Enhancing gear pitting fault classification accuracy through improved multiscale permutation entropy-driven thresholding and NA-MEMD

This work presents a statistically validated and data-driven framework for early fault diagnosis of multistage industrial gearboxes, with particular emphasis on pitting defect detection using multidimensional vibration measurements acquired under varying operating conditions (0–30 Nm loads and 1000–1500 RPM speeds). Noise-Assisted Multivariate Empirical Mode Decomposition (NA-MEMD) is employed for adaptive, non-stationary signal decomposition to obtain a set of Intrinsic Mode Functions (IMFs). To improve decomposition robustness and mitigate mode mixing, the influence of multiple white Gaussian noise (WGN) channel configurations (1–3 channels) is systematically investigated. Subsequently, the primary contribution of this study involves the quantitative estimation of an optimal IMF selection criterion. Five distinct thresholding strategies are comparatively evaluated, and the most effective approach is determined through maximization of the Signal-to-Noise Ratio (SNR), enabling retention of fault-sensitive components while suppressing noise-dominant IMFs. The resulting effective IMFs are further analyzed in both time and frequency domains, where Fast Fourier Transform (FFT) facilitates identification of characteristic fault frequencies and spectral signatures. Statistical descriptors, including root mean square (RMS), peak value, and standard deviation, are extracted to construct discriminative feature vectors. Classification performance is comparatively assessed using Decision Tree (DT), Weighted k-Nearest Neighbors (W-kNN), and Tri-layered Neural Network (TNN) models. Finally, Analysis of Variance (ANOVA) is conducted to determine the optimal noise configuration, thresholding strategy, feature set, and classifier. Experimental results demonstrate enhanced SNR and superior diagnostic accuracy, confirming the effectiveness of the proposed methodology for reliable gearbox condition monitoring and predictive maintenance.