Effective Variable‐Speed Bearing Fault Diagnosis From Motor Current Signals Using Kurtosis‐Guided VMD and Multi‐Branch Convolutional Neural Network

ABSTRACT Bearing fault diagnosis in induction motors under variable load and speed conditions remains a challenging task due to the complexity of fault‐induced transients in current signals. This study presents a novel deep learning‐based fault classification framework utilising Variational Mode Decomposition (VMD) for adaptive feature extraction and a Multi‐branch Convolutional Neural Network (1D‐MCNN) architecture for classification. The VMD hyperparameters were optimised based on kurtosis to ensure the extraction of the most informative Intrinsic Mode Functions (IMFs), significantly enhancing feature quality. Experimental validation under fixed, variable and noisy operating conditions demonstrated the superior performance of the proposed approach. The 1D‐CNN multi‐branch model consistently outperformed conventional artificial neural network (ANN) and single‐branch convolutional neural network (CNN) architectures, achieving 99.85% accuracy in fixed‐speed conditions and 99.75% in variable‐speed operations. Moreover, t‐SNE visualisations revealed improved class separability, confirming the robustness of the extracted features. These results highlight the efficacy of VMD‐guided deep learning architectures in accurately detecting bearing faults across diverse operational scenarios, reinforcing their potential for industrial predictive maintenance.