Parallel Convolutional Transfer Network for Bearing Fault Diagnosis Under Varying Operation States

Abstract

Rolling bearings commonly operate under varying operation states. Deep transfer learning (DFL) models are commonly used for fault diagnosis in such cases. Balancing their accuracy and efficiency is challenging because the more complex models tend to have higher accuracy, while the simpler models are more efficient. However, in industry applications, efficiency is as important as accuracy. To improve fault diagnosis accuracy and efficiency in unsupervised transfer learning (TL) scenarios, a lightweight DFL model, called parallel multiscale convolutional transfer neural network (PMCTNN), is developed in this article. PMCTNN optimizes its structure rather than increasing its depth, thus enhancing accuracy while maintaining efficiency. First, a novel parallel structure with parallel 1-D convolutional layers, a feature concatenation layer, and a 2-D pooling layer is constructed to implement feature fusion and filtering, which effectively reduces the number of extracted features. Then, a zero-padding strategy is proposed to unify the output sizes of the convolutional layers, which is necessary for the feature concatenation layer to integrate 1-D feature maps into 2-D feature maps. The combination of the parallel structure and zero-padding strategy decreases the number of trainable parameters of PMCTNN, thereby improving its efficiency. The multikernel maximum mean discrepancy (MMD) with edge distribution is employed to decrease the domain discrepancy and transfer the network. Finally, a cross-domain fault diagnosis method based on PMCTNN is proposed to diagnose bearing faults under varying operation states. The experimental results show that the PMCTNN exhibits higher diagnosis accuracy and efficiency on Case Western Reserve University (CWRU), Jiangnan University (JNU) datasets, and our actual fault dataset.