A hybrid multi-domain feature ranking framework for gear fault severity classification using vibration signal analysis and machine learning approach

Gear fault diagnosis is a vital part of condition monitoring and predictive maintenance for rotating machinery, particularly in automotive applications where early detection of fault severity can avoid unexpected failures and cut down maintenance costs. However, the accurate classification of progressive fault levels is still challenging, due to the non-stationary nature of vibrations, the overlapping of fault characteristics, and the redundancy of features. The existing diagnosis methods usually adopt single feature extraction or single feature selection methods, which are unable to adequately describe the complex dynamics of gear degradation. Moreover, many papers focus on the identification of the presence of a fault rather than its severity level, which limits their use for decision making during maintenance. In this study, a hybrid multi-domain feature ranking framework for classifying the severity of gear fault using vibration signal analysis and Machine Learning (ML) to overcome these challenges is proposed. A total of 50 feature representations were extracted from the time domain, frequency domain, wavelet domain, Hilbert envelope, and nonlinear signals. A hybrid ranking strategy was used to determine the most informative features, which combined Random Forest importance, mutual information, principal component analysis, and Recursive feature elimination. The experimental results on four gear health conditions (healthy, 25%, 50%, and 100% fault) showed that the proposed framework effectively reduces the feature dimensionality with high diagnostic performance. The highest classification accuracy was achieved by Random Forest (99.88%) followed by K-Nearest Neighbors (99.07%). The analysis also demonstrated that the frequency domain features, especially Spectral Spread and Spectral Centroid have the best diagnostic capability to classify the severity of the fault. The proposed framework provides an interpretable and computationally efficient solution for intelligent gearbox condition monitoring and predictive maintenance applications.