Research On Evaluation Of Vehicle Sound Quality Based On GA-BP Neural Network

With the development of socio-economic technology and the improvement ofliving standard, the popularity of cars is increasing and people’s request on vehicleride comfort is continuously improving. Characteristic of the car noise is an importantfactor affecting the vehicle ride comfort, while the engine is one of the main sourcesof interior noise, interior noise caused by the engine has become a hot research in theautomotive industry industry, but for the study of interior noise caused by C-class carengine is not documented.The topic comes from the key projects of Jilin Province Development Plan-theintelligent evaluation model of car sound quality and control technology. This paperfocuses on interior noise caused by C-class car engine, and takes psychoacousticparameters calculation and sound quality subjective evaluation test, and analyses ofthe correlation between them. Through the application of BP neural network based ongenetic algorithm(GA-BP), this thesis sets up the sound quality prediction model, andsuccessfully establishes the relationship between the objective and quantitative andsubjective feeling. Comparing the GA-BP prediction model and the BP predictionmodel in convergence, stability, and prediction accuracy to verify the reliability andaccuracy of the GA-BP neural network model.The main contents of this article is divided into four parts:（1）Get the interior noise, takes psychoacoustic parameters calculation andsound quality subjective evaluation test, and analysis of the correlation between them.The test conducted in semi-anechoic chamber of faw technology centre, usingrotating hub test to collect interior noise caused by the engine of five uniformconditions in four cars. Using level score method to score the irritability on the soundsamples, calculating7objective psychoacoustic parameters of sound samples,analysing correlation analysis of the value of subjective evaluation and objectiveparameters, and selecting the psychoacoustic parameters with a larger correlation withthe subjective evaluation value. The selected psychoacoustic parameters are loudness, sharpness, roughness and AI index. Psychological acoustic parameters as soundquality prediction model of input, subjective evaluation value as a sound qualityprediction model output.（2）Establish C-class car sound quality database.Establishment of C-class car sound quality database, which makes theinformation of experiment, operating conditions and model stored in the database, andsaves access time and provides data support for the C-class car lower noise.（3）Establish GA-BP neural network prediction model of sound qualityirritability.Determine the structure of BP neural network, including input, output layernumber of neurons, implied layer, hidden neurons number and the transfer function.Use genetic algorithm to code the weights and thresholds of BP network, anddetermine the fitness function to get the smaller error of the network. Finally usingchoice, cross and variation operation to search for the optimal solution, and the resultof genetic algorithm as weights and thresholds of BP network, then, the GA-BP modelof sound quality prediction can be gotten. And training the GA-BP model in order toachieve the required accuracy.（4）Comparing the GA-BP prediction model and the BP prediction model inconvergence, stability and prediction accuracy to verify the reliability and accuracy ofthe GA-BP neural network model.Through experiment，when the goal of network training error is same, GA-BPprediction model’s convergence rate is5times higher than the BP prediction model’sconvergence rate. Since the randomness of BP prediction model’s initial weights andthresholds, it may lead a greater difference when testing the same sample, but GA-BPprediction model using genetic algorithms which optimized the initial weights andthresholds of BP neural network, this ensure the stability of the network and thehigher consistency of the sound quality predictions.