Bearing fault anti-noise diagnosis method based on inverted pyramid depthwise separable convolution sequence and hierarchical vision transformer using shifted windows

Abstract To address the huge vibration-noise interference of bearings under complex operating conditions, which makes the bearing fault signals suffer from the problems of missing features and multi-feature coupling, we propose a fault diagnosis method that fuses inverted pyramid depthwise separable convolutional sequences (IPDSCS) with Swin Transformer, named IPDSCS-SwinT. Firstly, the noise of the one-dimensional vibration signals is analyzed in time–frequency domain by wavelet transform, and then the red–green–blue three-channel segmentation of the time–frequency domain image is performed by group convolution. The dynamic separable block and the residual depth separable block are combined to extract the time–frequency features of the fault, and then the Swin Transformer is used to capture the long-distance dependence of the noise. Finally, the faults are classified by the fully connected layer and the softmax output of the fault classification. IPDSCS-SwinT effectively extracts frequency and time domain features of vibration signals, and the effectiveness of the proposed method is verified on Case Western Reserve University and Paderborn University bearing datasets. By comparing the advanced classification methods, experimental results show that the proposed method has excellent performance under different signal to noise ratio conditions. Especially in the low signal to noise ratio and compound noise environment, its fault diagnosis accuracy and robustness are significantly better than other methods.