Sample distribution-aware parallelepiped-based method for class imbalance fault diagnosis

Uncertainty factors are ubiquitous in various engineering practices, and the underestimation of these uncertainties limits the accuracy of traditional data-driven fault diagnosis methods. To address the above issue, this study proposes a sample distribution-aware parallelepiped-based fault diagnosis (SD-PFD) method to achieve robust fault classification for imbalanced dataset. Firstly, the influence of imbalanced data on fault diagnosis is systematically analyzed, and an uncertainty-based framework is introduced to overcome limitations in conventional deterministic approaches. By partitioning the measured signals, the data preprocessing procedure including the local similarity-based identification method and the Bayesian Gaussian mixture model is first introduced to adaptively procure representative features. Subsequently, the parallelepiped-based uncertain feature is established to capture the inherent variability in statistical features. Under the evidence theory framework, an improved volume ratio-based pattern matching strategy integrating sample distribution is developed to deliver quantitative and intuitive classification results. The proposed SD-PFD method is compared with six baselines (including two resampling-augmented methods) across CWRU and Rotor bearing datasets, achieving leading average diagnostic accuracy of approximately 98.32 % and 97.92 % under varying imbalance ratios. By quantitatively capturing the uncertainty in statistical features and innovatively eliminating the pre-training restrictions in traditional classifiers, SD-PFD proves inherently suitable for class-imbalance fault diagnosis scenarios, achieving superior robustness in noisy environments.