在线自适应神经网络算法及参数鲁棒性分析  

An online adaptive neural network algorithm and its parameters robustness analysis

作　　者：王涛[1,2,3] 翟绪恒 孟丽岩[2] WANG Tao;ZHAI Xuheng;MENG Liyan(Key Laboratory of Earthquake Engineering and Engineering Vibration of China Earthquake Administration,Institute of Engineering Mechanics,China Earthquake Administration,Harbin 150080,China;School of Civil Engineering,Heilongjiang University of Science & Technology,Harbin 150022,China;State Key Laboratory of Reduction in Civil Engineering,Tongji University,Shanghai 200092,China)

机构地区：[1]中国地震局工程力学研究所中国地震局地震工程与工程振动重点实验室,哈尔滨150080 [2]黑龙江科技大学建筑工程学院,哈尔滨150022 [3]同济大学土木工程防灾国家重点实验室,上海200092

出　　处：《振动与冲击》2019年第8期210-217,共8页Journal of Vibration and Shock

基　　金：黑龙江省省属高等学校基本科研业务费科研项目(2017);黑龙江科技大学青年才俊培养计划(2017);国家自然科学基金(51408157);黑龙江省青年科学基金(QC2013C055)

摘　　要：为了提高传统BP神经网络在线预测精度和计算效率,提出一种在线自适应神经网络算法。该算法在传统BP网络的输入层和隐含层之间增加一个反馈层,通过存储内部状态增强网络动态映射能力,以提高算法自适应性;同时,在学习阶段采用递推形式在线训练算法权值和阈值,以提高算法计算精度和计算效率。基于两组防屈曲支撑构件拟静力试验数据,在线预测防屈曲支撑恢复力。研究表明:与传统神经网络算法相比,在线自适应网络算法具有更好的在线预测精度和计算效率;通过对网络结构中的输入变量、输入和观测样本、隐含层激活函数等算法参数进行鲁棒性分析,找到算法参数对算法性能的影响规律,给出算法应用时参数选择的建议。In order to improve the on-line prediction accuracy and computational efficiency of the traditional BP neural network, a novel online adaptive neural network algorithm was put forward. A feedback connection layer between the input layer and hidden layer in the proposed algorithm was added on the basis of the BP network to improve the adaptiveness of the algorithm by storing internal state information which can enhance the dynamic mapping capability of network. Meanwhile, weights and thresholds were online trained using the recursive form in the learning phase to improve the calculation precision and calculation efficiency. Then, restoring forces of the buckling-restrained brace (BRB) were online predicted based on two groups of BRBs pseudo-static test data. Results show that the proposed online adaptive neural network algorithm has better on-line prediction accuracy and computational efficiency compared with the traditional BP algorithm. Finally, the robustness of algorithm parameters including the input variable, samples of input and observation and activation function of the hidden layer in the network structure were analyzed.The influence rules of algorithm parameters on the algorithm performance were revealed and the parameter selections suggestions in the algorithm application were given.

关 键 词：在线预测 神经网络算法 恢复力模型 鲁棒性分析 防屈曲支撑(BRB) 

分 类 号：TU317[建筑科学—结构工程]