Pass-by noise prediction of motor vehicles using gradient boosted models and physical calculations of longitudinal driving dynamics

Abstract

In the prior decades, automotive manufacturers have steadily improved the vibroacoustics performance of their products. This process is highly influenced by the continuously rising requirements on regulations of noise emissions of motor vehicles, also known as pass-by noise. The legal requirements, demanding the reduction of the exterior noise emissions of motor vehicles, are caused by the impact of noise on human health. The latest legal norm of the Economic Commission for Europe (ECE) is the ECE R51.03 test, including continuous reductions of the allowed pass-by noise sound pressure level of motor vehicles. Additionally, the need for faster development cycles results in more intensive use of digital models to predict the expected physical behavior of future products. These models extend the so-called digital twin, which has gained a high importance in automotive manufacturing. Acoustic state-of-the-art simulation and analysis methods fail at the complexity of the over-all pass-by noise, consisting of multiple partial sound sources. These multiple sources are partly correlated, resulting in the absence of trivial deconvolution methods. This paper presents a digital representation of motor vehicles regarding their noise emissions by using machine learning techniques combined with physical calculations of longitudinal driving dynamics. Gradient boosted models are used to predict the exterior sound pressure levels of motor vehicles. The developed algorithm permits the possibility to predict the expected pass-by noise of future cars in early stages of the digital development process. This allows original equipment manufacturers to detect necessary changes to concepts and construction models of future cars. The prediction results are validated for three engines.