Continuously Indexed Domain Generalization for Fault Diagnosis Under Continuously Varying Working Conditions

Deep learning-based fault diagnosis often suffers from significant accuracy degradation under varying working conditions. While domain generalization (DG) methods can ensure consistency in fault prediction when the unavailable test data is out of domain, existing research typically involves discretizing continuous operating conditions, which fails to fully utilize the continuity of common factors such as speed and load. This oversight limits improvements in generalization capability. To address this, we propose a fault diagnosis framework suitable for continuously varying working conditions, utilizing continuously indexed DG methods. This framework builds on wavelet transform and convolutional neural networks, incorporating continuous domain mix (CDMix) data augmentation and continuous mutual information minimization (CMIM) loss constraints to achieve cross-DG. Specifically, CDMix determines mixing probabilities by measuring the distance between domain labels of samples, enhancing the reliability of the generated out-of-domain data. Meanwhile, CMIM estimates the mutual information between features and continuous domain labels using kernel methods and then minimizes it to guide the network in extracting fault features that are invariant to continuous working conditions. The proposed methods improve the generalization capability and cross-domain fault classification accuracy of fault diagnosis models. Extensive experiments on gearbox and bearing fault datasets validate the effectiveness of the proposed CDMix–CMIM framework, demonstrating significant superiority over existing methods and good robustness.