Deep learning-based thermal motion estimation and lay-up reconstruction framework towards machine-independent real-time AFP process monitoring and inspection

Official URL: https://dx.doi.org/10.1016/j.compositesb.2025.112951

Automated Fiber Placement (AFP) continues to advance composite manufacturing, yet real-world throughput and quality assurance remain constrained by labor-intensive inspection and the absence of automated, in-situ monitoring solutions. Existing methods are partial–confined to local, frame-level analysis lacking global motion context required for comprehensive lay-up inspection, or reliant on machine-coupled data that introduces synchronization errors and hinders generalizability. We present a novel, machine-independent framework for real-time, motion-aware AFP monitoring and inspection. We introduce ThermoRAFT-AFP, a custom deep learning-based motion estimation core, tailored with AFP-specific augmentations and process-aware runtime optimizations to enable stable and precise thermal flow tracking. These estimates power a two-stage reconstruction pipeline that first stitches course-wise thermal mosaics, then assembles them into ply-level, high-fidelity, and interpretable laminate visualizations–recovering global motion context. We validate the framework on a large-scale, diverse AFP thermal dataset comprising over 13,000 frames with varying lay-up conditions, speed profiles, and defect types. A comprehensive analysis of motion accuracy, runtime efficiency, and deployment robustness shows that ThermoRAFT-AFP achieves state-of-the-art subpixel accuracy with a mean RMSE below 5 mm/s and relative cumulative drift under 0.1%, all while operating at 25 fps on a commodity CPU. The system maintains robust performance under severe thermal noise and reliably generalizes across diverse process conditions. Qualitative evaluation against realistic AFP case studies highlights the framework's capabilities for thermal anomaly visualization and tracking, inter-layer thermal behavior propagation analysis, and enabling operator-informed decision-making. These findings establish a reliable foundation for next-generation intelligent AFP process monitoring and quality inspection systems.