臺灣博碩士論文加值系統

目前國內大部分的建築物都是由鋼筋混凝土構成。混凝土經歷了多年的地震與天災，表面時常產生裂縫，若裂縫過深，使得鋼筋曝露於空氣中，則會造成嚴重的安全性問題。因此，如何對於結構物進行安全性的檢測，成為當前重要的課題之一。而其中，非破壞檢測是常用於檢測混凝土結構物安全的技術，該技術中的時間域暫態彈性波法則有易於現地直接量測的優點。然而，使用此技術測量後，仍需要經由專業人員針對訊號進行處理及分析。除此之外，在量測時難精確地抓取到縱波到達時間，加上敲擊時間原點無法完全準確地被記錄，誤差難以避免。
深度學習是人工智慧中的一個分支，被廣泛地應用於電腦視覺領域。其以卷積類神經網路為基礎，可從大量的資料中提取重要的特徵，並學習解決問題的規則。為解決暫態彈性波法偵測裂縫時使用上的限制，本研究依循上述人工智慧在圖像辨識上的成功經驗，結合暫態彈性波法與深度學習技術，以進行混凝土結構之波速、鋼筋與裂縫的偵測。
本研究以有限差分法模擬波傳行為，建立大量資料庫，再藉由切取式窗、正規化、加入隨機直線等前處理，來預訓練卷積類神經網路與殘差網路。接著，將實驗量測所得之訊號加入模型訓練、微調其參數，並逐一探討各項結構與設定對類神經網路的影響，進而設計出最佳化的兩個模型。最後將所得兩個模型進行加總，以減少偏差。
以論文中所設計的最佳類神經網路架構進行之實驗來預測混凝土結構，以裂縫深度4 cm、角度90度與鋼筋直徑13 mm、保護層厚度2 cm之試體進行測試，在量測訊號的裂縫角度誤差為5.9度、裂縫深度誤差為0.38 cm、鋼筋保護層厚度誤差為0.35 cm、鋼筋半徑誤差為0.47 mm、縱波誤差為291 m/s、橫波誤差為194/s。同時，因為前處理中切取視窗的手法，可確保當視窗起點落在10 μs的範圍內時，系統必定能得出穩定、正確的答案，成功克服了傳統量測時的限制。而本文也嘗試以數據可視化的方式，呈現出資料特徵的分布以及模型的可解釋性，使人工智慧理論與波動理論可互相應和。
本研究以暫態彈性波理論為基礎，建立一套使用人工智慧來精準地探測鋼筋混凝土中各項安全相關結構資訊的方法，不僅使操作者能夠更簡便地、即時地對建築品質進行評估與監控，其建立人工智慧模型的方法及最佳化流程也可作為其他領域對於訊號分析之參考。

To date, most of the buildings in Taiwan are made of reinforced concrete. After years of earthquakes and natural disasters, there are often cracks on the surface of the structures. If a crack goes too deep and the reinforcement steels underneath are exposed, it can cause major safety issues. Thus, the evaluation of structures becomes increasingly significant. The non-destructive detection (NDT) technologies are often used for safety inspections of concrete structures. Amongst all the NDT technologies, the time-domain transient elastic wave method has the most straightforward and convenient way of implementation, however, it is difficult for most people to tell the accurate arrival time of the pressure wave. As a result, this method requires largely on the experts’ analyses, and like many other manual methods, accuracy is limited to the capabilities of human beings. In addition, when conducting this method, the time origins are often falsely recorded due to environmental factors.
Deep Learning is a branch of Artificial Intelligence which is widely used in the field of computer vision. Composed of convolutional layers, the Deep Learning model can extract relevant features from a sea of data to achieve its ultimate goal. To improve the transient elastic wave method in the detection of cracks, this research utilizes both methods to carry out the detection of wave velocity, reinforcements, and cracks in the concrete structure.
The Finite Difference Method (FDM) is applied to simulate wave propagation behavior, and to build a large database. Afterwards, such data went through pre-processing using techniques like cutting windows, normalization, and adding random straight lines in order to pre-train two kinds of both Convolutional models and Residual models. Next, these models are trained with experimental signals to fine-tune their parameters. During this process, this research examines the influence on the neural networks caused by different structures and settings, for the purpose of obtaining the two most optimal models, and ultimately add them together to minimize bias.
