Feature Adaptive Fusion Network With Interdomain Hybrid Weighting Mechanism: A Domain Adaptation Method for Cross-Domain Bearing Fault Diagnosis

Domain adaptation (DA) has achieved significant advancements in cross-domain bearing fault diagnosis. However, the variations in working conditions and environmental factors lead to nondiscriminative features. These nondiscriminative features not only contaminate the realistic representation of discriminative features but also cause negative transfer during the alignment of features between source and target domains. To address the problem, this article proposes a method termed the feature adaptive fusion network with interdomain hybrid weighting mechanism (FAFNIHW). First, a feature adaptive fusion network (FAFN) employs parallel convolutional branches to capture multiscale contextual features, followed by a dynamic weighting mechanism that adaptively fuses the parallel features, thereby enhancing the diversity of critical discriminative features. In addition, an interdomain hybrid weighting module (IHWM) is introduced to refine the alignment of features by considering the importance of features across both the channel and spatial dimensions; ultimately, this process prioritizes critical discriminative features while assigning lower priority to nondiscriminative ones. Comprehensive experiments on cross-machine fault diagnosis for gearbox bearings between the trolley and hoisting mechanisms of a quayside container crane (QCC) demonstrate the effectiveness and superiority of FAFNIHW compared with state-of-the-art methods.