Directed energy deposition (DED) is attractive for localized repair and dissimilar-material deposition; however, its quality is highly sensitive to history-dependent thermal behavior, which can readily lead to defects. Since quality assurance is still...
Directed energy deposition (DED) is attractive for localized repair and dissimilar-material deposition; however, its quality is highly sensitive to history-dependent thermal behavior, which can readily lead to defects. Since quality assurance is still largely performed after fabrication, there is a strong need for in-situ monitoring and AI-based defect detection grounded in measurable thermal information. To address this, this study proposes an integrated framework combining a thermal analysis model with process monitoring experiments to elucidate the fundamental heat transfer characteristics of the DED process, classify process states based on thermal features, and detect defects using object detection.
First, a bead width prediction model based on conduction-mode thermal analysis was used to theoretically predict bead formation behavior as a function of process variables. To experimentally validate this, single bead deposition experiments were conducted with powder feed rate, scan speed, and laser power as the primary variables. Based on the acquired temperature data, the average bead temperature (ATB) and average cooling rate (ACR) were derived to analyze correlations between process parameters and thermal characteristics. The results confirmed that lack-of-fusion, normal, and excessive melting states were clearly distinguishable based on their thermal behavior. Specifically, discontinuous bead formation or no deposition was observed under low heat input conditions, whereas continuous and stable bead formation was confirmed under appropriate heat input conditions. The experimental results showed a consistent trend with the predictive model, thereby verifying that the acquired dataset accurately reflects the underlying physical phenomena of the DED process.
In the cube deposition experiment, position-specific and layer-wise temperature data were extracted from the multi-layer deposition process. The average layer max temperature (ALMT) and average layer cooling rate (ALCR) were calculated to evaluate heat accumulation behavior in response to changes in the powder feed rate and layer count. A correlation analysis between thermal indicators and morphology, microstructure, and mechanical properties confirmed that variations in material properties could be quantitatively explained by changes in thermal characteristics.
Furthermore, an object detection AI model was trained using IR images acquired during single bead and basic geometry deposition. The training dataset was labeled with classes defined based on thermal characteristics and melt pool morphology. Specifically, a YOLOv5 model was employed to detect process states during deposition.
Performance evaluation results indicated mean average precision (mAP) scores of 0.719 and 0.966 for the bead and basic geometry datasets, respectively, demonstrating sufficient reliability for process state classification and defect detection. Consequently, the integrated framework proposed in this study provides a foundation for real-time detection and visualization of melt pool states and major defects in the DED process by combining thermal history-based indicators with object detection technology.