A novel intelligent system based on WPA-SVMs for fault signals recognition

Conventional wavelet transform (WT) omits some useful details information of fault signals since it only decomposes low frequency band in a higher scale. In this paper, a novel intelligent system is presented for real-time detection and diagnosis of the fault signals. The model consists of wavelet packet analysis (WPA) unit and support vector machines (SVMs) unit. When signals are decomposed in wavelet packet space, different frequency bands are processed from original signals adequately. WPA can improve abilities of feature extraction than conventional WT. We use a large number of samples to compare the accuracy rate of three kinds kernel function of SVMs, the results indicate that accuracy of Gaussian kernel is higher than polynomial kernel and multilayer perceptron (MLP) kernel. No matter whether the data set is small or huge, accurate classification rate of SVMs is better than RBF and BP neural network methods for normal and exceptional subjects.