IoT Biometric Sleep Tracker

An end-to-end IoT system for sleep analysis, featuring a custom-designed wearable, a secure BLE protocol, and on-device Machine Learning inference.

Engineering Goal: Bridge the gap between low-level hardware constraints and high-level software scalability by creating a full-stack pipeline from raw sensor acquisition to user-facing analytics.

---

🏗 System Architecture

The system follows a data streaming architecture. A low-power wearable device handles raw data acquisition, while a companion iOS application manages the heavy lifting: signal processing, ML inference, and visualization.

High-Level Workflow

1. Sensing: Accelerometer (FXOS8700) & Heart Rate sensor capture raw data.
2. Transmission: ESP32 microcontroller streams data packets via Bluetooth Low Energy (BLE).
3. Processing: iOS App aggregates data into rolling windows (Feature Engineering).
4. Inference: A Core ML model (PulseClassifier) classifies the state as Awake or Asleep.

---

📂 Project Structure

├── hardware/              # PCB Design files (Altium/KiCad exports)
│   ├── schematics/        # 2D/3D Renders and Schematics
│   └── assembly/          # Assembly drawings and BOM
├── ios-app/               # iOS Companion App (Swift, SwiftUI, CoreML)
├── src/                   # Python Codebase
│   ├── data/              # Data loading and cleaning scripts
│   ├── models/            # XGBoost training logic
│   └── scripts/           # Pipeline orchestration scripts
├── models/                # Serialized models (.joblib, .onnx)
├── notebooks/             # Jupyter Notebooks for EDA
└── images/                # Documentation assets

---

🛠 Component Breakdown

1. Wearable Device (Hardware & Firmware)

Role: Raw Data Acquisition & Transmission Status: Design complete, Ready for Fabrication.

• Microcontroller: ESP32-WROOM-32E. Selected for its dual-core architecture and deep sleep capabilities tailored for battery-operated IoT.

• Sensors:

• FXOS8700CQR1: 6-axis Accelerometer/Magnetometer for precise motion tracking.

• Heart Rate Sensor: Dedicated photoplethysmography (PPG) IC for pulse wave sensing.

• Power Management: USB-C charging circuit with dedicated voltage regulation (LDOs) to ensure stable power delivery.

🔧 PCB Design & Architecture

The PCB utilizes a two-sided component placement strategy to optimize for size and signal integrity:

1. Top Side (Digital Domain): Houses the ESP32, USB-C connector, and power management ICs. This isolates high-frequency digital noise.
2. Bottom Side (Analog/Sensor Domain): Dedicated to the Heart Rate sensor to ensure direct skin contact and clean analog signal acquisition.

3D Render (Top)	Electronic Schematic

Design Choice: The physical separation of the analog sensor (Bottom) from the noisy digital logic (Top) minimizes signal interference, crucial for accurate biometric readings.

2. Machine Learning Pipeline (Data Science)

Role: Classification Algorithm Stack: Python, Scikit-Learn, XGBoost

The core logic is a supervised learning model trained on the MMASH dataset.

1. Feature Engineering: Raw data is converted into 11 statistical features over rolling windows (1, 5, and 15 minutes).

• Examples: hr60_mean, vm15_std (Vector Magnitude Standard Deviation).

2. Model Selection: Compared Logistic Regression, Random Forest, and XGBoost.
3. Final Model: Tuned XGBoost Classifier.

• Performance: 93% Accuracy, 0.96 F1-Score (Awake).
• Optimization: Hyperparameters tuned via RandomizedSearchCV.

3. iOS Companion App (Software)

Role: Client, Processor & Visualizer Stack: Swift, SwiftUI, Charts, Core ML

• Real-time BLE: Handles scanning and parsing of the standard Heart Rate Service (UUID 0x180D).
• On-Device Inference: The Python-trained XGBoost model was converted to .mlpackage to run locally on the iPhone, ensuring privacy and offline capability.
• UI/UX: "Deep Night" aesthetic with interactive, scrubbable charts.

---

🚀 Getting Started

Prerequisites

• Hardware: Pulse PCB or ESP32 dev kit.
• Software: Xcode 15+, Python 3.9+.

1. Python Environment (ML Pipeline)

To reproduce the model training or explore the notebooks:

pip install -r requirements.txt
cd src/models
python train_final.py

2. iOS Application

1. Open ios-app/Pulse app.xcodeproj in Xcode.
2. Ensure PulseClassifier.mlpackage is linked in the Build Phases.
3. Run on a physical iPhone (Bluetooth required).

---

📊 Performance Metrics

              precision    recall  f1-score   support

   False           0.99      0.93      0.96      4893
    True           0.76      0.95      0.85      1182

accuracy                               0.93      6075

---

👤 Author

Noé Cornu - Computer Engineering Student @ EPITA

GitHub | LinkedIn