Key Takeaways
1. Each industrial 3-D printer produces a unique surface pattern specific to that machine.
2. A convolutional network can accurately identify the printer that produced a part with 98.5% accuracy.
3. The model can determine manufacturing process and material with up to 100% accuracy.
4. Image resolution and crop size influence model accuracy; smaller crops are sufficient for some techniques, while others require larger areas.
5. This method can enhance supply chain monitoring by verifying machine use and detecting unreported process changes.
Researchers at the University of Illinois have discovered that every industrial 3-D printer creates a unique surface pattern that is specific to that machine. By using a convolutional network trained on these patterns, they can accurately identify which printer produced a part.
Research Overview
The research team created 9,192 parts using 21 different commercial machines that utilized four types of additive-manufacturing processes: digital light synthesis, multi-jet fusion, stereolithography, and fused-deposition modeling. Each part was scanned at a resolution of 5.3 µm per pixel on a flatbed document scanner, resulting in a comprehensive high-resolution image library for both training and testing the model.
Model Performance
Employing an EfficientNet-V2 architecture along with a voting mechanism based on various random image crops, the model achieved an impressive 98.5 percent accuracy in identifying the source printer for previously unseen parts. Additionally, it could determine the manufacturing process and material with up to 100 percent accuracy, and even estimate the position of build trays for digital-light-synthesis parts within about 5 cm (~1.97 in).
Implications for Industry
The study explored how the accuracy of the model is influenced by image resolution and crop size. For techniques like digital light synthesis, a crop size of 200 µm was sufficient, while for fused-deposition parts, larger areas (around 3 mm) were necessary, although these could work with lower resolutions. This makes the method suitable for standard cameras and scanners.
Apart from basic classification, this method serves as a valuable tool for monitoring supply chains. It can verify whether a contractor used the correct machine, detect unreported changes in processes, and assist in tracing defective or counterfeit parts even without embedded labels or the need for supplier cooperation.
Source:
Link