以機器學習與主成分分析進行敲擊回音本質模態函數之分類

Abstract

敲擊回音法通常用於混凝土的缺陷檢測，為一種常見的非破壞性檢測，在傳統的檢測過程需要仰賴技術人員的知識才能區分缺陷的回音。有鑒於近幾年機器學習領域的突破性發展，本研究應用機器學習，對敲擊回音訊號經由經驗模態分解(EMD)所得到本質模態函數(IMF)進行自動判讀並找出缺陷回音。 本研究採用的機器學習有二種：類神經網路(ANN)及卷積神經網路(CNN)。模型的輸出是各個本質模態函數的類別：表面波、回音或振動。模型的輸入資料有二種：1. 敲擊回音訊號經由自適應噪聲之完整集合經驗模態分解法(CEEMDAN)所得到的前5個IMF；2. 將IMF進行主成分分析(PCA)得到的主成分。其中PCA在本文又分為PCA1與PCA2，PCA1是將所有IMF放在一起進行PCA；PCA2是對表面波IMF、回音IMF及振動IMF分別進行PCA。本研究亦嘗試將敲擊回音訊號的表面波移除之後再進行CEEMDAN。 本研究共發展六種模型，各模型之輸入資料與判讀實驗訊號的IMF準確率如下： (i) ANN-IMF為以IMF做為輸入之ANN模型，為本文表現最差的模型，其回音的準確率僅有33.62%。 (ii) CNN-IMF為以IMF做為輸入之CNN模型，因為模態混疊的問題在淺裂縫試體預測表現不佳，其回音的準確率為77.58%。 (iii) CNN-SWR_IMF為以移除表面波後的IMF做為輸入之CNN模型，因將淺裂縫回音訊號也一併切除些許，因此在淺裂縫試體預測表現不如預期，其在回音的準確率為89.66%。 (iv) CNN-PCA1為以IMF進行PCA1之主成分做為輸入之CNN模型，因PCA1之特徵向量沒有將三種類型的波完美分開，因此表現不如預期，其回音的準確率為60.34%。 (v) CNN-PCA2為以IMF進行PCA2之主成分做為輸入之CNN模型，為本文表現整體平均表現最佳的模型，其在回音的準確率為88.79%。 (vi) CNN-SWR_PCA2為以移除表面波後的IMF進行PCA2之主成分做為輸入之CNN模型，模型因為沒有考慮到表面波的準確率，在本研究列為表現第二佳的模型，其回音準確率為99.14%。 經過比較，可發現CNN-PCA2與CNN-SWR_PCA2對於判讀回音IMF均表現良好，且泛用性高，可以做為工程師進行敲擊回音檢驗時的有力輔助工具。

The impact echo (IE) test is often used to detect defects in concrete structures. The conventional approach for analyzing the IE data requires users’ expertise to differentiate the echo of defects from the other signals received. Given the breakthroughs in machine learning (ML) in recent years, this study applied ML to automatically classify intrinsic mode functions (IMFs) obtained by using empirical mode decomposition (EMD) and identify the echo IMF. This study proposes two types of machine learning: artificial neural network(ANN) and one-dimensional convolutional neural network (CNN). The outputs of the model were the category of each IMF: surface wave, echo, or vibration. The inputs of the model were two types of data: 1. the first 5 IMF of the IE signals obtained by using the complete ensemble EMD with adaptive noise (CEEMDAN); 2. the principal components of the first 5 IMFs obtained by principal component analysis (PCA). Two types of PCA were conducted in this study, namely PCA1 and PCA2. PCA1 applied PCA to all IMFs with the three types of waves put together; PCA2 applied PCA to surface wave IMFs, echo IMFs, and vibration IMFs separately. To deal with the problem associated with shallow-crack signals, this study tried removing the surface waves from the IE signals before using CEEMDAN. In this research, six models were developed, and their performances are as shown in the following: (i) ANN-IMF, with IMFs as input, performed 33.6% echo accuracy and was the worst model in this paper. (ii) CNN-IMF, with IMFs as input, performed 77.5% echo accuracy and predicted not well in shallow crack because of mode mixing. (iii) CNN-SWR_IMF, with surface wave removed and IMFs as input, performed 89.6% echo accuracy. Since some shallow crack echoes were removed as well, the performance in the shallow crack was not as good as expected. (iv) CNN-PCA1, with the principal components of IMFs obtained in PCA1 as input, performed 60.3% echo accuracy, which was not as good as expected due to the fact that the PCA1 eigenvector didn't ideally separate the three types of waves (v) CNN-PCA2, with the principal components of IMFs obtained in PCA2 as input, performed 88.7% echo accuracy. It is rated as the best model because it has the best overall average performance. (vi) CNN-SWR_PCA2, with surface wave removed and the principal components of IMFs obtained in PCA2 as input, performed 99.1% echo accuracy. This study rated it as the second-best model since it didn't consider surface waves. The results indicated that CNN-PCA2 and CNN-SWR_PCA2 performed well in identifying IMF echoes, which can be used as a reference model for engineers to inspect defects.