Bearing fault diagnosis based on multiscale dilated convolutional neural network

In recent years, the fault diagnosis of rolling bearings has become a significant research direction. This paper proposes a new intelligent diagnosis method for the fault diagnosis of rolling bearings. In this paper, a 2D-MDCNN model structure is proposed. Firstly, the 2D grayscale map of the original vibration signal and the wavelet transform spectrogram are fused, and the fused 2D hybrid feature image is imported into the model using the multiscale and dilation convolution. It has two advantages: 1) the wavelet transform and 2D gray map can combine the signal’s time domain and frequency domain, while the 2D gray map can fully ensure the information extraction. 2) Using multiscale and dilation convolution, the information can be extracted at different scales, significantly improving the feature extraction capability and expanding the perceptual field by dilation convolution.