摘要
小样本情况下神经网络模型泛化能力不足的缺陷限制了其在涡喷发动机风车工况建模中的应用。在十组风车工况实验数据的基础上建立了涡喷发动机风车工况的神经网络模型, 并且利用人们对涡喷发动机动静态、相似参数以及剩余功率与加速度的关系等先验知识不断对神经网络的输入变量进行变换, 逐次减少神经网络的训练样本数目, 最终只用一组训练样本就可以训练出泛化能力较强的神经网络模型, 大大提高了小样本情况下神经网络的泛化能力。仿真结果表明, 该方法简单有效。
The deficiency of weak generalization ability in the case of small sample size has restricted neural network's application in modeling of wind milling. Based on ten samples experimental data of wind milling, a neural network's model of wind milling is built. By incorporating priori knowledge of dynamic and static state of rotor, similar parameters and the relationship between residual power and acceleration, the training samples numbers can be decreased step by step. Finally a neural network model of wind milling, which has a good generalization ability, can be set up in the case of just one training sample. The incorporation of priori knowledge greatly improves neural network's generalization ability. Results of the simulation prove that the method is simple and effective.
出处
《推进技术》
EI
CAS
CSCD
北大核心
2005年第2期162-166,共5页
Journal of Propulsion Technology
