Fast nonlinear cross-sparse filtering for rolling bearings compound fault diagnosis

Abstract

The investigation of faults in rotating machinery has been thoroughly examined. Among the different methods under exploration, sparse optimization-based techniques have arisen as a highly desirable approach. However, in real industrial environments, the collected bearing signals often contain a random impact component resulting from changes in working conditions and load mutations. When a machine malfunctions, it can readily induce and generate new faults, resulting in composite faults. To address this challenge, this paper proposes a novel multidimensional blind deconvolution method named fast nonlinear cross-sparse filtering (FNCr-SF). The FNCr-SF aims to separate weak compound faults under random impact interference. Various preprocessing techniques, including Z-score normalization and nonlinear sigmoid activation function, are employed to amplify the faint characteristics of compound faults and minimize the influence of random interference. Furthermore, the FNCr-SF method enables adaptive decomposition of fault components without the need for prior knowledge or pre-processing. This approach effectively reduces random interference and accurately detects compound faults in bearings. Experimental and simulation signals validate the effectiveness of the FNCr-SF method in compound fault detection, demonstrating its high accuracy and robustness.