CONTENTS

6.1. Process Optimisation

The application of optimisation techniques is not restricted to the design of predictive constrained controllers. Process optimisation is a task in its own right. Unlike local controllers, which seek to maintain unit operating conditions at desired levels, the plant optimiser utilises a model of the plant to adjust operating conditions of the process so as to minimise raw material usage and maximise profits [Edgar and Himmelblau, 1989]. The outputs of the optimiser therefore set the targets for the local controllers, taking into consideration the operational limits of the plant. This effectively bridging the gap between the plant's true business objectives and its actual operations [Latour, 1979a, 1979b]. Figure 8 shows a generic configuration of a process optimisation scheme.

Figure 8. Structure of Optimisation Scheme

Due to the complexity and the scale of this type of optimisation problem, the model used is normally a steady-state description to enable a tractable solution. As with control algorithms, adaptive on-line optimisation is also feasible [e.g. Bamberger and Isermann, 1978; Kambhampati et al, 1992; Willis and Tham, 1989b, 1990].

6.2. Process Monitoring, Fault Detection, Location and Diagnosis

Fault diagnosis has become an area of primary importance in modern process automation. It provides the pre-requisites for fault tolerance, reliability or security, which constitute fundamental design features in complex engineering systems. The system under consideration is monitored and the data is passed to fault detection algorithms or procedures. The basic task of a fault detection scheme is to register an alarm when an abnormal condition develops in the monitored system. Once a fault is detected, procedures may also be subsequently used to identify or diagnose the cause of the abnormality.

Fault detection and diagnosis techniques are again based upon the use of process models. In addition to the mathematical models used in controller design, statistical as well as qualitative models are increasingly being employed [Isermann, 1984; Patton et al, 1989]. Mathematical models are normally used to develop state-estimators or state-observers. Data from the monitored plant is input to these algorithms and the outputs compared with the corresponding plant outputs. If there are discrepancies, then it is an indication that at least one fault has occurred. The next task is to determine the locations of these faults. Again a representative model, not necessarily the one used in fault detection, is employed. In some instances, the location of the fault may be deduced by the type of fault. Here genetic algorithms and rule induction systems can be used to classify the fault.

6.3. Process Supervision via Artificial Intelligence Techniques

Human beings are able to make judgements in the face of subtle nuances and ambiguities. These knowledge processing capabilities cannot be matched by number crunching data processing algorithms, such as those described above. Although, the human decision system may not be precise, the result is often of sufficient accuracy for quick and effective problem solving. It has been the goal of computer scientists for many decades to build systems that mimic the decision making powers of human beings, i.e. artificial intelligent (AI) systems. AI techniques are also model based. Some would regard neural network based techniques to fall into the AI category. However, we tend to consider neural networks as numerical function approximators. Although AI techniques can make use of mathematical and statistical models, including neural networks, much of their utility is based upon the use of qualitative models.

Perhaps the most well known AI process supervisory schemes is based upon the use of expert systems [Efstathiou, 1986]. Expert systems are made up of three components. The rule or knowledge base holds information and logical rules for performing inference between facts. Next, there is the inference engine which controls the operation of the system and carries out the logical inference by processing the information in the knowledge base. The user interface makes up the final component, enabling communication between the user and the computer. Thus, an expert system is a collection of computer programs which operate upon the knowledge of experts in a particular application domain. Its purpose is to enable a novice to solve a problem with the benefits of the expert's knowledge.

When the inference engine and the user-interface are packaged as a single entity, this is known as an expert system shell. Software for procedures that can be combined together to form such a shell are known as expert system tools. The increasing availability of expert system shells and tools is a major reason for the proliferation of expert systems, where all that remains to be done is the compilation of the knowledge base. The extraction of rules that govern the operation of a process is called knowledge elicitation. This is performed via question and answer sessions between the extractor of knowledge, the so called knowledge engineer, and the provider of knowledge, the domain expert. There are also systems that are able to generate rules for expert systems when presented with data collected from a process. These are either based on rule induction techniques or genetic algorithms. However, the knowledge base could comprise any of the other qualitative models described previously and in any combination, including mathematical and statistical models.

When the system is presented with a collection of facts or a process scenario, the inference engine moves through the knowledge base in a 'forward' manner to come up with 'expert' advice or suggestions. However, unlike the implementation of 'IF-THEN-ELSE' constructs in conventional programming languages, expert systems have the ability to traverse the knowledge base in a backward direction. Backward chaining is invoked when the expert system is presented with a final result, and it is asked to provide a line of reasoning as to the events that led to the given result. Thus, another distinguishing factor of expert systems is that they are also able to provide explanations as to why a particular piece of advice or suggestion has been made. Expert systems have therefore found use in providing operator advice and as a process simulator for operator training [Kaemmerer and Christopherson, 1985].

Expert systems can be operated in two ways. The most common is the consultative mode where the expert system asks the user a series of questions. Alternatively, the data required by the expert system is provided directly by interfacing to plant instruments. There is a growing number of expert system shells that can reason in real-time [Shaw, 1988]. Such Real-Time Knowledge Based Systems (RTKBS) have been used to tune controllers, supervise the performance of adaptive controllers, perform fault detection and diagnosis, perform alarm management and even provide direct on-line process control.