應用深度學習判讀裂縫與鋼筋之敲擊回音小波譜

Abstract

敲擊回音法屬於非破壞性檢測的應用，並經常使用於檢測混凝土結構。此方法在混凝土試體表面以鋼珠敲擊激發應力波，再量測待測試體表面之位移響應可得其時間域訊號，再對訊號進行傅立葉轉換，故從原先的時間域轉為頻率域，判讀頻譜中回音尖峰的頻率，推測出試體中缺陷的深度。然而只透過頻譜判斷缺陷是裂縫或鋼筋相當不易，在頻譜上裂縫和鋼筋的回音尖峰頻率相同，若將鋼筋的回音訊號誤判為裂縫，對結構的安全性會產生影響，因此正確判斷出回音訊號為裂縫或鋼筋至關重要。 小波轉換是一種將時間域訊號轉換至時頻域的方法，小波譜可以表現出訊號中各種頻率分量隨時間的變化。由於小波譜中的裂縫與鋼筋回音的強度與能量消散速度不同，因此本研究期望由小波譜辨識出裂縫與鋼筋的回音。不過這種辨識的標準也不易定義，即使對富有經驗的檢測人員也是極大的挑戰。由於深度學習中的卷積神經網路(Convolutional Neural Network, CNN)具有極強的圖像辨識能力，因此本研究擬建立一個CNN 分類模型，判別裂縫與鋼筋的敲擊回音小波譜，再建立一個CNN 迴歸模型，以預測裂縫或鋼筋之深度。 本研究之CNN 分類模型與迴歸模型均是以半無限域、內含裂縫或鋼筋的數值模擬訊號進行訓練，再以數值模擬訊號及實驗訊號進行測試，其中數值模擬訊號是無限平板內含裂縫或鋼筋的敲擊回音訊號，實驗則在多個混凝土試體上進行。訓練分類模型時考慮輸入影像色階、模擬訊號加入雜訊的方式及模型架構之影響。訓練迴歸模型則直接採用最佳分類模型之架構，但考慮不同訓練集的影響。 經測試後，最佳的 CNN 分類模型架構如下：CNN 架構為5 層卷積層、輸入影像採用 Jet 色階、訓練集訊號加入標準差為 5%表面波振幅的白雜訊。CNN 分類模型辨識模擬數據的準確率為86.9%，辨識實驗數據的準確率是99.7%。 NN 迴歸模型的準確率與訓練集有關，只要在訓練集加入足夠多種深度的資料，迴歸模型便能由模擬數據準確預測出的裂縫或鋼筋深度，但實驗測試結果還是存在較大的誤差。CNN 迴歸模型在數值測試中的平均絕對誤差皆在 1.1cm以內；在實驗測試中，預測內含深度6、10、12、25cm 水平裂縫之實驗數據之平均絕對誤差分別為0.13、3.42、0.46，3.82cm，預測內含深度8~12cm 傾斜裂縫之平均絕對誤差為3.76cm，預測深度6cm 鋼筋為0.65cm。 綜上所述，本研究建立的CNN 分類模型能有效辨識裂縫與鋼筋的敲擊回音小波譜，CNN 迴歸模型可能需要在訓練集中加入更多數據以提高其準確率。

The impact-echo method is a non-destructive testing application and is often used to test concrete structures. In this method, the surface of the concrete specimen is struck by a steel ball to generate stress waves, and the displacement response of the surface of the specimen is measured to obtain the time domain signal. Then, the Fourier transform is applied to convert the signal from the time domain to the frequency domain. Since the existence of inclusion would result in echo waves, one may use the frequency of the echo peak in the spectrum to estimate the depth of the inclusion in the specimen. However, it is difficult to determine whether the echo comes from a crack or rebar merely by examining the Fourier spectrum. Since cracks and rebars have opposite effects on structural safety, it is essential to find a way to differentiate crack and rebar echoes. Wavelet transform is a method of converting a signal from the time domain to the time-frequency domain, and the wavelet spectrum can show the variation of various frequency components in the signal over time. Since the intensity and energy dissipation rate of the crack and rebar echoes in the wavelet spectrum are different, it is expected that the crack and rebar echoes can be recognized from the wavelet spectrum in this study. However, the criteria for such recognition are not easy to define, which is a great challenge even for experienced inspectors. Since the Convolutional Neural Network (CNN) in deep learning has a strong image recognition capability, this study proposes to build a CNN classification model to recognize the wavelet spectrum of crack and rebar echoes, and then build a CNN regression model to predict the depth of cracks or rebars. The CNN classification model and the CNN regression model are trained with a numerical simulation of a half space domain with cracks or rebars, and then tested with a numerical simulation signal and an experimental signal, where the numerical simulation signal is an infinite plate with cracks or rebars, and the experiments are carried out on a number of concrete specimens. In the training of the CNN classification model, the effects of the color scale of the input image, the way of adding noise to the simulated signal, and the model structure were considered. The training of the CNN regression model is directly based on the structure of the best CNN classification model, but takes into account the effect of different training sets. After testing, the best CNN classification model architecture is as follows: the CNN architecture is a 5-layer convolutional layer, Jet color scale is used for the input image, and white noise with a standard deviation of 5% of the surface wave amplitude is added to the training set signals. the CNN classification model recognizes the simulated data with 86.9% accuracy, and recognizes the experimental data with 99.7% accuracy. The accuracy of the CNN regression model is related to the training set, as long as the training set contains enough data of various depths, the CNN regression model can accurately predict the depths of cracks or rebars from the simulated data, but the experimental test results still have a large error. the mean absolute error of the CNN regression model in numerical tests is within 1.1cm; the mean absolute error of the CNN regression model in experimental tests are 0.13cm, 3.42cm, 0.46cm, and 3.82cm for predicting horizontal cracks of 6cm, 10cm, 12cm, 25cm, and the mean absolute error for predicting inclined cracks of 8~12cm is 3.76cm respectively, and for the predicted 6cm rebar, the mean absolute error is 0.65 cm. In summary, the CNN classification model developed in this study is effective in recognizing the impact-echo wavelet spectrum of cracks and rebars, and the CNN regression model may need to enhance more data in the training set to improve its accuracy.