Machine Vision for inspection of Neptunium Oxide and Aluminum pellets
This project supports the NASA Plutonium-238 Supply Program, which utilizes Pu-238 in radioisotope thermoelectric generators (RTGs) for deep space missions. To produce this fuel, Neptunium oxide and aluminum pellets, fabricated at ORNL, must be inspected before they are irradiated in the High Flux Isotope Reactor (HFIR). The traditional inspection process was highly inefficient, relying on operators to manually inspect each pellet one at a time. This approach resulted in low throughput and exposed glovebox workers to high radiation doses. This project aimed to replace this manual, time-consuming process with an automated machine vision solution that could perform simultaneous inspections, adhering to ALARA (As Low As Reasonably Achievable) principles, maximizing throughput, and generating detailed documentation.
To validate the core concepts of the proposed machine vision process, I designed and built a robust, modular test bed that simulated the glovebox environment. This automated and repeatable platform utilized 80/20 aluminum extrusion and incorporated a Vention ball-screw linear actuator with 360mm of travel for precise scanning. The structure was designed to be modular and easily changeable, allowing for future improvements. This initial test bed proved that the proposed process could simultaneously perform visual defect and metrology inspections on a batch of up to 52 pellets. This capability established the project's viability to drastically reduce radiation exposure and increase throughput compared to the manual, single-pellet inspection method.
A critical step involved hardware evaluation, where I determined the initial Cognex Insight 3D L4050 camera's resolution was insufficient for the required 0.003 inch pellet diameter tolerance gap, only allowing for approximately 1.5 pixels. This led to the specification of the Keyence LJ-X 8060 as the alternative, which is capable of providing ≈19 pixels for the same tolerance gap. To prevent project delays, I strategized and implemented Dynamic Similitude by 3D-printing scaled mock pellets to match the new camera's resolution, allowing Python code development and testing to continue immediately. I developed custom Python scripts that utilize Photogrammetry and OpenCV to convert the laser-scanned point cloud data into 2D height maps for automated dimensional analysis and defect detection.
The project successfully proved the viability of the automated process, yielding a 98% success rate for dimensional measurements during initial mock pellet tests. Beyond dimensional checks, the solution provides increased quality assurance through AI/ML integration. This included collaborating on applying an Autoencoder to identify anomalies and testing a commercial Large Language Model (LLM) (4o-mini) for defect classification, proving the process could be leveraged for future AI applications. All project findings, design details, and process flows were meticulously documented and presented at both the AGS Conference and a dedicated AI/ML conference at Los Alamos National Laboratory. Current efforts have shifted to designing a more robust system to be integrated into the current radiological glovebox in the near future.
Differential Swerve Drive