☕ Coffee Bean Classifier – Gas Sensing + Embedded ML

This project demonstrates how embedded machine learning can be used to classify different coffee brands using gas sensing technology. By leveraging the BME688 4-in-1 environmental sensor and Microchip’s PIC32CX-BZ2 / WBZ451 microcontrollers, we built a complete workflow from data collection → model training → deployment on hardware.

The project highlights how low-cost sensors and TinyML techniques can be applied to real-world use cases such as food authentication, air quality monitoring, and industrial predictive maintenance.

---

🔎 Project Summary

• Problem: Coffee authenticity and quality control are often expensive and require lab-grade instruments.
• Solution: Use gas sensor readings (VOC/VSC, carbon monoxide, hydrogen signatures) + embedded ML to classify coffee brands in real time.
• Approach:
  1. Capture sensor data from different coffee brands.
  2. Train an ML classifier using Microchip’s MPLAB ML Development Suite.
  3. Deploy the model on a low-power PIC32CX-BZ2 MCU for on-device inference.

---

🛠️ Hardware Setup

• 2 × PIC32CX-BZ2 / WBZ451 Curiosity Boards
  • One board + BME688 sensor inside a sealed jar with coffee (sensor node).
  • Second board connected to PC via USB (host node).
• BME688 Environmental Sensor – measures temperature, humidity, pressure, and gas resistance.
• 3.7V Li-Po Battery Pack – powers the sensor node for portable operation.

---

📂 Repository Contents

├── firmware/
│ ├── sensor_node/ # Firmware for data collection setup with BME688
│ ├── host_node/ # Firmware for USB host board
│
├── user_guide/ # PDF guide (detailed setup, usage, ML workflow)
│
├── README.md # Project overview & documentation

---

📊 Data Collection & Training

• Warm up sensor for 20 minutes before recording.
• Capture 30-minute sessions for each coffee brand.
• Use MPLAB Data Collector to log sensor data.
• Import datasets into ML Model Builder for training.
• AutoML pipeline used to find optimal features + model.
• Best model achieved ~97% accuracy, with small memory footprint (<20 KB).

---

🤖 Deployment

• Exported model as a Knowledge Pack.
• Integrated into MCU firmware with simple API calls (kb.h, kb_model_init()).
• Flashed onto PIC32CX-BZ2 board using MPLAB X IDE.
• Real-time predictions streamed to PC via MPLAB Data Visualizer.

---

🌟 Key Features

• Fully embedded ML workflow (no cloud dependency).
• Works on resource-constrained MCUs.
• Portable, battery-powered setup.
• Generalizable to multiple applications:
  • Food authentication (spices, tea, wine).
  • Environmental monitoring (indoor air quality, VOC detection).
  • Industrial gas sensing and predictive maintenance.
  • Healthcare (VOC-based breath diagnostics).

---