Data Augmentation Methods for Electric Automobile Noise Design from Multi-Channel Steering Accelerometer Signals

Yongwon Jo, Keewon Jeong, Sihu Ahn, Eunji Koh, Eunsung Ko, Seoung Bum Kim

Abstract: Noise, vibration, and harshness (NVH) of electric automobiles is important because the loud NVH can reduce the satisfaction of automobile drivers and passengers. Therefore, the effective machine learning models to alleviate NVH is required. Although a huge amount of data is needed to construct the reliable models, the number of training data is very scarce in practice. In this paper, we propose a deep learning model combined with data augmentation methods (dropout and SpecAugment) that predicts interior noise levels from steering accelerometer signals when only a small number of training data is available. The effectiveness of the proposed framework was demonstrated using steering automobile accelerometer signals and noise levels from real automobiles.