1.4 Fault detection: structures and methodologies

Nowadays, the interest for supervision is increasing due to the growing demands for quality, safety, reliability, availability and cost efficiency in industrial processes. As systems grow in size and complexity, the possibility of misbehaviour increases. Thus, the call for fault tolerant systems is gaining more and more importance. Fault tolerance could be achieved either by passive or active techniques [Frank and Köppen-Seliger, 1995] :

• The passive approach makes use of robust control techniques to ensure that the closed-loop system becomes insensitive with respect to faults. This solution allows small faults be tolerated without control system reconfiguration.
• The active approach provides fault accommodation, i.e., the reconfiguration of the control system when a fault has occurred. Reconfiguration can be thought at various degrees, i.e. set point changes, parameters re-tuning or structural changes. The aim of this approach is to avoid a fast degradation of the whole system due to this fault. The majority of actual solutions involve human decision.


In this work, only active approach is taken into account. This chapter overviews different methodologies that can be applied in this sense to remark the importance of human knowledge about process behaviour and how it can be used to implement supervisory structures. Finally, it is concluded that the use of both, analytical and knowledge-based methods, together can improve the results of supervisory structures. In such cases, the main problem is about integration of methods provoked by differences in data representation (numeric, qualitative, symbolic).

Fault detection: structures and methodologies.

Supervision is essentially the set of techniques used with the goal of assuring the integrity of a system. The definition given in the last paragraph assigns to supervision the role of detecting (to recognise and to indicate) in real time abnormal behaviour of a process taken benefit of all information available about the process (measures, models, history, experience and so on).

According to these goals, the main part of supervision of a complex system is focused on to detect and isolate occurring faults and provide information about their size and source, [Frank and Köppen-Seliger, 1995]. The most important and difficult task is centred on fault detection and diagnosis where difficulty increases with the real time constraints and complexity of systems (non-linear, coupled dynamics, time dependencies, etc.). The core of the fault diagnosis methodology is the so-called model-based approach, where either analytical or knowledge-based models or combination of both are used in combination with analytical or heuristic reasoning [Frank and Köppen-Seliger, 1995]. The classical procedure of a fault diagnosis system is depicted in Fig. 2.2. This is achieved in three basic steps, residual generation, evaluation and analysis, not always clearly separable.

Fig. .2 Schematic representation of the procedure of fault diagnosis.

 
Hence, the methodology used in fault detection is clearly dependent on the process and the sort of available information. [Pal, 1995] gives an extensive classification and description of these methods (Fig. 2.3). A distribution of fault detection methods depending on applications is summarised in [Isermann and Ballé, 1996], from the main contributions presented in the domain main conferences from 1991 till 1995. The evaluation of fault diagnosis and supervision methods is more difficult because of little data, although rule-based reasoning methods are increasingly used and also the number of fuzzy rule-based applications is growing [Isermann and Ballé, 1996].

Fig. .3 Classification of fault detection methods [Pal, 1995].

 
Classification in Fig. 2.3 shows the diversity of methods available for fault detection using analytical models and the shortage of methodologies described in the case of the knowledge-based approach. The reason of this difference is due to the tools used in both situations. Most of the model-based fault detection methods are based on comparing measures and simulations for obtaining a residual as is depicted in Fig. 2.4. The application of existing signal processing and statistical methods to deal with numerical data boosted the used of such methods. The major inconvenient of those methods is the necessity of a model. The knowledge-based model offers an alternative to that situation in which an accurate model is difficult to be obtained.

In the following subsections, fault detection methods are briefly described and classified in model, signal and knowledge-based methods. Model-based methods group analytical methods based on comparing process output with theoretical model response to the same input applied real process. The next subsection, signal based methods, is restricted to those methods that only use process output variables in the fault detection task and no model is used. Finally, the knowledge-based method subsection, introduces the problematic of dealing with knowledge representation to monitor and supervise dynamic systems.

