Robust feature selection for cross-domain fault diagnosis using deep reinforcement learning

Recent research on deep learning-based fault diagnosis has made significant progress. However, over-parameterized models often introduce redundant features, leading to degraded cross-domain generalization and limited deployability on resource-constrained devices. To address these challenges, this paper proposes a reinforcement-learning-based robust feature selection (RFS) framework for cross-domain fault diagnosis. Unlike existing methods that treat network pruning and domain adaptation as separate processes, the proposed approach integrates them at the decision level via a domain-aware reward formulation. Specifically, channel pruning is formulated as a sequential decision-making problem, where a deep reinforcement learning (DRL) agent learns an adaptive pruning policy. Meanwhile, domain alignment loss is incorporated into the reward function, enabling the pruning policy to directly optimize cross-domain feature robustness during the decision-making process rather than through post-hoc feature alignment. This mechanism promotes the retention of features that are both discriminative and transferable, while eliminating redundant or domain-specific components. This framework offers a principled approach for jointly optimizing model compactness and cross-domain generalization, rather than treating them as independent objectives. Extensive experiments on two gearbox datasets under various domain shift scenarios demonstrate that the proposed framework consistently improves cross-domain diagnostic performance while reducing model complexity. In particular, the method achieves up to 19.59 % improvement in target-domain accuracy and reduces computational cost by up to 73.87 % in FLOPs, with reduced inference latency. These results demonstrate the effectiveness of the proposed approach in jointly achieving robust cross-domain generalization and efficient model deployment in practical industrial scenarios.