A Multi-Scale Invariant Feature Alignment Framework for Domain Generalization in Rotating Machinery Fault Diagnosis

Variations in operational conditions can cause distribution shifts between training data and real-world field data, leading to degraded model performance. Existing Domain Generalization (DG) methods typically focus on aligning global features, often neglecting intra-class feature aggregation. As a result, samples belonging to the same class tend to scatter within the feature space, while inter-class samples may be incorrectly aligned, ultimately reducing classification accuracy. To address these challenges, this paper proposes a Multi-scale Invariant Feature Alignment (MIFA) framework. At both the feature-space and decision-space scales, MIFA explicitly constrains feature vectors and decision vectors of intra-class samples from source domains to cluster around their corresponding class centers. This promotes effective intra-class aggregation and enhances interclass separability, significantly improving the generalization of the model under unseen operating conditions.