A Deep Transferable Convolution Neural Network for Unsupervised Cross-Domain Diagnostics of Rolling Bearings in Industrial Internet of Things

Abstract

In recent years, deep learning-based intelligent diagnosis approaches have showcased broad application prospects in the field of rolling bearing fault identification in Industrial Internet of Things (IIoT). Notwithstanding, the implementation effect of most existing deep diagnostic architectures can be severely restrained when there are distinct distribution discrepancies between the source and target domain, which could mainly be attributed to dissimilar working conditions, various fault severity levels, different machine structures, nonidentical complex noise interferences and so on. Aiming at ameliorating the effectiveness of unsupervised cross-domain fault diagnosis of rolling bearings, a deep transferable convolution neural network (DTCNN) based upon cooperative domain alignment (CDA) and partial-contractive layers is proposed in this article. On the one hand, the CDA in DTCNN features the conjoint match of both marginal and conditional distributions of the first and second order statistical characteristics of data to enhance domain adaptation. On the other hand, the partial-contractive layers can prompt DTCNN to learn internal factors and capture more robust domain-invariant features through penalizing the Frobenius norm of the Jacobian matrix of hidden activations with respect to the inputs. Furthermore, a modified network learning strategy is leveraged to facilitate the efficient adjustment of weights and biases at different training stages, which is beneficial for convergence during the network optimization process. Experiments on different transfer tasks tangibly authenticate the ascendancy of the proposed DTCNN in cross-domain diagnostics of rolling element bearings, compared with other contrastive methods. The results reveal that the proposed DTCNN model not only attains high-precision unsupervised transfer diagnosis of rolling bearings across a range of circumstances, but also exhibits superior noise resistance and computational efficiency, underscoring its substantial practical significance for IIoT applications.