Dictionary Learning-based Intelligent Recognition Method Towards Machinery Fault Diagnostics

Intelligent fault diagnostic techniques are essential to ensure the operational safety and reduce maintenance costs for complex mechanical systems, enabling intelligent operation and maintenance in the golden era of smart manufacturing. This paper presents a novel dictionary learning-based intelligent recognition (DL-IR) methodology towards machinery fault diagnostics. DL-IR learns representative dictionaries for sparse representation of vibration data under various health states and implements health state identification via a minimal sparse approximation error criterion. In detail, we apply an overlapping segmentation strategy at first to implement data augmentation for training and testing dataset preparation. Then, K-singular value decomposition is exploited to learn class-oriented dictionaries associated with distinct health states from training dataset. These learned class- oriented dictionaries can exhibit strong representation ability and share the consistent class attribute for test datasets. Finally, the health state identification for test samples is implemented via sparse approximations with respect to class-oriented dictionaries and a minimum reconstruction error criterion. Through the comprehensive experiment validation and comparison studies on public gearbox fault datasets, it is shown that our DL-IR method can obtain superior diagnostic accuracy of 99.9% and outperform three well-known sparse representation classification approaches for mechanical fault diagnostics.