Reinforced open-set domain adaptation with a trainable generalized extreme value model for fault diagnosis

Accurate fault diagnosis under varying operating conditions is crucial for reliable and safe maintenance of rotating machinery. In practical industrial scenarios, diagnosis models are often required to transfer knowledge from a labeled source domain to an unlabeled target domain with domain shifts, while previously unseen fault modes may appear. This cross-domain open-set fault diagnosis (COSFD) problem is challenging since conventional closed-set models tend to misclassify unknown target samples into known source classes. Existing COSFD methods still face two limitations. First, extreme value theory-based domain adaptation (DA) methods typically rely on numerical parameter estimation, making them difficult to update the extreme value model together with the evolving deep feature distribution during network training. Second, transferability-based sample weighting strategies may produce suboptimal sample-selection decisions. To address these challenges, this paper proposes a reinforced open-set DA (ROSDA) method. Specifically, a parameter-trainable generalized extreme value (GEV) model is integrated into the DA framework to estimate open-set probabilities, with its constrained parameters optimized by gradient descent. A reinforcement learning-based sample selector is constructed to progressively remove target unknown-class samples through long-term cumulative rewards, thereby improving sample-selection reliability and alleviating negative transfer. The retained target samples are then incorporated into adversarial DA to enhance shared-class alignment while preserving unknown-class separability. Experimental results demonstrate that the proposed method outperforms state-of-the-art DA methods.