The best model designed in this research is the one that assembles the 2D Convolutional model and the 2D Residual model. The error of the crack angle is 5.9 degrees, the crack depth error is 0.38 cm, the protection layer thickness error is 0.35 cm, the steel bar radius error is 0.47 mm. The pressure wave error is 291 m/s, and the shear wave error is 194/s. According to the cutting window algorithm, it can be certain that if the starting point of the window falls within the range of 10 μs, this measurement system can always obtain a stable and correct answer, which means this research successfully overcomes the obstacles of the traditional measurement methods. In addition, we also attempts to show the performance by dimensional reduction and model visualization, which help us learn more about the properties of those AI models.
Based on the theory of transient elastic waves, this research establishes a set of methods using the Deep Learning technology to accurately detect various structural information in reinforced concrete. Not only does it enable operators to monitor the quality of various establishments in real time with ease, the processes of optimizing AI models can also be used as a reference for signal analysis in other fields.

致謝 i
中文摘要 ii
ABSTRACT iv
目錄 vi
圖目錄 ix
表目錄 xii
第一章　導論 1
1.1 研究動機 1
1.2 文獻回顧 2
1.3 章節介紹 4
第二章　時間域暫態彈性波法 7
2.1 時間域暫態彈性波法原理 7
2.2 彈性波波速之量測 9
2.3裂縫與鋼筋之量測 10
2.4 有限差分法模擬波波傳行為 11
第三章　人工智慧模型設計 24
3.1 訓練資料的數量與分布 25
3.2 訊號前處理 26
3.2.1 實驗訊號前處理 27
3.2.2 模擬訊號前處理 27
3.2.3 訊號二維化 30
3.3 預訓練卷積類神經網路與殘差網路 31
3.3.1 特徵擷取 32
3.3.2 多任務學習 33
3.4 遷移學習 34
第四章　深度學習模型偵測鋼筋與裂縫之結果 55
4.1 特徵擷取與降維分析 55
4.1.1 t-SNE降維分析之原理 55
4.1.2 以降維分析評估特徵擷取效果 57
4.2 集成學習之最佳模型 58
4.3 擷取視窗起點討論 59
第五章　結論與未來展望 66
5.1 結論 66
5.2 未來展望 68
參考文獻 69

[1]T. Shiotani, M. Shigeishi, and M. Ohtsu, “Acoustic emission characteristics of concrete-piles,” Construction and Building Materials, vol. 13, no. 1-2, pp. 73-85, 1999.
[2]Y. Shigenori, O. Takahisa, S. Mitsuhiro, and O. Masayasu, “Quantitative evaluation and visualization of cracking process in reinforced concrete by a moment tensor analysis of acoustic emission,” Materials Evaluation, vol. 53, no. 6, 1995.
[3]J. S. Popovics and J. L. Rose, “A survey of developments in ultrasonic NDE of concrete,” IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, vol. 41, no. 1, pp. 140-143, 1994.
[4]J. R. Leslie and W. J. Cheesman, “An Ultrasonic Method of Studying Deterioration and Cracking in Concrete Structures,” ACI J. Proc., 46(1), pp.17, 1945.
[5]T. M. Proctor, “An improved piezoelectric acoustic emission transducer,” Acoustical Society of America , vol. 71, no. 5, 1982.
[6]T. M. Proctor, “More recent improvement on the NBS conical transducer,” Journal of Acoustic Emission , vol. 5, pp.134-142, 1986.
[7]M. Sansalone and N. J. Carino, “Impact-Echo Method: Detecting Honeycombing, the Depth of Surface-opening Cracks, and Ungrouted Ducts,” Concrete International, vol. 10, No. 4, pp. 38–46, 1988.
[8]N. J. Carino and M. Sansalone, “Pulse-Echo Method for Flaw Detection in Concrete,” National Bureau of Standards, 1984
[9]C.-C Cheng and M. Sansalone, “Determining the minimum crack width that can be detected using the impact-echo method Part 1: Experimental study,” Materials and Structures , vol. 28, no. 176, pp. 74-82, 1995.
