基于动态递归神经网络的HCCI发动机燃烧相位辨识模型  

A Gasoline HCCI Combustion Phasing Observer Based on a Dynamically Recurrent Neural Network

作　　者：谢辉[1] 孙艳辉[2] 夏超英[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]天津大学自动化学院,天津300072

出　　处：《内燃机学报》2007年第4期352-357,共6页Transactions of Csice

基　　金：国家自然科学基金资助项目(50476064);国家重点基础研究规划资助项目(2001CB209204)

摘　　要：为了实现HCCI汽油机闭环反馈控制,提出了一种利用动态递归神经网络从气缸压力信号在线辨识燃烧相位CA50(燃烧50%累积放热量的曲轴转角)的方法。该方法采集上止点附近40°CA范围的气缸压力信号,经过归一化和主元素法降维处理后,得到一个由9个特征数构成的时间序列。一个Elman动态递归神经网络以该序列为输入,计算出燃烧相位CA50。以基于全可变气门机构的汽油HCCI发动机为对象,选取了台架试验中4个典型的HCCI动态变负荷过程数据,其中一个作为训练样本,另外3个作为测试样本。测试结果表明:该方法对HCCI动态过程的燃烧相位CA50预测误差小于0.25°CA;与BP网络和RBF网络相比,具有更低的误差和更强的泛化能力;与直接热力学计算方法相比,具有突出的抗干扰性和容错能力。In order to implement close-loop feedback control of HCCI gasoline engine, a new observer model, called Elman observer, based on a dynamically recurrent neural network for on-line detecting combustion phase CA50 is presented. The cylinder pressure signal over 40° CA near TDC is collected. After being normalized and deducted based on principal component analysis, a time series consisting of 9 scores is acquired, which are inputted into an Elman network to calculate combustion phase CA50. 4 samples from different HCCI load transient procedures of a HCCI gasoline engine, which is equipped with a fully variable valve actuating system, are used for training and testing the Elman Observer. One sample is used as the training sample set, and the other three are used as the test sample sets. The results show that the detecting error of combustion phase is less than 0. 25° CA. Compared with BP network and RBF network, Elman observer gives the lower error and stronger generalization ability. Compared with the method of thermodynamic calculation, the Elman observer shows the stronger ability of anti-disturbing and fault toleralice.

关 键 词：HCCI汽油机 燃烧相位观测 动态递归神经网络 

分 类 号：TK412.2[动力工程及工程热物理—动力机械及工程]