조건부 생성적 적대 신경망을 활용한 폭발 하중 하 RC 기둥의 손상 분포 예측 프레임워크

초록

This paper presents a conditional generative adversarial network (cGAN)-based framework for predicting the spatial damage distribution of reinforced concrete (RC) columns under blast loading. The proposed framework employs a dual-output generator that simultaneously produces effective plastic strain (EPS) distribution images and binary spalling masks from a four-dimensional condition vector comprising reinforcement details and blast parameters. Adaptive instance normalization (AdaIN) is adopted to adaptively modulate feature maps in response to continuous physical condition variables. Training data were generated through finite element simulations using LS-DYNA, and a systematic ablation study on normalization strategies and loss function combinations confirmed that the AdaIN with L1+Perceptual combination achieved the best performance. The final model achieved an SSIM of 0.899 and L1 error of 0.128 for strain distribution prediction, and IoU of 0.843 and Dice coefficient of 0.874 for spalling region prediction. The results demonstrated that the proposed framework effectively reproduces spatially distributed blast damage patterns, confirming its applicability to post-blast damage assessment and repair strategy planning.

키워드