Satellite Fleet Health Management System
Real-time ISS telemetry monitoring with ML-powered diagnostics

Continuous Operations Network for Satellite Telemetry Evaluation, Life-cycle Analysis, Tracking, Intelligence, Operations, and Notification

Project Overview

CONSTELLATION monitors the International Space Station's attitude control and communications subsystems using real-time telemetry from NASA's public Lightstreamer feed. The system provides:

1. Data Ingestion Layer

Real-Time Telemetry Processor

• AWS Lambda function subscribing to NASA Lightstreamer
• Filters for attitude control and communications subsystems
• Writes to DynamoDB for real-time access
• Archives to S3 for historical analysis

Parameters Monitored:

Attitude Control (Reaction Wheels & CMGs):

• USLAB000084: Reaction Wheel Assembly (RWA) speed
• USLAB000085: RWA bearing temperature
• USLAB000086: RWA current draw
• USLAB000087: CMG (Control Moment Gyroscope) momentum
• Attitude quaternions (pitch, roll, yaw)
• Rate gyro outputs

Communications:

• S-band transponder power levels
• Ku-band signal strength
• Antenna pointing accuracy
• Data throughput metrics
• Ground station contact windows
• Communication link quality indicators

Cross-Cutting:

• Power system voltage/current (affects both subsystems)
• Thermal readings (reaction wheel bearings, transmitter temps)
• Time-on-orbit (cumulative degradation tracking)

2. Feature Engineering Pipeline

Time Series Features:

• Rolling statistics (mean, std, min, max) over multiple windows (1hr, 6hr, 24hr, 7day)
• Rate of change calculations
• Autocorrelation features
• Fourier transform for periodic patterns
• Lag features (t-1, t-6, t-24 for hourly data)

Domain-Specific Features:

• Reaction wheel friction coefficient (derived from speed vs. current)
• Thermal cycling count (number of orbital day/night transitions)
• Momentum accumulation rate
• Communication link budget margin
• Signal degradation trends
• Anomaly persistence scores

Engineering Calculations:

• Power efficiency ratios
• Thermal dissipation rates
• Bearing wear indicators
• Transmitter efficiency
• Pointing error accumulation

3. ML Model Suite

Model 1: Anomaly Detection (Isolation Forest + LSTM Autoencoder)

Purpose: Real-time detection of unusual telemetry patterns

Approach:

• Isolation Forest for fast, lightweight anomaly flagging
• LSTM Autoencoder for complex temporal anomaly detection
• Ensemble voting for final anomaly score

Training Data:

• Nominal operational periods (confirmed healthy operation)
• Labeled anomalies from NASA incident reports

Metrics: Precision, Recall, F1-Score, False Positive Rate

Model 2: Degradation Forecasting (Temporal Fusion Transformer)

Purpose: Predict subsystem performance degradation over time

Targets:

• Reaction wheel bearing temperature trend
• Solar panel output decline
• Battery capacity fade
• Communication signal strength degradation

Features: Time series telemetry + orbital mechanics (radiation exposure, thermal cycling)

Output: Forecasted parameter values with confidence intervals (7, 30, 90 days ahead)

Model 3: Survival Analysis (Cox Proportional Hazards)

Purpose: Estimate time-to-failure for critical components

Approach:

• Cox model for component-level survival curves
• Censored data handling for components still operational
• Hazard ratios for risk factors (high temps, usage patterns)

Output: Probability of failure within time windows (30d, 60d, 90d, 180d)

Model 4: Fault Classification (XGBoost)

Purpose: Diagnose root cause when anomalies occur

Classes:

• Thermal stress
• Mechanical wear (bearings, gimbals)
• Electrical fault
• Software/command error
• External disturbance (debris impact, space weather)
• Normal operational variation

Features: Anomaly signatures, subsystem interactions, environmental context

Output: Ranked list of probable causes with confidence scores

4. Maintenance Optimization Engine

Constraint Satisfaction Problem:

Variables:

• Maintenance task list (derived from predictions)
• Available maintenance windows
• Crew availability (for ISS; ground station access for unmanned satellites)
• Orbital position constraints
• Mission priority levels

Constraints:

• Ground station contact requirements
• Crew schedule conflicts
• Tool/equipment availability
• Task dependencies (some maintenance requires others first)
• Safety margins (don't defer critical items)

Objective Function:

• Minimize risk-weighted maintenance delay
• Balance urgency vs. operational disruption
• Optimize crew time utilization

Algorithm: Mixed Integer Programming (MIP) using PuLP or Google OR-Tools

┌─────────────────────────────────────────────────────────────────┐
│                     DATA INGESTION LAYER                        │
├─────────────────────────────────────────────────────────────────┤
│  NASA Lightstreamer → Lambda (Real-time) → DynamoDB             │
│  Historical Archive → S3 Data Lake                              │
└─────────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────────┐
│                   FEATURE ENGINEERING LAYER                     │
├─────────────────────────────────────────────────────────────────┤
│  • Time series windowing                                        │
│  • Statistical feature extraction                               │
│  • Subsystem correlation analysis                               │
│  • Degradation rate calculation                                 │
└─────────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────────┐
│                      ML MODEL SUITE                             │
├─────────────────────────────────────────────────────────────────┤
│  Anomaly Detection    → Isolation Forest / Autoencoder          │
│  Degradation Forecast → LSTM / Temporal Fusion Transformer      │
│  Survival Analysis    → Cox Proportional Hazards / Weibull      │
│  Fault Classification → Random Forest / XGBoost                 │
└─────────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────────┐
│                   OPERATIONAL LAYER                             │
├─────────────────────────────────────────────────────────────────┤
│  Health Scoring → Maintenance Scheduling → Alert Generation     │
│  Dashboard (Streamlit) → CloudWatch Monitoring                  │
└─────────────────────────────────────────────────────────────────┘

### Model Training Strategy

Training Infrastructure

Local Development:

• Jupyter notebooks for experimentation
• GPU-enabled local training for initial model development
• Small data samples for rapid iteration

Production Training:

• AWS SageMaker for full dataset training
• Hyperparameter tuning with SageMaker Automatic Model Tuning
• Distributed training for large models
• Model versioning with MLflow

Evaluation Metrics

Anomaly Detection:

• Precision, Recall, F1-Score
• False Positive Rate (critical for operational systems)
• Detection latency
• ROC-AUC, PR-AUC

Degradation Forecasting:

• RMSE, MAE, MAPE
• Prediction interval coverage
• Directional accuracy (did we predict the trend correctly?)
• Forecast horizon performance (7d vs 30d vs 90d)

Survival Analysis:

• Concordance index (C-index)
• Brier score
• Calibration plots (predicted vs observed survival)
• Time-dependent AUC

Fault Classification:

• Accuracy, Precision, Recall per class
• Confusion matrix
• Top-k accuracy (are correct diagnoses in top 3 predictions?)

Quick Start

Installation

# Clone repository
git clone <your-repo-url>
cd constellation

# Create virtual environment
python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Collect Telemetry Data

# Start telemetry collection
python -m src.ingestion.collect_telemetry

Data will be saved to data/raw/ in date-partitioned Parquet files.

LET'S CONNECT!

This project demonstrates production-grade ML engineering capabilities including distributed training infrastructure, experiment management, and systematic research methodology. All code and documentation available for technical review.

Michael Gurule

Data Science | ML Engineering

---

_{Demonstrating institutional-quality quantitative finance capabilities through production-grade portfolio optimization}