類神經網路於含裂縫簡單鋼架檢測之應用

Abstract

摘 要 裂縫為結構物破壞之初期徵兆，在未知構體何時發生破壞的情況下，對結 構物進 行檢測為一種防患未然的重要工作。 由於傳統非破壞檢測方法為結構物局部檢測，基於經濟、快速與全面性結 構物檢 測考量，本研究以有限元素法建立不同裂縫型式之鋼架模型，應 用ANSYS50A套裝軟體 作數值分析，並以真實鋼架進行振動實驗，以驗證 電腦模型之正確性。 本研究將數值分析之大量動態資 料經裂縫特徵化處理以後，分成學習與測試兩組 ，利用此學習資料，以 倒傳遞類神經網路(Back-Propagation Network,BPN) 經多次 試誤及調 整參數後，使類神經網路得到良好之學習收斂效果，再以測試資料驗證其 準 確性。本研究以振態網路辨識裂縫位置；以頻率網路辨識裂縫深度， 其中振態網路測 試最大誤差為5.10%；頻率網路測試最大誤差為2.78%， 藉由此良好之測試結果可作為 鋼架結構物之破壞診斷模式。 ABSTRACT Cracks are the signal to show that structures had been damaged. The routineinspection of structures is a necessary approach to prevent the calamity happen. Due to the traditional Nondestructive detecting methods for structure inspection are used only for local area, the establishment of the global inspection method is the goal of this research. The numerical modal of dynamic behavior for simple steel frame was setup by the finite element package ANSYS50A and verified by the modal testing. The numerical data of the dynamic behavior were processed to signify its crack characteristics and classified into learning data group and verifying data group. The Back Propagation Neural Network was utilized to learn how to identify the depth and position of the crack from the learning data by the try and error method and to adjust the parameters of the network. The verifying data were used to test the accuracy of the system. Two networks were used in this research, one for the depth of the crack by the network of the frequency that have a 2.78% maximum error and the other for the position of the crack by the network of the modal shape that have a 5.10% maximum error. The above excellent results show that this system can identify the depth and position of the crack with an acceptable accuracy.