Improving of pure electric vehicle sound and vibration comfort using a multi-task learning with task-dependent weighting method

Abstract

The improvement of pure electric vehicle (PEV) vibro-acoustical comfort quality has currently taken precedence in development. The acoustic and vibration characteristics of PEVs exhibiting a wide frequency distribution and encompassing numerous objectives, thereby presenting challenges in its predicting and optimizing. Traditional simulation and experimental methods, though capable of simultaneous analysis of noise and vibration information, grapple with intricate modeling and reduced efficiency. This paper proposes a solution through a multi-task learning approach with homoscedastic uncertainty interpretation as task-dependent weighting (TDW), effectively balancing losses across distinct regression tasks. This method rectifies biased predictions among tasks, ensuring accurate interior vibro-acoustical comfort prediction across diverse features and frequency distributions. Furthermore, a fusion method, integrating TDW-based multi-task learning with a vibro-acoustical comfort knowledge graph (VACKG), is introduced. The proposed method is validated through PEV noise and vibration prediction and optimization, demonstrating its accuracy and robustness in enhancing interior vibro-acoustical comfort in real vehicular experiments.