Improved Depthwise Separable Convolution for Transfer Learning in Fault Diagnosis

Convolution neural network (CNN)-based transfer learning (TL) has been investigated for knowledge transfer in machinery fault diagnosis. However, there remain certain challenges that need to be addressed: first, the optimization of deep network models is hindered by the substantial number of parameters involved; and second, the limitation of training data impacts the diagnostic accuracy of the model. In this work, a multilevel residual CNN based on dynamic feature fusion (MRCNN-DFF) is proposed for TL in machinery fault diagnosis. In MRCNN-DFF, depthwise separable (DS) convolution is introduced to reduce the number of trainable parameters, which offers the advantages and enhances the efficiency of model optimization. Meanwhile, to fully mine the useful information, parallel MRCNN channels are proposed to extract high-dimensional features. Then, the DFF is designed to fuse the bichannel information according to the feature importance, which effectively incorporates the key information contained in the limited data. Finally, the parameter transfer is utilized to transfer domain-shared information. The MRCNN-DFF is validated employing two bearing datasets, and the results demonstrate that the proposed MRCNN-DFF surpasses the comparison methods in terms of performance.