Jun Yoshida, NEC Corporation; Taichiro Sano, NEC Corporation; Katsuaki Taya, NEC Corporation; Daiki Mori, NEC Corporation; Moe Otagiri, NEC Corporation
Keywords: Unsupervised learning, Anomaly detection, Satellite attitude estimation, Synthetic Aperture Radar (SAR), Space Situational Awareness (SSA), Space Domain Awareness (SDA), Machine Learning.
Abstract:
In recent years, the demand for remote sensing satellites has significantly increased, and these satellites have played a crucial role in military intelligence gathering. For instance, satellite imagery has been used to confirm military actions of specific countries, demonstrating their utility for military purposes. In such context, countries possessing remote sensing satellites must carefully consider the feasibility of conducting observation activities, given their capability to observe other nations.
While it is somewhat possible to infer observation areas from satellite position information, confirming whether a satellite was observing a specific area is vital information. Therefore, by estimating the satellite’s attitude, we considered a method to infer the observation behavior by understanding which direction the sensor was facing. Particularly for Synthetic Aperture Radar (SAR) satellites, maintaining a charged state over an extended period is necessary, and distinguishing between normal (charging) and abnormal (imaging) states is a critical challenge. By leveraging the ground truth data from our SAR Earth observation satellite, ASNARO-2, and using ground-based observation data (Range, Az, El, RCS, etc.) as input, we developed an AI model to identify whether attitude control was being performed. This capability allows the determination of whether the satellite was in an imaging state.
This research introduces an unsupervised learning algorithm designed to facilitate the identification and classification of data patterns without the necessity of pre-labeled abnormal data. While general unsupervised anomaly detection only outputs an anomaly score, the proposed technique not only provides an anomaly score but also maps the satellite’s state into a feature space, enabling quantitative measurement of its proximity to previous states. We applied pseudo-labeling to normal data, labeled the time-series data, and used an LSTM (Long Short-Term Memory) model to extract features from the time-series data. Based on these features, we trained using triplet loss, a loss function commonly used in deep learning for tasks such as face recognition, image retrieval, and identity verification. The purpose of this loss function is to train the neural network to position similar data points close to each other, while placing different data points further apart. By using pseudo-labeling on normal data and training with triplet loss, we constructed a learning process to obtain the embedding space for the “charging state.” By considering the average distance to the top K points in the embedding space as the anomaly score, we identify the “imaging state” from the “charging state” embedding space. In the case of the “charging state,” it is expected to be mapped close to the embedding space learned from normal data, while the “imaging state” is expected to be mapped far from the embedding space learned from normal data.
By estimating the satellite’s attitude based on ground radar observation data and verifying with actual data from ASNARO-2, we confirmed the effectiveness of this technique by observing an increase in anomaly scores during the “imaging state.” This technology is expected to contribute to the enhancement of Space Situational Awareness (SSA) and Space Domain Awareness (SDA). Prospects include applying this technology to other satellite systems for broader SSA. Additionally, improvements in the accuracy of attitude estimation technology based on ground radar observation data and further refinement of anomaly detection algorithms are anticipated. This will enhance safety and reliability in space, opening new possibilities for space development.
Date of Conference: September 16-19, 2025
Track: Satellite Characterization