在线自适应神经网络算法及参数鲁棒性分析
|
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.
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.
引文
[1]WU B,WANG T.Model updating with constrained unscented Kalman filter for hybrid testing[J].Smart Structures and Systems,2014,14(6):1105-1129.
[2]王涛.基于模型更新的土木结构混合试验方法[D].哈尔滨:哈尔滨工业大学,2014.
[3]KWON O,KAMMULA V.Model updating method for substructure pseudo-dynamic hybrid simulation[J].Earthquake Engineering and Structural Dynamics,2013,42(13):1971-1984.
[4]HASHEMI M,MASROOR A,MOSQUEDA G.Implementation of online model updating in hybrid simulation[J].Earthquake Engineering and Structural Dynamics,2014,43(3):395-412.
[5]YANG W J,NAKANO Y.Substructure online test by using real-time hysteresis modeling with a neural network[Z].Advances in Experimental Structural Engineering,2005.
[6]白建方,白国良,史冠卿.基于人工神经网络的钢筋混凝土异型节点滞回模型[J].世界地震工程,2004,20(2):50-54.BAI Jianfang,BAI Guoliang,SHI Guanqing.Artificial neural networks based modeling of hysteretic behavior for RCabnormal joints[J].World Earthquake Engineering,2004,20(2):50-54.
[7]YUN G Y,GHABOUSSI J,ELNASHAI A S.A new network-based model for hysteretic behavior of materials[J].International Journal for Numerical Methods in Engineering,2008,73(4):447-469.
[8]张健.自适应子结构拟动力试验方法[D].哈尔滨:哈尔滨工业大学,2010.
[9]周大兴,闫维明,陈彦江,等.神经网络在振动子结构试验中的应用[J].振动与冲击,2011,30(12):14-18.ZHOU Daxing,YAN Weiming,CHEN Yanjiang,et al.Application of neural network in real-time substructure testing with shaking table[J].Journal of Vibration and Shock,2011,30(12):14-18.
[10]王涛,翟绪恒,孟丽岩,等.基于在线神经网络算法的混合试验方法[J].振动与冲击,2017,36(14):1-8.WANG Tao,ZHAI Xuheng,MENG Liyan,et al.Hybrid testing method based on an online neural network algorithm[J].Journal of Vibration and Shock,2017,36(14):1-8.
[2]王涛.基于模型更新的土木结构混合试验方法[D].哈尔滨:哈尔滨工业大学,2014.
[3]KWON O,KAMMULA V.Model updating method for substructure pseudo-dynamic hybrid simulation[J].Earthquake Engineering and Structural Dynamics,2013,42(13):1971-1984.
[4]HASHEMI M,MASROOR A,MOSQUEDA G.Implementation of online model updating in hybrid simulation[J].Earthquake Engineering and Structural Dynamics,2014,43(3):395-412.
[5]YANG W J,NAKANO Y.Substructure online test by using real-time hysteresis modeling with a neural network[Z].Advances in Experimental Structural Engineering,2005.
[6]白建方,白国良,史冠卿.基于人工神经网络的钢筋混凝土异型节点滞回模型[J].世界地震工程,2004,20(2):50-54.BAI Jianfang,BAI Guoliang,SHI Guanqing.Artificial neural networks based modeling of hysteretic behavior for RCabnormal joints[J].World Earthquake Engineering,2004,20(2):50-54.
[7]YUN G Y,GHABOUSSI J,ELNASHAI A S.A new network-based model for hysteretic behavior of materials[J].International Journal for Numerical Methods in Engineering,2008,73(4):447-469.
[8]张健.自适应子结构拟动力试验方法[D].哈尔滨:哈尔滨工业大学,2010.
[9]周大兴,闫维明,陈彦江,等.神经网络在振动子结构试验中的应用[J].振动与冲击,2011,30(12):14-18.ZHOU Daxing,YAN Weiming,CHEN Yanjiang,et al.Application of neural network in real-time substructure testing with shaking table[J].Journal of Vibration and Shock,2011,30(12):14-18.
[10]王涛,翟绪恒,孟丽岩,等.基于在线神经网络算法的混合试验方法[J].振动与冲击,2017,36(14):1-8.WANG Tao,ZHAI Xuheng,MENG Liyan,et al.Hybrid testing method based on an online neural network algorithm[J].Journal of Vibration and Shock,2017,36(14):1-8.