1.3 SCADA systems and expert supervision: an overview

Nowadays, automatic control applications in the domain of process supervision are restricted to decision making, including human operator, into the feedback loop [Millot, 1996] . In fact, the functionality of Supervisory Control And Data Acquisition (SCADA) packages, widely used in the industry, are restricted to monitoring and alarm generation tasks. Final decisions about the validity of these alarms and actions to perform in the process are done by expert operators in the plant. Design difficulties of supervisory systems increase with process complexity. Coupled systems, high order dynamics, and non-linearity are common situations in industrial process. In such situations, it is difficult to obtain accurate models for applying traditional model-based techniques for designing supervisory systems and knowledge-based approaches can be applied taking benefit of process operators and process engineers’ experiences to identify specific situations. Therefore, Expert Supervision is an active research line oriented to take advantage of expert knowledge of process engineers and plant operators to automatically decide about process behaviour and to propose adequate actions or changes in the set points, controllers parameters or reconfiguring strategies. Artificial intelligence (AI) technologies are applied to deal with this expert knowledge inside computers because of its capabilities for:

• Knowledge representation. This is the interpretation of human knowledge and translation into computers. AI proposes several methods for structuring and organising expert knowledge into knowledge bases (KBs). Rule-based systems, also called expert systems (ES), are the most common example of applications used to represent exert knowledge.
• Reasoning. This is the capability of using the expert knowledge to infer conclusions as a consequence of specific situations.
• Manipulation of heterogeneous data. The use of AI technologies allows to manage qualitative and symbolic information mixed with numerical data and methods.


Those capabilities are extremely useful in expert supervision because of the different origin of information to be used . Expert supervisory applications are designed to reason about process variables, i.e. signals that represent physical variables in the process and provide information about process behaviour. These signals can be measures, numeric data obtained directly from real process or controllers, or estimations, obtained from models or relations. According to the kind of relationship or models used to estimate a process variables we can differentiate between qualitative, or symbolic, and numeric estimations. These signals can be manipulated or processed to isolate significant information, i.e. abstracted information.

The complexity of designing, testing and validating the expert supervisory systems is basically due to the variety of tools and data to be used. Difficulties in integration are the main reason of failure in supervisory strategies implementation. This pitfall will decrease if all the available facilities to be used are available together in a framework, avoiding integration inconvenient, especially oriented to assist expert supervisory systems design.

General Problem Description.

Expert Supervisory Systems design involves manipulation of heterogeneous information such as signals, knowledge, qualitative data and so on. All of them are representations of process behaviour that must be taken into account when designing supervisory systems. Numerical data comes basically from direct measures of process variables or from estimations supplied by numerical models or equations. Usually, numerical data is the kind of information supplied from process periodically, at every sampling time. This kind of information can be saved, processed and represented in several ways to be analysed. On the other hand knowledge and qualitative appreciation of process dynamics come from heuristics or observations of representative situations done by process engineers (Fig. 1.1). In industrial applications the existing knowledge representation procedures for assisting experts in translating and structuring this knowledge into computers are basically reduced to production rules systems. For example, fuzzy based controllers or industrial ES are in this line. Reasoning about dynamic systems involve working with temporal restrictions as expiring data validity and limitations in the response time. Periodically (at each sampling time), ES input data is actualised and deductions about them must be performed.
 
 







Fig. .1 Representation of information supplied to a supervisory system.


 





Therefore, design of a knowledge-based system to reason about dynamic data involves an important task on data analysis and features extraction to match description of situations given by experts. Main drawbacks are in obtaining perceptual information from process variables, in order to interface expert KBs, defined from human perception of process behaviour, and measures, supplied by instruments and numerical methods. Consequently, the design of expert supervisory systems consists of an iterative procedure until specifications are reached.

The definitive application sometimes involves the election of adequate description of input and parameters tuning by trial an error. This iterative procedure, represented in Fig. 1.2, is necessary, not only in the designing step, but also in previous tasks for defining adequate input and output of expert supervisory system.

 
Input must be defined in number, type and quality. The number of input can be restricted to according to process limitations (location, observability, transmission) or economical restrictions. Then, estimation capabilities (indirect measures, modelling) can be used as alternative. Data typology is always submitted to KB to be used, due to human subjectivity in the observations. These observations of process behaviour are strongly time dependent and, consequently, process variables supplied to inference engines must incorporate this dependency. Thus, the use of abstracted significative time representations of process variables at different levels will be very useful and, therefore, the exact shape of those trends must be defined to match expert interpretations. Certainty and imprecision is another factor to be taken into account in some kind of measures or estimations that are subject to unknown factors (perturbations, unknown parameters dependencies, deviations, time dependency, failures, and so on). Similar specification of output must be performed if they are used to as a high level feedback loop to propose actions to improve global system performances.

It is evident that different tools will be used to deal with this variety of information. As a consequence, interfacing problems between applications and data type mismatch will occur at the same time that users (supervisory systems designers) will operate with distinct user interfaces and applications front-end. This work is centred on providing a set of tools integrated into a framework to facilitate management of information when designing expert supervisory systems.

