This repository features the development of a resilient, AI-driven electronics subsystem for the European Space Agency's (ESA) OPS-SAT CubeSat. The project's core mission is to create an Autonomous Attitude Determination and Control System (ADCS) Monitoring System that employs predictive maintenance algorithms to autonomously detect and correct anomalies while operating in space.
By running complex diagnostics directly on-board, the system significantly enhances mission resilience and reduces reliance on ground control, directly mitigating the high risk of failure often associated with CubeSat ADCS subsystems.
The solution leverages a dual-AI approach optimized for low-power edge computing.
| Feature | Description |
|---|---|
| Autonomous Anomaly Detection | An Artificial Neural Network (ANN) model analyzes real-time sensor data to flag faults and deviations from normal operating conditions. |
| Predictive Signal Correction | Nine specialized Random Forest models are utilized—one for each monitored ADCS channel—to autonomously calculate and apply corrections to abnormal signals. |
| Edge AI Optimization | Models are deployed onto the STM32F429 microcontroller for highly efficient, low-latency inference, respecting crucial CubeSat constraints (power, volume, radiation tolerance). |
| Detection Performance | The core ANN model validation yielded high results on the testing set: 95.65% Accuracy and an 89.1% F1 Score. |
The project code and resources are logically separated into three major development areas: Machine Learning (ML), Data Processing, and Embedded Implementation.
├── ML_models(detection) # Anomaly Detection (ANN) Model Notebooks and Training Data
├── ML_models(correction) # Random Forest Correction Model Notebooks (Model1 to Model9)
├── edge_impulse # Data Acquisition Scripts and Edge Impulse Project Configurations
├── STM32f429 # Final Embedded Firmware (High-Accuracy CubeAI & Edge Impulse Deployments)
├── schemaelectrique # Electrical Schematics and Hardware Interface Design Files
├── .gitignore # Specifies files and directories ignored by Git
├── LICENCE # Project's Proprietary License File
└── README.md # Project overview, features, and setup instructionsThe ML subsystem is divided into anomaly detection and predictive correction, covering 9 ADCS sensor channels.
Directory: ML_models(detection)
Detects anomalies in real-time based on magnetometer and sun sensor data.
Monitored Sensor Channels:
- Magnetometer:
X,Y,Z - Sun Sensor / Photodiodes:
PD1–PD6
Performance (Testing Set)(ANN model):
| Metric | Value |
|---|---|
| Accuracy | 95.65% |
| F1 Score | 89.10% |
| AUC | 97.63% |
Embedded Deployment (Cortex-M4F)(ANN model):
| Parameter | Value |
|---|---|
| Latency | ~2 ms |
| RAM Usage | ~0.2 KB |
| Flash Usage | ~34.7 KB |
Directory: ML_models(correction)
After the ANN detects an anomaly, 9 dedicated Random Forest models (one per sensor channel) predict corrected values, enabling autonomous recovery without ground control.
Directory: STM32f429
Implements the ML models on the STM32F429 Discovery Kit (ARM Cortex-M4F).
| Mode | Description | Purpose |
|---|---|---|
| CubeAI (Optimized) | Converts ANN to C for real-time inference | Flight-ready implementation |
| Edge Impulse | Rapid experimentation (lower accuracy) | Comparison + development |
- STM32F429 MCU → On-board processing + inference
- MPU-9250 IMU → 9-axis attitude sensing
- NanoCom AX100 → Communication subsystem simulation
Designed for CubeSat constraints: low power, radiation tolerance, thermal stability, compact integration.
Directory: edge_impulse
Contains:
- Scripts for uploading ADCS telemetry data
- Feature extraction / DSP processing pipelines
- Edge Impulse neural network training configuration
Used for:
- Data preparation
- Prototype model experimentation
- Performance comparison with CubeAI deployment
This project utilizes the following key tools and platforms:
- Microcontroller: STM32F429ZIT6 (Arm® Cortex®-M4)
- Embedded IDE: STM32CubeIDE
- ML Frameworks: TensorFlow/Keras, Scikit-learn, STM32Cube.AI, Edge Impulse
download the code as a ZIP file (required)