Residual joint adaptation adversarial network for intelligent transfer fault diagnosis

Although deep networks based diagnostic methods have been increasingly studied and acquired certain achievements in recent years, most of them suppose that the training and test data share similar probability distribution. The data distribution discrepancy is common and inevitable in practical industry due to the change of working conditions, equipment wear and environment interferences, which will lead to significant performance degradation of models. To address this problem, an unsupervised transfer learning framework named Residual Joint Adaptation Adversarial Network (RJAAN) is proposed in this paper for more effective intelligent fault diagnosis. In this framework, one-dimensional residual network is designed to directly process raw mechanical signal for adaptive feature learning, in which the joint maximum mean discrepancy (JMMD) and adversarial adaptation discriminator are introduced to simultaneously reduce the shifts of joint distribution and marginal distribution across different domains. Consequently, the proposed method can learn category-discriminative and domain-invariant features information for cross domain fault diagnosis. Eighteen transfer fault diagnosis tasks based on two experimental platforms, i.e. the planetary gearbox and the rolling bearing, are conducted to evaluate the effectiveness of the proposed method. Meanwhile, five popular methods are selected for comprehensive analysis and comparison. The results show that the robustness and superiority of the proposed approach under various diagnostic tasks.