Anomaly Detection in Spacecraft Telemetry Data Based on Transformer-LSTM

The development of the modern aerospace industry has led to increasingly complex communication satellite structures, resulting in a continuous increase in faults. Detecting faults in communication satellites has become a key issue in the current aerospace field. Currently, fault detection is primarily based on simple threshold-based methods, which can only detect specific types of faults. This paper proposes a satellite telemetry data anomaly detection method based on a Transformer-LSTM encoder-decoder structure. This model fully leverages the advantages of both Transformer and LSTM. The Transformer models historical time series data using the Multi-Head Attention mechanism, learning their patterns of change, while LSTM excels at capturing dependencies within local time series. The article uses satellite periodic events as anomaly data, and experiments show that the Transformer-LSTM model, compared to standalone Transformer and LSTM models, achieves higher precision and can accurately locate abnormal intervals in satellite telemetry data. This research provides an effective approach for satellite fault detection based on deep learning.