satellite-anomaly-detection

Project Overview

I built this repository to support my paper, Hybrid Satellite Telemetry Anomaly Detection: A Dataset, Fault Taxonomy, Recurrence Plot Computer Vision Method, and Comparative Machine Learning Study. My goal is to make every experiment reproducible while keeping sensitive local environment details out of version control.

I study satellite telemetry anomaly detection across multiple simulated spacecraft so I can identify both known and subtle failure signatures. Fault type labels matter because operators need to know what failed, not only whether a sample is anomalous.

I include a dedicated WHEEL_OSCILLATION class because reaction wheel instability can cascade into attitude control failure. I use the Kepler reaction wheel failure as a motivating historical example that shows why early detection of oscillation behavior is critical.

Dataset Description

My dataset contains approximately 100,000 timestamped telemetry readings sampled across ten simulated satellites and four telemetry channels:

• power_w
• temp_c
• voltage_v
• wheel_rpm

My fault taxonomy has five classes:

• POWER_SPIKE
• THERMAL_DRIFT
• VOLTAGE_DROP
• WHEEL_OSCILLATION
• SENSOR_DROPOUT

Models Implemented

I evaluate the four architectures from my paper:

• LSTM Autoencoder
• CNN on Recurrence Plots
• Standard Autoencoder
• Isolation Forest

Key Results

These are the headline numbers I report in the paper:

• CNN recurrence plot model achieves 0.91 F1 on WHEEL_OSCILLATION
• LSTM autoencoder achieves 0.84 overall F1
• Hybrid hardware noise improves generalization by approximately 6.5 percent F1

Live Embedded Demo

My embedded demonstration pipeline is:

Arduino sensor array → Raspberry Pi 4 → recurrence plot encoding → model classification

I measured end to end inference latency under 150 milliseconds in the demo setup.

Reproducibility

I provide scripts that regenerate dataset processing, model training, and evaluation. If Kaggle data is not downloaded yet, my replication script creates a local simulation fallback so the full workflow still runs end to end. I skip figure generation by default in the one-command script so setup is faster.

1) Download dataset from Kaggle

kaggle datasets download aryantputta/hybrid-satellite-telemetry -p data --unzip

2) Set up environment

bash replication/bootstrap_environment.sh
source .venv/bin/activate

Manual option:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Or with Conda:

conda env create -f environment.yml
conda activate satellite-anomaly-detection

3) Train models

python training/train_cnn.py
python training/train_lstm.py
python training/train_autoencoder.py
python training/train_isolation_forest.py

4) Evaluate models

python evaluation/evaluate_models.py
python evaluation/compute_metrics.py

5) Regenerate figures (optional)

python figures/generate_fig5_architecture.py
python figures/generate_fig8_results.py
python figures/generate_fig9_fault_type.py
python figures/generate_fig10_noise_effect.py

One command reproduction (figures skipped by default)

bash replication/reproduce_paper_results.sh

To include figures in the one-command run:

SKIP_FIGURES=0 bash replication/reproduce_paper_results.sh

Figure Reproduction

I fixed layout issues in my figure scripts:

• Figure 5 output labels are not cropped.
• Figure 7 Arduino Uno label fits correctly inside its diagram box.

Evaluation Protocol

I use a per satellite split to prevent memorization of a single simulated satellite.

• Training satellites: SAT_01 to SAT_08
• Testing satellites: SAT_09 to SAT_10

Environment and API Security

I keep local secrets in .env and I do not commit secrets, keys, or private config files.

I provide .env.example as a safe template.