Development of system identification algorithm for health monitoring of infrastructures

通过学习用系统动力学模型表示的非线性动态结构，提出了一种自关联滞回特性的系统辨识技术。对此，以神经网络的学习为手段。该准则函数由输出误差平方和定义，并考虑了输入结构的地震运动能量。为了获得有规律的定向学习，有效地利用结构学习作为参数和系统辨识问题的一种替代方法。提出了一种全局和局部迭代的方法，以保证循环网络学习的稳定性和快速性。提出了一种应用小波变换识别非退化型滞后恢复系统参数的方法。利用该方法，可以根据模型参数稳定收敛到最优阶段的参数，识别出等效于任何滞后系统的非退化型非线性模型。数值算例对三自由度模型结构的动力参数辨识取得了满意的结果。提出了利用重要采样和抑制采样滤波器对滞回退化多自由度剪力梁结构进行有效识别的方法。通过与常规方法的比较，分析了该方法的效果和精度。利用收敛参数，将模拟响应和滞回力特性与已知的真实响应进行比较。得到了稳定解，并快速收敛到最优解。数值算例表明，该方法是一种有效的参数辨识工具。

System identification technique to autoassociate hysteretic characteristics is proposed by learning a nonlinear dynamic structure which is represented by a system dynamics model. As this regard, neural network's learning serves as the means. The criterion function is defined by the sum of squared output errors which takes into account the energy of earthquake motion input to the structure. In order to obtain a regularly directed learning, effective use of the structural learning is considered as an alternative method for parameter and system identification problems. A global and local iteration procedure is proposed to obtain stable and fast convergency in the learning of recurrent networks.A method is developed to identify parameters on a hysteretic restoring system of non-degrading type by applying the wavelet transform. By this method, a nonlinear model of non-degrading type equivalent to any hysteretic system may be identified in terms of the model's parameters at the stage of their stable convergency to optimal ones. Numerical examples give satisfactory results to identify dynamic parameters of model structure with 3-degree-of-freedom.Effective identification scheme for hysteretic, degrading multi-degree of freedom shear beam structures is developed using importance sampling and rejection sampling filters. The effects and the accuracy of this procedure are analyzed by comparing with the conventional methods. Using converged parameters resimulated responses and hysteretic force characteristic are compared with known true ones. Stable solutions as well as their fast convergency to the optimum ones are obtained. It is found by numerical examples that the proposed method is a powerful tool for parameter identification.