Machine Fault Diagnosis Based on Wavelet Packet Coefficients and 1D Convolutional Neural Networks

Deep neural networks are becoming popular to automatically extract discriminative features from vibration signals for the purpose of recognizing machine running state. The fact that little attention has been paid to the nature of signals as to whether the combination of preprocessing can help to achieve better diagnosis performance motivated this investigation. A fault diagnosis method based on wavelet packet coefficients and 1D convolutional neural networks (1D-CNN) is proposed in this paper. Firstly, the signal features involved in both low- and high-frequency components are extracted based on wavelet packet decomposition, and combined without changing their output sequence directly for the sake of retaining all the characteristics. Secondly, the combined features are taken as input of a 1D-CNN, which consists of several convolutional layers, to further learn the abstract features and improve classification performance. Finally, testing data were fed into the trained model to recognize the machine running state. The efficacy of the developed model and its anti-noise performance under different simulated noise levels were evaluated by analyzing the bearing vibration signals.