Neural Net Based Machine Learning Algorithms for Automated NDT of Military and Other Hardware

NOVI proposes to develop a prototype system capable of automating NDT of munitions using deep learning techniques. The system will be designed to process a stack of radiography scans for a 6.5” diameter, 36” tall munition with 0.002” – 0.020” defects in under one minute, and largely utilize Government-provided images (appropriately annotated) to create the necessary training data set. Specific activities will include: 1. Developing a unique set of training data with labeled defects in radiographic scans of munitions; 2. Customizing and training a LinkNet Convolutional Neural Net (CNN) model for the defect detection task; 3. Developing a framework for assembling 2D results into a volumetric 3D output for end-user analysis; 4. Testing the trained model on out-of-sample scans, and providing qualitative and quantitative performance analysis for detection accuracy and speed; and 5. Providing recommendations for continued enhancement of the approach, along with a preliminary plan for integration into production-class radiography inspection systems.