Bearing diagnosis for unknown working conditions: mixed enhancement and adaptive triplet learning

To address the issues of generalization and robustness in bearing fault diagnosis models under unknown working conditions, this article proposes a dual-stage adaptive mixed domain generalization network as a novel domain generalization approach. This method employs a dual-stage domain mixed enhancement strategy at both the primary-level and feature-level to generate diverse intermediate domain, aiming to efficiently minimize the distributional differences between the source domain and the target domain. At the primary-level, a convex combination of multisource domain raw vibration signals is employed to construct virtual domain samples with smooth transition features. At the feature-level, a more discriminative feature representation space is constructed by performing linear interpolation on the extracted domain-invariant features. Meanwhile, an adaptive triplet loss function is designed to optimize the relative distance relationships between sample pairs, thereby guiding the feature space to automatically form a discriminative geometric structure. Experiments on two bearing datasets show that the proposed method achieves diagnostic accuracies of 92.80 and 94.39%, respectively, demonstrating its cross domain generalization performance and robustness.