4.16 Artificial Intelligence in control systems

Since its beginning, AI has been present in control and process engineering. At that time AI was focused on learning architectures using connective systems called artificial neural networks. Some structures such as, Perceptron, by Rosenblatt in 1962, and ADALINE, by Widrow in 1962, are two representative examples of these epoch. From then until now, several approaches have succeeded in the AI domain to improve control systems according to two perspectives ; on one hand the techniques that try to mimic the expertise of humans and on the other hand the techniques based on machine learning [Årzen, 1995b]. In spite of the fact that this thesis is centred on the first approach (using expert knowledge to improve control actions and supervise process behaviour), a brief description of these approaches is given :

• Learning control, refers to control systems that are able to estimate unknown information during their operation and determine an optimal action from the estimated information. It is conceptually similar to classical adaptive control. Some learning algorithms are based on pattern classification, Bayesian estimation and stochastic approximation.
• Neural Control. This terminology refers to the use of artificial neural networks for learning process behaviours by training them. It has had an increasing popularity since the eighties. Various learning strategies have been developed for learning input/output relationship (supervised learning) or only sets of input (unsupervised learning) or delayed output (reinforcement learning) data. Multiple strategies can be found in the literature.
• Fuzzy control. This is the use of fuzzy logic for implementing direct control strategies. The idea is to describe control actions by means of simple rules applied to the input variables to decide about control action. It involves three steps : fuzzyfication, and defuzzyfication. Fuzzy logic is maybe the most extended technology, from the AI domain, in the industry. Nowadays, fuzzy logic is used in control systems where classical PID does not offer satisfactory solution (for instance, in non-linear or time dependent systems) as industrial controller or fuzzy logic unit into programmable logic computer (PLC).
• Expert control. The notion of expert control was originally introduced in [Åström, Anton and Årzén, 1986]. The aim is to use AI-base systems to supervise numeric control algorithms. These are systems designed to deal with both signal and symbols. The idea is to separate signal processing algorithms and logic. The implementation of logic is typically by means of rule-based ES. Some applications in this domain are auto-tuning controllers or fault diagnosis systems and therefore, fall within the scope of this thesis. Example of an expert control architecture is depicted in Fig. 3.5.

Fig. .5 Expert controller structure proposed in [Årzen, 1995b]

All of these AI technologies have been tested and applied with success in direct control applications. The difference between applying such technologies in a control strategy or in a supervision structure, is basically the kind of data to be used. Control is performed close to process and, therefore, input and output of controllers are numerical sampled data coming from, usually, one source. Simple numerical features (derivatives, filtering) are obtained from these signals before being processed by the controller. In supervisory systems more significant information must be obtained to be useful for automatic use by AI tools, because they try to identify those signals with structural changes, faults or localised missfunctions. Moreover, qualitative representations of these signals must be managed together with numerical values.

Expert systems were originally developed to solve static problems (situations where the premises do not change with time). Consequently , the introduction of ES in the control domain had some difficulties because of the importance of temporal dependencies of dynamic systems where control strategies are applied. Then, normally they are used in a higher level where time dependency is less than in direct control. Expert control falls within this direction and represents an example of how ES can be used for supervision.