SOFT COMPUTING APPROACHES TO FAULT DIAGNOSIS FOR DYNAMIC SYSTEMS: A SURVEY R J Patton, F J Uppal & C J Lopez-Toribio Control and Intelligent Systems Engineering, Faculty of Engineering and Mathematics, The University of Hull, Cottingham Road, Hull Hu6 7RX, United Kingdom Email: r.j.patton@eng.hull.ac.uk Abstract: Recent approaches to fault detection and isolation for dynamic systems using methods of integrating quantitative and qualitative model information, based upon soft computing (SC) methods are surveyed. In this study, the use of SC methods is considered an important extension to the quantitative model-based approach for residual generation in FDI. When quantitative models are not readily available, a correctly trained neural network (NN) can be used as a non-linear dynamic model of the system. However, the neural network does not easily provide insight into model behaviour the model is explicit rather than implicit in form. This main difficulty can be overcome using qualitative modelling or rule-based inference methods. For example, fuzzy logic can be used together with state space models or neural networks to enhance FDI diagnostic reasoning capabilities. The paper discusses the properties of several methods of combining quantitative and qualitative system information and their practical value for fault diagnosis of real process systems. Keywords: Soft computing methods, fault-diagnosis, FDI, computational intelligence, AI methods 1. INTRODUCTION There is an increasing demand for man-made dynamical systems to become safer and more reliable. These requirements extend beyond normally accepted safety-critical systems of nuclear reactors, chemical plants or aircraft, to new systems such as autonomous vehicles or fast rail systems. The early detection of faults can help avoid system shut-down, breakdown and even catastrophes involving human fatalities and material damage. A system which includes the capacity of detecting, isolating, identifying or classifying faults is called a fault diagnosis system. During the last two decades many investigations have been made using analytical approaches, based on quantitative models. The idea is to generate signals that reflect inconsistencies between nominal and faulty system operation. Such signals, termed residuals, are usually generated using analytical approaches, such as observers (Patton et al 2000, Chen & Patton, 1999), parameter estimation (Isermann, 1994) or parity equations (Gertler, 1998) based on analytical (or functional) redundancy. Considerable attention has been given to both research and application studies of real processes, using analytical redundancy as this is a powerful alternative to the use of repeated hardware (hardware or software redundancy). The monitoring of faults in feedback control system components has come to be known as fault detection and isolation (FDI). The procedure of generating control action which has a low dependency on the presence of certain faults is known as fault-tolerant control. The FDI unit provides the supervision system with information about the onset, location and severity of any faults. Based on system inputs and outputs together with fault decision information from the FDI unit, the supervision system will reconfigure the sensor set and/or actuators to isolate the faults, and tune or adapt the controller to accommodate the fault effects. Early detection and isolation of small, incipient (rather difficult to detect) faults can be achieved with model-based processing of all measured variables, using either qualitative or quantitative modelling. Neural networks and fuzzy logic techniques are now being investigated as powerful modelling and decision making tools, along with the more traditional use of non-linear and robust observers, parity space methods and hypothesis- testing theory. Requirements for precise and accurate analytical model imply that any resulting modelling error will affect the performance of the resulting fault detection and isolation (FDI) scheme. This is particularly true for non-linear systems, which represent the majority of real processes. To circumvent this precision problem (at least in part) more abstract models, based on qualitative physics (de Kleer & Williams, 1987 Shen & Leitch, 1993, Kuipers, 1994, Lunze et al., 1999) may be used. Alternatively fuzzy-logic rules may be developed to either assist or replace the use of a

model for diagnosis (Dexter, 1995). The key advantage of fuzzy logic is that it enables the system behaviour to be described by ���if-then��� relations. Some research has been based upon neural networks which can be trained to reproduce a specified system behaviour from the data sets alone. Neural networks can, indeed, provide an excellent framework for dealing with non-linear systems (Naidu, Zafirou & McAvoy, 1990). The main feature of neural networks are their ability to model any non-linear function, given suitable weighting factors and an appropriate architecture. However, whilst such a configuration can be well trained on numerical data, heuristic knowledge from experts cannot easily be incorporated. It is also argued that, due to their ���black box��� characteristics, conventional neural networks do not give an insight into the behaviour of the system which is sufficiently comprehensible by the operator. Another drawback of substitution the operator���s ���intelligence��� by an automated analytical approach is that the operator���s expertise, built up over several years, is simply not used. This is mainly due to the inability of analytical methods to represent symbolic information. In the authors��� opinion a robust FDI system should combine both numerical (quantitative) and symbolic (qualitative) information. Some investigators tackled this problem by combining parameter estimation or observers with fuzzy logic (Frank & Kuipel, 1993 Isermann, 1994). The main idea has been to generate residuals using either parameter estimation or observers, and allocate the decision- making to a fuzzy-logic inference engine. In so doing, it has been possible to include symbolic knowledge with the quantitative information and, thereby, minimise the false alarm rate. Indeed, the key benefit of fuzzy-logic is that it lets the operator describe the system behaviour or the fault-symptom relationship with simple if-then rules. Here we use the term ���soft computing��� (SC) for all methods employing computational intelligence algorithms, e.g. fuzzy logic, neural networks, neuro-fuzzy schemes, evolutionary programming, etc. This paper gives an outline of SC methods which are considered a powerful extension to quantitative/analytical approaches to fault detection and isolation (FDI) for dynamic systems. One approach is to use a fuzzy rule-base to select the dynamic model which is most appropriate for a particular operating point (Wang et al., 1995 Tanaka et al., 1996). This is the so-called fuzzy inference multiple-model approach. The idea has recently been applied to FDI problems by Lopez- Toribio et al. (1998). It is important to structure a quantitative model in a way that qualitative knowledge about the process could be included as well as extracted. These can be achieved using a neuro-fuzzy approach. The underlying concept is to structure a neural network, which can model highly non-linear systems efficiently, in a fuzzy-logic format the network could therefore be trained more rapidly and will also provide a linguistic description about the causes of faults. The B-spline network can be a suitable network architecture for this problem due to an interesting equivalence relation with the function of fuzzy rule sets (Brown & Harris, 1994a). The difficulty with this approach is the rapidly increasing complexity of the rule base with system order and complexity. 2. MODEL-BASED FDI PRINCIPLES The aim of a quantitative model-based fault diagnosis is to generate information about the location and timing of a fault, using the measurements available in that system, as well as the precise mathematical relationships that relate them. Fig. 1 illustrates the conceptual structure of a model- based fault diagnosis system, which comprises the following main stages. System G(s) Hy(s) Hu(s) Residual Generator u(s) y(s) r(s) Decision Making fault information 1. Residual Generation: This is an algorithm which processes the measurable inputs and outputs of the system to generate the residual signal. 2. Decision making: The residuals are, then, examined for the likelihood of faults, and a decision rule is then applied to determine if any fault has occurred. A decision process may be based on a simple threshold test, on the instantaneous values or moving averages of the residuals, or it may consist of methods of statistical decision theory, e.g. likelihood ratio testing or sequential probability testing. The successful detection of a fault is followed by the fault isolation procedure whose aim is to locate the fault. ��� Observers: The underlying idea is to estimate the system outputs from the available inputs and outputs (Patton, 1997). The residual will then be a weighted difference between the estimated and the actual outputs. The flexibility in selecting the observer gain has been fully exploited in the Fig. 1 Model-based Fault Diagnosis Residual Signal: r(s) = Huu(s) + Hyy(s) (1) Objectives: choose Hu & Hy so that r(s) = 0 when no fault occurs r(s) ��� 0 when a fault occurs