Adaptive Mixup-Based Domain Adaptation Method for Intelligent Fault Diagnosis

Recent years have witnessed the successful application of domain adaptive methods to tackle intelligent fault diagnosis of rotating machinery under variable working conditions. However, existing work always ignores the learning of feature discriminability when developing transferable models based on domain-invariant representation learning strategies. In addition, they have difficulty handling the knowledge transfer between domains with significant differences. To address these problems, an adaptive mixup-based adversarial network (AMAN) is proposed in this paper. It develops an inter-domain mixup method based on the sample adaptive screening strategy to generate high-quality virtual samples to guide domain adaptation while improving the learned feature representations’ discriminability. The comprehensive results of numerous DA diagnosis tasks built on the gearbox dataset validate AMAN’s effectiveness and application prospect.