Feature Analysis for Gear Growl in Electrified Powertrains

Electrified powertrains generate less noise and vibration than internal combustion engines, making secondary sources like gear-induced noise more noticeable. One such issue, gear-growl, is increasingly concerning for passengers and manufacturers. However, its signal properties and perceptual thresholds are not well defined. This study aims to extract and select informative features from vibration and acoustic signals to analyze gear-growl phenomena better. Root-mean-square and spectral entropy effectively distinguish gear-growl in vibration signals, while acoustic signals require a broader feature set due to background noise and nonlinearity. Feature selection techniques are used to improve interpretability and reduce dimensionality. Supervised machine learning, specifically a kernel support vector machine, evaluates the selected features’ effectiveness. A multiclass model, informed by subjective listening tests, categorizes gear-growl severity levels with a test accuracy of 0.89. These classification results support the feature selection process but are not the study’s primary focus. Further listening studies are suggested to improve labeling consistency and reliability. Though this work centers on gear-growl, the feature-based framework may be applicable to other noise and vibration diagnostics.