王飞扬,王贵勇,王煜华,等.基于深度学习神经网络的柴油机NOx瞬态排放预测[J].内燃机工程,2026,47(1):115-125.
基于深度学习神经网络的柴油机NOx瞬态排放预测
Diesel Engine NOx Transient Emission Prediction Based on a Deep Learning Neural Network
DOI:10.13949/j.cnki.nrjgc.2026.01.012
关键词:柴油机  NOx排放  卷积神经网络  门控循环神经网络  混合专家  多头注意力
Key Words:diesel engine  NOx emission  convolutional neural network(CNN)  gated recurrent unit(GRU)  mixture of experts(MoE)  multi-head attention
基金项目:云南省重大科技专项计划项目(202402AE090009)
作者单位E-mail
王飞扬* 昆明理工大学 动力工程系昆明 650500 805303736@qq.com 
王贵勇* 昆明理工大学 动力工程系昆明 650500 wangguiyong@kust.edu.cn 
王煜华 昆明理工大学 动力工程系昆明 650500 2869290956@qq.com 
彭云龙 昆明理工大学 动力工程系昆明 650500 2654341247@qq.com 
汪志远 昆明理工大学 动力工程系昆明 650500 1253112561@qq.com 
何述超 昆明理工大学 动力工程系昆明 650500
昆明云内动力股份有限公司昆明 650500 
heshuchao.130@163.com 
摘要点击次数: 91
全文下载次数: 43
摘要:针对传统的静态模型和单一神经网络模型在捕捉柴油机NOx瞬态排放复杂动态变化方面存在局限的问题,提出了一种基于卷积神经网络(convolutional neural network, CNN)、门控循环神经网络(gated recurrent unit, GRU)、混合专家神经网络(mixture of experts, MoE)、多头注意力机制(multi-head attention, MHA)融合的深度学习神经网络模型。通过世界统一瞬态循环(world harmonized transient cycle, WHTC),收集柴油机运行的关键参数并采用数据预处理和特征选择技术得到数据集;然后利用CNN神经网络提取数据集的特征;再使用GRU神经网络时间序列处理能力拟合数据;最后利用MoE神经网络的动态权重分配和MHA机制的多角度特征关注提高模型的预测精度和泛化能力。试验结果表明:CNN-GRU-MoE-MHA神经网络模型的平均绝对误差(mean absolute error, MAE)为21.53 mg/L,均方根误差(root mean squared error, RMSE)为26.91 mg/L,与GRU、CNN-GRU、CNN-GRU-MoE模型相比显著降低,同时其R2更高,说明CNN-GRU-MoE-MHA模型具有较高的预测精度和良好的稳定性。
Abstract:In response to the limitations of traditional static models and single neural network models in capturing the complex dynamic changes of diesel engine NOx transient emissions, a deep learning neural network model was proposed based on the fusion of convolutional neural network (CNN), gated recurrent unit (GRU), mixture of experts (MoE), and multi-head attention (MHA) mechanisms. Key parameters of diesel engine operation were collected using the world harmonized transient cycle (WHTC), and a dataset was obtained through data preprocessing and feature selection techniques. The CNN neural network was utilized to extract features from the dataset, and the GRU neural network’s capability in time series processing was employed to fit the data. The MoE neural network’s dynamic weight allocation and the MHA mechanism’s multi-angle feature focus were used to enhance the model’s predictive accuracy and generalization ability. Experimental results indicate that the CNN-GRU-MoE-MHA neural network model achieved a mean absolute error (MAE) of 21.53 mg/L and a root mean squared error (RMSE) of 26.91 mg/L, which are significantly lower than those of GRU, CNN-GRU, and CNN-GRU-MoE models, and the CNN-GRU-MoE-MHA model exhibits a higher coefficient of determination (R²), suggesting superior predictive accuracy and stability.
查看全文  HTML   查看/发表评论