压气机特性的高斯变换多层感知器预测方法  

Compressor performance map prediction with a multilayer perceptron neural network based on Gaussian transformation

作　　者：费景洲[1] 赵冬来[1] 冯永明[1] 

机构地区：[1]哈尔滨工程大学动力与能源工程学院,黑龙江哈尔滨150001

出　　处：《哈尔滨工程大学学报》2012年第8期996-1000,共5页Journal of Harbin Engineering University

基　　金：中央高校基本科研业务费专项资金资助项目(HEUCF110302;HEUCFZ10006)

摘　　要：使用BP(back propagation)神经网络进行压气机特性预测时,容易出现外插精度低和系统稳定性差的问题.根据压气机特性图中等转速线线形分布特点,提出一种新型的多层感知器神经网络,利用高斯函数对数据样本进行变换,提取特性图中各条等转速线间的相似度关系.分析结果表明,该神经网络在预测精度、网络稳定性和逼近能力等方面优于BP神经网络.利用该神经网络对某型压气机特性进行预测,结果表明该神经网络对于样本数据内插值和外插值预测都有比较理想的精度.The main problem in compressor performance map prediction with a BP（back propagation）neural network is the poor extrapolation prediction accuracy and calculation stability.The Gaussian transformation method was proposed to extract the similarity relationship between speed line shapes in a compressor performance map.A new kind of multilayer perceptron neural network was presented in which the data samples were transformed by the Gaussian function.The results show that the neural network is superior to a BP neural network in prediction accuracy,system stability,and approximation ability.The compressor performance map was predicted with the neural network,and satisfactory prediction accuracy was obtained for both interpolation prediction and extrapolation prediction.

关 键 词：压气机 高斯变换 多层感知器 特性预测 

分 类 号：TK47[动力工程及工程热物理—动力机械及工程]