Synthetic E-Nose Data Generation Using Transformer

📌 Project Overview

This project focuses on generating realistic synthetic Electronic Nose (E-Nose) time-series data using a Transformer-based model. The goal is to improve data availability, generalization, and downstream classification performance for coffee quality assessment.

This work was carried out as part of an internship project.

---

🎯 Objectives

• Generate realistic synthetic E-Nose sensor data
• Capture and distinguish spectral patterns of different coffee quality types
• Evaluate the usefulness of synthetic data in downstream machine learning tasks

---

📊 Dataset Description

The dataset is based on E-Nose measurements for Colombian coffee quality control, aimed at detecting defects during cup tests.

Key details:

• 58 coffee samples
• 3 quality labels:
  • High Quality (HQ)
  • Average Quality (AQ)
  • Low Quality (LQ)
• Time-series data sampled at 1 Hz for 300 seconds
• 8 gas sensors per sample:
  • SP-12A, SP-31, TGS-813, TGS-842
  • SP-AQ3, TGS-823, ST-31, TGS-800
• Sensor readings are resistance values (kΩ)

---

🧠 Model Architecture

A Transformer Encoder–Decoder architecture was used for time-series reconstruction and synthetic data generation.

Key components:

• Self-Attention to capture long-range temporal dependencies
• Multi-Head Attention for diverse feature learning
• Positional Encoding to preserve time-step order

---

⚙️ Training Setup

• Framework: PyTorch
• Environment: Google Colab
• Version Control: GitHub

Preprocessing & Training:

• Data normalization
• Loss function: Mean Squared Error (MSE)
• Train–validation split for generalization testing

---

📈 Results

Reconstruction Quality

• Low MSE between real and synthetic data
• Strong sensor-wise similarity
• Real vs. Synthetic plots show close alignment

Training Behavior

• Smooth decrease in training and validation loss
• No overfitting observed
• Stable and well-regularized training process

---

🔍 Downstream Evaluation

Synthetic data was evaluated using an LSTM-based classifier.

Classification Accuracy:

• Real data only: 0.6667
• Synthetic data only: 0.7500
• Hybrid (real + synthetic): 0.7083

✅ Synthetic data improves classification performance.

---

⚠️ Challenges

• Imbalanced dataset
• Overfitting risks
• Long training times
• Handling time-series data with Transformer models

---

📚 Key Learnings

• Importance of synthetic data in machine learning
• Transformers are effective for time-series generation
• Visualization and downstream evaluation are critical
• Synthetic data enhances model robustness

---

🚀 Conclusion & Future Work

• Successfully generated realistic synthetic E-Nose data
• Demonstrated usefulness for coffee quality classification
• Future directions:
  • Statistical comparisons (PCA, t-SNE) between real and synthetic data
  • Exploring GANs and Informer models for improved performance