A novel approach for adaptively separating and extracting compound fault features of the in-wheel motor bearing

For compound fault detection of in-wheel motor bearings, this paper proposes a novel approach to adaptively separate multi-source signals and extract compound fault features. Building upon blind source separation (BSS), this approach integrates blind deconvolution to address the challenge of extracting weak features. To resolve the undetermined condition of BSS and enhance feature expression, an adaptive signal reconstruction strategy based on local mean decomposition is proposed. Non-negative matrix factorization, a commonly used BSS method, is refined to suit practical applications by adopting the Itakura-Saito distance and the sparse constraint. Then, fault source signals are adaptively identified based on the proposed envelope spectrum peak factor. By introducing a new waveform extension strategy to effectively reduce the endpoint effect, multipoint optimal minimum entropy deconvolution adjusted is improved and used to enhance and extract weak features. Simulation and experimental results validate the effectiveness and robustness of the proposed approach across various stable working conditions and different types of compound faults.