A hybrid machine learning Hurdle-based framework for quantitative surface defect sizing in eddy current testing

Abstract

Eddy Current Testing (ECT) is a non-destructive testing method for identifying irregularities in electrically conductive materials using electromagnetic induction. While ECT is precise at detecting flaws, accurately determining their size remains challenging. The two main obstacles are obtaining high-quality ground-truth data for training machine learning models and handling class imbalance, since defect-free regions greatly outnumber those with defects, leading to biased predictive models. To address these challenges, this work presents a hybrid classification-regression framework based on a Hurdle model formulation to quantify defect size. The framework is tested on an unbalanced laboratory database. In this approach, defect morphology is reconstructed as the product of defect probability and estimated depth. The probability is obtained from a classifier model, while depth estimation is obtained from a regression model. The algorithms considered are a gradient-boosted classifier with probability calibration and a random-forest regressor. To evaluate the proposed framework, its performance is assessed in two ways: (1) spatial reconstruction–predicting the shape of a physically connected (contiguous) region held out from the training data–with a Mean Absolute Error (MAE) between 0.32 and 0.54 mm over all scan points of the held-out region at lift-off distances between 0.25 and 1.0 mm, and (2) Leave-One-Experiment-Out (LOEO) cross-validation across 21 different flaws (defect locations), yielding a median MAE of 0.29 mm over all scan points. Restricted to defective points and grouped by lift-off distance, the median of the MAE ranged from 0.82 to 1.16 mm. The LOEO strategy also examines the effects of flaw size (defect diameter), vertical size (defect depth), and sensor distance (lift-off) on the model’s predictive ability. Results show that deeper defects are associated with lower regression accuracy, whereas defect diameter and lift-off are not statistically significantly related to framework performance.