The Method And Application Of Fault Diagnosis Based On Granular Computing And Neural Network

In recent years, many scholars in the world has researched the theory and its application of Granular Computing, found that the theory of granular computing not only provides a powerful tool for the analysis, reasoning of incomplete data, and the discovery relationship between data, extract useful properties, simplify information processing, but also provides a new approach for representation, learning, induction of imprecise and uncertain knowledge. It has achieved fruitful results in various fields. Granular computing theory presents the method of reduction and finding core on knowledge and data, and offers the ability of extracting rules from decision-making information system. So the theory of granular computing has advantages in describing attribute importance and reduction of knowledge representation system, which can effectively deal with consistent or inconsistent decision table. But granular computing theory also keeping some shortcomings, for instance in tolerance and promotion ability is relatively weak, only handle quantitative data, in dealing with pathological data on often proved powerless. Artificial neural network use parallel processing methods, so it can quickly associate to learning samples of similar situation, gets a quick decision. Neural network can solve the problem of self-learning and knowledge acquisition, but when the quantity and distribution space of the samples are numerous and complex, the neural network training is difficult to converge and the processing of ANN lacks of transparency. If the granular computing can combine with neural network, which not only good use their advantages, but also make up each other mutual shortage.Therefore this paper proposes a loose coupling fault diagnosis method based on granular computing and neural network, which uses attribute reduction advantages of granular computing theory and self-learning and knowledge acquisition ability of neural network. Its core idea is:put granular computing theory as the front-end processor of the neural network, namely simplify primitive information making use of powerful ability of granular computing reduction; based on this, in this paper, the concepts of relative granularity and significance of attributes based on binary granular computing are proposed to select reasonable input variables to form the most simplified rules, and are used as the heuristic information of reduction algorithm, thereby reducing solving scale; And then construct neural network based on the minimum attribute sets, using BP neural network to model and parameter identify; finally through the training samples to train. Such not only reduces the ANN training time, but also improves the diagnostic accuracy. Compared with the result of fault diagnosis for one distribution network, it shows that this method is feasible and effective, and has better tolerance and on-line fault diagnosis potential.