[10]C.-C Cheng and M. Sansalone, “Determining the minimum crack width that can be detected using the impact-echo method. Part II: Numerical fracture analyses,” Materials and Structures , vol. 28, pp. 74-82, 1995.
[11]Y. Lin and M. Sansolone, “Detecting Flaws in Concrete Beams and Columns Using the Impact-Echo Method,” Materials Journal, vol. 89, no. 4, pp. 394-405, 1992.
[12]吳政忠, 方金壽, “暫態彈性波在混凝土品質與裂縫偵測之應用,” 台灣大學應用力學所, 1996.
[13]T.-T. Wu, J.-S. Fang, and P.-L. Liu, “Detection of the depth of a surface-breaking crack using transient elastic waves,” The Journal of the Acoustical Society of America, vol. 97, no. 3, pp. 1678–1686, 1995.
[14]T.-T. Wu, J.-S. Fang, G.-Y. Liu, and M.-K. Kuo, “Determination of elastic constants of a concrete specimen using transient elastic waves,” The Journal of the Acoustical Society of America, vol. 98, no.4, pp. 2142–2148, 1995.
[15]G. E. Hinton and R. R. Salakhutdinov, “Reducing the Dimensionality of Data with Neural Networks,” Science, vol. 313, no. 5786, pp. 504-507, 2006.
[16]A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks,” Advances in Neural Information Processing Systems 25, pp. 1097-1105, 2012.
[17]K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition,” Computer Vision and Pattern Recognition, pp. 1-14, 2015.
[18]C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich, “Going Deeper with Convolutions,” The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1-9, 2015.
[19]K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770-778, 2016.
[20]N. Tajbakhsh et al., "Convolutional Neural Networks for Medical Image Analysis: Full Training or Fine Tuning?," in IEEE Transactions on Medical Imaging, vol. 35, no. 5, pp. 1299-1312, May 2016, doi: 10.1109/TMI.2016.2535302.
[21]S. J. Song and L. W. Schmerr, “Ultrasonic flaw classification in weldments using probabilistic neural networks,” Journal of Nondestructive Evaluation, vol. 11, no. 2, pp. 69-77, 1992.
[22]M. Lorenz and T. S. Wielinga, “Ultrasonic characterization of defects in steel using Multi-SAFT imaging and neural networks,” NDT & E International, vol. 26, no. 3, pp. 127-133, 1993.
[23]A. A. Shah, S. H. Alsayed, H. Abbas, and Y. A. Al-Salloum, “Predicting residual strength of non-linear ultrasonically evaluated damaged concrete using artificial neural network,” Construction and Building Materials, vol. 29, pp. 42-50, 2012.
[24]K. Virupakshappa and E. Oruklu, "Multi-Class Classification of Defect Types in Ultrasonic NDT Signals with Convolutional Neural Networks," 2019 IEEE International Ultrasonics Symposium (IUS), Glasgow, United Kingdom, 2019, pp. 1647-1650, doi: 10.1109/ULTSYM.2019.8926027.
[25]吳政忠, 朱彥達, “以人工智慧預測混凝土之彈性波波速,” 台灣大學應用力學所, 2019.
[26]吳政忠, 洪浚譯, “以人工智慧偵測混凝土結構中之裂縫,” 台灣大學應用力學所, 2019.
[27]Z. Zhang, “Impact force measurement by signal deconvolution. Cornell University, March, 1983.
[28]Y. V. Lange, "The local method of free elastic vibrations and its applications to testing in industry,” NDT international, vol. 18, no. 5, pp, 256-260,1985
[29]M. Sansalone and N. Carino, "Impact-Echo: A Method for Flaw Detection in Concrete Using Transient Stress Waves', NBSIR 86-3452, National Bureau of Standards, Washington, DC," PB 87-10444/AS (National Technical Information Service, Springfield, MA, 1986)1986.
[30]K. Gra, "Wave motion in elastic solids," Ohio State University, 1975.
[31]T.-T. Wu and J.-S. Fang, “A new method for measuring in situ concrete elastic constantsusing horizontally polarized conical transducers,” The Journal of the Acoustical Society of America , vol. 98, no. 4, pp. 2142-2148, 1995.
[32]吳政忠, 童建樺, “彈性波混凝土品質檢測系統之研製與應用,” 台灣大學應用力學所, 2001.