Entropy-weighted manifold-adjusted transfer learning for cross-condition fault diagnosis with imbalanced and missing labels

Abstract

In industrial fault diagnosis, data imbalance, missing labels, and cross-condition scenarios increase complexity and challenges. While existing research has made progress in these areas, gaps remain in addressing cross-condition fault diagnosis with imbalanced and incomplete labels. To tackle this, we propose the entropy-weighted manifold alignment (E-WMA) method. First, we use sparse filtering techniques for manifold alignment to extract features with good separability from the source and target domains. Next, we adopt an entropy-weighted maximum mean discrepancy strategy to dynamically adjust sample weights based on label information entropy, which reduces distribution differences and mitigates data imbalance. Finally, we build a softmax regression classifier to train and evaluate the fault diagnosis model using the aligned features and adjusted sample weights, enhancing diagnostic accuracy and robustness. Extensive experiments on wind turbine planetary gearbox and bearing fault datasets validate our approach. The results show that our method effectively addresses cross-condition fault diagnosis amid data imbalance and missing labels. This can lead to more accurate fault detection in real-time operations, minimize unplanned downtime, and significantly reduce maintenance costs in industrial environments.