Reseach On Intelligent Evaluation Method For Nonstationary Vehicle Noise Based On Psychoacoustics

Sound quality evaluation (SQE), as a new evaluation method has been used in vehiclenoise evaluations because its result is much closer to people’s feeling. However, at present,the SQE theories are not mature enough and have still being discussed. Many researchresults have not been generally published, due to the technical and commercial secrecyreasons. Thus, studying auditory-perception-based SQE approaches for nonstationaryvehicle noise has significant theoretical and practical values in engineering.Based on the previous studies, this paper presents a new energy-based neural networkSQE model, considering the auditory feature of human. The SQE related phycoacousticalconcepts and parameters are investigated and their influences to the indices, such asloudness, sharpness and roughness, etc., are determined. Based on the measured low-frequency vehicle accelerating noise signals, some time-frequency analysis techniquescommon used in digital signal processing, such as the short-time Fourier transform (STFT),Wigner-Ville distribution (WVD), wavelet transform (WT) and Hilbert-Huang Transform(HHT) are compared. The results show that, the wavelet packet method can decompose anoise signal into some component signals with the same specific frequency bands as thoseof human auditory system, is more suitable for time-frequency feature extraction ofnonstationary vehicle noises.According to the specific frequency partition, a wavelet packet with db35function inMatlab software is selected and the vehicle noise signals are decomposed into twenty-onefrequency bands. The time axis is divided into a set of subsections with a50ms timeinterval. The energy features of each decomposed signal and its envelope are extracted,respectively. Taking the energy features as inputs and the phycoacoustical parametersobtained from the traditional algorithms as outputs, the intelligent evaluation models for loudness, sharpness and roughness of nonstationary vehicle noise are established by usingthe BP neural network. Finally, the intelligent evaluation models are further optimized byadjusting the network topological structure and their parameters and verified by tests.The verification results show that, for the nonstationary vehicle noises, the errors ofthe loudness, sharpness and roughness calculated from the intelligent SQE models are allbelow10percent, even the loudness and sharpness can reach to5percent, which suggest agood accuracy of the proposed energy-based neural network models. The proposed modelscan be directly applied in SQE of vehicle interior noise. The modeling approach presentedin this thesis may be extended to other sound related field for sound quality modeling,analyzing and evaluations.