以人工智慧輔助電磁超聲波在鐵條缺陷檢測之研究

A Study on the Crack Detection of Steel Rod Using AI-assisted Electromagnetic Acoustic Wave Measurement

張必昀(Pi-Yun Chang)

指導教授 ： 吳政忠

共同指導教授 ： 劉佩玲(Pei-Ling Liu)

國立臺灣大學/工學院/應用力學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202103807

摘要

在螺絲製作時，會發現某些螺絲強度特別低，問題來源為製作的鐵條本身就存在一些細小的缺陷，導致利用到包含缺陷的材料時，會有螺絲強度不足的狀況產生。因此，要如何及時且快速的，將包含缺陷的鐵條篩檢出來，成為了生產者最重要的課題。此時利用電磁聲波檢測法此種非破壞檢測的方式，便成為最佳的選擇，此種探頭製作便利且成本低廉，操作方式簡便易學，透過探頭在鐵條上激發出彈性波訊號，使其在鐵條中傳遞，再利用探頭接收訊號，便可從中判斷出其中有無包含缺陷反射訊號，然而，低信噪比的缺點，一直都是此探頭的硬傷。
因此，本研究透過結合深度學習，達到實現人工智慧的要求，透過機器的自我學習，歸納出一系列判讀缺陷反射訊號的方法。其中，又以卷積式類神經網路，在辨識圖像上，有最好的表現結果，透過其特有的卷積以及池化層，能夠先將圖片特徵提取出來，以利全連階層進行歸納統整。透過此項技術結合電磁聲波換能器，並探討最好的資料處理方法，以及最佳的卷積式類神經網路模型架構，最後設計出一個最完善的判讀模型。
本研究所量測的鐵條直徑為3.5mm，量測缺陷深度從1mm~0.6mm，準確率可以到達97.4%，並且在判斷缺陷準確率為92.31%，判斷沒有缺陷的準確率為98.44%。同時，在確保有無缺陷的比數不會相差過於懸殊的狀況下，利用訊號相減的方法，模擬出無邊界鐵條時，應該要量測到的訊號模式，在受限於實驗試體的有限長度的狀況下，盡可能地貼近實際應用情形。
經透過硬體與軟體的配合，建構出一套最佳的判斷缺陷反射訊號模型，在即時、且簡便的前提下，解決判讀訊號時，因量測訊號信噪比過低，需要依賴專業知識人員的困境，讓此檢測模式可以廣泛的應用於鐵條缺陷的檢測上，減少螺絲後續的品管成本。

關鍵字

非破壞檢測 ； 電磁聲波換能器 ； 人工智慧 ； 卷積式類神經網路

並列摘要

During the production of screws, some screws are found to have very low strength. The source of the problem lies in small defects found in steel bars, resulting in inadequate screw strength when materials with defects are used. Therefore, how to timely and quickly detect steel bars with defects turns out to be the most important issue for producers. At this time, the electromagnetic acoustic detection method, a non-destructive detection approach, has become the best choice. This type of probe production is convenient and cost-effective. The operation method is easy to learn, which involves the use of a problem to excite elastic wave signals on an iron bar, which are transmitted through the iron bar. A probe is then used to receive signals, from which whether there are defect reflection signals can be determined. However, the disadvantage of a low signal-to-noise ratio has always been a major flaw for this probe.
Therefore, in this study, the artificial intelligence requirements can be achieved by combing depth learning. Through machine learning, a series of defect reflection signal interpretation methods have been summarized. Among them, convolutional neural networks achieve the best performance results in image recognition. Through a specific convolution layer and a pooling layer, image features are first extracted to facilitate summarization and compilation through the fully connected layer. Through this technology coupled with an electromagnetic acoustic transducer, the best data processing method and the best convolutional neural network model structure were explored. Finally, the most comprehensive interpretation model was designed.
In this study, the iron bar had a diameter of 3.5mm. The depth of the measured defect ranged from 1mm~0.6mm, with 97.4% accuracy. In addition, defect determination had 92.31% accuracy, while defect-free determination had 98.44% accuracy. At the same time, under the condition that the defect and defect-free ratios showed no great disparity, the signal subtraction method was adopted to simulate the measurable signal model for borderless steel bars. Limited by the experimental sample’s limited length, the actual applications were reproduced as much as possible.
Through the conjunctive use of hardware and software, the best reflection signal model for determining defects was constructed. Under the premise of instantaneity and simplicity, the plight of dependence on professionals during signal interpretation due to the excessively low signal-to-noise ratio of measured signals was solved. This detection model can be widely applied in detecting iron bar defects, thereby reducing the subsequent quality control cost of screws.

並列關鍵字

Nondestructive testing ； Electromagnetic acoustic transducer ； Artificial intelligence ； Convolutional neural network

參考文獻

M. Afzal, S. Udpa, "Advanced signal processing of magnetic flux leakage data obtained from seamless gas pipeline," NDT E Int., 35 (7) (2002), pp. 449-457.

S. Saha, S. Mukhopadhyay, U. Mahapatra, S. Bhattacharya, G.P. Srivastava, "Empirical structure for characterizing metal loss defects from radial magnetic flux leakage signal," NDT E Int., 43 (6) (2010), pp. 507-512.

L. Udpa, S. Udpa, A. Tamburrino, "Adaptive wavelets for characterizing magnetic flux leakage signals from pipeline inspection," IEEE Trans. Magn., 42 (10) (2006), pp. 3168-3170.

J. Garcia-Martin, J. Gomez-Gil, E. Vasquez-Sanchez,"Non-destructive techniques based on eddy current testing," Sensors, 11 (2011), pp. 2525-2565.

V. Sundararaghavan, K. Balasubramaniam, N.R. Babu, N. Rajesh," A multi-frequency eddy current inversion method for characterizing conductivity gradients on water jet peened components," NDT E Int., 38 (7) (2005), pp. 541-547.