Nathan Highsmith, Modern Technology Solutions, Inc.; Jorge O’Farrill, MTSI
Keywords: Digital Twins, Transformers, Adaptive Optimization, Adversarial Testing
Abstract:
We will demonstrate that artificial intelligence can be used to emulate complex physical models for the purpose of generating precision estimation of target attributes from data collected by non-resolved remote EO/IR data. Our approach relies on both generative and inferential deep neural networks that are trained on data generated by high-fidelity physics-based models. Our presentation will provide technical details on data conditioning, designing, training, and testing of our AI model. We then show how it can be used as a kernel for estimating changes in a target’s geometric configuration using EO/IR data collected by ground-based sensors.
Physics-based models provide representative data from which concepts can be developed constructively or evaluated against realistic environments. These models typically have computation times that scale with fidelity. This computational overhead makes it hard to use such models as kernels for estimation engines, especially over large dimensional domains.
Characterization of space object attributes can be used to predict future behavior and deviations from this expected behavior may involve estimating small changes in target attributes with limited data. Meaningful estimation of changing target attributes with paucity of data requires representative high-fidelity models. The computational burden of these models, however, is prohibitive for two reasons: proliferation assets in the near-earth space environment and lack of precise physical representation of targets of interest. To bridge the gap between fidelity and run time we have developed narrow digital twins (NDTs) that encode the physical response, such as observed infrared signature of a specific target configuration, into deep neural networks. These machine learning models provide a high-fidelity response without the need for large volumes of training data. This reduction of data is also beneficial due to the large run-times of the physics-based models. To make our models robust, we use a reinforcement learning test environment called Harnessing Artificial Intelligence to Develop and Evaluate Systems (HADES) to perform adversarial attacks on our NDT. This testbed maps out the relationship between the input physical model and the outputs of the NDT, specifically where the NDT is failing or producing large errors. This paradigm of training limited but precise models and testing with adversarial reinforcement learning is extensible and can provide a path for the Space Community to perform precision estimation of changes which will facilitate Space Situational Awareness.
We create our NDT by training generative and inferential deep neural networks to predict the temperatures on the faceted representation of two classes of satellites: cube satellite and cylindrical satellite with panels. The algorithm consists of a base model and an upsampler. The base model predicts the temperatures on a carefully selected subset of the facets, roughly 5% of the total facets. We then use transformers to upsample these subset facet temperatures to the rest of the body. The complete thermal profile of the target is then used to generate an EO/IR signature from the perspective of a user defined sensor. This model generates EO/IR signatures at rates of 1000-10000x of the physics-based models. The algorithm is then used as the kernel of an optimization engine that ingests a trajectory estimate as well as a time series of EO/IR signatures of known wavebands and performs a least squares minimization on predicted and ingested radiometric signal. The optimizer searches target geometry space: size scaling, stretching, and segment removal for the case of the satellite with panels for the best fit. If the attitude of the target satellite is known this solution will determine the temperature profile and shape that best fits the incoming data set. This technique is shown to work for one sensor and improves when more sensors are added. If we have sufficient viewing diversity the attitude of the target can be estimates alongside the geometry and thermal profile. In this case the optimizer must now search a space that includes five motion parameters associated with the attitude of the body: angular momentum vector and precession rate and angle about this vector.
We will provide examples of extracting geometric and attitude solutions and determine limits on precision as a function of sensor viewing diversity. This presentation can be delivered orally or as a poster.
Date of Conference: September 17-20, 2024
Track: Satellite Characterization