Feature Analysis for Fault Diagnosis in Electrified Powertrains

Electrified powertrains generate lower levels of noise and vibration compared to internal combustion engines, making secondary sources such as gear-induced noise more perceptible. One such phenomenon, gear-growl, has become an increasing concern for both passengers and manufacturers; however, its signal characteristics and perceptual thresholds remain poorly defined. Additional noise contributions arise from axle components, including the gearset and bearings. This study aims to extract and select informative features from vibration and acoustic signals to facilitate improved analysis of these noise phenomena.Root-mean-square (RMS) and spectral entropy were found to effectively discriminate faults in vibration signals, while acoustic signals required a broader feature set due to the presence of background noise and signal nonlinearity. Feature selection techniques were employed to enhance model interpretability and reduce dimensionality. A supervised machine learning approach, specifically a kernel-based support vector machine (SVM), was used to evaluate the effectiveness of the selected features. For gear-growl detection, a binary classifier achieved a test accuracy of 94.8%. A multiclass model, informed by subjective listening tests, categorized gear-growl severity levels with a test accuracy of 89.1%. On a secondary dataset focused on axle-related noise, classification achieved 100% test accuracy using a majority voting scheme across segments.While classification results support the relevance of the selected features, the primary goal of this study is not the development of a diagnostic tool but the exploration of a feature-based framework for analyzing NVH phenomena. Further listening studies are recommended to enhance labeling consistency and perceptual reliability. Although this work focuses on noise from electrified powertrains, the proposed methodology may be broadly applicable to other noise and vibration diagnostic tasks.