以卷積自動編碼器進行敲擊回音深度頻譜斷層掃瞄之裂縫偵測

Abstract

敲擊回音法為廣泛應用於混凝土結構之非破壞檢測技術，主要是藉由外力敲擊待測物產生應力波後，將結構物表面反應的時間域訊號轉換成頻率域訊號，並根據頻譜中的尖峰頻率計算出潛在缺陷之對應深度。當量測範圍擴大時，深度頻譜斷層掃瞄法不僅能夠同時對比測線上的多筆訊號，也能更直觀地了解結構內部情況。然而，敲擊回音訊號通常會摻雜雜訊以及來自邊界的反射訊號，訊號頻寬也會受敲擊源影響，使深度頻譜斷層掃瞄影像出現雜訊、干擾與暗帶，即使具經驗的檢測人員在判讀影像時都會遭遇困難。 本研究嘗試發展一套能夠使斷層掃瞄影像中的裂縫區域更明顯之深度學習模型，並由輸出影像進一步標記出裂縫位置、尺寸及走向。本研究用以強化影像的深度學習模型為卷積自動編碼器(Convolutional Autoencoder, CAE)，CAE憑藉著非監督式學習之演算法特性，對於影像的異常檢測具有良好的效果。本研究先以有限元素分析軟體模擬無裂縫之半無限域，將受6mm鋼珠敲擊之回音訊號繪製斷層掃瞄影像後用以訓練CAE模型，使該模型具備還原無裂縫斷層掃瞄影像之能力。將內含裂縫之斷層掃瞄影像輸入訓練完成之CAE模型後，該模型便會盡可能地還原無裂縫區域的影像。將輸入影像扣除輸出之重構影像所得到的殘差影像便可清楚顯示裂縫區域。接著，本研究先描繪出殘差影像中各亮帶之邊界，再由各亮帶範圍內之像素亮度加總，以選出最可能的裂縫區域，再透過加權線性迴歸以一條線段標示出裂縫位置及走向。 為提升CAE模型之實用性與偵測裂縫之精度，本研究分別對訓練影像添加不同大小的雜訊、選擇不同激勵函數及改變CAE模型架構深度訓練模型，並以殘差影像之背景/缺陷亮度比做為評比指標，以找出較優的CAE模型。經過數值模擬影像測試，當編碼器和解碼器各為3層，以整流線性單位(Rectified Linear Unit, ReLU)函數作為激勵函數，並以無雜訊及添加3%雜訊之影像進行訓練而成的CAE模型表現最佳。 在標記裂縫影像部分，本研究分別模擬不同尺寸、深度及走向之裂縫斷層掃瞄影像。大多數裂縫影像所標記之裂縫深度誤差皆在1cm以內，在水平跨度誤差則大部分在5cm以內，惟有深度6cm和16cm之水平裂縫的水平跨度誤差較為不規律。在標記實驗裂縫影像部分，深度10cm之水平裂縫的水平跨度為32cm，迴歸標記的裂縫水平跨度為26cm，誤差為6.0cm，深度誤差為0.1cm；深度12~8cm之傾斜裂縫的水平跨度亦為32cm，迴歸標記的裂縫水平跨度為36cm，誤差為4.0cm，深度誤差則為0.2cm。 為測試模型的泛用性，本研究嘗試以不同大小鋼珠作為敲擊源進行試驗。測試結果發現當鋼珠直徑不為6mm時，只需將斷層掃瞄影像之強度及垂直軸以鋼珠直徑等比例調整，CAE模型便能有效地消除不必要雜訊，將斷層掃瞄影像中的裂縫區域明顯標示出來。 綜上所述，本研究訓練之CAE模型能夠有效消除敲擊回音斷層掃瞄影像中的背景訊號及雜訊干擾並凸顯裂縫，而線性迴歸模型也能大致偵測並標記出裂縫的位置及走向，可作為檢測人員判讀混凝土內部情況的有效工具。

The impact-echo method is a widely used nondestructive testing technique for concrete structures. It involves generating stress waves in the test object by applying an impact force and converting the resulting surface response from the time domain into the frequency domain. The depth of a potential defect is then calculated based on the peak frequency in the spectrum. By conducting a series of impact-echo tests along a test line, one may use depth spectral tomography to produce a more intuitive view of the internal structure. However, depth spectral tomogram often contain noise, interferences, and dark zones. Even experienced inspectors may have difficulty interpreting the images. This study aims to develop a methodology to enhance the crack image in the impact-echo depth spectral tomograms. A convolutional autoencoder (CAE) is adopted to this end because it can detect anomalies in images.The finite element analysis has been used to simulate the response of a perfect half-space due to the impact of a 6mm steel ball. By training the CAE with simulated tomograms, the model learns their features and is able to reconstruct them. As the tomogram of the specimen that contains cracks is input into the CAE, the model tries to reproduce the crack-free image. After subtracting the reconstructed tomogram from the input tomogram, the unwanted noise, interferences, and dark zones are diminished in the residual tomogram, but the crack image remains intact and noticeable. Apply edge detection to find the bright zones in the residual tomogram. Then, use the weighted linear regression to draw a line segment that denotes the crack. The influence of several factors on the performance of CAE models has been studied, including the noise level in the training tomograms, choice of activation function, and depth of the CAE model architecture. The CAE models are compared based on the residual images' background-to-defect brightness ratio. The CAE model with three layers for both encoder and decoder, using rectified linear unit (ReLU) as the activation function, trained with noise-free and 3% noise-added tomograms, yields the best results. Labeling cracks by weighted linear regression has been conducted on the residual tomograms of specimens with various crack sizes. For the simulated tomograms, depth errors are primarily within 1cm, and the errors in crack horizontal spans are generally within 5cm for 32cm span cracks. For the experimental tomograms, labeling of a 10cm depth, 32cm span horizontal crack exhibits depth and span errors of 0.1cm and 6.0cm, respectively; labeling of a 32cm span inclined cracks with depth varying from 8 to 12 exhibits depth and span errors of 0.2cm and 4.0cm, respectively. To examine the versatility of the CAE model, tomograms generated by various impact balls have been tested. It is found that the results are not satisfactory as the diameter of the steel ball changes. However, if the intensity and vertical scale of the tomograms are scaled according to the diameter, the model can effectively enhance the crack images in the residual tomograms. In summary, this study's CAE model can effectively eliminate unwanted noise, interference, and dark zones from the impact-echo depth spectral tomogram. As such, it highlights the crack image. The weighted linear regression can roughly label the cracks in the residual tomogramss. It can assist inspectors in interpreting the internal conditions of concrete structures.