Title Performance Analysis of CNN-Based Magnesia-Chrome Refractory Crack Prediction According to Input Data Characteristics
Authors 박재한(Park, Jae-Han) ; 김동환(Kim, Dong-Hwan) ; 김주석(Kim, Joo-Seok) ; 강상구(Kang, Sanggoo)
DOI https://doi.org/10.5659/JAIK.2026.42.4.359
Page pp.359-367
ISSN 2733-6247
Keywords Magnesia-Chrome Refractory Brick; Non-Destructive Testing; Impact-Based Signal; Deep Learning
Abstract This study presents a deep learning-based non-destructive testing method for detecting internal cracks in magnesia-chrome refractory bricks. Five datasets were created using different signal representations, including time-domain, frequency-domain, time-frequency domain, and their combinations. CNN models with identical structures were independently trained for each input type, with key hyperparameters optimized using the Optuna framework. Time-domain models achieved high precision with relatively lower recall, while frequency-domain inputs improved recall. The Time?Spectrogram Domain CNN (TSDC) achieved the best overall performance across all metrics. These results demonstrate that fusing time and frequency information enhances feature learning and classification performance. This approach offers an efficient framework for early crack detection in refractories. However, limitations due to dataset size and single-sensor input remain, indicating the need for future work focused on data expansion and multi-sensor system integration to improve robustness and generalization.