稀燃汽油机LNT神经网络模型的建立与应用  

Establishment and Application of Artificial Neural Network Based Model of Lean Burn Gasoline Engine with LNT

作　　者：李志军[1] 常庆[1] 张洪洋[1] 刘磊[1] 陈韶舒[1] 曹曼曼[1] 岳东鹏[2] 

机构地区：[1]天津大学内燃机燃烧学国家重点实验室,天津300072 [2]天津职业技术师范大学汽车与交通学院,天津300222

出　　处：《天津大学学报（自然科学与工程技术版）》2015年第3期234-239,共6页Journal of Tianjin University：Science and Technology

基　　金：国家高技术研究发展计划(863计划)资助项目(2008AA06Z322);国家自然科学基金资助项目(50276042;50776062;51276128)

摘　　要：建立了关于稀燃汽油机LNT(lean-NOx trap)催化器的NOx排放量、比油耗和NOx转化效率的人工神经网络(ANN)预测模型.模型所需的训练及测试样本通过一台改制的CA3GA2三缸12气门电控稀燃汽油机的台架试验获得.采用标准的误差反向传播(back propagation,BP)神经网络.网络经过训练,再由测试样本进行测试.测试结果表明,其绝对分数方差(absolute fraction of variance)R2均接近于1,且均方根误差(root mean squared error,RMSE)及平均相对误差(mean relative error,MRE)均在可接受范围内.以确定最佳稀燃时间为例,说明了利用神经网络的泛化能力可对稀燃汽油机进行优化和控制.An artificial neural network（ANN）model of a lean burn gasoline engine with lean-NOx trap（LNT）was built to predict the NOx emission,brake specific fuel consumption and NOx conversion efficiency of LNT. The data for training and testing the proposed ANN were obtained from a number of experiments performed with a 3-cylinder, 12-valve,electronic controlled CA3GA2 lean burn gasoline engine. A standard back propagation ANN was adopted. After the training,the performance of the ANN predictions was measured by testing data. It is found that the R2（absolute fraction of variance）values are close to 1,and root mean squared error（RMSE）and mean relative er-ror（MRE）values are in the acceptable range. The determination of the best lean mode period exemplifies the applica-tion of the generalization ability of ANN in the optimization and control of lean burn gasoline engine.

关 键 词：人工神经网络 稀燃汽油机 催化转化效率 

分 类 号：TK411.5[动力工程及工程热物理—动力机械及工程]