The past decade has seen scores of academic studies investigating the use of machine learning to automatically find anomalies in spacecraft telemetry streams. Despite the tremendous progress in this area, use of machine-learning at actual ground stations is not widespread. Much of this lack of adoption is due to the fact that training machine-learning models requires subject-matter expertise. Moreover, once the models have been trained, there exists no accepted and straightforward protocol for (a) integrating these models into existing workflows at ground stations and (b) ensuring that spacecraft operators are able to configure and deploy those models with ease. In an attempt to help ameliorate some of these issues, this paper introduces the AMBASSADOR framework and analysis tool. AMBASSADOR allows ground-operations personnel to train semi-supervised novelty-detection algorithms without the help of machine-learning experts. An arbitrary number of models may be trained at the same time and each of those models can be dependent on an arbitrary number of telemetry streams. Once the models have been trained, the AMBASSADOR framework allows them to be used for inference - either on spans of stored telemetry or on live telemetry as it arrives in real time. In order to ensure scalable deployment and ease-of-use, the AMBASSADOR framework has been implemented within a microservices architecture using a containerized Python API. This API is capable for retrieving data from actual telemetry archives such as COSMOS DART. User interaction with the API is facilitated via a modern, intuitive, web-based frontend. Taken together, this means that multiple users can simultaneously train custom anomaly-detection models and use them to analyze actual telemetry streams without ever having to interact with the underlying machine-learning code.
Automated Training and Deployment of Machine-Learning Models for Anomaly Detection in Telemetry
2022-03-05
4231087 byte
Conference paper
Electronic Resource
English
British Library Conference Proceedings | 2018
|EVALUATING ANOMALY DETECTION IN SATELLITE TELEMETRY DATA
TIBKAT | 2022
|