A Comparative Evaluation of LSTM- and Transformer-Based Deep Learning Architectures for Satellite Telemetry Anomaly Detection
Abstract
In this study, five deep learning architectures are systematically evaluated on the Mission 1 lightweight subset (6 channels, channel_41–46) of the European Space Agency Anomaly Detection Benchmark dataset released by the European Space Agency. As the primary contribution, a Transformer Decoder combination with a dynamic, input-conditioned query derived from the Long Short-Term Memory (LSTM) encoder output is proposed for the first time on this dataset, to the best of the authors’ knowledge. In the four-seed evaluation, the LSTM-Transformer Hybrid achieved a Corrected Event-wise F0.5 of 0.810±0.027, outperforming the reference method Telemanom-ESA-Pruned (Corrected Event-wise F0.5 = 0.786). Multi-channel experiments (58–76 channels) demonstrate that a voting strategy effectively prevents false positive explosions, and that telecommand integration attempts yielded no improvement due to the lightweight subset’s design. The ablation study indicates the decisive advantage of LSTM sequential memory over Transformer attention mechanisms in satellite telemetry.