The Study On Theory And Method Of Fault Intelligent Diagnosis Based On Support Vector Machine

To effectively acquire, transfer, deal with, regenerate and utilize diagnostic information has been the core of intelligent fault diagnosis that exhibits an ability to precisely identify the patterns of faults and predict future faults of the system which is diagnosed. Recently, the dominating difficulties that fault intelligent diagnosis system faces, are terrible lack of typical fault samples and finding problem of diagnosis knowledge, both of which badly prohibit the development of machinery fault intelligent diagnosis. In recent years, the research of support vector machine (SVM), which is basing on statistical learning theory (SLT), is becoming gradually one of the most important branches in the field of machine learning. Mainly according to the key problems of support vector machine need to resolve in fault intelligent diagnosis system, this paper makes more systemic and thorough researches in feature extraction based on wavelet packet, model building of fault classifiers, fault feature selection based on kernel principal component (KPCA), parameters optimization of kernel function, incremental learning, engineering application, and so on.Aiming at the characteristics of fault diagnosis with finite samples and based on difficulties traditional mode identifying method of gradual-close theory faces in fault pattern classifier, the important significant is discussed about applying systemically support vector machine to fault intelligent diagnosis system, and a novel research method is put up for fault intelligent diagnosis. The key problem is studied in applying support vector machine to fault diagnosis, and the foundational realizing steps are brought out for support vector machine applying to fault diagnosis According to that standard support vector is deduced from two class classification problem, which can't be applied directly to resolving multi-class problem just like fault diagnosis, multi-class arithmetic is put forward which adopts decision directed acyclic graph, and multi-fault classifier is set up. Taking typical fault, which rotor experiment platform simulates, as diagnosed object, the method of extracting fault features by applying wavelet packetdecomposition is studied, and also fault classifier model is set up and fault detection and diagnosis comes true successfully. Finally, all kinds of factors, which influence classifying property of fault classifier, are analyzed deeply.In order to decrease the calculating complexity of classifier, increase the separability of fault models, it is necessary to select the feature vectors. Aiming at some shortage of principal analysis in fault feature selection, a kind of effective nonlinear feature selection method, which bases on kernel principal analysis, is put out and realized. By means of calculating the integral operator kernel functions of original feature space, this method actualizes nonlinear mapping from original feature space to high dimensional feature space, and gets nonlinear principal components of original feature data by making principal analysis on mapping data in high dimensional feature space. Through the selection of feature vectors, dimensions of feature data and calculating complexity of classifier are decreased effectively, and experiments show that kernel principal analysis is sensitive to nonlinear features of machinery faults, and is fit for the selection of nonlinear features in fault signals than principal analysis.Classifying performance of fault classifiers has great relations with the parameters of kernel function in support vector machine. Based on studying existent kernel parameters optimization, the theory of kernel function optimization of taking Fisher linear discrimination as object function, is discussed. Moreover, self-optimization algorithm, which is based on combination of Fisher linear distinguish function and improved genetic algorithm, is put forward. The new algorithm makes use of genetic operator to realize parameter optimization of kernel function and needn't calculate grads. Parameter optimization of two-class fault classifier comes true according to the new algorithm.After deeply analyzing the features of support vector set, Kuhn-Tucker Conditions (KKT), geometrical distributing of sample dots and change rule of support vector set after increment learning are researched, then a novel increment learning method of support vector machine is put forward based upon exceeding margin technique and Kuhn-Tucker Conditions. With the field knowledge included in gradually accumulating sample space dataduring increment learning. It becomes possible to discard sample selectively. With this algorithm, simulated data and practical fault data are tested. It makes clear that while guarantying classification precision, this algorithm can obtain effectively field knowledge contained in sample data, decrease quantity of training sample and occupied save space and increase speed of following training.Combined with practical conditions of factory, the mode of system development, which applies combination of Client/Server (C/S) and Blower/Server (B/S), is put forward. Based on this, the whole structure of distributed, networked condition monitoring and fault diagnosis system is put forward, and several key technique problems in system realization are studied. Then an on-line condition monitoring and fault diagnosis system for gas blower was developed. After integrating above research achievements, intelligent diagnosis system based on support vector machine is achieved. Experiment results prove that the system is satisfied to requirements on the real-time, exactness, reliability and security. Difficulty, which bothers corporation's manufacture, is resolved.