E-nose

Project Overview

E-nose system capable of differentiating perfume, coffee, and clean air with 98% accuracy

This project uses an array of six MQ-series gas sensors combined with a lightweight machine learning model trained using Edge Impulse.
The model is deployed on the Adafruit Feather M4 Express, which supports TensorFlow Lite Micro for efficient ARM-based inference.

Hardware Setup

Wiring Diagram (Prototype on Feather M4 Express)

Sensor Array Used

The following sensors make up the current gas-detection array:

Sensor	Primary Sensitivity
DF-NH3	Ammonia
MQ-136	Hydrogen sulfide
MQ-135	VOCs / CO₂
MQ-8	Hydrogen
MQ-4	Methane
MQ-2	Smoke / LPG

These sensors together form a basic but effective VOC fingerprint for classification.

MCU: Adafruit Feather M4 Express

Reasons:

• ARM Cortex-M4F with hardware floating-point accelerators
• Fully supported by TensorFlow Lite Micro
• Supported directly by Edge Impulse Arduino libraries
• Higher performance vs AVR for ML workloads

---

Model Training (Edge Impulse)

Project (public link):
https://studio.edgeimpulse.com/public/661994/live

Data Types Collected

• 6 analog voltage channels
• Sampling frequency ~20 Hz
• 5-second windows (~100 sample frames)
• Labels: Perfume, Coffee, Natural Air

Model Performance

• 99% accuracy on GPU (training)
• ~97% estimated accuracy on Feather M4
• Latency: ~5 seconds
• Low overfitting risk (consistent validation performance)

---

Data Collection Pipeline

1. Upload firmware to Arduino for sensor reading

Source file:
collect-sensor-array.ino
:contentReference[oaicite:0]{index=0}

2. Log sensor values using Python

Script:
record-sensor-array.py

Run:

python record-sensor-array.py

This script:

• Reads serial output
• Cleans & parses voltage values
• Saves into a CSV dataset

---

Edge Impulse Workflow

Step 1 — Clone the project

Go to studio.edgeimpulse.com → Clone Project

Step 2 — Import your own data (Optional)

Data Acquisition → Upload Data
Upload CSV files generated from your Python logger.

Step 3 — Build an impulse

Impulse Design → Create impulse

Recommended:

• Window size: 5000 ms
• Frequency: 20 Hz
• DSP Block: Raw Data
• ML Block: Classification (Neural Network)

Step 4 — Train the model

Impulse Design → Classification

Settings:

• Epochs: 100
• Learning rate: 0.001
• Train/Test Split: 80/20

---

Deployment to Adafruit Feather M4

Step 1 — Export Arduino library

Go to:

Deployment → Arduino Library → EON Optimizer (uint8)
Download the ZIP and place it under:

Documents/Arduino/libraries/

Restart Arduino IDE. Click on Sketch -> Include Zip Library -> Install the library.

Step 2 — Use the included inference sketch

Inside the exported folder:

examples/arduino/feather-m4/feather-ei-inferencing.ino

This code:

• Reads analog data
• Formats a feature vector
• Passes it to the model
• Prints probability scores

Typical Output

Coffee: 0.98
Perfume: 0.01
Natural Air: 0.00

---

License

MIT License

---

Contact

[email protected] December 2025
E-Nose Project