基于EEMD与GA-小波神经网络的传动系声品质预测

Sound metric prediction of a power train system based on EEMD and GA-wavelet neural network

为了进行车辆传动系声品质预测,实施了传动系整车转鼓试验,并结合主、客观分析量化了影响传动系噪声烦恼度的主要异响指标;同时,通过相关分析揭示了心理声学客观参量与主观评价的内在关系。引入聚合经验模态分解(Ensemble Empirical Mode Decomposition,EEMD)方法与本征模态函数(Intrinsic Mode Function,IMF)样本熵值对传动系噪声特征进行了提取;在此基础上,以Morlet小波基函数作为隐含层节点的传递函数构建小波神经网络(Wavelet Neural Network,WNN),同时运用遗传算法(Genetic Algorithm,GA)优化小波神经网络的层间权值和层内阈值,构造出GA-小波神经网络模型并用于传动系声品质预测;为了对比所提取的传动系噪声特征性能,将心理声学参量也作为模型输入以进行预测,同时,为了对比GA-小波神经网络模型的预测效果,引入了传统的GA-BP神经网络模型。分析结果表明:GA-小波神经网络较GA-BP神经网络能更准确、有效地对传动系声品质进行预测,并且以本征模态函数样本熵值作为预测模型的输入特征其预测结果较心理声学参量效果更佳。

To predict the sound quality of a vehicle power train system's noise, a complete vehicle drum test was conducted. Combining subjective analysis and objective one, six sound indices for the power train system's noise annoyance level were quantified. Via the correlation analysis, internal relations between the six psychoacoustic objective parameters and the subjective evaluation were revealed. Features of the power train system's noise were extracted by using the ensemble empirical mode decomposition (EEMD) method, and the intrinsic mode function sample entropy. The Morlet wavelet basis functions were used as the transfer function of hidden layers to develop a wavelet neural network. In addition, the genetic algorithm was applied to optimize weights between layers and thresholds within layers so that a GA-wavelet neural network was developed to predict the sound quality of the power train systems's noise. In order to validate the the newly extracted power train system's noise features, the psychoacoustic parameters were also taken as the model's inputs to predict the sound quality. Meanwhile, the conventional GA-BP neural network was also introduced to compare its performances with those of the GA-wavelet neural network. The results showed that the GA-wavelet neural network can predict the sound quality of the power train system's noise more accurately and effectively than the GA-BP neural network can; the intrinsic mode function sample entropy values are better than the psychoacoustic parameters to be taken as the input features of the prediction model. © 2017, Editorial Office of Journal of Vibration and Shock. All right reserved.