Jorge O’Farrill, Modern Technology Solutions Inc.; Tracy Mims, Modern Technology Solutions Inc.; Tad Janik, Modern Technology Solutions Inc.; Ivan A. Fernandez, Modern Technology Solutions Inc., Mississippi University
Keywords: EO/IR data fusion, characterization of unresolved objects
Abstract:
Advances in computing power and Artificial Intelligence have led to new areas of research with application to space situational awareness (SSA). We propose to use reinforcement learning coupled with a fast signature generator to characterized space objects by ingesting their EO/IR signals and estimating their physical configuration. We will further determine if that physical configuration has changed since the last observation. This characterization will be performed using unresolved imagery and realistic targets. We will characterize the shape and material properties of space objects by fusing unresolved infrared and visible data collected from remote sensors. This will be done by leveraging previous work performed by our team at MTSI: SBIR for multi-band IR characterization of signal modulation, multiple IRAD and SBIR projects for faster than real-time high fidelity EO/IR data generation, IRAD for the use of reinforcement learning to augment test and evaluation (TE) and MTSI support of MDA/SS’s (and SDA) Hypersonic and Ballistic Tracking Space System (HBTSS). We will show that an object’s physical configuration can be estimated and that information can be used to characterize changes in that configuration. The characterization of changes based on target physical configuration should be more informative and robust than using the EO/IR signals alone
Our approach relies on the efficient generation of high-fidelity EO/IR signatures. MTSI has developed a fast-running EO/IR signature generator called HIRTSS (High Rate Thermal Solver). It is being used to support both threat data generation and algorithm verification for ground-based missile defense in the Missile Defense Agency, MDA/GM. This signature solver will be used to fit the signals collect by EO/IR sensors to a class of predefined space object geometries: cubes, cylinders, and various paneled configurations. These objects are characterized by a mesh grid with surface materials on each facet. We will use a simple reinforcement learning algorithm, a multi-armed bandit (MAB), to morph the grid and choose between materials to best fit the incoming EO/IR signals, The MAB will return the physical configuration of the target that minimize the residuals between the collected EO/IR signals and those predicted by HIRTSS. Unlike standard optimization techniques, the MAB does not get stuck in local minima, and it is possible that more than one target configuration will yield viable results. This information will be valuable for the purposes of detecting changes.
We will create two groups to perform this research. The red team will generate EO/IR signatures data for simple objects (cones, spheres, cubes) made of various materials using the Optical Signatures Code (OSC). OSC has been used for decades to produce high fidelity EO/IR signatures in support of Ballistic Missile Defense (BMD). These target configurations will not be known to the black team. The black team will run these signals through representative sensor models to generate unresolved imagery. This data will stimulate the HIRTSS+MAB algorithm for target configuration estimation. Once we show that the MAB can estimate the physical configuration of simple targets, we will introduce changes in their physical configuration and assess the ability of our algorithm to detect and characterize these changes. This phase of the research will allow us to determine the morphing parameters and appropriate material properties databases needed to characterize simple targets. The red team will then generate EO/IR signatures for complex targets with articulating components, including solar panels using HIRTSS. We will then repeat the analysis and leverage lessons learned from the simple targets to estimate the physical configuration of complex targets.
Our minimum viable product is an algorithm that can estimate the shape and materials of simple targets and determine if that configuration has changed. We will estimate their shape and material properties. Certain classes of changes will be characterized: size change, heating, even possibly reorientations. We will be able to quantify data collection requirements: sensor radiometric accuracy and collection time for the different physical configurations We have benchmarked HIRTSS against various signature generation codes employed by MDA therefore we expect our results to generalize to any set of target configurations the space community desires. We expect our method to outperform those that do not try to estimate the target’s physical configuration at the expense of possibly oversimplifying this physical configuration to avoid fine-tuning.
Date of Conference: September 27-20, 2022
Track: Non-Resolved Object Characterization