Fault diagnosis method for imbalanced and unlabeled data based on bayesian graph balanced learning

Abstract

Abstract In fault diagnosis, it is crucial to address the combined challenges of imbalanced sample sizes and unlabeled data. Traditional methods often generate pseudo-samples or pseudo-labels. These can lead to inaccurate diagnostic outcomes if they are not representative of the original data. To address these challenges, this paper proposes an innovative fault diagnosis method based on bayesian graph balanced learning (BGBL). Firstly, a balancing strategy was developed to tackle sample imbalance by assigning and optimizing weights for samples in imbalanced categories. Graph theory techniques were then used on unlabeled data to establish and update category beliefs. Following this, posterior estimates of samples were derived within the bayesian neural networks framework. This led to the training of a fault diagnosis model. Finally, fault diagnosis was conducted using this trained model. Three sets of experiments were conducted on the planetary gearbox fault dataset. The results showed that the proposed BGBL method significantly improved the accuracy of fault diagnosis. Specifically, under conditions of imbalanced data and missing labels, the BGBL method increased the accuracy by over 26% compared to existing methods. This demonstrates its effectiveness in these challenging scenarios.