以CNN自動編碼器辨識敲擊回音試驗之異常訊號

The Detection of Anomaly Impact-Echo Signals Using CNN Auto-encoder

林昀儒(Yun-Ru Lin)

指導教授 ： 劉佩玲

國立臺灣大學/工學院/應用力學研究所/碩士(2020年)

https://doi.org/10.6342/NTU202002895

摘要

敲擊回音法為廣泛應用於檢測混凝土結構之非破壞檢測技術。進行混凝土敲擊回音試驗時，在敲擊過程中常常發生因不當敲擊或混凝土表面劣化等因素造成量測到的回音訊號異常。檢測人員常須在檢測位置敲擊多筆訊號，以目測方式檢視訊號是否正常。此作法相當耗時，且要倚賴操作人員之經驗方能進行判讀，若經驗不足的檢測人員可能誤判而存取了無效訊號，導致錯誤的檢測結果。因此判斷訊號是否正常是敲擊回音試驗成功的第一步。
本研究發展一套人工智慧深度學習網路，以自動判別敲擊回音訊號是否正常。本研究所使用之深度學習網路為卷積神經網路自動編碼器模型，該網路對於二維影像辨識能力十分優越。由於敲擊回音訊號為一維時間域訊號，須先轉換為二維影像，本研究嘗試之轉換方式包括: 1. 將一維時間域原始訊號圖視為二維影像。2. 一維時間域訊號經由Gramian Angular Difference Field(GADF) 轉換為二維影像。3. 一維時間域訊號經由Gramian Angular Summation Field(GASF)轉換為二維影像。將二維影像輸入至卷積神經網路自動編碼器模型進行訓練，此模型內部將提取影像特徵，並還原出與原始輸入影像相近的圖像。
本研究的訓練資料為1100筆的正常訊號，測試資料為100筆異常訊號與80筆正常訊號。因為用以訓練自動編碼器模型的1100筆訊號皆為影像皆為正常訊號，故當測試資料為正常訊號，則重建影像會與原圖很相似，但若測試資料為異常訊號，則重建影像會與原圖有很大的差異。據此，本研究建立了一個分類器來分辨正常與異常訊號。
為提升辨識之準確率，本研究探討五種產生輸入影像的方式：1. 訊號長度3 msec之原始訊號圖；2. 訊號長度3 msec訊號轉為GADF二維影像；3. 訊號長度0.08 msec訊號轉為GADF二維影像；4. 訊號長度依敲擊源大小調整再轉為GADF二維影像；5. 訊號長度依敲擊源大小調整再轉為GASF二維影像。第4、5種方式中的訊號長度非固定，當敲擊鋼珠直徑為6mm，長度設為0.08 msec，當鋼珠直徑變化，訊號長度則依比例調整。此調整之目的在確保表面波出現於輸入影像中。以前述5種影像格式分辨正常與異常訊號的準確率分別為65％、71%、94%、100%及99%，其中以第4種格式的準確率最高。因此，將訊號長度依敲擊源大小調整再轉為GADF二維影像，再輸入本研究所開發之人工智慧辨識系統，可自動偵測異常敲擊回音訊號，有效提升敲擊回音試驗之效率與可靠度。

關鍵字

敲擊回音法 ； 非破壞檢測 ； 異常訊號偵測 ； 深度學習 ； 自動編碼器 ； 卷積神經網路

並列摘要

The impact-echo method is effective in non-destructive testing of concrete structures. However, if anomalous signals are measured in the impact-echo test, owing to improper impact or uneven strength of the concrete surface, the follow-up analysis would lead to the wrong result. Therefore, it is a common practice that inspectors conduct multiple tests at the same location to assure that the acquired signals are normal. This is time-consuming and experience-dependent. Worse yet, inexperienced inspectors may make wrong judgments.
The objective of this research is to develop a deep learning neural network to automatically identify anomalous impact-echo signals. Since the convolutional neural network is effective in image recognition, a convolution-based auto-encoder neural network is adopted in this study. To apply the neural network, the one-dimensional time-domain signal is firstly converted into an image using one of the following approaches: 1. use the graph of the original curve directly as the input image; 2. convert the time-domain signal to an image by the Gramian angular difference field (GADF) method; 3. convert the time-domain signal to an image by the Gramian angular summation field (GASF) method. When the image is input into the auto-encoder neural network, the max pooling layers generate down-sampled feature maps, which are then upsampled by the upsampling layers to restore the image. The neural network is trained such that the difference between the output and input images is minimized.
The training dataset constitutes of 1100 normal signals, and the testing dataset constitutes of 100 anomalous signals and 80 normal signals. The auto-encoder model is trained using only normal signals. Hence, if the input image comes from a normal signal, the output image looks similar to the input image. On the other hand, if the input image comes from an anomalous signal, the output image looks different. Finally, a classifier was constructed based on the difference between the input and output images to judge whether the image came from a normal signal or an anomalous signal.
To improve the accuracy of the auto-encoder neural network, 5 ways of generating the input image were discussed in this study: 1. use the graph of the original curve directly as the input image, signal duration = 3 msec; 2. apply GADF to the signal, signal duration = 3 msec; 3. apply GADF to the signal, signal duration = 0.08 msec; 4. apply GADF to the signal, signal duration adjusted according to impact duration; 5. apply GASF to the signal, signal duration adjusted according to impact duration. In the 4th and 5th approaches, the signal duration is set to 0.08 msec when the diameter of the impact steel ball is 6mm. As the diameter varies, the signal duration is adjusted proportionally. The adjustment is made to ensure that the surface wave appears in the image. The accuracies of these approaches are 65％, 71%, 94%, 100%, and 99%, respectively. The fourth approach yields the best result.
In conclusion, the auto-encoder neural network developed in this study can detect anomaly of impact echo signals effectively. This neural network can serve as a powerful tool to improve the efficiency and reliability of impact echo tests.

並列關鍵字

Impact-echo Method ； Non-destructive Test ； Detection of Anomalous Signals ； Deep Learning ； Auto-encoder Network ； Convolution Neural Network

參考文獻

[1] Sansalone, M. and N.J. Carino, Impact-Echo:A Method for Flaw Detection in Concrete Using Transient Stress Waves. NBSIR 86-3452. 1986, Gaithersburg, MD:National Bureau of Standard 222.

[2] Lin, Y., Sansalone M. Carino N. J., Finite Element Studies of the Transient Response of Plates Containing Thin Layers and Voids. J. Nondestructive Evaluation, 1990. 9(1): p. 27-47.

[3] Lin, Y. and M. Sansalone, Transient Response of Thick Circular and Square Bars Subjected to Transverse Elastic Impact. J. Acoustical Society of America, 1992. 91(2): p. 885-893.

[4] Lin, Y. and M. Sansalone, Detecting Flaws in Concrete Beams and Columns Using the Impact-Echo Method. Materials Journal of the American Concrete Institute, 1992(July/August): p. 394-405.

[5] Cheng, C. and M. Sansalone, The Impact-Echo Response of Concrete Plates Containing Delaminations: Numerical, Experimental and Field Studies. Material and Structures, 1993. 26: p. 274-285.