Label-free Fault Detection via Adversarial Inverse Reinforcement Learning

A novel approach utilizing adversarial inverse reinforcement learning (AIRL) for detecting machinery faults eliminates the necessity for fault labels. In contrast to supervised learning, which faces challenges due to limited labels, and existing reinforcement learning techniques that revert to static classification, AIRL derives an inherent "health" reward from observed state transitions. This method frames fault detection as an offline inverse reinforcement learning challenge, negating the need for manual reward crafting or fault labels. AIRL stands out as the sole technique that consistently performs well across three run-to-failure benchmarks (HUMS2023, IMS, XJTU-SY). The research has been published on arXiv with ID 2607.22987.