A Metal Classification System Based on Eddy Current Testing and Deep Learning
Bangda Cao, Zhijie Zhang, Wuliang Yin, Dong Wang, Zexue Zhang
Abstract
This article proposed a novel system for classifying different metal samples based on a double-coil sensor. Traditional classification methods often require a known lift-off for the sensor, which can be a challenge due to the vibration of the conveyor belt. Therefore, to overcome the adverse effects of this disorderly change of lift-off, the article innovatively combines eddy current testing (ECT) with deep learning, utilizing the neural networks’ ability of nonlinear fitting to distinguish five materials of metal: aluminum, zinc, tin, brass, and titanium. First of all, the article designed a coaxial double-coil probe, which can minimize the influence of asymmetry of targets’ shape and posture. Then, we constructed a driver model for deep learning, including deriving the theory of ECT and selecting features of categorization by <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${K}$ </tex-math></inline-formula> -Means clustering. Subsequently, the two-tier classification networks were created, and we designed an impedance collector to get data from the double-coil sensor. After inputting the impedance data into our neural networks, it could output the classification results finally. In the tests, a series of flat-bottom specimens with inclination and elevation were used to examine our algorithm, and the conveyor belt we used with about 1 mm vibration amplitude and 0.3 m/s transmission speed. Upon comparison, it is concluded that the deep learning method and cluster center features have significant gains for classification, the accuracy of the classification system can reach 94.3%.