Detecting anomalies in spacecraft telemetry using evolutionary thresholding and LSTMs

Detecting anomalies in telemetry data captured on-board satellites is a pivotal step towards their safe operation. The data-driven algorithms for this task are often heavily parameterized, and the incorrect hyperparameters can deteriorate their performance. We tackle this issue and introduce a genetic algorithm for evolving a dynamic thresholding approach that follows a long short-term memory network in an unsupervised anomaly detection system. Our experiments show that the genetic algorithm improves the abilities of a detector operating on multi-channel satellite telemetry.