A comprehensive study on statistical prediction and reduction of tire/road noise

Abstract

Promoting safe tires with low external rolling noise increases the environmental efficiency of road transport. Although tire builders have been striving to reduce emitted noise, the issue’s sophisticated nature has made it difficult. This article aims to make the problem straightforward, relying on recent significant improvements in statistical science. In this regard, the prediction ability of new methods in this field, including support vector machine, relevance vector machine, and convolutional neural network, along with the new architecture of the neural network is compared. Tire noise is measured under the coast-by condition. Two training strategies are proposed: extracting features from a tread pattern image and directly importing an image to the model. The relevance vector method, which is trained using the first strategy, has provided the most accurate results with an error of 0.62 dB(A) in predicting the total noise level. This precise model is used instead of experimentation to analyze the sensitivity of tire noise to its parameters using a small central composite design. The parametric study reveals striking tips for reducing noise, especially in terms of interactions between parameters that have not previously been shown. Finally, a novel two-stage approach for reducing noise by tread pattern optimization is proposed, inspired by two regression models derived from statistical investigation and variance analysis. Changes in tread pattern specifications of two case studies and their randomization have resulted in a reduction of 3.2 dB(A) for a high-noise tire and 0.4 dB(A) decrement for a quieter tire.

A Fast Approach to Optimize Tread Pattern Shape for Tire Noise Reduction

“…In terms of optimization time, it takes only 5.5 s to obtain the first optimization solution and only 76 s to produce the final result. Different from regression methods, which include the support vector machine (SVM), relevance vector machine (RVM), ANN [57], and convolutional neural network (CNN) [58] (the RVM, ANN, and CNN take 66,297, 32,428, and 205,694 s for data training [59]), the optimization method proposed in this paper does not need to collect and train data. In addition, there are fewer parameters for pattern parameterization [31].…”