Abstract
To tackle the difficulties in cross-domain fault diagnosis for rolling bearings, researchers have devised numerous domain adaptation strategies to align feature distributions across varied domains. Nevertheless, current approaches tend to be vulnerable to noise disruptions and often neglect the distinctions between marginal and conditional distributions during feature transfer. To resolve these shortcomings, this study presents an innovative fault diagnosis technique for cross-domain applications, leveraging the Attention-Enhanced Residual Network (AttenResNet18). This approach utilizes a one-dimensional attention mechanism to dynamically assign importance to each position within the input sequence, thereby capturing long-range dependencies and essential features, which reduces vulnerability to noise and enhances feature representation. Furthermore, we propose a Dynamic Balance Distribution Adaptation (DBDA) mechanism, which develops an MMD-CORAL Fusion Metric (MCFM) by combining CORrelation ALignment (CORAL) with Maximum Mean Discrepancy (MMD). Moreover, an adaptive factor is employed to dynamically regulate the balance between marginal and conditional distributions, improving adaptability to new and untested tasks. Experimental validation demonstrates that AttenResNet18 achieves an average accuracy of 99.89% on two rolling bearing datasets, representing a significant improvement in fault detection precision over existing methods.