Research Of Fault Diagnosis Based On Clustering

After 1980s, the large-scale automatic equipment in which embed micro-electronics and computer has been widely applied, meanwhile the fault diagnosis of large-scale automatic is becoming more and more complex, so the research of fault diagnosis is very important. In this thesis the method of combining clustering with neural network is used to improve the fault diagnosis system's intelligent degree.Clustering algorithms, for example K-MEANS,DBSCAN,CURE,STING,ART(adaptive resonance theory), have been researched in this thesis, a conclusion that ART is more suitable for fault diagnosis was got. Because ART not only inherits the fast processing speed,strong learning ability,association ability of neural network, but also has the unsupervised feature of clustering, ART could improve the fault diagnosis system's intelligent degree in some extent. ART1 and ART2, which are two algorithms of ART, will be studied in this thesis mainly.Because ART1 only could process binary data, the data conversion algorithm based on threshold was selected to convert the data into the form that ATR1 could process, then the 0-1 feature selection algorithm based on generalized matrix was used to reduce the dimensions of the input data of ART1. Since ART2 could process analog data, the Min-max normalization algorithm was used to transform the data into the form that ART2 could process; the feature selection algorithm based on consistency was used to reduce the dimensions of the input data of ART2.This thesis develops a method which through changes the weights of ART2 network to improve the inaccurate cluster phenomenon occurring when ART2 rapidly self-organize pattern recognition categories in response to low-dimension input patterns. An ART2 progressed algorithm which is composed of ART2 algorithm and K-Means algorithm is developed in this thesis, and it can restrain the drifting of cluster centers efficiently. Using gray theory, the forecasing model was built to avoid the problem that the feature selection algorithm based on consistency is difficult to determine the value range of numerical feature. At the end of the thesis, through the program of Visual C++ and SQL Server proved the feasibility of above algorithms.