AIIRL Enables Label-Free Industrial Fault Detection

Machinery fault detection (MFD) traditionally relies on supervised learning, which struggles with the common scarcity of fault labels in industrial settings. While reinforcement learning (RL) offers a suitable framework for modeling sequential degradation, many "RL-based" MFD methods simplify the problem into a static contextual bandit, ignoring crucial state transitions and temporal factors. This paper introduces an Adversarial Inverse Reinforcement Learning (AIRL) framework that re-frames MFD as an offline Inverse RL problem. Unlike reconstruction-based methods that depend on static error margins or contextual bandits that overlook dynamics, AIRL directly infers an intrinsic "health" reward from observed state transitions. This innovative approach removes the need for manual reward engineering and, critically, for explicit fault labels. Evaluated on three run-to-failure benchmarks (HUMS2023, IMS, XJTU-SY), AIRL demonstrated superior performance, being the only method to achieve non-saturated post-detection consistency across all datasets. In contrast, contextual bandit baselines failed to detect gradual degradation, and reconstruction models often collapsed into perpetually anomalous states. The code and data for this research are publicly available.