Integrating multiscale entropy feature extraction with lightweight classifiers for robust bearing fault diagnosis in industrial systems
Year: 2026
Type: article
Abstract: Context—Rolling-element bearings are among the most critical components in industrial machinery, and their failures may result in unexpected downtime, economic losses, and safety risks. As Industry 4.0 applications continue to expand, there is an increasing demand for reliable, computationally efficient, and real-time fault diagnosis systems capable of operating under varying industrial conditions.Objective—This study aims to develop a robust and lightweight bearing fault diagnosis framework by integrating multiscale entropy-based feature extraction methods with conventional machine learning classifiers. The objective is to systematically evaluate the effectiveness, robustness, and parameter sensitivity of different entropy measures for accurate bearing fault classification under diverse operating conditions while maintaining suitability for edge-computing and real-time industrial monitoring applications.Method—The proposed framework combines three multiscale entropy methods—Multiscale Sample Entropy, Multiscale Fuzzy Entropy, and Multiscale Permutation Entropy—with six lightweight classification algorithms: k-Nearest Neighbors, Support Vector Machine, Logistic Regression, Backpropagation Neural Network, Extreme Learning Machine, and Softmax Regression. Experimental evaluations were conducted using two benchmark datasets: the Case Western Reserve University (CWRU) drive-end bearing dataset and the CISAR three-axis industrial vibration dataset collected under realistic conditions. Hyperparameter selection was performed through stratified 10-fold cross-validation to ensure fair model comparison and prevent overfitting. In addition, sensitivity analysis was conducted by varying the entropy scale factor between 6 and 18 to investigate its effect on classification performance and feature stability.Results—Experimental findings demonstrate consistently high classification performance across different fault types, fault diameters, and operating conditions. Among the investigated feature extraction methods, MFE showed comparatively high classification accuracy and stable performance across the experimental settings, frequently exceeding 99% classification accuracy across multiple classifier combinations. Several classifier-feature pairs achieved perfect classification performance on the CWRU dataset. The sensitivity analysis revealed that the optimal scale-factor range lies between τ = 12 and 14. Compared with MSE and MPE, MFE showed descriptively lower sensitivity to scale-factor variation, indicating stronger robustness against operational variability and signal noise.Conclusion—The proposed framework provides an accurate and lightweight approach for bearing fault diagnosis, with potential applicability to resource-constrained monitoring environments. The integration of multiscale entropy features with lightweight classifiers enables reliable predictive maintenance under varying industrial conditions. The results highlight the empirically higher performance of MFE-based feature extraction and provide practical guidance for developing robust Industry 4.0 fault monitoring systems.
Language: Turkish
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7213332338