Project information

  • Client: European Space Agency (ESA)
  • Categories: Anomaly Detection, Explainable AI (XAI), Deep Learning, Data Engineering
  • Main technologies: Python, PyTorch, SQL

Summary

DL4Space is a pioneering initiative led by the European Space Agency (ESA) to modernize spacecraft operations and testing using advanced Artificial Intelligence. My contribution focused specifically on the Assembly, Integration, and Testing (AIT) phase, where I was responsible for building an intelligent system capable of identifying anomalies in satellite time-series data and automatically pinpointing their root causes.

Faced with a massive and unstructured database of satellite telemetry, I architected a robust data engineering pipeline to clean, process, and structure the data for machine learning. I then designed a custom Deep Learning model using PyTorch to accurately classify anomalous events, replacing legacy manual processes with a highly scalable AI solution.

The Challenge: Moving Beyond Rule-Based Systems

Traditionally, anomaly detection in satellite systems relied on rigid, predefined sets of rules. As satellite architectures have grown exponentially more complex, this legacy approach became tedious, time-consuming, and prone to blind spots.

More importantly, when a legacy system flagged an anomaly, engineers still had to manually hunt for the root cause. The primary goal of DL4Space was to bridge this gap: not just to detect when something went wrong, but to explain exactly why.

Technical Approach & Explainable AI (XAI)

To ensure the Deep Learning model was transparent and actionable for ESA engineers, I heavily integrated Explainable AI (XAI) techniques into the pipeline. By utilizing frameworks like:

  • SHAP (SHapley Additive exPlanations): To calculate the exact contribution of each satellite sensor or telemetry feature to the model's anomaly prediction.
  • LIME (Local Interpretable Model-agnostic Explanations): To provide localized, human-readable explanations for individual anomalies.

The Result: The final system transformed a massive, unstructured database into a streamlined PyTorch and SQL-backed diagnostic tool, increasing anomaly detection accuracy by 15% compared to previous rule-based methods. It successfully allowed engineers to move away from black-box AI, giving them precise, data-backed insights into the factors triggering spacecraft anomalies.