Abstract
Bearings are critical components whose failures in industrial machinery can lead to catastrophic breakdowns and costly downtime; yet, accurate early-stage diagnosis remains challenging due to the non-stationary, nonlinear nature of vibration signals and noise interference. This study proposes a multidataset-integrated bearing fault diagnosis methodology incorporating variational mode decomposition (VMD), refined composite multiscale weighted permutation entropy (RCMWPE) feature extraction, and whale optimization algorithm (WOA)-optimized support vector machine (SVM). Addressing the non-stationary and nonlinear characteristics of bearing vibration signals, raw signals are first decomposed via VMD to effectively separate intrinsic mode functions (IMFs) carrying distinct frequency components. Subsequently, RCMWPE features are extracted from each IMF component to construct high-dimensional feature vectors. To address visualization challenges and mitigate feature redundancy, the t-distributed stochastic neighbor embedding (t-SNE) algorithm is employed for dimensionality reduction. Finally, WOA optimizes critical SVM parameters to establish an efficient fault classification model. The methodology is validated on two public bearing datasets: PRONOSTIA and CWRU. For four-class fault diagnosis on the PRONOSTIA dataset, the model achieves 96.5% accuracy. Extended to ten-class diagnosis on the CWRU dataset, accuracy reaches 99.67%. Experimental results demonstrate that the proposed method exhibits exceptional fault identification capability, robustness, and generalization performance across diverse datasets and complex fault modes. This approach offers an effective technical pathway for early bearing fault warning and maintenance decision making.