類神經網路及機器學習應用於渦電流檢測金屬管缺陷訊號分析

Artificial Neural Network and Machine Learning on the Analysis of Eddy-Current Signals of Metal Tubes’ Defects

高永浩(Yong-Hao Gao)

指導教授 ： 郭茂坤

國立臺灣大學/工學院/應用力學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202203343

摘要

本研究藉由類神經網絡的方法對渦電流檢測資料做有效的濾波，消除支撐板對管件的訊號之非線性影響，以便對缺陷進行辨識。本研究的管件包含銅鎳、鈦和不鏽鋼的標準管，和其對應的支撐環，管件上的支撐板為鐵磁性的材料，在判別過程中往往很難去判讀缺陷資訊，藉由類神經網絡的方法可以濾除支撐環非線性的訊號並完整保留其底下的缺陷訊號，且能將阻抗圖漂移的訊號做有效的定位。第二步藉由人工的方式對阻抗圖(電阻與電感二維做圖)提取特徵，常見的特徵為八字圖的半徑、寬度、角度等。除了人工提取的特徵外，本研究利用卷積神經網路來加強人為的特徵的判別，提取神經網路卷積和池化後的係數。
藉由不同特徵彼此的相關圖和單一特徵的訓練準確度可以對特徵做有效的縮減。往往人工提取特徵需要對此領域相關的背景知識，本研究使用PCA(主成分分析)，對渦電流的特徵也能在維度降低下，保留三個主軸特徵下，得到約99%的準確度。
最後在支撐環的非線性的影響下，支撐環與缺陷的相對位置對訊號的影響是很巨大的，使用缺陷種類和相對位置的2個特徵資訊，利用模型可以模擬出缺陷對應相對位置範圍內的渦電流雙頻訊號。
本研究主要藉由機器學習的方法，在原本的專家系統的判斷依據下，提供輔助的系統，以便與專家系統做更精確的判讀。

關鍵字

熱交換管 ； 渦電流訊號 ； 類神經 ； 主成分分析 ； 濾波 ； 類神經網絡 ； 特徵提取

並列摘要

In this study, the artificial neural network (ANN) filter and simulator are developed to process data of eddy-current test (ECT) of a heat-exchange tube. Additionally, a machine learning of the random-forest classifier is used to identify various defects in tube. ANN filter is used to effectively process the data of two-frequency ECT to eliminate the nonlinear effect of the support plate on the ECT signals. Three standard tubes of copper-nickel, titanium and stainless steel with various defects are studied. The support plate is made of carbon steel, which is a ferromagnetic material. ANN model is used to filter out the nonlinear effect of a support ring on the ECT signal to display the signal of various defects under the support ring. Subsequently, we manually extract several features from the signals in the impedance plane (resistance and inductance); the pattern of defect’s signal is an 8-profile. Main features are the radius, the width, and the angle of the 8-profile. Except these defined features, the convolution neural network (CNN) is also used for features extraction to assist the discrimination of defects. Finally, the random-forest method is used to classify the various defects, according to code.
Through the analysis of the correlation map of different features and the training accuracy of each single feature, the number of features can be effectively reduced. Normally, the feature extraction requires the domain knowledge. In this study, PCA (principal component analysis) is also used to reduce the dimension of the features of eddy-current test; the accuracy is about 99% to remain the most important features of three principal axes. For ANN simulator, we use two features (the defect type and its relative position to a support) to simulate a ECT signal of a specific defect under the support ring.
Based on the original expert system, this research uses the ANN, CNN and random-forest method to identify and classify the ECT signals of various defects of tube. It shows that artificial intelligence can provides an auxiliary method for the identification of various defects.

並列關鍵字

Heat exchange tube ； Eddy current signal ； Artificial Neural Network ； Principal Component Analysis ； Filter ； CNN ； Simulator

參考文獻

[1] P. Kot, M. Muradov, M. Gkantou, G. Kamaris, K. Hashim, and D. Yeboah, “Recent advancements in non-destructive testing techniques for structural health monitoring,” Applied Sciences, 11(6), 2750, 2021.

[2] M. I. Jordan, T. M. Mitchell, “Machine learning: Trends, perspectives, and prospects,” Science, 349(6245), 255-260, 2015.

[3] S. Chuang, “Eddy current automatic flaw detection system for heat exchanger tubes in steam generators,” Iowa State University Digital Repository, 1997.

[4] J. García-Martín, J. Gómez-Gil, and E. Vázquez-Sánchez, “Non-Destructive Techniques Based on Eddy Current Testing,” Sensors, 11(3), 2525-2565, 2011.

[5] A. N. AbdAlla, M. A. Faraj, F. Samsuri, D. Rifai, K. Ali, and Y. Al-Douri, “Challenges in improving the performance of eddy current testing,” Measurement and Control, 52(1-2), 46-64, 2019.