IMSF-Diagnosis: intelligent multivariate signal fusion for rolling bearing fault diagnosis

Abstract To address the limitations of shallow machine learning methods in rolling bearing fault diagnosis, particularly their heavy reliance on manual prior knowledge and sensitivity to noise, this study proposes a novel intelligent fault diagnosis method, referred to as IMSF-Diagnosis. The method synergistically integrates three modules: adaptive signal denoising and preprocessing, fault feature extraction, and fault identification and classification. Specifically, an improved artificial bee colony algorithm is developed to optimize the decomposition of multivariate vibration signals. The improved artificial bee colony algorithm incorporates an information exchange mechanism and a dynamic step size updating strategy among bee colonies, allowing for adaptive hyperparameter configurations across different decomposition stages. Furthermore, a refined composite multivariate multiscale fuzzy entropy metric is employed in the module of multivariate fault feature extraction to quantify the nonlinear dynamic characteristics embedded in the decomposed multivariate modal functions, thereby constructing a high-dimensional feature set that comprehensively reflects bearing health states. Subsequently, a supervised isometric mapping is applied for manifold dimensionality reduction to extract low-dimensional representations of these features, which are then fed into an elaborate support vector machine classifier for fault diagnosis. Experimental validations using rolling bearing datasets demonstrate that the proposed IMSF-Diagnosis effectively integrates fault related information from multivariate vibration signals, exhibiting superior diagnostic performance with enhanced robustness and noise resistance. This study establishes a new paradigm for multivariate signal physics information guided measurement in rolling bearing fault diagnosis under noisy environments.