應用小波神經網路於橋樑實測地震反應之損壞偵測

Abstract

本計畫係應用小波神經網路來發展以實測之地震反應資料偵測橋樑損壞的方法。小波轉換（Wavelet Transform）改善了傅立葉轉換(Fourier Transformation)只能觀察到頻率域資料的特性，其二維(時間和頻率)分析的功能使得信號在時間和頻率域內的變化能同時被偵測到，因此非常適合應用於結構損壞偵測。小波神經網路（Wavelet Neural Network）是根據類神經網路（Artificial Neural Network）架構所發展出來的，由小波轉換取代類神經網路中神經元（Neuron）的轉換函數（Transfer Function）而建構出小波元（Wavelon）。實際之大維度（Dimension）問題，由於可得到的資料很少，若應用小波轉換，則需要大量的小波來分解（Decompose）及合成（Synthesis）訊號。小波神經網路的訓練採用類似於類神經網路常用的BP（Back-Propagation）學習演算法，根據工程問題的輸入與輸出訊號自動調整小波元中的參數，因此可大量減少分解及重建訊號所需的小波數目。並快速而有效地分解及重建訊號。此外，這個計畫的研究成果可提供工程界參考，做為將來結構健康診斷之基本資料。本計劃之流程如下：（一） 利用地震反應訓練建立小波神經網路；（二） 進行數值模擬確認程序可行性；（三） 應用至實測橋 樑地震反應。

The study is an application of the Wavelet Neural Network (WNN) in bridge damage detection after excitation of strong earthquakes based on the recorded structural responses measured in site. The drawback of Fourier transform is that only signal properties in frequency domain can be observed. Wavelet transform improves this drawback, its analysis abilities in both time domain and frequency domain make it possible to detect, at the same time, the signal changes in both time domain and frequency domain. Thus, wavelet transform is highly promising for structural damage detection. WNNs are developed based on the architecture similar to the Artificial Neural Network (ANN). Wavelet transform is then used as the activation function for the ?§Wavelons?? in the wavelet neural network (WNN) instead of ?§neuron?? in the ANN. In most practical situations of large dimension, the available data are sparse. Therefore, considerable wavelons are needed to decompose and synthesize signals by applying wavelet transform. The forward networks and back-propagation based learning algorithm are adopted to converge the WNNs during training, and the parameters of wavelons can be adjusted automatically according to input and output signals of a certain engineering problem. Thus, the number of wavelons for decomposing and synthesizing signals can be drastically reduced, and signals can be decomposed and synthesized rapidly and effectively. Besides, the results of this study can offer to civil engineers as the basic data for structural health monitoring. The procedures of this project is as follows: (1) using structural responses excited by earthquakes to train and construct the WNN; (2) verifying the feasibility of the proposed approach by numerical simulations; and (3) applying the proposed approach to in-situ measured structural responses excited by earthquakes.