A bearing fault diagnosis method based on adaptive residual shrinkage network

Convolutional neural networks have been widely applied in the fault diagnosis domain of rolling bearings. However, as the number of network layers increases and the amount of input data is large, the problem of overfitting, gradient disappearance may occur; sometimes only one Fully-Connected layer can not solve nonlinear problems; additionally, there is the problem of poor anti-noise ability of the network. To address these problems, an algorithm based on the adaptive residual shrinkage network is proposed. The algorithm adopts the adaptive residual shrinkage module to solve the problem of gradient vanishing; soft thresholding is used to remove noise; additionally, two Dropout layers and two Fully-Connected layers are utilized to enhance the nonlinear fitting ability and generalization ability of the model; the adaptive parameter rectified linear unit is utilized to adapt the nonlinear transformation. Experimental results show that the proposed algorithm can effectively enhance fault diagnosis accuracy and anti-noise ability.