Sound Quality Evaluation Of Vehicle Interior Noise Based On Deep Learning

Interior sound quality is a key factor in the user’s overall impression of the car,which plays a vital role in the evaluation of the overall quality of the car and straightly influences the consumers’ purchasing decision.Therefore,it is very important to evaluate the interior sound quality.The commonly used sound quality evaluation methods at home and abroad all adopt listening test,which is time-consuming and laborious,so it is of great practical significance to establish an objective sound quality evaluation model.However,in the existing research,the features used in the sound quality prediction model are manually extracted and selected,which highly depends on the previous acoustic theory and empirical knowledge.In order to take full advantage of the initial data included in the original interior noise samples and further improve the accuracy of the model,in this paper we established an interior noise’s sound quality evaluation model on the basis of deep convolution neural network.Firstly,according to the relevant provisions of the national standard,this paper designs and completes the interior noise samples collection,collects the interior noise signals of five different brands of vehicles under different speed and uniform conditions,and completes the preparation of sound samples for subjective evaluation of sound quality.The subjective evaluation listening test of sound quality was organized and completed.The evaluation results were tested by data,and the noise samples were marked according to the evaluation results,which provided the data basis for the follow-up research work.One dimensional convolutional neural network is used to extract the time-frequency features from the original audio,which is used to compose the fusion features with the artificial logarithm Mel spectrum,then,the shallow feature extraction module of vehicle interior noise signal is constructed.Based on two-dimensional convolutional neural network and fully connected layer network,the deep feature extraction module and classifier module are constructed respectively.Finally,the interior sound quality evaluation model based on deep convolution neural network is established.Deep learning relies on abundance data to train the model.However,under objective conditions,interior noise data is shortage in number.As a result of the reasons just mentioned,two data enhancement methods for audio data sets are adopted,and three transfer learning training strategies are used to complete the training of the model based on the pretraining model.The ability of this model in the task of vehicle interior noise evaluation is verified and evaluated by using the previously built vehicle interior noise data set.The results show that the prediction accuracy rate of the model obtained by fine tune strategy in the sound quality evaluation task reaches 96.6%,and it is the best among the three models trained by different strategies,the results are in good agreement with those obtained from human jury test,and can be used in the establishment of vehicle interior noise quality evaluation system.