Satellite-Telemetry-Anomaly-Detection

Project Overview

This project is an end-to-end Machine Learning and Reliability Engineering pipeline designed to monitor the physical health of spacecraft subsystems. The goal is to computationally detect early-onset anomalies in satellite thermal sensors before they lead to critical failure. By engineering a robust Digital Twin simulation, this project validates the ability of unsupervised learning to distinguish between harmless sensor noise and genuine physical degradation.

Key Achievements:

Scale: Architected a Digital Twin generator capable of simulating thousands of orbital cycles with stochastic failure injection.

Precision: Achieved a 0.982 F1 Score in detecting system instability, successfully distinguishing high-variance vibration from Gaussian background noise.

Engineering: Implemented a modular Champion/Challenger evaluation system, proving Isolation Forest superiority over density-based methods (LOF) for high-frequency telemetry.

How do we Monitor Spacecraft?

Distinguishing Signal from Noise is incredibly difficult.

1. The Concept: Orbital Thermal Cycling

A satellite in Low Earth Orbit (LEO) experiences extreme temperature swings every 90 minutes (Day/Night cycle).

$$T_{sensor} = T_{base} + A \cdot \sin(\omega t) + \epsilon$$

• $T$: Temperature Reading.
• $\sin(\omega t)$: The periodic heating from the Sun.
• $\epsilon$: Sensor Noise (Gaussian).

2. The Data Engineering Challenge

In a vacuum, heat transfer is slow. Therefore, temperature data should always be "smooth" (high autocorrelation). However, raw telemetry is noisy due to radiation and sensor imperfections. A simple threshold alert (e.g., "If Temp > 100") is useless because it triggers false alarms on noise spikes and misses subtle drifts that indicate localized failure.

3. The Solution: Physics-Based Feature Extraction

To solve for health, I implemented a Feature Engineering engine that converts raw time-series data into a physical context. We don't just look at the value; we look at the behavior.

• Volatility ($\sigma$): If the standard deviation spikes, the system is vibrating, or the processor is unstable.
• Velocity ($\Delta$): If the rate of change exceeds physical thermal limits, it is a voltage surge, not heat.

This pipeline filters for these violations using rolling window statistics: $$Z_{score} = \frac{x - \mu_{rolling}}{\sigma_{rolling}}$$

---

Results

After processing the telemetry data and running a benchmark tournament between three anomaly detection algorithms, the pipeline produced the following results:

Metric	Value
Target F1 Score (Production Requirement)	> 0.90
Calculated F1 Score (My Model)	0.982
Precision	0.99

---