Study on fault detection using wavelet packet and SOM neural network

Abstract

Successful fault detection is based on effective feature exaction and selection processes. Feature map is one of the current fault diagnosis methods. By continuously tracking the trajectories, degradation trend in feature space can be detected. The challenge is how to construct a feature space that can consistently exhibit the degradation pattern. Self Organizing Map (SOM) neural network can map any high-dimensional input into a low-dimensional space, remaining its original topological structure. In this paper, the energy values of different frequency channels of acquired vibration signal are extracted as feature vector by wavelet packets decomposition. SOM based method is proposed to address the problem of feature space construction. Fault detection can be achieved by Minimum Quantization Error calculation (MQE), which can also be transformed into normalized Confidence Value(CV). Finally, the proposed method was also verified to be effective and pragmatic for fault detection via a hydraulic pump test.