一种研究逆向卸荷膜片式减压器稳定性的BP神经网络改进算法  

An improved BP neural network algorithm for researching on stability of reverse unloading diaphragm pressure reducing regulator

作　　者：刘文英[1] 王拥军[1] 王宝山[1] 陈阳[2] 陈经禄 

机构地区：[1]北京航空航天大学数学与系统科学学院,北京100191 [2]北京航空航天大学宇航学院,北京100191

出　　处：《航空动力学报》2017年第5期1241-1249,共9页Journal of Aerospace Power

基　　金：国家自然科学基金(11371044;11101023)

摘　　要：为更好地研究多结构参数耦合变化下减压器PPR(pressure reducing regulator)的稳定性,使用BFGS(Broyden-Fletcher-Goldfarb-Shanno)拟牛顿法替换梯度下降法,实现了基于Wolfe条件的一维线搜索变步长BP(back propagation)算法.结果表明:改进的BP算法使迭代次数减少了1~2个数量级,且易于收敛到最小点.该算法用于逆向卸荷膜片式减压器时,能适应2~3个结构参数的耦合,可预测大于106个数据点的数据集.多结构参数同时变化时,更容易找到使得减压器稳定的结构参数组合.更重要的是这些结构参数同时变化时减压器的稳定性比仅其中一个参数变化时更好.To investigate the stability of PRR （pressure reducing regulator） with adjusting multiple structure parameters simultaneously,BFGS （Broyden-Fletcher-Goldfarb-Shanno） quasi-Newton method and line search with Wolfe conditions were applied to optimize the BP （back propagation） algorithm.Results showed that the improved BP algorithm reduced number of iterations by 1-2 orders of magnitude,making it easy to reach a global minimal point.When the improved BP algorithm was used for reversc unloading diaphragm PRR,it could adapt to coupling of 2-3 structure parameters and predict the data set with more than 106 data points.It was easy to find a combination of structure parameters to stabilize PRR when multiple structure parameters change simultaneously.More importantly,when these parameters changed together,the stability of PRR was much better than the case with change of just one of the parameters.

关 键 词：BP神经网络 拟牛顿法 变步长 减压器(PRR) 稳定性 结构参数 

分 类 号：V434[航空宇航科学与技术—航空宇航推进理论与工程]