Sparsity-assisted Supervised Classification Approach with Application to Data-driven Machine Diagnostics

Data-driven machine diagnostics is vital to intelligent maintenance and operation of various modern complex industrial systems. However, classical data-driven fault diagnosis techniques depend upon manually designed features or the expert knowledge, which greatly limits their applicability in automatic and intelligent diagnostic system. In this study, a new sparsity-assisted supervised classification (SASC) approach is proposed to learn discriminative features automatically for robust fault diagnosis. Our proposed SASC approach involves two stages for dictionary design and fault recognition. First, SASC designs the structured whole dictionary in a data-driven way considering the self-similarity information of vibration data within and across multiple health states. Second, SASC implements fault recognition using a sparse classification strategy based on the minimal reconstruction errors. Compared to classical data-driven fault diagnosis methods, SASC can get rid of the manual feature design and selection, enabling automatic and robust fault diagnosis. The SASC approach has been verified with gearbox datasets in the 2009 IEEE PHM data challenge. Besides, comparison results with several traditional sparse representation-based classification approaches prove the superiority of the SASC approach for data-driven machine diagnostics.