MADAN: A Multi-Angle Domain Adversarial Network for Robust Cross-Condition Rolling Bearing Fault Diagnosis

Abstract Rolling bearing fault diagnosis under varying operating conditions remains challenging due to significant domain shifts in vibration‐signal distributions. To address this, we propose a Multi‐Angle Domain Adversarial Network (MADAN) that unifies multi‐angle perception and multi‐view utilization within an adversarial adaptation framework. First, a dual‐branch feature extractor captures both time‐domain and frequency‐domain representations via multi‐scale convolutions, augmented by channel and temporal attention, and fuses them into a concise 512-dimensional embedding. Second, bidirectionally complementary discriminators impose “source vs. non-source” and “target vs. non-target” adversarial tasks, yielding finer‐grained domain confusion. Third, a structurally complementary dual-head classifier—comprising a locally robust, high-dropout head and a globally oriented, low-dropout head—provides diversified decision boundaries, further regularized by an inter-head consistency loss. Extensive experiments on the PU and MCDSP bearing datasets demonstrate that MADAN consistently outperforms other models, achieving superior transferability and classification accuracy across diverse operating scenarios.