Artificial Nose — Spice Identification using Machine Learning

Implementation of the research paper:

Artificial Nose: Machine Learning model for categorization & identification of spices
Kunal Bhujbal, Prof. Asha Rawat — IEEE Conference, 2023

---

Overview

An Artificial Nose system that identifies three spices — Cinnamon, Clove, and Nutmeg — using an array of 8 MOS gas sensors and a Random Forest classifier (10 decision trees). A Multilayer Perceptron (MLP) is also trained for comparison.

Model	Cinnamon	Clove	Nutmeg
Random Forest	100%	100%	100%
MLP Neural Network	~96%	~97%	~97%

---

System Architecture

Odour Sample → Gas Sensor Array (×8) → Signal Conditioning
             → PIC Microcontroller (10-bit ADC) → Serial USB
             → PC Data Acquisition (Java / Excel)
             → Python ML Pipeline → Spice Classification

Sensors Used

Sensor	Type
MQ 135	Air quality / ammonia
TGS 813	Combustible gases
TGS 800	Air contaminants
TGS 823	Organic solvents
TGS 2610	LP gas / butane (most important feature)
TGS 2602	VOCs / odorous gases
TGS 2620	Organic solvents
TGS 2611	Methane

---

Project Structure

Artificial Nose/
├── data/
│   ├── raw/                    # Raw sensor CSVs (real or synthetic)
│   └── processed/              # Reserved for future use
├── src/
│   ├── data_generator.py       # Synthetic data / real CSV loader
│   ├── preprocessor.py         # Clean → encode → scale → split
│   ├── evaluator.py            # Metrics: Table I, Table II, confusion matrix
│   ├── visualizer.py           # Figures 6, 7, 8 + comparison chart
│   └── models/
│       ├── random_forest.py    # Primary model (paper's main contribution)
│       └── mlp.py              # Comparison model
├── notebooks/
│   └── analysis.ipynb          # Full walkthrough with inline figures
├── results/                    # Auto-generated figures (PNG)
├── main.py                     # Entry point
└── requirements.txt

---

Quickstart

# 1. Install dependencies
pip install -r requirements.txt

# 2. Run with synthetic data (no hardware required)
python main.py

# 3. Run with real PIC sensor data
#    CSV must have columns: MQ135, TGS813, TGS800, TGS823,
#                           TGS2610, TGS2602, TGS2620, TGS2611, label
python main.py --data data/raw/sensor_data.csv

# 4. Run without opening plot windows (CI / headless server)
python main.py --no-plots

---

Using Real Hardware Data

If you have the physical Artificial Nose setup (PIC microcontroller + sensor array):

1. Run the Java acquisition software — it saves readings to an Excel file.
2. Export as CSV with these exact column names:

   MQ135, TGS813, TGS800, TGS823, TGS2610, TGS2602, TGS2620, TGS2611, label
   

   where label is one of: cinnamon, clove, nutmeg.
3. Place the file at data/raw/sensor_data.csv and run python main.py.

---

Output — Paper Figures

File	Corresponds to
results/figure6_sensor_responses.png	Figure 6 — Avg sensor voltage per spice
results/figure7_confusion_matrix_random_forest.png	Figure 7 — RF confusion matrix
results/figure7_confusion_matrix_mlp_neural_network.png	Figure 7 — MLP confusion matrix
results/figure8_feature_importance.png	Figure 8 — TGS2610 most important
results/algorithm_comparison.png	Table II — RF vs MLP bar chart

---

Key Results (Paper)

Table I — Mean accuracy across 9 test runs (Random Forest):

Spice	Accuracy
Clove	90.65%
Cinnamon	90.06%
Nutmeg	99.09%

Table II — Final test accuracy:

Algorithm	Cinnamon	Clove	Nutmeg
Random Forest	100%	100%	100%
MLP	96.87%	97.38%	97.32%

---

Notebook

For an interactive walkthrough with inline figures, open:

jupyter notebook notebooks/analysis.ipynb

---

Citation

Bhujbal, K., & Rawat, A. (2023). Artificial Nose: Machine Learning model
for categorization & identification of spices. IEEE Conference.