On-line fault diagnosis of industrial processes based on artificial intelligence techniques
In this research the application of artificial intelligence techniques for on-line process control and fault detection and diagnosis are investigated. The majority of the research is on using artificial intelligence techniques in on-line fault detection and diagnosis of industrial processes. Several on-line approaches, including a rule based controller and several fault detection and diagnosis systems, have been developed and implemented and are described throughout this thesis. The research results obtained demonstrate that rule based controllers can be an alternative in situations where conventional mathematical modelling fails to give a high level of automation. The research on on-line fault detection and diagnosis emphasises the use of deep knowledge based approaches. Therefore, two on-line fault detection and diagnosis systems based on qualitative modelling have been implemented. For the first one only single abrupt faults have been considered while the second one can cope with single and multiple simultaneous abrupt faults. In order to overcome the problems associated with the inherent ambiguity of qualitative reasoning, a fuzzy qualitative simulation algorithm, which allows a semiquantitative extension to qualitative simulation, has been investigated. The adoption of fuzzy sets allows a more detailed description of physical variables, through an arbitrary, but finite, discretisation of the quantity space, and also allows common-sense knowledge to be represented rough the use of graded membership. Further research concerning self-reasoning has been one for qualitative model based diagnosis approaches. A self-learning system which can find any inappropriate settings of fault detection and diagnosis parameters and also learn fault symptoms from on-line sampled data, has been developed. Through machine learning techniques, the system can adjust fuzzy membership functions of the process variables automatically, as well as build the knowledge base on-line very efficiently. In order to cope with incipient faults and transient behaviour of the process under concern, a distributed online fault detection and diagnosis system, consisting of a knowledge based approach coupled with a fuzzy neural network, has been developed. The fault detection task is performed through the knowledge based approach. A systematic methodology for formulating fault detection heuristic rules from knowledge of system structure and component -functions has been investigated. Since structural decomposition corresponds to plant topology, such a method could be easier to implement. A fuzzy neural network approach has been used for fault diagnosis. This system combines the advantages of both fuzzy reasoning and neural networks. In order to speed up the fuzzy neural network training task, an extension of the classical backpropagation learning algorithm is also investigated. The research results achieved with this fault detection and diagnosis system reveal a very good performance and reliability provided that the training data is available.