Tire noise prediction based on transfer learning and multi-modal fusion

Abstract

Vehicle traffic noise has become an important factor in urban noise pollution. With the increase in the number of new-energy vehicles, the current situation of tire/road noise as one of the main noise sources of automobiles is becoming increasingly prominent. Tire noise prediction is the basis for optimizing the design of low-noise tires, providing a reference basis for tire design with low-noise performance. This paper used deep learning methods to predict the noise performance of TBR radial tires. We obtained descriptive statistical features of tire structure and tread pattern images as input to the model, and used measured tire/road noise as output, then we constructed a multi-modal tire noise dataset. Comparing the prediction performance of three pre-trained transfer models such as Resnet18, VGG16, and Inception V3 on the tire pattern images, the Resnet18 had the best prediction effect. Resnet18 was selected as a feature extractor to extract image features, which fuse with the tire structural features at the feature level. The experiment constructed the TLMF-TRNP model to predict tire noise. The experimental results of the TLMF-TRNP model indicated that RMSE, MAE, and R2 were 0.1337, 0.0948, and 0.9864 respectively, achieving ideal prediction accuracy on a small-scale tire noise dataset and controlling the absolute error of the test tires within ±0.4 dB effectively. An accurate tire noise prediction model will provide a theoretical basis for tire design with low noise tires.