The Research Of Automobile Sound Conversion System Based On Sparse Autoencoder And Feedforward Neural Network

With the rapid development of automobile industry,the vehicle NVH performance has got more and more attention.Vehicle NVH performance has a great impact on vehicle quality.Nowadays,the artificial head acquisition system as the main tool is used for the collection of the automobile interior noise.However,the equipment is expensive and complex to operate.The research team independently builds a portable noise acquisition system.But the portable noise acquisition system does not take the characteristics of human hearing into account.It does not meet the requirement of industrial application.This has the potential for further research.In order to improve the accuracy of portable noise acquisition system,this article takes the noise of electric vehicles as the data sample.From the perspective of sound signal processing,the sound conversion system based on sparse autoencoder and feedforward neural network is established.The sound signal collected by the portable noise acquisition system is converted into the sound signal collected by the artificial head acquisition system.The converted sound signal passes the subjective evaluation and the objective evaluation based on the sound quality parameters.The evaluation results show that the portable noise acquisition system can approximately replace the artificial head acquisition system,and it can be used to collect interior noise and calculate the sound quality.The main research contents of this project are as follows:(1)Under the same working condition,two kinds of equipment are used to collect the noise samples of the car.Then the collected sound signals is performed a preliminary analysis.(2)The WORLD vocoder is used to extract the feature parameters of original sound signals and target sound signals.The extracted feature parameters are: the fundamental frequency,the spectral envelope parameter and the band aperiodic parameter.The feature parameters of each group of sound signals is integrated into a parameter matrix.Then parametric cell groups of original sound signals and target sound signals are estabilished.(3)Due to the larger dimension of the feature parameters of sound signals,the sparse autoencoder is used to transform high-dimensional parameters into low-dimensional parameters.Then the feedforward neural network is used to fit low-dimensional parameters.A conversion model of sound signal characteristic parameters is established.After the characteristic parameters of the original sound signal are converted by the model,the WORLD vocoder synthesizes the converted parameters into a new sound signal.Finally,the sound conversion model is established.(4)The subjective and objective evaluation of the conversion performance of the sound conversion model is carried out.The subjective evaluation mainly uses the ABX and the MOS test method.Through tests,the synthesized sound signal is similar to the target sound signal in the aspect of subjective hearing.Objective evaluation is mainly based on time-domain,frequency-domain and sound quality.In the time-domain waveform and the frequency-domain waveform,the waveform of the synthesized sound signal is similar to the waveform of the target sound signal.Through the comparison of loudness,sharpness and roughness,the synthesized sound signal is similar to the target sound signal.The average error meets the industrial requirements.