ScholarWorks@성균관대학교

초록

This study investigates the systematic measurement and interpretation of flux leakage (FL) through a data-centric learning framework for assessing ferromagnetic structural components. A comprehensive simulation-based analysis was conducted to determine the governing FL parameters, considering defect sizes ranging from 1 to 5 mm and inspection velocities between 3.6 and 10 km/h. The developed simulation model successfully captured the directional behaviour and flux density variations associated with defect morphology and motion dynamics. An experimental setup was subsequently designed to detect FL in steel pipelines containing artificially introduced curved defects, employing a newly developed smart magneto-sensor array integrated with a Raspberry-Pi control system. Through iterative magnetic flux enhancement and data acquisition, a generalized extreme-value distribution was applied for high-confidence flux thresholding. By synthesizing multi-velocity FL datasets and applying time-frequency representation (TFR) based feature extraction, this study examines how magnetic flux distortion and motion effects influence defect detection. Classical ML models, including DT, RF, SVM, and Ensemble classifiers, achieved 72.4-81.2 % accuracy, reflecting sensitivity to flux-leakage variation across defect sizes. A proposed 1D-CNN improved accuracy by nearly 10 %, demonstrating stronger spatial-temporal feature extraction. With domain-adapted multi-style FL image augmentation, proposed YOLOv5 achieved 91.9 % precision, 92.3 % recall, and 92.1 % F1, confirming TFR optimization's crucial role in robust MFL-based inspection.

키워드