摘要
针对参数辨识过程运算时间长的问题,本文提出一种基于卷积神经网络的参数辨识方案。该方案避免了对数值模型的大量迭代,能够根据多个连续时间步的实测系统状态对多参数进行快速估计,实现参数辨识;同时,为了帮助神经网络更好地提取特征,还引入一种双向标准化的方法对数据进行处理。以Lorenz63为实例,对其参数进行了分析、实验。实验结果表明:该方案能够有效地估计当前物理场状态对应的模型参数,并且计算时间仅为传统方法的4%,大大提升了计算效率。
To reduce the time cost in parameter estimation processes,a new parameter estimation scheme based on a Convolutional Neural Network is proposed.This scheme can quickly estimate multiple parameters based on a temporal sequence of system states by avoiding time-consuming iterations of numerical models.At the same time,a two-way standardization method is used to help the neural network extract features better.The scheme has been tested in a Lorenz63 nonlinear system.Results show that it can effectively estimate model parameters corresponding to the current physical state with a calculation time only 4%of the particle swarm optimization method.
作者
武频
常旭婷
郎佳林
潘凯凯
龚思泉
WU Pin;CHANG Xuting;LANG Jialin;PAN Kaikai;GONG Siquan(School of Computer Engineering and Science,Shanghai University,Shanghai 200444,China;State Key Laboratory of Aerodynamics of China Aerodynamics Research and Development Center,Mianyang 621000,China)
出处
《空气动力学学报》
CSCD
北大核心
2021年第4期69-76,共8页
Acta Aerodynamica Sinica
基金
空气动力学国家重点实验室基金资助(SKLA20180303)
上海市自然科学基金(19ZR1417700)
TSP牛顿基金(TSPC1086)。