以深度學習對敲擊回音時頻圖擷取之時間域訊號進行分類

Abstract

敲擊回音法是一種非破壞檢測技術，常用於檢測混凝土中之缺陷。鋼珠敲擊混凝土表面，量測到的時間域訊號經傅立葉轉換後可從頻譜中判讀回音尖峰，但僅透過傅立葉頻譜並無法判斷尖峰是由鋼筋或裂縫回音所造成。雖然過去曾有研究顯示以經驗模態分解或時頻分析對鋼筋與裂縫回音作判別，但仍存在一些限制。 為了解決鋼筋與裂縫回音判別的困難，本研究提出一個結合時頻分析與深度學習的判別方法。首先，將時間域訊號進行Reduced Interference Distribution (RID)轉換，並從時頻圖中擷取特徵訊號，再將特徵訊號輸入訓練過的深度學習模型進行分類，以判別特徵訊號屬於鋼筋回音、裂縫回音或其他。 特徵訊號的擷取是將時頻圖各頻率的時間訊號振幅加總，以找出尖峰頻率，以試體底部反射頻率作為分界點，在試體底部反射頻率以上的頻率區間中，取最高的尖峰頻率作為回音頻率，試體底部反射頻率以下的頻率區間中，則選取最高與次高尖峰作為其他類頻率，再將這些特徵頻率對應的時間域訊號擷取出來作為深度學習模型的輸入。 本研究發展兩個卷積神經網路(Convolutional Neural Network, CNN)分類模型，分別以單筆訊號輸入及三筆訊號作為輸入之模型，並對兩種模型進行比較。CNN分類模型皆以內含鋼筋或裂縫之無限平板的數值模擬訊號進行訓練，訓練完成後，以數值模擬和實驗數據訊號進行測試。在訓練過程中，考慮在模擬訊號中添加不同幅度的雜訊以及使用不同深度的卷積層之影響。 經過測試後，最佳的模型是以三筆訊號作為輸入，採用4層卷積層，並在訓練集數據添加標準差為0~5%表面波振幅之雜訊，其對於辨識整體模擬數據的準確率為98.8%，對於實驗數據辨識準確率為100%。 本文研究也對模型進行泛化性測試，儘管模型在模擬數據表現良好，但在辨識實驗訊號時，若未能擷取正確的特徵訊號模型仍然會發生誤判。 總而言之，本研究創立的分類模型能夠有效識別鋼筋、裂縫及其他類的敲擊回音時頻特徵訊號，但若實驗雜訊過多以致選取尖峰頻率有困難，仍可能失效。因此，在實際工程應用中，該模型的預測結果應僅作為輔助工具，供工程師參考。

The impact echo method is a non-destructive testing technique and it is used to detect defects within concrete structure. As a steel ball strikes the concrete surface, the time-domain signals recorded can be analyzed through Fourier transform to identify echo peaks in the frequency spectrum. However, interpreting the presence of rebars or cracks solely based on these frequency peaks can be challenging. In previous work, our research group applied Empirical Mode Decomposition (EMD) directly to time-domain signals of impact echoes and successfully identified defect echoes. Nonetheless, this approach often results in signals containing unexpected features due to mode mixing, where the decomposition yields components that do not align with the original signal characteristics.

To address the issue of mode mixing, attempting to transform the time-domain signals into the time-frequency domain using a Reduced Interference Distribution (RID), and extract feature signals from the time-frequency representation. Since the energy decay rates of rebar and crack echoes differ over time, and the modal vibration responses persist throughout the entire sampling period, classification based on signal differences poses significant challenges for inspectors. Therefore, this study aims to develop deep learning techniques for classification, distinguishing between features associated with rebars, cracks, or other modes(including concrete bottom reflections and modal vibrations).

We developed two Convolutional Neural Network (CNN) classification models, using both single-signal input and multi-signal (three signals) input approaches for training. They are trained with numerical simulations of an infinite plate containing rebars or cracks. After training, the models were tested using both numerical simulations and experimental data. The numerical simulation data included specimens with rebars or cracks in an infinite plate, while the experimental data contained measurements from several concrete specimens. The training process considered the impact of different noise added to the simulated signals and varying convolutional layer depths on model performance.

Following a comprehensive evaluation, the best classification model architecture was determined to be a model with four convolutional layers, training data augmented with noise having a standard deviation of 0~5% of the surface wave amplitude, and inputting three signals simultaneously. This model achieved an accuracy of 98.8% for identifying simulated data and 100% for experimental data.

Although the model is highly advanced, the generalization tests revealed some risk of misclassification for experimental data. Therefore, in practical engineering applications, the model's predictions should be used as a auxiliary tool for engineers, and not as the good decisions, as it cannot entirely act for the professional engineers.