发动机动态特性组合神经网络建模新方法  

A New Modeling Method for Engine Dynamic Characteristics Based on Assembled Neural Networks

作　　者：刘树成[1] 魏巍[1] 杨阳[2] 闫清东[1] 

机构地区：[1]北京理工大学机械与车辆学院,北京100081 [2]中国北方车辆研究所,北京100072

出　　处：《北京理工大学学报》2014年第11期1130-1134,共5页Transactions of Beijing Institute of Technology

基　　金：国家部委基金资助项目(VTDP3101;40402050202);国家自然科学基金资助项目(50905016)

摘　　要：针对现有发动机组合神经网络建模方法对不同数组结构的样本数据泛化能力较差的不足,提出一种多步线性插值法的组合神经网络建模方法.该方法基于有限元建模思想,以具有丰富样本数据的某一维输入量构造网格线,对多维输入样本空间进行划分.在网格线上,样本数据按照BP算法对网络模型进行训练,得到高精度神经网络函数,而在网格线中间,所求输出根据相邻的两条网格线的神经网络函数进行多步线性插值.与传统组合神经网络建模方法的对比结果表明,在处理不同数组长度的多维发动机动态特性试验数据方面具有很好的适应能力.Focusing on the defects of current assembled artificial neural network（ANN） models, its weak generalization ability for engine experiment sample data of different array structure, multi-step linear interpolation method （MLIM for short）, a new assembled ANN modeling method, was put forward, which was based on finite element method. In MLIM, using one- dimensional input vector with abundant sample data, some mesh lines were set up to make a division of the input space. The sample data on these mesh lines was brought in BP neural model training process, from which some high-precision artificial neural network functions were obtained. Output of sample data between m^shing lines was multi-step linearly interpolated by the most two neighboring mesh line ANN function value. Compared with traditional assembled neural network modeling methods, MLIM has good adaptability in processing multi-dimensional engine dynamic characteristic testing data with different input array length.

关 键 词：发动机 组合神经网络 多步线性插值法 动态特性 

分 类 号：U27[机械工程—车辆工程]