Implementation of expert supervisory systems.

Despite implementation of supervisory systems is not an extended topic in literature, it is an important stage. Especially, when dealing with knowledge-based methods because of the limitations of available shells. Nowadays, knowledge-based methods are thought to be useful in all supervisory tasks. The easy use of rule-based ES for classifying or the use of graphs or fault trees as analysis tool, are simple tasks where expert knowledge can improve fault detection. Moreover, the use of fuzzy logic to deal with imprecision and uncertainty, or neural nets for classification, are other representative examples of how AI is coming into supervisory schemes.

Nowadays, final actions are restricted to human decision, then the implementation of intelligent supervisory systems is far from the actual implementations in industrial process and restricted to simple controller reconfigurations or rule-based controllers (fuzzy controllers). A reason for these difficulties in implementing expert supervisory systems is the complexity in knowledge representation and validation of such applications.

Assisting expert supervision design.

A possible block representation of supervisory system, is depicted in Fig. 2.10. A model-based (analytical or qualitative) fault diagnosis system is used to reconfigure the control system. Actual shells and existent frameworks used for industrial monitoring tasks (SCADA systems as CITECT, InTouch and so on.) permit such kind of reconfiguration according simple conditions. These are well dotted of numerical methods and tools, but they lack of knowledge-based methods. In fact, the application domain of these applications is reduced to simple monitoring, representing, alarm generation and registration tasks. On the other hand, actual shells conceived to deal with knowledge-based systems are designed to manage specific kind of knowledge representation and there exist some difficulties in integrating numerical capabilities. This is the example of the shell G2, the actual state-of-the-art in real-time knowledge-based process diagnose tasks. It is provided with an object-oriented graphical user interface that offers several knowledge representations tools, such as rule bases, frames, tables. G2 inference engine is able to deal with KB in several ways (forward and backward chaining, focusing, invoking). Despite of its capabilities of knowledge representation, it presents some drawbacks in qualitative modelling and uncertainty (and imprecision) management, which are necessary in a complete knowledge-based framework. Another inconvenient is detected when dealing with multi-level representation of process variables. In such cases it is necessary to abstract significant information from numerical measures, and those frameworks do not provide sufficient tools for numerical management and signal processing.

Another important factor to be taken into account in the design of supervisory systems is the application domain. Different tools and signals must be used for process supervision than for controllers supervision. It also applies to continuous, discrete or hybrid systems and to distributed or centralised systems.

Management of both numerical and qualitative data and the use of analytical and knowledge-based models and tools is necessary in actual supervisory systems. Taken into account that the supervisory loop is closed by means of reconfiguration of the control systems, a supervision shell must also include control systems design capabilities and some evaluation and graphical representation tools. The actual Computer Aided Control Systems Design (CACSD) frameworks incorporate such capabilities and a great number of numerical algorithms, including modelling and simulation capabilities. On the other hand, these frameworks are not provided with knowledge-based capabilities. Consequently, a possible approach for assisting expert supervisory systems design based on control reconfiguration, is to add following knowledge-based capabilities to CACSD frameworks:

• Numeric/Qualitative interfaces. Process variables are the main link between the process and the supervisory system. Therefore, different observations of them must be possible and, consequently, a multi-level representation of them must be performed. Numeric features of process measures, as its actual value, derivatives, mean or trend estimation, must be available together with qualitative representations such as qualitative tendency, labels or landmarks, to be useful for working together with both analytical and knowledge-based supervisory methods. Then a set of different methods for obtaining accurate qualitative representation of signals are needed to be used as numeric to qualitative interfaces. i.e. abstraction methods. The object oriented approach can be used to encapsulate the multi-level representation of these variables (object-variables) as it is explained in chapter 5.
• Knowledge representation. Several tools can be used to deal with expert knowledge at several abstraction degrees. Knowledge about process can be about process variables, the relationship between them, the description of situations (faults) related to physical elements, functionality, behaviour and so on. All of these possibilities require some specific tools with different representation capabilities.
• Knowledge processing. Usefulness of knowledge-based systems reside in the capability of deduction using knowledge representation. Then those inference engines must also be present in the framework to deal with qualitative relationship (qualitative reasoning), rules (ES), logic, and other possible knowledge representations. Capabilities of certainty and imprecision management must also be present, for example, by using fuzzy reasoning.


Some tools are selected and presented in chapter five to assist in the design of control systems supervision. With this goal a specific commercial open package from the control domain, MATLAB/Simulink, has been used as a platform where knowledge representation capabilities and qualitative reasoning tools have been integrated by means of an object oriented approach [Melendez et al., 1996b]. The goal is to take advantage of the existent representation and analysis tools used in control systems design, extending its capabilities with knowledge-based techniques to assist engineers in such designs, avoiding the always difficult problem of interfacing separated tools reasoning on dynamic systems. The work has been centred on solving main drawbacks in knowledge representation and processing. With this aim, a tool for dealing with simple qualitative representations, ALCMEN, and an ES shell, CEES, have been added. Other possibility is pointed by [Rengasamy, 1995], using CIM models to integrate several functionalities in order to assist supervision.