Wearable Health Device for Body Temperature Monitoring and Fall Detection

This repository contains the source code, implementation details, and documentation for a Wearable Health Device (WHD) that monitors body temperature and detects falling events. The system is designed using an STM32 Nucleo-L432KC microcontroller and employs machine learning (CNN) for accurate fall detection. This project is my final Embedded System Design course assignment.

---

📌 Features

• Body Temperature Monitoring 🌡️ using the TMP102 sensor.
• Fall Detection 🏃 using the ADXL345 accelerometer.
• LED and Buzzer Alerts 🚨 for abnormal temperature or falls.
• Machine Learning with CNN 🧠 to improve detection accuracy.
• Real-Time Data Visualization 📊 via serial communication.

---

🚀 Getting Started

1️⃣ Prerequisites

• Hardware

  • STM32 Nucleo-L432KC
  • TMP102 Temperature Sensor
  • ADXL345 Accelerometer
  • 9 LEDs & Passive Buzzer
  • Breadboard and Jumper Wires

• Software & Libraries

  • PlatformIO (for STM32 development)
  • Mbed OS Library
  • Python 3.10 (for visualization)
  • torch, matplotlib, pyserial (for CNN training & visualization & Print information)

2️⃣ Folder Description

• Project/: Source code to download into L432KC.
• Epoch_size_study/: Experiment files for finding the best epoch size.
• History/: Test files, just omitted.

---

🏗️ System Architecture

Hardware & Sensor Connection

The device consists of three main modules:

1. Temperature Sensor Module 🌡️ (TMP102)
2. Fall Detection Module 🏃 (ADXL345 Accelerometer)
3. Alert & Communication Module 📢 (LEDs & Buzzer)

TMP102 Sensor Register Structure

ADXL345 Functional Overview

---

📊 Machine Learning - CNN Model for Fall Detection

📌 Why CNN for Fall Detection?

• Traditional threshold-based methods are unreliable.
• CNN learns complex patterns in acceleration data.
• Provides higher accuracy in real-world scenarios.

📜 CNN Architecture

• Input: Acceleration data in $x$, $y$, $z$ axes (over 5-time steps).
• Layers: Convolutional, MaxPooling, Fully Connected.
• Output: Probability of a fall event.

📈 Model Training & Performance

• Dataset: 167 manually labeled samples.
• Epoch Selection: 750 epochs found optimal.
• Accuracy: 76.9% for normal falls, 90% for slow falls.

Validation Loss Over Different Epochs

---

🎛️ Real-Time Data Visualization

The device supports real-time monitoring of temperature and acceleration data.

Accelerometer Data - Before and After Max Acceleration Processing

Real-Time Temperature and Fall Detection

python visualization.py

This will open a live plot showing:

• Temperature changes over time.
• Detected fall events.
• System status updates.

---

🔍 Results & Discussion

✅ Fall Detection Accuracy

Scenario	Total Falls	Correct Detections	False Positives	Success Rate
Normal Falling	30	30	9	100%
Slow Falling	22	9	1	40.9%

Accelerometer-Based Fall Detection Results

✅ Temperature Monitoring

Temperature (°C)	LED Status	Alert
35.5 - 36.5	Normal	❌
37.0 - 38.3	Warning	⚠️
38.5 - 42.0	Danger	🚨

Temperature-Based Alerts

LED Status Under Different Conditions

---

🔮 Future Improvements

• Improve CNN accuracy using more training data.
• Reduce false positives by tuning model parameters.
• Implement wireless communication (Bluetooth/WiFi).
• Enhance power efficiency for prolonged battery life.

---

📜 References

• TMP102 Datasheet: Texas Instruments
• ADXL345 Datasheet: Analog Devices
• Wearable Health Monitoring Devices - IEEE

---

📬 Contact

📧 Email: marcobisky@outlook.com
🔗 GitHub: github.com/marcobisky