Research of Support Vector Machine Based Fault Diagnosis System
|Course||Electronics and Communication Engineering|
|Keywords||SLT SVM fault diagnosis kernel functions multi-fault classifier|
Knowledge based diagnosis method is a promising method, especially in the application of nonlinear system. Its intelligent technology and expert knowledge provides a simple and reliable system to the users. However, the absence of fault samples restricts the practicability of this technique. Support Vector Machine (SVM) is a machine-learning algorithm based on statistical learning theory, and has excellent performance even under a small number of fault samples. SVM can mine the intrinsic characteristics contained in the datum and extract the classification information furthest. So it’s very suitable for the application of fault diagnosis.This thesis first studies the set condition of empirical risk minimization principle, the relationship of empirical risk and expected risk under finite samples, and how to use these theories to find out new learning method. Then the control methods of confidence range of VC dimension are presented. The structural risk minimization principle founded on VC theory can control the real risk boundary by controlling the empirical risk and confidence range, which provides a solid theoretical foundation for building a learning mechanism with good generalization performance.This thesis introduces the construction of the optimum hyper plane and the corresponding solution. For linearly separable data set, classification hyper plane can be constructed in the input space. While the data set isn’t linearly separable, classification hyper plane can be constructed in feature space by kernel function mapping. Taking the noise data into account, classification hyper plane can be constructed by introducing slack variables.Two types of data are constructed on the basis of the Iris standard data set, and corresponding classification algorithm is simulated. Employing three widely used kernel functions, an in-depth study on model selection of SVM is carried out by Matlab simulation, which shows the selection of the optimal parameters, the effect of these parameters and the comparison between different kernel functions. In the end, the optimal SVM with different kernel functions are obtained by model selection.Finally, a fault diagnosis method based on SVM is proposed, using different kernel functions and parameters. This method is applied to the actual grinding machine fault diagnosis, which illustrates that different algorithm, different kernel functions and different parameters, will impact the support vector machine performance. Parallel diagnostic network is proposed based on the support vector machine, to improve the accuracy of fault classification when the number of samples is limited.