以人工智慧預測混凝土之彈性波波速

Abstract

混凝土在大部分的土木結構中扮演不可或缺的角色，非破壞檢測技術即提供了對於新建、現有、老舊或災後建築的混凝土品質監控及評估有效且可行的解決方案。其中時間域暫態彈性波法對於彈性波波速的量測具有直接且易於現地量測之特性，然而通常需要有經驗的專業操作者進行訊號處理及分析，且間接求得橫波波速的方法會導致計算材料常數時有誤差產生。 深度學習為人工智慧的其中一項技術，可以自行提取資料之重要特徵，學習解決問題的規則。本研究結合時間域暫態彈性波法之訊號量測方法與人工智慧預測混凝土彈性波波速。將一筆訊號以包含縱波、橫波及表面波資訊的視窗的方式輸入模型供其學習，首先設計各種類神經網路，包括深度類神經網路、卷積類神經網路與遞迴類神經網路之最佳架構，其不論與過去之多層感知器亦或是傳統量測方法相比，於彈性波波速及材料常數之計算都有較好的表現。而其中卷積類神經網路局部擷取訊號特徵的模式為最適合預測彈性波波速之模型，訓練準確率與測試準確率分別為99.84%與99.80%。 實際量測混凝土試體後，對模型進行調整，增加訊號前處理以過濾不重要資訊，避免模型學習過多無關特徵，有助於辨識實驗訊號，使得實驗之平均預測準確率由92.08%上升至96.06%。比較理論及實驗訊號發現，當視窗抓取數量為25個以上時，類神經網路模型之預測波速和視窗位置無關，另外還需根據模型的使用環境來建立訓練的理論訊號數據庫。由於一筆位移訊號事實上包含了三種彈性波資訊，將人工智慧的訓練過程以人類的學習角度來思考，若能夠告訴人工智慧模型越多知識，其學習效果就能越好。本研究設計之最佳類神經網路模型為卷積類神經網路模型，其對於實驗訊號彈性波波速預測之縱波準確率為96.38%、橫波準確率為97.28%、表面波準確率為94.52%，平均準確律為96.06%。

Concrete plays an important role in civil structures. The nondestructive test provides an effective method to monitor and evaluate the quality of concrete in new, existing, or post-disaster building. Transient elastic wave method is one technique of the nondestructive tests by measuring elastic wave velocities of concrete. It can be applied directly and simply in site, but it usually needs experts to process and analyze signals. Moreover, Lamé constants are estimated by wave velocities and some errors appear because transverse velocity is calculated indirectly. Deep learning is a technique of artificial intelligence (AI). It can extract important features by learning the rules of data and can solve many kinds of problem. In this thesis, AI models were developed to predict elastic wave velocities from results of transient elastic wave method. Wave signals selected from a window were input into the model to train the AI algorithm to analyze information of longitudinal wave, transverse wave and Rayleigh wave. First, the deep neural network, convolutional neural network and recurrent neural network were designed and optimized. The AI algorithms performed better than the traditional method in calculating elastic wave velocities and Lamé constants. The convolutional neural network has the best performance, whose training accuracy and testing accuracy are 99.84% and 99.80%, respectively. The AI models were modified further after measuring real concrete specimen. Wave signals with pre-processing can remove unnecessary information and avoid the model learning unrelated features. Pre-processing helps AI model identifying experimental signals effectively. The experimental accuracy raised from 92.08% to 96.06%. After comparing theoretical and experimental signals, we found that the prediction of wave velocities is converged on more than 25 different starting-positions of windows. We also designed the training dataset from the real operational condition. We got a better result if we trained the AI model to solve three kinds of elastic wave information in the displacement signals at the same time. In this study, the best AI model is a convolutional neural network, whose experimental accuracy of longitudinal wave velocity, transverse wave velocity, and Rayleigh wave velocity are 96.38%, 97.28% and 94.52%, respectively, and the average experimental accuracy is 96.06%.