Joint Distribution Alignment via Mutual Information for Cross-Device Fault Diagnosis

Abstract

Current data-driven fault diagnosis methods suffer from poor transferability. It is challenging to apply a model effective on one device directly to another. Many methods now employ domain adaptation algorithms to align their fault distributions for model transferability. However, most methods focus only on aligning either marginal or pseudo-labels-based conditional distributions, ignoring cases where both label and conditional distributions change, along with the unreliable nature of pseudo-labels. This oversight can lead to transfer failures. To tackle this, this article introduces an information theory-based joint distribution alignment model. The algorithm starts by maximizing mutual information between predicted categories and input samples for conditional alignment without pseudo-label involvement. Simultaneously, the model introduces virtual adversarial training with a penalty term to improve the robustness of prediction results. When label distribution changes, the model uses entropy values to assign data in categories unique to the target domain to “outliers,” thus preventing misalignment of these data. In experiments, this algorithm outperformed other domain adaptation-based methods.