- 學位論文
應用深度學習分析敲擊回音時間域訊號
Analysis of Impact-Echo Time Signals by Deep Learning
摘要
敲擊回音法是用於評估混凝土結構缺陷的非破壞檢測技術。透過敲擊產生應力波之位移響應具有週期性，利用傅立葉轉換後的頻率尖峰檢測缺陷深度。然而，淺裂縫的反射波容易受到表面波和模態干擾，導致頻譜中的尖峰頻率難以識別。隨著機器學習的發展，本研究利用深度學習模型直接判讀人工難以識別的時間域訊號，期望以新的方式分析裂縫深度。 本研究以數值模擬產出裂縫深度6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26cm之半無限域敲擊回音時間域訊號，作為深度學習模型之訓練集數據，輸入訊號時間長度320μs（取樣間距1.6μs，序列長度200）。測試數據則是使用有底部反射影響之無限平版，其中包含不同試體厚度（30cm與20cm）裂縫深度與無裂縫試體。 本研究先進行多層感知器（MLP）、全卷積網路（FCN）、殘差網路（ResNet）和時間卷積網路（TCN）四種深度學習模型的比較，發現TCN模型能夠有效提取敲擊回音訊號特徵。 在進一步調整TCN模型架構時發現，卷積層過多或過少都無法獲得良好結果。卷積層太少無法有效提取訊號特徵，過多又會造成過擬合現象。卷積核大小會直接影響感知域，我們選擇的卷積核大小需要確保感知域至少大於輸入長度以獲得較好的結果。此外，增加卷積核數量可以提高模型的穩定性。最終我們的TCN模型採用5個殘差區塊、卷積核大小8且卷積核數量25作為最優模型架構。 由於實驗訊號不同於數值模擬訊號，會受雜訊、時間原點偏移及敲擊鋼珠大小的影響，故本研究嘗試在訓練集訊號加入雜訊、時間原點偏移，並增加不同尺寸的鋼珠敲擊源。經過數值模擬訊號測試後，發現在訓練集訊號中加入最大標準差為表面波振幅3%的雜訊，對時間原點作最大30μs之平移，並使用4, 6, 10mm等多種鋼珠直徑進行敲擊，對於模型穩定性和泛用能力之提升表現最佳。此外，為處理實驗模型與數值模型波速不同所造成的波形差異，本研究也對實驗訊號依試體波速進行時間軸縮放，使深度學習模型可適用於不同波速試體。 本研究以有限平板的模擬訊號對優化過的TCN模型進行測試，發現該模型裂縫對深度經過訓練的訊號均表現優異，平均絕對誤差均（MAE）在0.2cm以下，對6至26cm的內插深度亦表現良好，MAE在0.3cm以下，但對4cm和30cm的外插深度誤差較大。以實驗訊號測試之結果顯示，對於裂縫深度6, 10, 12, 12-8, 14-6, 25cm的判斷均表現良好，MAE在0.6cm以下。在未使用濾波的情況下，模型都能有效判斷裂縫深度。此外，本研究也嘗試在訓練集中加入振幅縮放（scaling）、振幅扭曲（magnitude warping）和時間扭曲（time warping）等資料增強技術，結果顯示這些技術對模型分析對實驗訊號的表現並無助益，尤其結果顯示尤其是時間扭曲會破壞訊號的時間關係，導致模型結果反而變差。 總結而言，本研究經過數值模擬訊號與實驗訊號的測試，TCN模型確實能有效地分析敲擊回音時間域訊號，尤其對淺層裂縫的偵測更是遠優於頻率域方法，為混凝土結構之非破壞檢測提供了一種全新且有效的方法。
並列摘要
The impact-echo method is a widely used nondestructive technique for detecting defects in concrete structures. It generates stress waves and analyzes the resulting surface displacement response using Fourier transform to identify defects from peak frequencies in the spectrum. However, shallow cracks often produce reflections that are obscured by surface waves and modal interferences, making it difficult to accurately detect frequency peaks. With advancements in machine learning, this study explores the use of deep learning models to directly interpret time signals, aiming to provide a novel approach for analyzing crack depths. This study utilizes numerical simulations to generate impact-echo time signals from a half-space with crack depths of 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, and 26 cm as training data for deep learning models. The input signal duration is 320μs (sampling interval 1.6μs, sequence length 200). The test data consists of impact-echo signals from finite plates with bottom reflection influences, including specimens with crack depths and thicknesses of 30 cm and 20 cm, as well as intact specimens. Initially, this study compared four deep learning models: multi-layer perceptron (MLP), fully convolutional network (FCN), residual network (ResNet), and temporal convolutional network (TCN). The findings revealed that the TCN model exhibited superior feature extraction capabilities for impact-echo signals. Further optimization of the TCN model revealed that both too many and too few convolutional layers resulted in suboptimal performance. An insufficient number of layers failed to capture signal features effectively, while an excessive number led to overfitting. The chosen kernel size significantly influenced the receptive field, and it was crucial for the receptive field to exceed the input length for optimal results. Additionally, increasing the number of filters enhanced model stability. Our optimized TCN model employs 5 residual blocks with a kernel size of 8 and 25 filters. Experimental signals, unlike numerical simulation signals, are influenced by noise, time shifts, and the diameter of the steel ball. This study attempts to enhance the training signals by incorporating noise, time shifts, and variations in steel ball diameter. Numerical simulation tests show that adding noise with a maximum standard deviation of 3% of the surface wave amplitude, time shifts up to 30μs, and using steel balls with diameters of 4, 6 and 10mm in the training data significantly improves the model's stability and generalization ability. Additionally, to address waveform differences caused by varying wave velocities between experimental and numerical models, this study scales the time axis of experimental signals according to the specimen's wave velocity, thereby making the deep learning model applicable to specimens with different wave velocities. The optimized TCN model was tested with finite plate simulation signals, showing excellent performance in predicting trained crack depths, with mean absolute errors (MAE) below 0.2cm. The model also performed well for interpolated depths of 6 to 26cm, with MAE below 0.3cm, but had larger errors for extrapolated depths of 4cm and 30cm. Testing with experimental signals showed good predictions for crack depths of 6, 10, 12, 12-8, 14-6, and 25cm, with MAE below 0.6cm. The model effectively determined crack depths even without filtering. This study also explored data augmentation techniques such as amplitude scaling, magnitude warping, and time warping in the training data. However, these techniques did not improve the model's performance in analyzing experimental signals; in fact, time warping particularly degraded the model's results by disrupting the temporal relationships in the signals. In summary, this study demonstrates through tests with numerical simulation and experimental signals that the TCN model can effectively analyze impact-echo time signals, particularly excelling in detecting shallow cracks and significantly surpassing frequency domain methods. This provides a novel and effective approach for the nondestructive testing of concrete structures.