梁凯，赵海军，宋伟志.基于卷积神经网络的内燃机声品质评价方法研究[J].内燃机工程,2019,40(2):67-75.
基于卷积神经网络的内燃机声品质评价方法研究
Research on Evaluation Method of Internal Combustion Engine Sound Quality Based on Convolutional Neural Network
DOI：10.13949/j.cnki.nrjgc.2019.02.010
关键词:内燃机  声品质  卷积神经网络  听觉谱  BP神经网络
Key Words:internal combustion engine  sound quality  convolutional neural network  auditory spectrum  BP neural network
基金项目:国家自然科学基金项目(U1604141)
作者单位
梁凯，赵海军，宋伟志 1.洛阳理工学院 信息化技术中心，洛阳 471023； 2.天津职业技术师范大学 汽车与交通学院，天津 300222 
摘要点击次数: 2706
全文下载次数: 1748
摘要:为解决内燃机声品质评价中人工效率低、成本高的问题，引入卷积神经网络(convolutional neural network, CNN)模型和声谱分析方法构建了CNN声品质预测模型；同时模型中设计了带通滤波器，可对噪声样本进行自动特征提取，并以此为输入数据，利用自适应时刻估计（adaptive moment estimation，Adam）算法优化网络中各层权重，并将模型用于声品质评价。为证明CNN模型预测的性能，构建了基于心理声学参量的后向传播算法(back propagation，BP)声品质评价模型，并用于对照试验，在样本标签值（人工评价值）处理时，分析了客观评价心理声学参数与评分值的相关性，选取与人工评价结果相关度最大的4个心理声学参量作为BP模型的输入值进行预测。试验结果表明，基于CNN的声品质评价模型能更精确地预测内燃机声品质，并且在CNN预测模型中基于听觉谱的输入评价值比基于时域的短时平均能量、频域的频谱通量输入评价值精度更高。
Abstract:In order to solve the problem of low efficiency and high cost in the evaluation of sound quality of internal combustion engines, the convolutional neural network(CNN) model and sound spectrum analysis method were used to construct a CNN sound quality prediction model(CNNSQP model) to evaluate the sound quality. Noise samples were extracted automatically by the Gammatone filter and used as input data. The adaptive moment estimation(Adam) algorithm was used to optimize the weights of the layers in the network. To prove the predictive performance of the CNNSQP model, a back propagation(BP) sound quality evaluation model based on psychoacoustics objective parameters was built and used for comparison experiments. The sample label value(manual evaluation) was processed, and four psychoacoustics objective parameters that are most related to the manual evaluation result were selected as the input values of the BP model for prediction. The experimental results show that the CNN-based sound quality assessment model predicts the sound quality of internal combustion engines more accurately than the assessment based on the BP model, and the input evaluation value based on the auditory spectrum in the CNNSQP model is more precise than that based on short-term average energy in time domain and spectral flux in frequency domain.
查看全文  HTML   查看/发表评论