- 學位論文
以深度學習方法判讀敲擊回音時頻圖
The Interpretation of Impact-Echo Spectrogram Using Deep Learning Methods
摘要
敲擊回音法為一種廣泛應用於檢測混凝土結構的非破壞檢測技術。以敲擊回音法量測之時間域訊號藉由傅立葉轉換可轉換為頻率域訊號，具有較高能量的頻率分量會在傅立葉頻譜中形成尖峰，檢測員可以據此來檢測混凝土中裂縫的位置。 然而，不僅來自裂縫的回波會在頻譜中形成尖峰，模態振動也是如此。表面波也在頻譜中也會形成一個駝峰。振動之尖峰和表面波之尖峰的存在有時會妨害時頻圖中裂縫回波的檢測，因為與振動和表面波相比，裂縫回波通常較弱。因此，本研究使用時頻分析將時間訊號轉換為時頻圖。由於回波、表面波和模態振動的持續時間不同，我們可以通過觀察時頻圖中亮帶長度的差異來更好地區分它們。 過去，時頻圖的解釋依賴於經驗豐富的檢測員手動完成。這是耗時且不可靠的，因為人類會犯錯。本研究的目標是開發一種基於深度學習方法的識別系統，可以在時頻圖中識別回波、表面波和振動，並自動計算裂縫的深度（如果有的話）。 卷積神經網路（CNN）為一種強大的深度學習方法。可以訓練它對圖像中出現的對象進行分類。 You only look once (YOLO) 為另一種深度學習方法，它可以在單個圖像中進行多個對象檢測。本研究同時使用了 CNN 迴歸模型和 YOLO V4 模型來判讀敲擊回音時頻圖。 CNN迴歸模型的輸入為時頻圖，輸出為裂縫深度。 YOLO V4 模型的輸入也是時頻圖，輸出則是帶註釋的邊界框，包括“裂縫”框、“表面波”框和“振動”框，以及裂縫的深度，其中裂縫深度是使用“裂縫”框之位置進行計算。 CNN 回歸模型是使用給定裂縫深度的時頻圖進行訓練。 YOLO V4 模型則是使用預先標註之時頻圖進行訓練，該圖像已預先標註了裂縫、表面波和振動，並給予裂縫深度。 在訓練模型時，考慮了幾個問題，包括色階、正規化方法、時頻分析方法、雜訊的加入、圖像解析度、訓練數據集的組成以及混凝土試體的邊界條件。產生最佳結果的模型如下： 1. 使用 YOLO V4 模型， 2. 使用灰階 RID 時頻圖作為輸入， 3. 時頻圖自動正規化， 4. 圖像的解析度為 ， 5. 訓練集包括含有雜訊之數據和實驗數據； 6.訓練集中包含半無限空間模擬數據。對於模擬數據，裂縫檢測率為99.7%，深度預測誤差為2.3%。對於實驗數據，裂縫檢測率為95.6%，深度預測誤差為2.8%。 最佳模型的裂縫檢測率和深度預測誤差是令人滿意的。期望藉助該模型，可以大大地減少檢測誤差和檢測時間，並同時提高判讀結果的準確性。
並列摘要
The impact-echo method is effective in the non-destructive testing of concrete structures. The time-domain signal measured in the impact echo test can be converted into the frequency domain by the Fourier transformation. The frequency components with higher energy would form peaks in the Fourier spectrum, based on which the inspector may detect the location of cracks in the concrete. However, not only echo waves from cracks form peaks in the spectrum, so do modal vibrations. Surface waves also form a hump in the spectrum. The existence of vibration peaks and surface wave hump sometimes jeopardize the detection of crack echo in the spectrum because the crack echo is usually weak, compared with vibrations and surface waves. Therefore, this study uses time-frequency analysis to convert time signals into spectrograms. Because the duration of echo, surface wave, and modal vibrations are different, we can better distinguish them by observing the difference in the length of the bright band in the spectrogram. In the past, the interpretation of spectrograms relies on experienced inspectors to do the job manually. That is time-consuming and unreliable because human makes mistakes. The goal of this research is to develop an identification system based on deep learning methods such that the echo waves, surface waves, and vibrations can be identified in the spectrograms, and the depth of the crack, if any, can be determined automatically. The convolutional neural network (CNN) is a powerful deep learning method. It can be trained to classify the object appearing in an image. You only look once (YOLO) is another deep learning method, which can conduct multiple objects detection in a single image. This research uses both the CNN regression model and the YOLO V4 model in the interpretation of the impact-echo spectrogram. The input of the CNN regression model is the spectrogram, and the output is the crack depth. The input of the YOLO V4 model is also the spectrogram, and the output is annotated bounding boxes, including “crack” box, “surface wave” box, and “vibration” box, and the depth of the crack, calculated using the location of the “crack” box. The CNN regression model was trained using spectrograms with crack depth given. The YOLO V4 model was trained using spectrograms, with crack, surface wave, and vibrations pre-annotated and crack depth given. While training the models, several issues were considered, including color scale, normalization method, time-frequency analysis methods, the addition of noise, image resolution, the constitution of training data, and boundary conditions of the concrete specimens. The model that yielded the best result is as follows: 1. using the YOLO V4 model, 2. using the gray-scale RID spectrogram as input, 3. the spectrogram normalized automatically, 4. the resolution of the image being , 5. Including noise and experimental data in the training set; 6. including semi-infinite space simulation data in the training set. For the simulation data, the crack detection rate is 99.7% and the depth prediction error is 2.3%. For the experimental data, the crack detection rate is 95.6% and the depth prediction error is 2.8%. The crack detection rate and depth prediction error of the best model are satisfactory. With the aid of this model, it is hoped that detection errors and inspection time can be reduced greatly, and the accuracy of interpretation results can be improved at the same time.