Dan Shen, Intelligent Fusion Technology, Inc; Carolyn Sheaff, Air Force Office of Scientific Research (AFOSR); Genshe Chen, Intelligent Fusion Technology, Inc; Mengqing Guo, Intelligent Fusion Technology, Inc; Nichole Sullivan, Intelligent Fusion Technology, Inc.; Erik Blasch, Air Force Office of Scientific Research (AFOSR); Khanh Pham, Air Force Office of Scientific Research (AFOSR)
Keywords: Space situational awareness, GANs, ResNet, satellite characterization, sensor models, machine learning, general-sum games, 3D-CNN
Abstract:
Perhaps the biggest obstacle to adopting machine learning (ML) techniques and evaluating their analytics in space situational awareness (SSA) application is the lack of large labeled data sets for training and validation. In the literature, there are several techniques designed to deal with small data sets including boosting, transfer learning, and simulation. Transfer learning enables users to construct a network for a small dataset without overfitting, however, the feature detectors need to be trained on a large dataset. An autoencoder is a deep neural network (DNN) whose output and input are the same. Thus, for an autoencoder, any dataset is labeled dataset, but the events detected by trained autoencoders are limited. Usually, other algorithms are used to classify the results of autoencoders. A training set of a small size can also be made to appear larger through data augmentation. One of the challenges to use data augmentation is that in the real-world scenario, the data is captured over a limited set of conditions while machine learning algorithms require training data over as many possible operation conditions.
In this paper, we present a game theory enabled data augmentation method for satellite behavior detection. The realistic sensor data (azimuth angle, elevation angle, range, range rate) are propagated using SGP4/SDP4 and the various maneuver strategies from the proposed space game model, which is played by on-ground radar and space objectives. We use two-player Markov game to investigate the sensor management for tacking evasive space objects. The Markov game to investigate the sensor management for tracking evasive space objects. The Markov game approach provides a method to solve SSA behavior detection problems, where the Resident Space Objects (RSO) will exploit the sensing and tracking model to confuse the SA observer by corrupting their tracking estimates, while SA observer wants to improve the tracking performance. The different cost functions will generate different maneuver strategies. In addition, to rich the training data, we exploited generative adversarial networks (GANs) to furtherly augment the simulated data. GAN models can generate more data from the simulated data to improve the robustness of our trained model as well as increase the inference accuracy of the trained networks for space behavior detection. In this way of simulated data plus GANs, various satellite behaviors are simulated and augmented to generate synthetic datasets with labels for use by ML methods. To demonstrate the proposed data generation method, we process catalog data (space-track.org) and obtain the tracks without maneuvers. Then, we simulate various satellite behaviors guided by game-theoretic maneuver strategies. The catalog data are modified based on the propagation results from SPG4/SDP4 with maneuver strategies and labeled based on the behavior simulated.
To evaluate the performance, a 3D convolutional neural networks (3D-CNN) is designed and implemented for satellite behavior classification. Python 3 and TensorFlow Keras are used in this implementation. The 3D-CNN is provided the generated synthetic and labeled data with 143 possible satellite behaviors (15 degrees for two-directions maneuvers). Over 72,000 generated tracks, 1/10 of the data is randomly selected as test set for validating the 3D-CNN performance. The convolutional filters are initially chosen arbitrarily, so the classification is done randomly. It is noted that as iteration increases, the training model converges nearly to 100% accuracy. The remaining dataset is used for evaluation for the well-trained CNN model. The trained machine learning model can efficiently and correctly classify the satellite behaviors with a 97% rate. Meanwhile, the ResNet improved GAN model architecture can accurately perform the data augmentation, which further augment the simulated data for building robustness 3D convolutional neural networks. Overall, our model-based game theoretic synthetic data can improve the training and validation performance of machine learning algorithms.
Date of Conference: September 15-18, 2020
Track: Machine Learning Applications of SSA