Layered Explainable Anomaly Detection in Satellite Telemetry Using Spectrogram-Based 2D-CNN Autoencoders with Event-Aware EWMA Filtering

Yildirim B. B., Dinc Y., Ozdemir M., Alremali A., ARSLAN H., Somuncuoglu A. N.

11th International Conference on Recent Advances in Air and Space Technologies, Conference Program, RAST 2026, İstanbul, Türkiye, 13 - 15 Mayıs 2026, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/rast69551.2026.11672527
  • Basıldığı Şehir: İstanbul
  • Basıldığı Ülke: Türkiye
  • Anahtar Kelimeler: anomaly detection, autoencoders, explainable AI, satellite telemetry
  • Ankara Yıldırım Beyazıt Üniversitesi Adresli: Evet

Özet

Satellite telemetry anomaly detection plays an important role in ensuring space mission reliability and operational safety. Although modern deep learning models achieve strong detection performance, their black box nature limits operator trust and complicates root cause analysis. This study proposes a layered explainable anomaly detection framework that combines spectrogram based representation learning with event aware post processing and multi level interpretability. One dimensional telemetry signals are transformed into two dimensional time frequency spectrograms using the Short Time Fourier Transform (STFT), allowing anomaly detection to be modeled using a 2D convolutional autoencoder. To reduce false alarms caused by transient disturbances and pointwise evaluation bias, an Exponentially Weighted Moving Average (EWMA)-based event pruning mechanism is introduced to support event level detection. The framework is evaluated on ESA's Anomaly Dataset (Mission 1) using telemetry channels 41-46. The experimental results indicate relatively consistent precision, recall, and corrected F0.5 performance within the evaluated threshold range, achieving a conservative event level F0.5 score of 0.55 on an unseen test set. Moreover, an explainable AI module is designed to offer diagnostic transparency in spacecraft telemetry monitoring by providing interpretable insights through the residual heatmaps, channel frequency attribution, as well as the Grad CAM based bottleneck visualization.