Speaker
Description
Laser-based additive manufacturing processes require precise monitoring of meltpool geometry to ensure process stability, dimensional accuracy, and metallurgical quality. Infrared thermography provides a non-contact, high-speed sensing solution capable of capturing thermal emission during laser–material interaction. However, extracting reliable geometric and energetic features from meltpool images remains challenging due to high dynamic range, sensor noise, blooming effects, and rapid temporal variations inherent to the process.
Conventional image processing techniques, including fixed global thresholding, adaptive thresholding, morphological filtering, contour extraction, and ellipse fitting, are commonly used for meltpool segmentation. Although these approaches are computationally efficient and suitable for real-time implementation, they are strongly dependent on threshold selection and local intensity normalization. Adaptive methods can artificially stabilize the segmented region, reducing sensitivity to subtle geometric changes, while fixed thresholds increase sensitivity but become unstable under background drift and fluctuating thermal conditions. As a result, small yet physically relevant variations in meltpool morphology may be masked or exaggerated, compromising the robustness of feature extraction for control loop integration.
To overcome these limitations, this work proposes a deep learning-based segmentation pipeline for real-time meltpool monitoring using thermographic image data acquired during a laser-based manufacturing process. A dataset of infrared images was manually annotated at pixel level to generate ground-truth segmentation masks. A U-Net++ architecture was selected due to its multi-scale feature aggregation capability and improved boundary reconstruction, enabling accurate delineation of the meltpool region under varying thermal conditions.
The trained network was exported to ONNX format and deployed within a real-time inference pipeline using ONNX Runtime with GPU acceleration, allowing integration into an industrial monitoring framework. From the predicted segmentation masks, physically meaningful descriptors were extracted, including meltpool width (via minimum-area rectangle fitting), segmented area, and energy-related metrics derived from baseline-subtracted intensity integration within the detected region.
The neural segmentation approach was quantitatively compared with classical image processing strategies under varying laser power levels. Results indicate improved spatial consistency, enhanced sensitivity to process-induced variations, and reduced dependence on manual parameter tuning. In addition, inference latency measurements confirmed compatibility with real-time industrial constraints. These findings demonstrate that deep neural segmentation, when coupled with physically grounded feature extraction, provides a robust and scalable framework for quantitative thermographic meltpool monitoring in industrial additive manufacturing environments.
