A joint discriminative domain adaptation fault diagnosis method for addressing extreme small-sample and imbalance issues in marine centrifugal pump bearings

Abstract The bearing is a key component of marine pump systems, and its condition is critical to vessel safety. Harsh engine room environments degrade sensor performance, causing signal loss with class-specific discrepancies, leading to insufficient samples and data imbalance, and making extreme small-sample problems more likely under high-speed, full-load conditions. To address the challenges of extreme small samples and imbalanced data, this paper proposes a Joint Discriminative Domain Adaptation Network (JDDAN). For the small-sample problem, a transfer learning approach is adopted. To alleviate sparse feature distribution in extreme small-sample scenarios, a Higher-order Correlation Alignment method (HO_CORAL) is proposed. This method combines Higher-order Moment Matching (HOMM) with Correlation Alignment (CORAL) and aligns cross-domain features using the first- to fourth-order moments (mean, covariance, skewness, and kurtosis), ensuring that multi-class fault features maintain clear discriminative properties even with limited samples.To address the problem of data imbalance, this paper proposes a Higher-order Conditional Domain Adaptation with Sample Weighting method(HO_CDA_S). Built upon traditional Conditional Domain Adaptation (CDA), the method introduces a weighting mechanism that dynamically adjusts alignment weights for each subdomain according to the number of samples in each class, preventing the model from overemphasizing large-sample subdomains while neglecting small-sample ones, thereby reducing the risk of overfitting. Meanwhile, HO_CDA_S employs HO_CORAL to perform higher-order moment feature alignment for each subdomain, enabling more precise cross-domain feature adjustment. To comprehensively validate the effectiveness of JDDAN, experiments were conducted in three typical industrial scenarios–balanced, imbalanced, and extreme small-sample–to evaluate its performance across different data distributions.