ESA Mission 1 Anomaly Detection

PyTorch-based anomaly detection on ESA telemetry using an autoencoder and reconstruction error thresholding.

Dataset

• Source: https://zenodo.org/records/15237121
• Scope used in this repo: Mission 1 only (ESA-Mission1/)
• Processed file used by the model: esa_anomaly.csv

Project Structure

• ESA_anomaly_detection_eda.ipynb — data loading, preprocessing, and label generation
• ESA_anomaly_detection_Model.ipynb — training, thresholding, and evaluation plots
• model/autoencoder.py — autoencoder architecture
• ESA-Mission1/ — Mission 1 telemetry and metadata

Method

1. Merge selected telemetry channels.
2. Downsample and clean missing values (ffill/bfill).
3. Build binary anomaly labels from interval annotations.
4. Create temporal sliding windows.
5. Train autoencoder on normal data.
6. Compute reconstruction error and classify anomalies using a threshold.

Metrics Reported

• Precision
• Recall
• F1-score
• Classification report
• Confusion matrix
• ROC curve (AUC)

How to Run

1. Create/activate a Python environment.
2. Install dependencies:
   • pandas
   • numpy
   • matplotlib
   • scikit-learn
   • torch
   • ipykernel (for notebook kernel)
3. Run ESA_anomaly_detection_eda.ipynb to generate/update esa_anomaly.csv.
4. Run ESA_anomaly_detection_Model.ipynb to train and evaluate.

Notes

• This repository intentionally focuses on Mission 1 for reproducibility and manageable compute.
• Threshold can be fixed or selected automatically from the precision-recall curve.