Research on Metallic Spheres Radius Classification Method Using Machine Learning With Eddy Current Testing

Abstract

Metallic spheres play a crucial role in industry and their accurate measurement is essential to ensure the safety of industrial production. Eddy current testing (ECT), which is non-contact and non-invasive, provides an efficient and precise approach for the parameter evaluation of metallic spheres. In this paper, we utilize machine learning (ML) methods to invert inductive signals in order to address the inverse problem of ECT, with the aim of reconstructing the radius of a metallic sphere. Datasets containing the radius information of the metallic sphere were constructed based on the simplified analytical solution. The datasets were divided into two parts based on the real part (RP) and imaginary part (IP) features, and the connection between the two features and the radius of the metallic sphere were compared by five classification models. While achieving accurate classification of aluminum and stainless steel spheres with different radius, the models are evaluated to ensure the reliability and validity of the models. The results show that the use of IP data as a classification feature has better accuracy as compared to RP. The K nearest neighbor (KNN) radius classifier has the highest accuracy of 95.5% in aluminum spheres and the random forest (RF) radius classifier has the highest accuracy of 95.9% in stainless steel spheres. In addition, all five classifiers are able to overcome the effect of lift-off on the classification results.