Intelligent Diagnosis of Motor Rolling Bearing Based on Wavelet Dropout Convolutional Neural Network With Asymptotically Trained Interference

In recent years, in practical industrial applications, the task of accurate and robust compound fault diagnosis of motor rolling bearings under huge noise is still the focus of current research. To solve these problems, an end-to-end framework of the wavelet Dropout convolutional neural network (ATIWD-CNN) with asymptotically trained interference was proposed. First, an innovative six-layer discrete wavelet Dropout attention layer (DWDA-Layer) with an asymptotic deactivation rate was proposed, in which the time-domain signal was mapped to the wavelet domain to simulate noise interference by the dropout operation, thereby improving the shortcomings of deep neural network models in signal analysis; the proposed local dual-scale frequency attention mechanism (LDSFAM) can focus on stronger fault-related information from the mixed frequency components. Second, a larger convolution kernel was used in the first learnable convolution block to suppress the interference of high-frequency noise and extract a wider range of time-domain features, in which the smoother Gaussian error linear unit (GELU) was used in all convolution blocks, and the group normalization (GN) technique and the setting of ultrasmall batch size were used to make the training of the model more refined. Compared with some other advanced methods, the average diagnostic accuracy of the ATIWD-CNN on two experimental datasets (with −10-dB noise level) was increased by 7.16% and 2.42%, respectively.