Session: Research Posters

Paper Number: 166380

Investigation into the reliability of Real-Time Defect Detection in DED Process Using CNN-Based Deep Learning with Infrared Image 

Directed Energy Deposition (DED) is an additive manufacturing (AM) process that employs a high-energy heat source to melt a supplement material on the substrate. The DED process is particularly effective for repairing and refurbishing mechanical components and enabling multi-material fabrication that provides enhanced wear and corrosion resistance. However, DED produced components often experience quality problems such as microstructural heterogeneity, porosity, and cracking. These defects, primarily caused by thermal Inhomogeneous, negatively impact the mechanical properties and overall reliability of the final products. Detecting defects early helps maintain process reliability and prevents quality deterioration. Hence, implementing real-time monitoring systems for detecting deposition defects is an effective method to overcoming deposition problems. This study aimed to investigate the reliability of Real-Time Defect Detection in the DED Process Using CNN-Based Deep Learning with Infrared Images. In order to develop a CNN-based artificial intelligence (AI) model for the DED process, experiments for learning data acquisition, pre-processing of collected data, training of AI model, and reliability test with the suggested AI model were sequentially performed. The experimental set-up for robot base DED with the off-axis in-situ monitoring system was designed. The off-axis system provides a global observation of the melt pool and substrate, allowing for a broader analysis of process stability. A robot-based DED system was used for the experiments. Ytterbium fiber laser was employed with a wavelength of 1,040 nm for the heat source of the DED system. KUKA industrial robot system was used to operate the deposition system. The experiment for learning data acquisition was performed to acquire Infrared image data for training. S45C steel substrates were used as the base material, and SUS316L metal powder, with particle sizes ranging from 53 to 150 µm, was deposited onto these substrates to create multi-layer test specimens measuring 50×50 mm. The laser power and scanning speed were used in the experiment parameters. The ranges of laser power and scanning speed were from 350 to 1,100 W and 1,000 to 1,500 mm/min, respectively. The acquired image data were classified into three conditions including insufficient energy input, excessive energy input, and normal energy input. These IR images were used as training data to ensure accurate defect detection under diverse process conditions. The collected IR images were labeled to indicate defect regions. The YOLO algorithm and data augmentation techniques were used for training. A suitable level of dataset size and resolution to improve detection accuracy was suggested. The performance of the trained AI model was tested by analyzing precision, recall, and detection accuracy. From the results of the test, the optimal confidence and Intersection over Union (IoU) thresholds were determined. The accuracy of the developed system was over than 70 %. Through the analysis of these performance metrics, the reliability of the suggested AI model with the monitoring system for the DED process was discussed.

Presenting Author: Yoo-Ri Lee Korea institute of Industrial Technology

Presenting Author Biography: Researcher Lee Yoo-ri is a master's degree student in the Department of Mechanical Engineering at Pusan ​​National University, South Korea, and is conducting research related to 3D printing application technology.

Authors:

Yoo-Ri Lee Korea Institute of Industrial Technology
Ho-Jin Lee Korea institute of Industrial Technology