An adaptive Gaussian-guided feature alignment network for cross-condition and cross-machine fault diagnosis of rolling bearings

In recent years, transfer learning-based fault diagnosis methods have achieved significant performance. However, most studies directly map the source and target features into the same space, posing substantial challenges when handling complex and diverse data in practical engineering. In contrast, indirect feature alignment methods simplify this process with intermediate distributions, but still suffer from negative transfer and rely on simplified distribution discrepancy measurement. To address these defects, an adaptive Gaussian-guided feature alignment network (AGFAN) is proposed for cross-condition and cross-machine fault diagnosis of rolling bearings. Specifically, to effectively reduce domain distribution discrepancy, a dynamic Gaussian-guided alignment method is developed to adaptively facilitate indirect feature alignment between the source and target features. Furthermore, to ensure the accuracy of distribution discrepancy measurement, a probability transformation estimation method is proposed for enhancing the feature alignment capability of the AGFAN. Fault diagnosis tasks on three bearing datasets are designed to validate the effectiveness of the proposed AGFAN. The research results indicate that the AGFAN outperforms related methods.