The main challenges for partial set cross-domain fault diagnosis problems, where the target label space is only a subset of the source label space, are to facilitate positive transfer between shared classes and avoid negative transfer caused by unrelated classes. To address the above challenges, a prototype-driven class-wise adversarial transfer network is proposed in this study. First, aiming to enhance the classification robustness, a fault prototype-based discrimination method without learnable parameters is designed to replace the traditional classifier for health state recognition. Then, based on the intrinsic similarity between the target samples and the fault prototypes, a novel prototype similarity based weighting mechanism is proposed to precisely measure the transferability of each source class, thus decreasing the contribution of unrelated source class samples. Finally, the proposed class-wise adversarial adaptation framework facilitates fine-grained knowledge transfer between shared classes and enhances domain adaptation performance. The experimental results show that the proposed method outperforms all the comparison methods, achieving over 10% improvement in average diagnostic accuracy on the two rolling bearing datasets and maintaining over 90% overall diagnostic accuracy.
30 March 2026 in IEEE Transactions on Instrumentation and Measurement
Under strong noise disturbance, local signals collected by a single sensor often fail to fully capture fault characteristics. In addition, diagnostic models commonly suffer from limited robustness and generalization under complex operating conditions. To address these problems, a multi-sensor information fusion fault diagnosis method is proposed in this paper. First, a time–frequency dual-branch feature extraction module is designed. It performs one-dimensional convolution on both the original time-domain signal and its frequency-domain amplitude. Then, a selectable convolution module is added to each branch. It dynamically fuses multi-scale information from different receptive fields. A bidirectional cross-attention mechanism enables deep interaction and complementary enhancement between time–frequency features. Finally, experiments are conducted based on the two datasets obtained from the metro bogie transmission system fault simulation test rig and the radial load bearing fault simulation test rig. Results show that the proposed method achieves high recognition accuracy and strong robustness under variable operating conditions and intense noise, which achieves promising performance over the comparison methods.