Sound quality prediction and improving of vehicle interior noise based on deep convolutional neural networks

Abstract

Interior sound quality plays a vital role in vehicle quality assessment because it forms users' general impressions of vehicles and influences consumers' purchase intentions. Thus, evaluating vehicle interior sound quality is important. Many researchers have developed intelligent prediction models to precisely evaluate vehicle interior sound quality. Deep convolutional neural networks (CNNs) can automatically learn features and many studies have applied deep CNNs to address noise and vibration issues. However, those studies suffer from two problems: i) the time and frequency characteristics of noise that influence interior sound quality have not been considered simultaneously; ii) the noise features that deep CNNs have learned need to be explored. Therefore, in this paper, to overcome the first problem, we develop a regularized deep CNN model that takes a noise time–frequency image as input. In addition, we introduce a neuron visualization algorithm for deep CNNs to solve the second problem. To verify the proposed methods, we establish an interior noise dataset through vehicular road tests and subjective evaluations. The sound quality of this recorded interior noise is evaluated through the developed deep CNN model, which reveals that deep CNNs that use a noise time–frequency image as input perform better than do those using time vector and frequency vector data as input. By analyzing feature maps extracted from the convolutional layers and the fully connected layer of the CNNs, we found that the deep CNN feature learning process can be regarded as color filter and Gabor filter processes applied to the noise time–frequency image. These results provide a new approach for evaluating vehicle interior sound quality and help in understanding which noise features deep CNNs learn.