A hierarchical adversarial multi-target domain adaptation for gear fault diagnosis under variable working condition based on raw acoustic signal

Acoustic-based diagnosis (ABD) relied on air-coupled measurement is received great attention in recent years due to its ability in overcoming the limitation of vibration-based diagnosis on contact measurement. However, most of the ABD approaches are only focus on a stable working condition for analysis and research. Main challenge of ABD task is the unstable working condition including variable load and speed in real-industrial scenarios, which involve variation in amplitude and phase of acoustic signal to cause domain shift problem for fault pattern detection. In this condition, to perform ABD task for meeting the requirement of scalability in dealing with multiple targets under continuously variable working condition, a novel hierarchical adversarial multiple-target domains adaptation (HAMT-DA) learning framework is proposed in this paper. Following the naturally solution idea, a target-compress adversarial adaptation (TCAA) mechanism is developed in the first-level of framework to compress multiple-target domains by minimizing the feature distribution discrepancy over all domains, and narrow the divergence between source and target domains concurrently. Then, domain invariant feature refining (DIFR) strategy is designed to take responsibility for second-level domain adaptation in the intermediate of framework by refining domain invariant attribute and encoding the domain invariant feature into high-level semantic representation space. Subsequently, a class-separate adversarial adaptation (CSAA) method is explored in the roof of the hierarchical framework for the exploration of fault-discriminative property by forming third-level domain adaptation. Thus, the entire HAMT-DA learning framework can be obtained for single source to multiple-target (SSMT) cross-domain ABD task. The experiment result in variable working conditions indicate that the proposed method has more than 94% average diagnosis performance for gear fault patterns in five cross-domain validation tasks and achieve the highest recognition accuracy among four verifications, which is about 7% average improvement than that of the optimal comparison method. It verifies the superiority of our method on robustness and generalization for acoustic-based gear fault diagnosis under SSMT setting.