Real-time Defect Detection in Wire-laser Directed Energy Deposition Process Using U-Net-based Semantic Segmentation

This study introduces a real-time monitoring system for wire-laser directed energy deposition (W-LDED) process, utilizing a U-Net-based semantic segmentation model. The system accurately identifies critical defective features in the process, such as the residual heat affected zone (HAZ) and dripping defects, at the pixel level. A dataset was collected using a high dynamic range camera, and the U-Net was trained for pixel-wise classification of process images. By fine-tuning hyperparameters of model and applying data augmentation, the segmentation performance was enhanced, enabling the precise extraction of positions and boundaries of distinct regions in real-time process images. Additionally, a pixel-based morphology measurement algorithm was developed to quantify the length of the residual HAZ and the area of the dripping defect. This system provides in-depth insights into the process and contributes to improved monitoring and control of key parameters, which ultimately enhances process stability and bead quality. The results suggest that this monitoring system could be integrated into future automated control frameworks for the W-LDED process, thereby enhancing productivity in the metal additive manufacturing industry.