Related Experiment Video
Updated: Jun 26, 2025
10:10
Three-Dimensional Particle Shape Analysis Using X-ray Computed Tomography: Experimental Procedure and Analysis Algorithms for Metal Powders
Published on: December 4, 2020
1.8K
Automated Porosity Characterization for Aluminum Die Casting Materials Using X-ray Radiography, Synthetic X-ray Data
Stefan Bosse1, Dirk Lehmhus2, Sanjeev Kumar3
1Department of Mathematics & Computer Science, University of Bremen, 28359 Bremen, Germany.
Sensors (Basel, Switzerland)
|May 11, 2024
Summary
This study addresses challenges in detecting material defects using X-ray imaging and machine learning. Synthetic data aids model training but real-world application requires careful consideration of noise and data quality for robust defect detection.
Area of Science:
- Materials Science and Engineering
- Non-Destructive Testing (NDT)
- Artificial Intelligence in Manufacturing
Background:
- Detecting hidden defects in materials like aluminum die castings and Fiber-Metal Laminates (FML) remains a significant challenge.
- Existing methods struggle with data variance, accurate feature labeling, and establishing ground truth for supervised machine learning.
- X-ray imaging techniques, including radiography and computed tomography (CT), are crucial for NDT but require advanced analysis.
Purpose of the Study:
- To discuss methods and challenges in data-driven modeling for automated damage and defect detection using X-ray imaging.
- To investigate the utility of synthetic data for training robust machine learning (ML) feature detectors.
- To explore the transition from high-quality CT to lower-quality single-projection radiography for defect analysis.
Main Methods:
- Utilized semantic pixel Convolutional Neural Networks for data-driven feature detection.
- Generated synthetic X-ray data to create ground truth datasets and labels for supervised ML.
- Acquired experimental data using diverse X-ray devices: low-quality radiography, industrial radiography/CT, and high-quality micro-CT (μ-CT).
Main Results:
- Synthetic data can provide ground truth and labels for ML model training but is insufficient for direct prediction of manufacturing defects.
- Noise significantly impacts feature detection accuracy, necessitating careful consideration in data-driven models.
- Trained robust and generalized ML feature detectors using only synthetic data, demonstrating potential for broader application.
Conclusions:
- Data-driven approaches, particularly with synthetic data augmentation, show promise for automated defect detection in materials.
- The quality of X-ray imaging devices impacts feature detection; transition from CT to single-projection radiography is feasible with robust models.
- The developed methods are applicable beyond aluminum die castings to various materials and defect characterization tasks.
