A New Universal Cross-Domain Bearing Fault Diagnosis Framework With Dynamic Distribution Adaptation Guided by Metric Learning

Bearing fault diagnosis is crucial for mechanical equipment as it ensures the reliable and efficient operation of machinery. Transfer learning (TL) based on maximum mean discrepancy (MMD) has been widely studied due to its strong performance and interpretability in bearing fault diagnosis using unlabeled datasets. Current research mainly aims to align the distributions between the source and target domains, improving model adaptability. However, bearings operate under complex conditions, making it difficult to distinguish fine-grained fault characteristics. In addition, substantial differences in data distribution present significant challenges to applying TL in fault diagnosis. This article presents a novel fault diagnosis method for cross-domain applications guided by metric learning to facilitate dynamic distribution adaptation. Initially, we implement a fine-grained feature extraction network (FGFEN), utilizing metric learning to improve the model’s capacity to identify specific fault characteristics. To address the indiscriminate adaptation of marginal and conditional distributions (CDs) across domains, we employ relative entropy (RE) to determine dynamic weights. It provides a quantitative assessment of different distributions’ relative importance in various tasks. Experimental results on six cross-operational fine-grained transfer fault diagnosis tasks and six cross-machine transfer fault diagnosis tasks demonstrated that the proposed fine-grained domain dynamic adaptive network (FDDaN) outperforms typical domain adaptation methods, achieving an average accuracy of 98.79% and 86.07%, respectively.