Endogenous Noise-Expanded Multivariate Empirical Mode Decomposition and Its Application to Mechanical Compound Fault Diagnosis

Abstract

Mechanical compound fault diagnosis is a great challenge of current multivariate signal parallel processing methods. In response, endogenous noise-expanded multivariate empirical mode decomposition (ENMEMD) is proposed. First, cyclic attractor tensor construction by the phase space reconstruction is designed to maximize feature information saturation degree in high-order space. Second, a multivariate noise synchronous estimation strategy is established to synchronously estimate the multivariate inherent noises from original signals by high-order singular value decomposition (HOSVD) with the dispersion entropy (DE) technique. Third, an endogenous noise-expanded model is proposed for the utilization of the estimated multivariate noises to improve the input data model of multivariate empirical mode decomposition (MEMD). The model enhances the estimation accuracy of a multivariate envelope mean, which achieves the purpose of reducing mode mixing, frequency scale alignment of multivariate intrinsic mode functions (IMFs), and denoising. Eventually, all multivariate kernel IMFs are output to achieve complete fault feature extraction. The effectiveness and feasibility of ENMEMD are verified by repeatable simulations and engineering cases with various comparison methods. Particularly, tensor construction methods with different information saturation for final feature extraction effects are discussed by simulations. The results show that ENMEMD is an effective method for mechanical compound fault diagnosis.