應用類神經網路於橋樑實測地震反應之模態識別

Abstract

由文獻回顧，知自從1943年McCulloch and Pitts提出類神經網路之數學模型後，於80年代開始類神經網路才逐漸被受到重視。由於類神經網路具有可訓練性及容錯性，其已成功地被應用於不同領域，包含自動控制、最佳化問題、語言及影像之判識、氣象預測、結構反應模擬。從文獻中可發現類神經網路應用於結構系統之模態識別則不多見。 本研究即擬以利用倒傳遞神經網路分析橋樑地震反應量測數據，利用訓練所得之權重矩陣直接估算結構系統之振態頻率、阻尼比及模態。本研究中擬分別利用batch learning及per-example learning兩模式訓練網路，並利用敏感度分析減少輸入層中不必要之神經元（減少系統識別中虛擬振態之產生）。此分析模式將先以數值模擬驗證，再應用於實測橋樑地震反應。

It is known from article review that McCulloch and Pitts first proposed a mathematical model of neural networks in 1943. However, the neural computing did not catch the attention of researchers until 80’s. Due to the capacity of training and the high tolerance to partially inaccurate data, neural networks have been successfully applied to various fields such as automatic control, optimization, speech and image recognition, weather prediction, and prediction of structural responses. Nevertheless, it is hardly found the applications of neural networks to determine the dynamic characteristics of structures in the published work. The main purpose of this project is to extend the application of neural network to identify the dynamic characteristics of bridges from their earthquake responses. The natural frequencies, modal damping, and mode shapes can be directly determined from the weighting matrices of a neural network. In this research, the neural network will be trained by using batch learning and per-example learning to investigate the effects of the two types of learning process on establishing an appropriate neural network. Furthermore, sensitivity analysis will also be carried out to cut out the unneeded neuron in the input layer. The proposed procedure will be verified by numerical simulation, then, applied to processing the measured responses of bridges.