Learning-Based Diagnosis of Multiple Faults in Bearings

Abstract

Reliable diagnostic framework must have the ability to deal with many diagnostic conditions, including the cases of multiple faults in bearings. Timely and reliable fault detection and assessment in such cases are of utmost importance for the prevention of missed detections, inadequate maintenance, and loss of profits due to failures. The problems of multiple fault diagnosis attracted relatively little attention in the literature on the background of common interest to improvements in single fault diagnosis. Multiple fault diagnosis has additional, in comparison to single fault diagnosis, challenges: submergence of the weak fault by the strong fault, overlap of vibration characteristics in time and frequency domains, coupling of frequency components and so on. To address these challenges, several solutions were proposed, including those based on artificial intelligence. However, majority of intelligent methods relied on manual feature extraction based on prior information of the faults and a new problem usually requires a new design of the feature extractor. Deep learning is a promising tool to cope with known challenges of commonly proposed intelligent methods. This paper presents a new learning-based framework to improve the efficiency of the fault diagnosis in the case of multiple faults of bearings. Deep learning integrated into the framework helps to overcome the challenges of manual feature engineering, while maintaining good diagnostic efficiency. Inputs to the classification stage are presented by versions of the dataset using generic signal processing techniques. Results from this method demonstrate promising outcomes in the detection and classification of multiple faults.