PhD Dissertation: Muhammed Zemzemoğlu

A Learning-Centric End-to-End Hybrid System for In-situ Thermographic Inspection in Automated Fiber Placement

Muhammed Zemzemoğlu
Mechatronics Engineering, PhD Dissertation, 2025

Thesis Jury

Prof. Dr. Mustafa Ünel (Thesis Advisor)

Prof. Dr. Bahattin Koç
Assist. Prof. Dr. Melih Türkseven

Prof. Dr. Haydar Livatyalı

Assoc. Prof. Dr. Ali Fuat Ergenç

Date & Time: July 22nd, 2025 – 10:30 AM

Place: FENS L067

 

Keywords: Automated Fibre Placement (AFP), In-Situ Process Monitoring, Thermographic Inspection, Defect Identification, Quality Assessment, Motion Estimation, Lay-up Reconstruction, Computer Vision, Deep Learning 

Abstract

Automated Fiber Placement (AFP) technology continues to lead and transform composite manufacturing, but its progress remains constrained by persistent quality assurance challenges. Emerging material and process defects compromise structural integrity, while inspection practices remain largely manual, reactive, and incapable of real-time feedback—leading to costly downtimes. Existing methods face two core limitations: partial, mostly offline frame-wise analysis and the lack of global, temporally consistent lay-up visualization. To address these challenges, this thesis proposes a dual-framework, real-time, learning-centric inspection system that unifies local defect intelligence with global lay-up traceability—operating machine-independently using only thermal imagery.

The first framework implements a hybrid, frame-wise defect analysis and quality assessment system composed of three synergistic modules—Dynamic Tow Identification, Hierarchical Defect Identification, and Lay-up Quality Evaluation—configured to run in parallel or conditionally for runtime efficiency. It begins with setup-independent spatial–temporal analysis that estimates tow boundaries with sub-pixel accuracy (mean error < 0.8 px), enabling tow-level reasoning. In parallel, high-level defect detection uses a Gabor-based SVM classifier exceeding 95% accuracy and recall. Defective frames are forwarded to a custom 12-layer deep convolutional neural network for fine-grained classification, achieving 96.4% multi-class accuracy across defect types. Upon detection, a seeded active contour model adapted to thermal textures yields interpretable segmentation masks with 93.2% pixel accuracy—outperforming baseline methods under noise. These outputs are fused with tow geometry to compute the novel Defect Area Percentage (DAP) metric, which quantifies severity at both tow and course levels and forms the core of the operator decision support system (AFP-DSS). Operating at 5 fps, the framework enables fully autonomous, real-time AFP quality inspection.

The second framework introduces ThermoRAFT-AFP, a machine-independent, deep learning-based thermal motion estimation core tailored to AFP. Built upon the RAFT optical flow, it incorporates AFP-specific augmentations and runtime optimizations (e.g., predictive initialization, drift correction, adaptive early exit) for stable, precise thermal flow tracking. It estimates dense inter-frame motion to drive a two-stage reconstruction pipeline that generates course-wise mosaics and assembles high-fidelity, ply-level laminate visualizations. This restores temporal consistency and global alignment across evolving lay-ups, enabling traceable defect analysis and mirroring industrial inspection workflows. ThermoRAFT-AFP achieves a velocity estimation RMSE of 4.83 mm/s and cumulative drift below 0.1%, while maintaining robustness down to SNR = 14.4 dB and sustaining real-time operation at 25 fps. Robust against setup variations without retraining and near-zero tuning, it produces interpretable, temporally aligned visualizations that support laminate-scale defect propagation analysis.

Together, the two frameworks form an integrated system that fuses frame-wise defect outputs with global reconstructions into a unified thermal quality view. This fusion links semantic analysis with temporal context in an interpretable visualization aligned with real-world inspection workflows. Evaluated on over 13,000 thermal frames spanning diverse speeds, geometries and defect types, the system meets aerospace-grade benchmarks, eliminates robot-coupled dependencies, and delivers scalable, real-time AFP quality inspection directly from thermal imagery.