Investigation of Inverse Analysis and Neural Network Approaches for Identifying Distributed Load using Distributed Strains
Investigation of Inverse Analysis and Neural Network Approaches for Identifying Distributed Load using Distributed Strains
复制标题
DOI：
10.2322/tjsass.62.151
复制
发表时间：
2019-01-01
影响因子：
1.1
通讯作者：
Nakamura, Toshiya
中科院分区：
文献类型：
--
作者：
Wada, Daichi;Sugimoto, Yohei;Nakamura, Toshiya
本文提出并研究了两种利用应变测量值识别平板载荷分布的方法。一种方法是利用载荷和应变关系的逆矩阵的逆分析，另一种方法是使用应变作为输入和载荷作为输出来训练神经网络的神经网络方法。对于这两种方法，我们提出了一种方法，使用压力离散化映射来表示作为一组离散压力值的负载分布。该方法使载荷识别适用于任意剖面的载荷分布。为了检验和验证性能，我们进行了数值模拟和实验。数值仿真结果验证了这两种方法的有效性，但当应变测量误差存在时，逆方法的识别结果是不稳定的。另一方面，神经网络方法通过使用包括人工应变误差的数据来训练神经网络，显示出对应变误差的高鲁棒性。基于计算结果，我们讨论了载荷识别方法的适用性。
We propose and investigate two approaches to identify load distributions on a fiat panel by using strain measurement values. One approach is an inverse analysis that utilizes the inverse matrix of the load and strain relationship, and the other is a neural network approach that trains a neural network using strains as input and loads as output. For both approaches, we propose a method using a pressure discretization map to represent the load distributions as a set of discrete pressure values. This method makes load identification applicable to load distributions with arbitrary profiles. In order to examine and verify the performance, we conducted numerical simulations and an experiment. Numerical simulation results verified both approaches; however, identification results using the inverse approach were unstable when the strain measurement error existed. On the other hand, the neural network approach showed high robustness to the strain errors by training neural networks with data including artificial strain errors. Based on the results, we discuss the applicability of the load identification approaches.