Graph modeling of singular values for early fault detection and diagnosis of rolling element bearings

Early fault detection and diagnosis plays an important role in reducing maintenance cost and ensuring reliability of rolling element bearings (REBs). Singular value decomposition (SVD) is considered as a promising method to achieve this end, but lacks of consideration of inter-correlation between resulting singular values leading to the loss of weak fault information hidden in specific components. This paper, motivated by recent advances in graph modeling of highly noisy vibration signals, presents a novel method, called graph-modeled singular values (GMSVs), that integrates graph theory and SVD with the purpose of inspection of dynamic REB health conditions. The method utilizes the singular values as inputs to construct the graph, as such it achieves a balance between sensitivity to early fault and robustness to noise; meanwhile, it brings a more powerful ability of fault discrimination. Taking merits of GMSVs, a common null hypothesis testing is performed to inspect whether a fault occurs or not during REB successive operations; the KNN classifier is used to identify the fault type. Experiments are conducted on two publicly-available data sets: XJTU-SY data set and CWRU data set. Comprehensive experimental results along with comparison of those state-of-the-arts demonstrate the priority and great potential of the method in real applications. (C) 2020 Elsevier Ltd. All rights reserved.