Imbalanced fault diagnosis based on semi-supervised ensemble learning

The imbalance of fault modes prevails in industrial equipment monitoring. Many methods were presented for imbalanced fault diagnosis only by resampling labeled fault dataset, which limited the diagnostic performance due to information loss from unlabeled fault dataset. To perfectly exploit the information from unlabeled and labeled datasets, this study proposed a semi-supervised ensemble learning method termed as SSTI for imbalanced fault diagnosis. First, the sample information was evaluated based on Mahalanobis distance, and a novel sample information-based synthetic minority oversampling technique (SI-SMOTE) was presented for balancing the labeled dataset. Second, the tri-training architecture-based imbalanced co-training technique (Tri-ImCT) was developed to exploit the information contained in the unlabeled dataset. In the Tri-ImCT, rebalancing the training subsets and variable weighted voting were utilized to improve the performance of proposed method for imbalanced fault diagnosis. To verify the performance of proposed method, several experiments were carried out on several imbalanced datasets derived from two bearing datasets and one subway wheel dataset. We utilized three indicators of G-mean, average precision, and average F-score for evaluating the performance of classifiers. Experimental results show that the performance of proposed method exceeds that of other methods, which is very close to the upper bound of fully-supervised performance. It substantially indicates that this study provides a very promising methodology for imbalanced fault diagnosis.