Defect detection in EBSM components through selective box fusion of modern object detection

Area of Science:

• Materials Science
• Manufacturing Engineering
• Computer Vision

Background:

• Additive Manufacturing (AM), specifically Electron Beam Selective Melting (EBSM), offers high precision and material properties for aerospace and automotive applications.
• Internal defects in EBSM parts significantly compromise their performance and reliability.
• Accurate detection of these defects is crucial for quality control in AM.

Purpose of the Study:

• To evaluate the effectiveness of modern object detection models in identifying internal defects in EBSM parts using X-ray computed tomography (CT) data.
• To propose and validate a novel ensemble method, Selective Box Fusion (SBF), for improved defect detection.
• To enhance the overall performance and reliability of EBSM components through advanced defect analysis.

Main Methods:

• Utilized X-ray computed tomography (CT) to generate cross-sectional images of Electron Beam Selective Melting (EBSM) parts.
• Trained and evaluated state-of-the-art object detection models, including Sparse R-CNN and YOLO series, on CT images for internal defect detection.
• Developed and implemented a Selective Box Fusion (SBF) ensemble approach, combining voting and weighted fusion of detection boxes.

Main Results:

• Sparse R-CNN showed strong overall performance, while YOLO models excelled in specific detection metrics.
• The proposed Selective Box Fusion (SBF) ensemble method demonstrated superior performance across various evaluation metrics compared to individual models.
• SBF effectively mitigated errors inherent in individual detection models by integrating their strengths.

Conclusions:

• Modern object detection models show promise for identifying internal defects in AM parts.
• The Selective Box Fusion (SBF) ensemble approach offers a robust and effective strategy for enhancing defect detection accuracy in EBSM.
• This work contributes to improving the quality and reliability of additively manufactured components through advanced computational defect analysis.