a real-time

application of artificial intelligence for fermentation processes,

Cont. Eng. Pract. 1(2) (1993) 315±321.

[41] H. Ojamo, P. Vaija, M. Dohnal, Intelligent assistance to the

fermentation operation, Acta Polytechnica Scandinavica, Chemical

Technology and Metallurgy Series, no. 218, 1994, 30 pp.

[42] M. Kishimoto, H. Suzuki, Application of an expert system to high

cell density cultivation of Escherichia coli, J. Ferm. Bioeng. 80(1)

(1995) 58±62.

[43] T. Siimes, P. Linko, C. von Numers, M. Nakajima, I. Endo, Real-

time fuzzy-knowledge-based control of Baker's yeast production,

Biotech. Bioeng. 45 (1995) 135±143.

[44] T. Morimoto, K. Hatou, Y. Hashimoto, Intelligent control for a plant

production system, Cont. Eng. Pract. 4(6) (1996) 773±784.

[45] E. Roca, J. Flores, I. Rodriguez, C. Cameselle, M.J. Nunez, J.M.

Lema, Knowledge-based control applied to fixed bed pulsed

bioreactor, Bioproc. Eng. 14 (1996) 113±118.

[46] J.-Ph. Steyer, I. Queinnec, F. Capit, J.-B. Pourciel, Qualitative rules

as a way to handle the biological state of a fermentation process: an

industrial application, Journal EuropeÂen des SysteÁmes AutomatiseÂs

RAIRO-APII 30(2±3) (1996) 381±398.

[47] T. Ohtsuki, T. Kawazoe, T. Masui, Intelligent Control System Based

on Blackboard Concept for Wastewater Treatment Processes,

Seventh IAWQ Workshop on Instrumentation, Control and Automa-

tion of Water and Wastewater Treatment and Transport System,

Brighton, UK, 6±11 July 1997, pp. 131±138.

[48] P.C. Pullammanappallil, S.A. Svoronos, D.P. Chynoweth, G.

Lyberatos, Expert system for control of anaerobic digesters, Biotech.

Bioeng. 58(1) (1998) 13±22.

A. Genovesi et al. / Biochemical Engineering Journal 3 (1999) 171±183 183