Abstract: Excitation level time-frequency spectrum can be used to establish the convolution neural network (CNN) model of vehicle sound quality evaluation (SQE). However, due to the discrepancy between the fluctuation characteristic of time-varying sound and the smooth characteristic of the instantaneous subjective evaluation curve of vehicle interior sound quality, the time-varying SQE model produces a fluctuating response to an input of fluctuating sound feature sequences. The performance of the CNN model of the overall annoyance evaluation of vehicle interior noise will be limited by directly using the fluctuating excitation level spectrum in time domain. In this paper, the Savitzky-Golay filter is used to smooth the excitation level spectrum in time domain, and CNN is used to build the mapping relationship between the overall subjective evaluation results of the comprehensive annoyance degree of vehicle interior noise and the time-series smoothed excitation level spectrum so that the overall annoyance CNN evaluation model based on the time-series smoothed excitation level spectrum is established. The leave-one-out cross-validation results indicate that compared with the excitation level spectrum CNN model, the time-series smoothed excitation level spectrum CNN model has better performance in overall annoyance evaluation of vehicle interior noise, with improvement in prediction accuracy (mean error decreased by 10.43%), stability (prediction variance decreased by 44.26%) and consistency (the Pearson correlation coefficient increased by 4.13%). The time-series smoothed excitation level spectrum can better represent the overall annoyance of vehicle interior noise than the excitation level spectrum

Key words: vehicle interior noise, overall annoyance evaluation, excitation level spectrum, CNN, Savitzky-Golay filter