Bearing Fault Diagnosis Based on Hermitian Wavelet and One-Dimensional Convolutional Neural Network

Abstract

The two-dimensional convolutional neural network (2D CNN) not only has numerous parameters and consumes long training time, but also is prone to overfitting. A bearing fault diagnosis method based on Hermitian wavelet and one-dimensional convolutional neural network is proposed (HW-1D CNN). The method combines the advantages of Hermitian wavelet that is sensitive to singularities caused by fault signals and the low computational complexity of 1D CNNs. Secondly, using the kurtosis index as the best scale selection criterion, the wavelet scale with the largest kurtosis value is chosen to extract the bearing fault transient impact features, and the 2D time-frequency matrix is reduced to a 1D vector. Finally, the reduced 1D vector is taken as the input of the 1D CNN, and the automatic extraction and fault classification of bearing fault features are achieved by 1D CNN. The experimental results demonstrate that the HW-1D CNN method has a high recognition accuracy for the bearing fault signals in the low signal-to-noise ratio background, which verifies the effectiveness of the method.