CubeSat-Health-Monitor

Industrial-grade MLOps pipeline for autonomous CubeSat telemetry anomaly detection.

Orbit-Q ingests noisy, high-frequency sensor streams from CubeSat-class satellites and runs a 3-model anomaly detection ensemble (Isolation Forest + PyTorch Autoencoder + LSTM) to identify hardware failures, data corruption, and operational drift in real time — all under 500 ms from anomaly onset to alert.

---

Mission Profile

• Orbit: 550 km circular LEO, 97.6° inclination, ~95-minute period
• Sensors: Panel temperatures (6 faces), battery voltage/current, reaction wheel speed, gyros, coarse sun sensors
• Tracked metrics: Thermal margins, state of charge, attitude stability (deg/s), RF link uptime
• Inference latency: ~15.72 µs median (63,622 events/sec throughput)
• False alarm rate: ~3–5% over a 24-hour simulated window

---

Architecture

Fault-Injection Simulator → Firebase Realtime DB
                                    ↓
                          ML Orchestrator (polling)
                                    ↓
                         Feature Processor (rolling window)
                                    ↓
          ┌─────────────────────────────────────────┐
          │           Anomaly Engine                │
          │  Isolation Forest | Autoencoder | LSTM  │
          └────────────────┬────────────────────────┘
                           ↓ Score Fusion
              Anomaly Decision → MLflow + Firebase Alerts
                                    ↓
                        Streamlit Command Center

Key components:

• src/orbitq/ensemble/ — Score fusion and voting logic
• src/orbitq/engine/models/ — PyTorch Autoencoder and LSTM detector
• src/orbitq/orchestrator/ — Polling loop and feature engineering
• src/orbitq/ingestion/ — Kafka client for high-throughput ingestion
• src/orbitq/simulator/ — Fault-injection telemetry generator
• src/orbitq/dashboard/ — 10-page Streamlit command center
• src/orbitq/security.py — HMAC-SHA256 authentication with TTL
• src/orbitq/mlflow_tracking/ — Model registry and retraining pipeline

---

Results

Metric	Value	Target	Status
Precision	0.942	> 0.90	✅
Recall	0.915	> 0.85	✅
F1 Score	0.928	> 0.88	✅
Throughput (EPS)	63,622	> 10,000	✅
Inference Latency	15.72 µs	< 1.0 ms	✅
False Alarm Rate (24h)	~3–5%	< 5%	✅

---

Quick Start

Prerequisites

• Python 3.10+
• (Optional) Firebase project for real-time features

Installation

git clone https://github.com/poojakira/CubeSat-Health-Monitor.git
cd CubeSat-Health-Monitor
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -e .

Run (Local Mode)

export MLFLOW_TRACKING_URI=sqlite:///mlflow.db
orbit-q benchmark

Run (Production Mode with Kafka)

docker-compose up -d
docker-compose logs -f ingestion

Configuration

cp configs/.env.example .env
# Set FIREBASE_DB_URL and SERVICE_ACCOUNT_PATH in .env

---

Project Structure

CubeSat-Health-Monitor/
├── .github/workflows/    # CI: linting, type-checking, pytest
├── assets/               # Screenshots and diagrams
├── configs/              # .env.example and config templates
├── models/               # Trained ensemble model (.pkl)
├── results/              # Benchmark logs and evaluation CSVs
├── src/orbitq/
│   ├── ensemble/         # Score fusion and voting
│   ├── engine/           # PyTorch models + C++ fusion kernels
│   ├── orchestrator/     # Polling loop + feature engineering
│   ├── ingestion/        # Kafka client
│   ├── simulator/        # Fault-injection telemetry generator
│   ├── dashboard/        # Streamlit command center (10 pages)
│   ├── mlflow_tracking/  # MLflow model registry + retraining
│   └── security.py       # HMAC-SHA256 auth
├── tests/                # 11+ unit and integration tests
└── pyproject.toml

---

Running Tests

pytest tests/ -v
pytest tests/ --cov=src --cov-report=term-missing

---

License

MIT — see LICENSE.

---

Author

Pooja Kiran & Rhutvik Pachghare

Built as graduate research at Arizona State University (2025–2026).

• GitHub: @poojakira
• LinkedIn: Pooja Kiran