Self-Supervised Learning for Vehicle Noise Prediction With Limited Labeled Data

With advancements in artificial intelligence, its application for developing quieter vehicles is increasingly researched in the automotive industry. Vehicle noise is highly impacted by the acceleration from electric power steering (EPS); therefore, determining the relationship between EPS-induced acceleration and vehicle noise is important. However, collecting labeled data for noise prediction is challenging because it requires expensive acceleration sensors attached to the EPS, and experts must measure the noise directly by ear. In contrast, obtaining unlabeled acceleration data from environments similar to those of the EPS is manageable. This study proposes an autoencoder-based self-supervised learning with information maximization (ASSIM) method to predict vehicle noise. ASSIM allows robust feature learning via pretraining to reconstruct unlabeled data with necessary augmentations. The experimental results demonstrate that ASSIM performs better than the other comparative methods in predicting vehicle noise in environments with less labeled data. The proposed method may aid in the design of quieter vehicles while reducing the data collection time.