Measurements for Combined Gamma-Ray and Video Modalities

Project # 23-024 | Year 2 of 3

Chris Burt^a, Rod Stabile^a, Roy Abbott^a, Johanna Turk^a, Ronald Wolff^b, Michael Willis^c

^aSpecial Technologies Lab (STL), ^bRemote Sensing Lab-Andrews (RSLA), ^cOak Ridge National Laboratory
_{This work was done by Mission Support and Test Services, LLC, under Contract No. DE-NA0003624 with the U.S. Department of Energy, the NNSA Office of Defense Programs, and supported by the Site-Directed Research and Development Program. DOE/NV/03624–1912.}

Abstract

Previously, the Nevada National Security Sites (NNSS) created a network of spectral gamma sensors in Northern Virginia (aka NoVArray), supporting Defense Nuclear Nonproliferation (DNN) NA-22. This current research effort is upgrading these aging, unused NNSS assets and exploring ways to improve detection of materials of interest through multimodal data analysis. The applicability space of the NoVArray is being extended from dataset generation to information generation by integrating the platforms with edge compute, context imagery, and communication hardware. Utilizing a novel data analysis pipeline that includes in-house gamma convolutional neural networks (CNN) for isotopic identification and object recognition via video, a robust platform for Special Nuclear Material (SNM) detection is being created. In the first two years of this Site-Directed Research and Development (SDRD) project, permits were extended, hardware was redesigned then updated, contextual sensors were added, and an initial implementation of real-time processing was performed. The third year of this SDRD will evaluate the performance of the systems at Oak Ridge National Laboratory (ORNL) with well-characterized radioactive sources, and complete upgrades for the other 18 NoVArray units that had been deployed in Fairfax County, VA.

Background

The original NoVArray provided a rich gamma dataset for the nonproliferation and data science communities but lacked ground truth for contextual information. When targets were identified in data as potential materials of interest, there was no way to verify. The original platforms did not do any on-board processing of the data. Large amounts of unlabeled, uncurated data were telemetered to the cloud, requiring expensive human-in-the-loop analysis. In the years since these platforms were deployed, the subcontract with a previous corporate collaborator on the project expired. The data were linked to their cloud infrastructure, preventing NNSS from accessing them without payment. Recently the communications hardware in the sensors became obsolete, so they had been unable to communicate at all. This SDRD project hopes to revitalize an unused asset.

Technical Approach

In Fiscal Year (FY) 2022, platform limitations were addressed by redesigning the hardware to utilize visible imagers, ingest meteorological data, and process gamma and visible imagery in real-time. The contextual information has potential to improve identification algorithms and the real-time analysis extends NoVArray’s applicability space from a dataset generating array to an information generating one. These efforts utilize custom NNSS software developed for previous projects: object recognition software labels targets in the video, and CNNs analyze gamma data for isotopic identification.

The Special Technologies Laboratory’s (STL’s) custom object recognition software leverages the popular open-source algorithm You Only Look Once (YOLO)¹ by adding in custom motion detection algorithms, object filtering, and thresholding.

In Hoteling et al. (2021)², an NNSS research team published a novel gamma analysis technique, utilizing machine learning tools typically leveraged for image analysis. By converting the list-mode gamma data (2D) into waterfall plots (3D), the proven power of CNNs can be used to classify the gamma spectra “images.”

Early efforts for this project included understanding how to implement in real-time previously developed code architectures, data flows, input/output formats, and applicability domains. The historical applicability domain for past work has been with static detector and mobile source scenarios, which is the same as the NoVArray detectors. The team modified the NaI detector data acquisition software for persistent acquisition and implemented the codebase at the edge on the NoVarray units.

In FY 2022 and early FY 2023, hardware builds for two prototype NoVArray units were completed. These units were deployed to ORNL where real-world gamma sources were utilized to evaluate system performance. At the end of FY 2023 the old units that had been deployed in Northern Virginia were removed and sent back to STL to begin the process of upgrading the hardware.

In FY 2024, testing at ORNL will continue with added complexity. Most of the deployed units will be upgraded to match the prototypes and re-deployed to Fairfax County. Algorithms will continue to be developed and optimized, and a detailed evaluation will be performed at ORNL. The AWS [Amazon Web Services] cloud-based infrastructure for data storage will be built so that data can be telemetered into an infrastructure controlled by NNSS. Orthogonal data modalities will be examined to see the feasibility of utilizing them to reduce uncertainty and improve classification results (such as open-source weather data scraped from the internet).

Potential transition partners for this work have been identified, and this technology will be discussed with these interested users, including partners at Fairfax County, Los Angeles City, and DNN.

• [1] Redmon, J., Divvala, S., Girshick, R., and Farhadi, A., “You Only Look Once: Unified, Real-Time Object Detection,” arXiv, 2015.
  
• [2] Hoteling, N., Moore, E., Ford, W., McCullough, T., and McLean, L., “An analysis of gamma-ray data collected at traffic intersections in Northern Virginia,” 2021 IEEE Symposium on

Results and Technical Accomplishments

In the first two years of this Research and Design (R&D) effort, an improved NoVArray sensor was designed, built, and tested. The crystal and digibase were integrated with two edge processors, a power-over-ethernet (POE) camera, and a Cradlepoint cellular modem for communications. A novel and proven CNN architecture for isotopic gamma identification that had been abandoned due to loss of staff was rebuilt. A new data collection and real-time data analysis pipeline for combining gamma-ray and video modalities was created. The contextual video object recognition software is complete, and the initial gamma anomaly detection algorithm appears to be working well. Issues with the CNN accuracy are being troubleshooted, but a workflow to utilize the architecture in real-time has been implemented. This is something the CNN architecture had not been originally designed for. Additionally, most of the original units deployed throughout Northern Virginia were pulled down, with the intention of upgrading them to the new design in FY 2024. Initial bench top testing of the units, with sources, has been completed at STL and at ORNL. This project has also allowed for the growth of close collaborations between STL and Remote Sensing Laboratory, Andrews, as well as ORNL.

Conclusions and Path Forward

In FY 2024 system testing and evaluation efforts at ORNL will continue. The NoVArray prototype platform will be moved to ORNL’s High Flux Isotope Reactor (HFIR). This will allow for a more practical real-world testing. Sources will be driven past the sensor at logical distances and speeds, similar to how the sensors operate in Northern Virginia. STL will utilize these data to build optimal calibration configurations for the sensors, as well as troubleshoot any issues with the CNN architecture. The remaining old units will be upgraded to include the new edge processing and communications hardware and integrated with the POE cameras. A sensor shake out will be performed for these devices at ORNL prior to having them redeployed in Northern Virginia. Also in FY 2024, the AWS cloud infrastructure for data storage will be developed. It is hoped that a way to utilize the object recognition results to reduce uncertainty in gamma CNN isotopic ID can be developed. In FY 2024, a transition partner for the technology and tools will be sought out, with contacts already made at DNN and Los Angeles City. Work from this SDRD research is already being leveraged in a DNN-funded project in FY 2024.

A new NNSS asset has been developed that has the potential to be an important tool for nuclear nonproliferation R&D. This platform allows for an in-house solution for a data collection, transmission, and curation architecture that utilizes NNSS developed hardware and software. The novel, real-time analysis tools developed are already being utilized by other projects and can be a useful tool in gamma source detection and isotopic characterization.

Back to Enabling Technologies for Autonomous Systems and Sensing