Eddy Current Sensor as In-situ Assessment and Characterization of In-process Surface Defects in Wires/Rods Using Machine Learning Algorithms

Abstract

In-situ inspection of naturally grown in-process surface defects such as sliver, fin, chip-off, entrapment, abrasion, etc. of high-speed drawn steel wires/rods is indispensable nowadays for the wire drawing industries. The present communication aims at investigating a real-time assessment of such kinds of defects using machine learning (ML) algorithms. Emphasis has been laid on the development of assessment protocol for evaluation of figures of merits of their occurrences including the severity of defect prediction, minimum predictable resolution, computation efficiency, and range of defect classifications. A single tone encircling coil eddy current (EC) sensor along with an advanced signal processing toolkit is implemented to get the unique EC signal patterns that correspond to various kinds of defects. A linear sampling approach has been adopted using an automated motorized linear scanning system wherein the EC encircling coil scans the parts (wires/rods) axially and the EC signals corresponding to each kind of defect are logged into the PC. A set of algorithms based on support vector machine (SVM) has been trained against various defects which have been implemented for the classification of those in-process surface defects of the wires/rods obtained from plants. Various kinds of SVM Kernel functions have been trained using the experimental EC data. Amongst them, SVM with fine Gaussian Kernel functions is more robust in terms of prediction accuracy (~100%) and speed (~ 4500 observations per second). Moreover, the proposed SVM model is immune to noisy signals with SNR above 30 dB. From the perspective of the application, these findings indicate the importance of achieving higher efficiency combined with reliable quality control of the final products.