Anomaly detection in satellite telemetry data is essential for ensuring the reliability and safety of space missions. While traditional Machine Learning approaches have shown promising results, their high computational cost and energy consumption present challenges for onboard deployment. This thesis explores Spiking Neural Networks (SNNs) as an efficient and biologically inspired alternative for real-time anomaly detection in resource-constrained environments. The proposed approach, ADStrobot, integrates Neuromorphic Computing with FPGA-based hardware acceleration, enabling a low-power, high-performance anomaly detection system. The study examines various spike encoding strategies, SNN architectures, and training methodologies to optimize performance while maintaining the temporal dynamics of satellite telemetry data. Experimental validation on the ESA Anomaly Detection (ESA-AD) dataset demonstrates that SNN-based models achieve competitive detection accuracy while significantly reducing power consumption compared to conventional Deep Learning methods. Specifically, for the first mission of the dataset, ADStrobot achieves an F0.5 score of 0.9240 for event-wise detection and 0.9238 for channel-aware detection. The FPGA implementation reveals a power consumption of 120 mW, meeting the timing requirements dictated by the dataset’s sampling frequency, with a simulation latency of 670 ns. These results confirm the feasibility of SNN-based anomaly detection for satellite systems, emphasizing the potential of Neuromorphic Computing for autonomous space applications.
Il rilevamento delle anomalie nei dati di telemetria satellitare è fondamentale per garantire la sicurezza e l’affidabilità delle missioni spaziali. Tuttavia, i metodi tradizionali di apprendimento automatico, pur mostrando risultati promettenti, presentano costi computazionali ed energetici elevati, rendendone difficile l’implementazione a bordo. Questa tesi propone l’uso delle Reti Neurali Spiking (SNN) come alternativa efficiente e biologicamente ispirata per il rilevamento delle anomalie in tempo reale in ambienti con risorse limitate. L’approccio sviluppato, ADStrobot, combina il calcolo neuromorfico con l’accelerazione FPGA, offrendo un sistema di rilevamento a basso consumo e alte prestazioni. Lo studio analizza diverse strategie di codifica degli spike, architetture neurali e tecniche di addestramento per ottimizzare il modello, preservando le dinamiche temporali dei dati di telemetria. I test sul dataset ESA Anomaly Detection (ESA-AD) dimostrano che le SNN possono raggiungere un’accuratezza di rilevamento competitiva riducendo significativamente il consumo energetico rispetto agli approcci di deep learning convenzionali. Per la prima missione del dataset, ADStrobot ottiene un punteggio F0.5 di 0,9240 per il rilevamento degli eventi anomali e 0,9238 considerando i canali coinvolti. L’implementazione su FPGA mostra un consumo di 120 mW, garantendo il rispetto delle esigenze di temporizzazione del dataset con una latenza di 670 ns. I risultati confermano la fattibilità dell’uso delle Reti Neurali Spiking per il rilevamento delle anomalie nei sistemi satellitari, evidenziando il potenziale della computazione neuromorfica per applicazioni spaziali autonome.
ADStrobot: a Spiking NN for Satellite Anomaly Detection on FPGA
Ritirato, Paolo
2023/2024
Abstract
Anomaly detection in satellite telemetry data is essential for ensuring the reliability and safety of space missions. While traditional Machine Learning approaches have shown promising results, their high computational cost and energy consumption present challenges for onboard deployment. This thesis explores Spiking Neural Networks (SNNs) as an efficient and biologically inspired alternative for real-time anomaly detection in resource-constrained environments. The proposed approach, ADStrobot, integrates Neuromorphic Computing with FPGA-based hardware acceleration, enabling a low-power, high-performance anomaly detection system. The study examines various spike encoding strategies, SNN architectures, and training methodologies to optimize performance while maintaining the temporal dynamics of satellite telemetry data. Experimental validation on the ESA Anomaly Detection (ESA-AD) dataset demonstrates that SNN-based models achieve competitive detection accuracy while significantly reducing power consumption compared to conventional Deep Learning methods. Specifically, for the first mission of the dataset, ADStrobot achieves an F0.5 score of 0.9240 for event-wise detection and 0.9238 for channel-aware detection. The FPGA implementation reveals a power consumption of 120 mW, meeting the timing requirements dictated by the dataset’s sampling frequency, with a simulation latency of 670 ns. These results confirm the feasibility of SNN-based anomaly detection for satellite systems, emphasizing the potential of Neuromorphic Computing for autonomous space applications.File | Dimensione | Formato | |
---|---|---|---|
2025_04_Ritirato_ExecutiveSummary.pdf
accessibile in internet per tutti a partire dal 13/03/2028
Descrizione: Executive Summary
Dimensione
1.01 MB
Formato
Adobe PDF
|
1.01 MB | Adobe PDF | Visualizza/Apri |
2025_04_Ritirato_Tesi.pdf
accessibile in internet per tutti a partire dal 13/03/2028
Descrizione: Thesis
Dimensione
5.66 MB
Formato
Adobe PDF
|
5.66 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/236315