An Adaptive Sampling Protocol for Real-Time Defect Assessment Using Eddy Current Sensor and Machine Learning Algorithm

doi:10.1109/TIA.2023.3284782

2023-09-01

PRE

DOI：10.1109/TIA.2023.3284782

原文链接

原文求助

摘要

En 中文

Eddy current (EC) sensors, in combination with machine learning methods have shown great potential for automatic detection and classification of defects in industrial components [1]. Machine learning (ML) tools can be effectively utilized for improving inspection accuracy and diminishing human interventions. There are reports on defect classification with Support Vector Machine (SVM) and EC sensor signal, utilization of kernel-based Principle Component Analysis (PCA) schemes for automatic detection and classification of surface and sub-surface defects. However, the reported works dealt with artificially grown synthetic defects/flaws only. The utility of these scheme on naturally grown in-process defects/flaws in industrial components need further attention. In this communication, we studied and implemented an adaptive sampling protocol based on in-process defect assessment using an EC sensor to optimise the prediction accuracy of the ML algorithms. An encircling coil EC sensor was developed in-house and implemented for extracting the EC signal patterns of various naturally grown in-process defects/flaws. The extracted signals were then marshalled using a suitable in-cognitive clustering technique. A set of algorithms based on cognitive SVMs were trained and subsequently, were tested on EC testing data. We found fine Gaussian kernel functions as the most robust kernel function in terms of prediction accuracy (similar to 99%) and speed (similar to 4500 observations per second). Moreover, the proposed SVM model is immune to noisy signals with SNR above 30 dB. From the application's perspective, these findings indicate the importance on achieving higher efficiency combined with reliable quality control of the final products culminating in further energy conservation.

Keyword:

Artificial neural network

Gaussian kernel functions

machine learning

single tones ECT

support vector machine