以人工智慧偵測混凝土結構之裂縫

Abstract

以彈性波為基礎之土木非破壞檢測技術可有效檢測建築物之混凝土品質，其中時間域暫彈性波法可易於直接現地量測，但需要配合專業人員對訊號進行判讀，才可得到精準的波傳時間進行下一步分析。因此本論文將針對此技術提出改善方法，使用人工智慧輔助時間域暫態彈性波法偵測混凝土結構之裂縫位置。 深度學習之類神經網路為人工智慧目前發展最快的領域，其中深度類神經網路與卷積類神經網路為本論文所使用的類神經網路模型。為產生類神經網路模型數據庫，研究中以時間域有限差分法計算大量理論訊號當作數據庫，並將其波動訊號正規化後再輸入到類神經網路模型中進行訓練與測試，實現加快梯度下降求最佳解及提升模型的收斂速度。再逐步探討類神經網路中的隱藏層層數與神經元數量、卷積層與池化層分布及濾鏡大小與數量對模型的影響，設計適合本論文數據庫的最佳類神經網路模型。 最後，以所設計的最佳類神經網路模型進行實驗預測混凝土結構之裂縫。由結果顯示對實驗訊號進行擷取視窗的前處理，可以解決時間原點與雜訊對模型的影響；增加新的感測器資料後，模型得到更多的特徵有助於模型辨識實驗量測訊號，使模型預測裂縫長度與角度誤差由0.51cm與10.79度降低至0.17與4.23度。 總結來說，本研究成功實現以深度學習之類神經網路判讀測混凝土結構之裂縫資訊，不僅提供操作者可以更即時的對建築品質進行監控與評估，亦可以讓沒有專業背景的人進行操作。而條件多樣的訓練數據庫可讓實際量測時有更大的量測範圍，應用上會更為便利。

Nondestructive testing (NDT) on the basis of elastic waves as a perfect solution to monitoring and testing of the quality of reinforced concrete (RC) structures, the transient elastic wave system with the strength of in-situ measurement requires a specialist in charge of signals reading on the basis of precise travel time of wave velocity before further assays can be done. Accordingly, this study proposes a new approach to improve the transient elastic wave system and the author utilized artificial intelligence (AI) to support the transient elastic wave system that detects where surface breaking cracks are in RC structures. Currently, artificial neural networks (ANNs) of deep learning (DL) develop the fastest in the domain of AI where this study employed the ANNs model based on deep neural networks (DNNs) and convolutional neural networks (CNNs). Firstly, to generate the database of the ANNs model, this study obtained the database via the finite-difference methods (FDM) that computed considerable signals in theory. Secondly, normalized the signals and input them in the ANNs model for training and testing, and managed to speed up gradient descent for obtaining the optimum and promoting the convergence rate. Thirdly, how the ANN model was affected by the count of hidden layers and neurons, the distribution of convolution and pooling layers and the size and quantity of filters was gradually investigated. Fourthly, the author developed the optimal neural networks model fitted to the database in this study. Lastly, the ANN model applied to trials that predicted surface breaking cracks in RC structures. Based on the findings, the trial signals processed with windows snapshot helped tackle the impacts of time origin and noise on the model. After adding the new sensor data, the model obtained more characteristics helpful to identify the trial measurement signals. By doing so, the length and angle errors predicted by the model dropped from 0.51cm and 10.79 degrees to 0.17cm and 4.23 degrees. Overall, this study employed the ANNs model of DL to detect surface breaking cracks in RC structures, which not only makes the operator monitor and assess the quality of buildings in a real-time manner but also lets non-professionals operate such system. As the diversified database is broadened to a wider measurement scope for in-situ measurement, such model can be more useful and handy in terms of its application.