Development of a machine learning model to predict engine noise perception considering regional and driving environment differences

In hybrid vehicles, the acoustic contrast between engine-off and engine-on states makes engine noise more annoying than in conventional vehicles. To support the development of engine control strategies for less annoying engine noise, subjective annoyance was quantified using machine learning models to predict in-cabin noise perception. In-cabin sound, vehicle signals, and three-level subjective ratings (“Not noisy”, “Noisy”, “Very noisy”) were synchronously recorded during on-road driving with compact series hybrid vehicles in Japan. A random forest classifier trained on five psychoacoustic metrics—Loudness, Sharpness, Roughness, Fluctuation Strength, and Tonality—achieved 57% accuracy for predicting subjective ratings. Including key vehicle signals (engine speed, driveshaft torque, and vehicle speed) improved accuracy to 67%, and further incorporating the history of engine speed over about 5.5 s raised it to 75%. To assess generality, similar evaluations were conducted in the United States and United Kingdom using battery-electric vehicles with synthesized engine sound. Although overall accuracy was lower than in Japan due to differences in driving style and traffic, the same improvement pattern was observed: adding vehicle parameters and the history of engine speed consistently improved predictive performance. These findings suggest that accounting for driving conditions and the history of engine speed over several seconds will help minimize engine noise annoyance in future hybrid vehicles.