Structured Prediction in Latent Subspace for Unsupervised Fault Diagnosis With Small and Imbalanced Data

Fault diagnosis techniques, through the collection and analysis of fault information, effectively prevent frequent equipment failures. The widespread adoption of sensors facilitates data-driven diagnosis using extensive device data. However, these data are often unlabeled, and fault data are scarce, which limits the generalization capability of data-driven models. To overcome this, we utilize fault synthesis to generate ample training data. Despite the discrepancies between the synthetic and real faults, unsupervised domain adaptation (UDA) offers a viable solution. Existing research primarily focuses on global adaptation but inadequately addresses the internal structures of data, posing challenges in imbalanced fault scenarios. This article introduces an innovative integration of domain adversarial training with structured prediction pseudolabeling, employing local preserving projection (LPP) to bridge the gap between the synthetic and real fault data. Our proposed framework, validated using a synthetic dataset derived from the Case Western Reserve University (CWRU) bearing dataset, uses healthy samples as a foundation and incorporates expert knowledge to create a balanced fault dataset, ensuring its reliability. The framework achieves identification accuracies of 80.52% on balanced data and 83.94% on imbalanced data, surpassing the current state-of-the-art by 5.71%. The effectiveness of structured prediction pseudolabels is further confirmed through ablation experiments.