Model-based fault detection.

Model-based fault detection methods are focused on residual generation, i.e. fault indicators (Fig. 2.4). Residuals are obtained as changes or discrepancies in special features of the process obtained from process variables (for example, output signals, state variables) or coefficients (for example, estimated parameters or other calculated ratios). To achieve this goal, data obtained from the process is compared to the data supplied by models representing normal operating conditions. For this, different change detection methods are applied. The next step consist in residual evaluation to decide about faults existence. Threshold logic, statistical decision theory, pattern recognition, fuzzy decision making or neural networks are actual methods used to decide whether and where a fault has occurred.

Different kinds of process models and methods can be used to generate residuals of state or output variables. A basic classification of them is represented in Fig. 2.5. Observer-based residual generation consists in using observers or Kalman filters to reconstruct the interest output of the system. Then, the error between real data and estimated data or a function of them are used as residual. In the parity space approach the process equations (for instance, state space equations) are modified with the aim of getting residuals decoupled from system states and different faults. The inconsistency of these parity equations represent the residuals. On the other hand parameter estimation methods are based on the assumption that the faults are provoked by changes in the physical system parameters (mass, friction, resistance, viscosity, etc.). Therefore, these methods use process measures to repeatedly estimate the parameters of the actual process. Estimations are compared with the parameters of the reference model, obtained under fault-free conditions. A wider description of those methods and other interesting variants of them are included in [Frank, 1996].

Fig. .5 Dependency between method and model to be used in a model-based fault detection.

Although these approaches have proved to be very effective in many applications, they have two major shortcomings : the complex technology or natural process are generally non-linear time-varying systems, which makes it particularly difficult to detect structural changes in the system and to obtain adequate models for this purpose. Secondly, the available model is often assumed to represent normal operating conditions, and the impact of a departure from these conditions on the model outputs is difficult to predict [Du, Elbastawi and Wu, 1995a]. Consequently, these methods are difficult to be applied to dynamic process submitted to repeated changes in the operation mode.

• Symptom based. It consists in organising expert knowledge into diagnosis ESs. Then, ESs deal with process variables to identify fault symptoms in them (See Fig. 2.7). When symptoms are considered in connection with process input, we speak of a symptom-model-based approach. In such implementations, major difficulties exist in the knowledge acquisition task and knowledge representation, for obtaining a rule base for example, and knowledge processing and interfacing, when running the ES application with real data.
• Qualitative model-based. This methodology is based on using process knowledge to represent systems structure in terms of rules and facts. The goal is to dispose of a rough model to be used as a model-based approach. The use of qualitative techniques implies a description of process variables given by short sets of labels or symbols ( low, normal, high). Consequently, the inevitable deviations from the exact model, (uncertainties and incompleteness) always present in these representations, requires the AI support.

Fig. .6 Basic scheme of qualitative model-based fault detection method

 
The use of one or the other method is only submitted to the knowledge about process behaviour or faults. Furthermore, there are not exclusive methodologies and a combination of both can be required. In fact, the best strategy tries to use all available knowledge and data for reaching the fault detection goal.

Additional problems, when dealing with knowledge representation, of process behaviour or faults description, occur when interfacing fault detection structure (linguistic representation of magnitudes and process variables) and process (numerical variables, measures). However, the use of all available information, i.e. numerical data and knowledge, can improve fault detection structures because they are complementary approaches.

Fig. 2.7 represents how both, numerical techniques and knowledge-based techniques, can be merged for fault detection. In fact, heuristic knowledge acquired from process observations is used in this proposal to reinforce analytical methods applied to numerical variables in the symptoms generation step. While numerical methods are used for features extraction of measured variables, i.e. signal processing techniques as filtering and analytical estimations, some drawbacks appear in extracting features from operator observations. Specialised knowledge processing techniques must be used with this aim. Taking into account that difficulties in such systems are presented at knowledge representation level, the use of those techniques for features extracting are not very extended.