Support vector machine optimized by multi-objective particle swarm and application in gear fault diagnosis

Support vector machine (SVM) is a classification method based on the structured risk minimization principle. The classification accuracy and generalization capability of SVM model depend on the selection of penalization coefficient, kernel function and kernel parameters. Currently, empirical method, error method and single-objective optimization are usually used in the process of parameters selection, which waste time and energy and cannot obtain the global optimal solution. In the paper, SVM misclassification rate and support vector occupation ratio are chosen as two objective functions, and SVM penalization coefficient and kernel parameters are optimized by multi-objective particle swarm optimization (MOPSO). MOPSO is utilized to produce many optimization solutions in training stage, then these solutions will be evaluated in testing stage. The validation of design method is conducted by the case of the gear fault classification. Firstly, the vibration signals are preprocessed; secondly, the data after reducing dimensions of the standard deviation of wavelet package coefficients are regarded as the input eigenvector; finally, four typical faults of gear are classified. The results of test show that the method proposed is particularly valid and SVM classifier with mixed kernel function has higher accuracy and stronger generalization capability.