Imbalanced fault diagnosis of rotating machinery via multi-domain feature extraction and cost-sensitive learning

Fault diagnosis plays an essential role in rotating machinery manufacturing systems to reduce their maintenance costs. How to improve diagnosis accuracy remains an open issue. To this end, we develop a novel framework through combined use of multi-domain vibration feature extraction, feature selection and cost-sensitive learning method. First, we extract time-domain, frequency-domain, and time-frequency-domain features to make full use of vibration signals. Second, a feature selection technique is employed to obtain a feature subset with good generalization properties, by simultaneously measuring the relevance and redundancy of features. Third, a cost-sensitive learning method is designed for a classifier to effectively learn the discriminating boundaries, with an extremely imbalanced distribution of fault instances. For illustration, a real-world dataset of rotating machinery collected from an oil refinery in China is utilized. The extensive experiments have demonstrated that our multi-domain feature extraction and feature selection can significantly improve the diagnosis accuracy. Meanwhile, our cost-sensitive learning method consistently outperforms the traditional classifiers such as support vector machine (SVM), gradient boosting decision tree (GBDT), etc., and even better than the classification method calibrated by six popular imbalanced data resampling algorithms, such as the Synthetic Minority Over-sampling Technique (SMOTE) and the Adaptive Synthetic sampling method (ADASYN), in terms of decreasing missed alarms and reducing the average cost. Owing to its high evaluation scores and low average misclassification cost, cost-sensitive GBDT (CS-GBDT) is preferred for imbalanced fault diagnosis in practice.