A SIGNAL PROCESSING AND OPTIMIZATION-BASED APPROACH FOR BEARING FAULT DIAGNOSIS USING SVMD, SVD, AND WO-SVM

Abstract

Roller bearing acceleration vibration signals are inherently non-stationary and noisy. However, benchmark datasets commonly used in pattern recognition often contain minimal noise, whereas real-world bearing vibration signals typically include both the bearing’s inherent vibration and additional noise from related machine components. To enhance the accuracy of fault diagnosis methods when processing real-world signals, Gaussian noise was intentionally added to the original signals. Subsequently, the Successive Variational Mode Decomposition (SVMD) method was employed to decompose these signals into multiple modes. The Singular Value Decomposition (SVD) technique was then applied to construct a feature matrix from these modes, which served as input to the Support Vector Machine (SVM) classifier. The SVM hyperparameters (C and γ) were optimized using the Walrus Optimizer (WO) algorithm. Experimental analysis on signals under both normal and faulty bearing conditions demonstrates that the proposed fault diagnosis method achieves higher classification accuracy and faster convergence compared to other methods.