Research On Evaluation And Application Of Vehicle Sound Quality Based On Artificial Neural Network

With the idea of vehicle sound quality design was proposed, automotive acoustic design closely links to consumer demand and make it into whole automobile and parts, which reflect the concept of "people-oriented" design concept. It is not the important research content of vehicle noise to establish a good, comfortable and satisfactory vehicle interior acoustic environment,but also the future direction of development. At present, the research on vehicle sound quality is not mature in domestic and international, and still have many problems need to be solved, such as: subjective evaluation of sound quality tests are time-consuming, wealth-consuming, and the heavy workload; evaluators are vulnerable to the impact of fatigue and external factors leading to inaccurate evaluation results; previous multiple linear regression models are poor to promote, and low-prediction accuracy; sound quality research lacks referential value and instructive significance for automotive acoustic design. This paper aims to study the evaluation model and practical application of vehicle sound quality, then, using artificial neural network technology to build model of sound quality for domestic vehicle, so that the model can reflect on the domestic consumer's subjective feeling of the sound quality, and get valid accurate evaluation of the results without the need for subjective evaluation tests, complex modeling and calculation.In this paper, the main line is evaluation technology and practical application of vehicle interior sound quality. The subjective evaluation tests of domestic cars were carried out for the index of preference and annoyance respectively. Based on the results of subjective evaluation tests, using artificial neural network approach, a BP neural network prediction model of sound quality preference and annoyance was established. Comparing with multiple linear regression analysis, the proposed model is more accurate and superior. A vehicle sound quality neural network evaluation system was established by the model and virtual Instruments. By the system, the tested vehicle interior sound quality was not only evaluated and analyzed, but also improved. This proved the effectiveness and the practical value of neural network prediction model of sound quality. The main contents include:First, for B-class cars, subjective evaluation tests on the sound quality preference in the uniform condition and sound quality annoyance in the acceleration condition were carried out. The tests used paired comparison and grade scoring method to evaluate, and the paired comparison was for the sound quality preference of sound samples, the scoring method was for the sound quality annoyance. Then the results were tested, and the substandard results were removed. According to the results of the remaining assessors' subjective evaluation, the subjective evaluation value of each sample was received. By using correlation analysis and multiple linear regression analysis, a mathematical model which can describe sound quality preference and annoyance of objective evaluation based on psychoacoustic quantification of objective parameters was establish. The results show that, for B-class cars in uniform conditions, preference of vehicle interior sound quality is mainly affected by the loudness and sharpness; in the acceleration conditions, annoyance is mainly affected by loudness,roughness and A-level. Those provide experiences and sample data for establishing neural network prediction model of sound quality.Secondly, BP neural network prediction model of sound quality preference and annoyance was established by using artificial neural network technology. The network structure and the training parameters of that model were received by analyzing and discussing Matlab results. Then, genetic algorithm was used to determine the optimal initial weights of the network, by doing this, the local minimum occurred during network training can be avoided, and the network training time was reduced. Using the LM learning algorithm to train the network, the connection weights and thresholds of each network was obtained. After error analysis and correlation analysis had been done for the training and testing the results of two BP neural network models, it shows that those models have strong learning ability and generalization ability. Neural network model and multiple linear regression model were applied to predict the sound quality of samples, and the results were compared. Compared with the values of subjective evaluation, the maximum relative errors of that two models are 18.9% and 11.3%, far lower than the multiple linear regression model which are 26.7% and 16.7%; performance index of the two models are 0.003 and 0.001, also lower than the multiple linear regression model which are 0.006 and 0.003. The results show that both BP neural network models are superior to multiple linear regression models, and also prove the accuracy and superiority of the neural network model.And then, to prove the effectiveness and the practical value of neural network prediction model of sound quality, a vehicle sound quality neural network evaluation system was established by virtual instrument technology based on the established neural network prediction model of sound quality preference and annoyance. In the acceleration condition of the tested vehicle, the established system was used to measure the vehicle interiors noise and analyze the sound quality. Improving the vehicle sound quality by effective noise reduction measures. Using LabVIEW and Matlab software, the system can acquire and process the vehicle interior noise in the conditions of uniform and acceleration. The system also has the function of calculating acoustic parameters, displaying graphical and evaluating sound quality indicators. The system used Matlab software to calculate the sound quality parameters, such as A-level, loudness, roughness and sharpness. Compared with the results of ArtemiS software, psychoacoustic parameters calculation model in this paper are more reliability and accuracy, and it laid the foundation for sound quality neural network evaluation system.Finally, in the acceleration condition of the tested vehicle, the established sound quality neural network evaluation system was used to measure the vehicle interiors noise and analyze the sound quality, and validate the practical value and significance of the sound quality evaluation model in improving vehicle interior sound quality. Improving the vehicle sound quality for engine speed at 3000r/min ~ 3500r/min. Through partical coherence analysis between psychoacoustic parameters and vibration signals, the sound source locations which greatest impacting the interior sound quality are instrument panel, vehicle roof and the floors of the driver and co-pilot positions, then targeted measures were taken to control them. Road testing was carried out on the noise reducted car, the results showed that, the sound quality improved more than 12% compared with noise not reduced. It confirmed that the partial coherence analysis for sound quality in noise source identification were feasibility and effectiveness.The result of the research shows that the artificial neural network theory is fully applicable to the vehicle interior sound quality evaluation, and the theory has a high theoretical significance and a good prospect.