We present PASS-Net, a physics-guided deep learning framework for automated defect segmentation in infrared non-destructive testing (NDT) of composite materials. The architecture uniquely integrates a U-Net backbone with Fourier-based physics-aware layers and bidirectional State-Space Models (SSMs) to capture both spatial patterns and temporal thermal dynamics. By incorporating thermal diffusion principles directly into the network architecture, the model ensures thermodynamically consistent predictions while maintaining computational efficiency. The SSM enables effective long-range dependency modeling with linear complexity, addressing limitations of traditional attention mechanisms. Moreover, the framework delivers a comprehensive uncertainty analysis by combining inference-time stochastic dropout with evaluation on multiple augmented input variants, decomposing total uncertainty into epistemic and aleatoric components for reliable decision-making in safety-critical contexts. Validated on aerospace-grade composites, such as CFRP and fiberglass, PASS-Net outperforms traditional U-Net models, achieving at least 6% improvement in mean Intersection over Union (mIoU). It demonstrates resilience to real-world challenges, including material heterogeneity and non-uniform heating, making it suitable for industrial-scale deployment. A comparative analysis further reveals a superior defect contrast-to-noise ratio, highlighting the model’s potential for adoption in industrial non-destructive testing (NDT) applications that require both accuracy and computational efficiency. The integrated physics-based loss ensures consistent performance across diverse materials and operational conditions, representing a significant step toward reliable, deployable deep learning solutions in non-destructive testing (NDT). The implementation, including code and datasets, is available in https://faculty.uestc.edu.cn/gaobin/zh_CN/lwcg/153392/list/index.htm.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号