林旭,梁兴雨,代鹏.基于深度学习的发动机声品质预测模型研究[J].内燃机工程,2024,45(5):19-27.
基于深度学习的发动机声品质预测模型研究
Research on Engine Sound Quality Prediction Model Based on Deep Learning
DOI：10.13949/j.cnki.nrjgc.2024.05.003
关键词:发动机  声品质  预测模型  改进麻雀搜索算法
Key Words:engine  sound quality  prediction model  improved sparrow searchalgorithm （ISSA）
基金项目:青海省自然科学基金项目（2022-ZJ-757）
作者单位E-mail
林旭 青海民族大学 土木与交通工程学院，西宁 810007
天津大学 先进内燃动力全国重点实验室，天津 300072 		linxuaaa@126.com 
梁兴雨* 天津大学 先进内燃动力全国重点实验室，天津 300072 lxy@tju.edu.cn 
代鹏 天津大学 先进内燃动力全国重点实验室，天津 300072 dai_peng@tju.edu.cn 
摘要点击次数: 449
全文下载次数: 255
摘要:为建立发动机辐射噪声品质深度学习预测模型，搭建试验台架采集发动机辐射噪声，计算噪声信号心理学客观参数并进行主观评价试验。采用卷积神经网络（convolution neural network， CNN）提取信号特征，引入长短期记忆网络（long short-term memory network， LSTM）模型捕获信号长期依赖信息，利用注意力 （Attention）机制使模型自动学习关键特征信息。以心理学客观参数为输入，主观评价得分为输出，建立CNN-LSTM-Attention声品质预测模型，引入改进麻雀搜索算法（improved sparrow search algorithm， ISSA）优化模型超参数，提高预测准确性。研究结果表明，ISSA-CNN-LSTM-Attention模型对发动机声品质具有良好的训练性能和预测能力，训练集和测试集的决定系数分别为0.988、0.981，训练集和测试集的平均绝对误差分别为0.204、0.241。该模型能够准确地反映客观评价参数与主观满意度之间的非线性映射关系，为发动机声品质预测提供了新的思路和方法。
Abstract:In order to develop a deep learning prediction model for the sound quality of engine radiated noise， a test bench was constructed to collect engine radiated noise. The psychological objective parameters of noise signals were determined， and subjective evaluation experiments were conducted. The psychological objective parameters of noise signals were determined， and subjective evaluation experiment was conducted. Signal features were extracted employing convolutional neural network （CNN）， long-term dependence information was acquired utilizing long short-term memory （LSTM） network， and essential feature information was automatically learned with the Attention mechanism. The subjective evaluation score was applied as the output， while the psychological objective parameters were used as inputs. A sound quality prediction model for CNN-LSTM-Attention was developed. To increase prediction accuracy， the improved sparrow search algorithm （ISSA） was employed to adjust the hyperparameters of model. The results show that the ISSA-CNN-LSTM-Attention model accurately predicts engine sound quality. For the training and test sets， the average absolute errors are 0.204 and 0.241， respectively， while the coefficient of determination are 0.988 and 0.981， respectively. The model presents new perspectives and methods for predicting engine sound quality since it can effectively represent the nonlinear mapping relationship between objective parameters and subjective satisfaction.
查看全文  HTML   查看/发表评论