Gear fault pattern recognition based on atomic decomposition and Support Vector Machines

In order to solve the problem of feature extraction in the gear fault pattern recognition, a method of feature extraction based on atomic decomposition was proposed. Signals are rapidly decomposed using matching pursuit with the constructed Gabor dictionary. The frequency parameters and respective correlation values of the selected atoms constitute the feature vector of signal. Binary Tree Support Vector Machine is used as the classifier. The kernel parameter was optimized through 5-fold cross-validation. Pattern recognition of six classes gear fault is conducted, and the result shows that the method is valid. Through comparison, it has been found that the method is better than statistical indices.