AI/ML for Visual Processing of Energetic Defects

Defects occurring in the energetic fills can alter a munitionÆs safety and performance characteristics. Small defects, such as gaps, voids, cracks, and foreign materials, in the chemical makeup of the energetic material can lead to un-foreseen chemical reactions that have lessened or increased the potency of the energetic material, that would lead to costly, hazardous, or even life-threatening misfires. Currently, energetic filled parts are manually inspected for critical defects during the manufacturing process using x-ray imagery. A similar manual inspection process is also employed to analyze and understand ambiguous cases post-production and in the design of the munition defect specification that defines critical defects. Manual assessment is not only costly, laborious, and time-consuming, but also prone to error and inconsistencies. In addition, x-ray imaging cannot adequately penetrate the metal shell surround-ing the energetic material to show the condition of the chemicals within. Therefore, the Army has a critical need for accurate and reliable detection of energetic defects to eliminate defective parts from leaving the production floor, thereby reducing the potential for hazardous incidents in the field. Ā During Phase I, Sky Park Labs demonstrated the feasibility of EnergeticAI, an AI system that analyzes radiographic images (CT, n-rays, and x-rays) to automatically identify and localize energetic defects. EnergeticAI uses ōweakly supervisedö machine learning to learn normal patterns exclusively from readily available, fault-free, radiographic images of munitions and highlights the location of deviations as potential defects. Then, the system automatically classifies the defect type and severity by clustering based on image similarity. During Phase I, we demonstrated these capabilities on CT, n-ray, and x-ray data, as well as a variety of publicly available industrial datasets, achieving accuracy, precision, and recall rates of over 97% for detection and classification tasks. The Phase II work will leverage the Phase I preliminary designs and prototype results to design and implement a full-scale software tool automatically identify and characterize defects in energetic fills of munitions.