Time Series Classification on E-Nose Data with ConvTran and Comparative Models

This repository provides a complete framework for classifying time series collected from gas sensors (e-noses) using a range of deep learning models, including the Transformer-based model ConvTran. Two curated datasets, resistance_values.csv and resistance_values7class.csv, are used to benchmark performance under both balanced and imbalanced class conditions.

---

📦 Datasets

Dataset File	Description
resistance_values.csv	Case Study 1 (CoffeePow-4): 4 classes (3 coffee types + ambient air)
resistance_values_7class.csv	Case Study 2 (Aroma-7): 7 classes (3 coffee type + ambient air + 3 fragranced creams)

Both datasets were collected using a Bosch BME688 sensor array, executing controlled 10-step temperature cycles to generate distinctive odor fingerprints.

📈 CoffeePow-4 (4 Classes)

• Classes: Ambient Air, Borbone Coffee, Lavazza Coffee, Primia Coffee
• Samples: 3583 sequences, class-balanced
• Task: Evaluate baseline accuracy

🌿 Aroma-7 (7 Classes)

• Classes: All from CoffeePow-4 + Almond Cream, Thyme Cream, Marigold Cream
• Samples: 4750 sequences (imbalanced, 3:1 coffee:cream ratio)
• Task: Test robustness under class imbalance and cross-domain volatility

Sensor Setup

• Hardware: Bosch BME688 ×8 array
• Features used: Gas resistance (10 readings per heater cycle)
• Preprocessing:
  • Sensor-specific standardization (zero mean, unit variance)
  • Stratified 80/10/10 train/val/test split

---

🧠 Models Implemented

Model	Type	Key Features
ConvTran	Transformer	tAPE + eRPE position encodings
ALSTM-FCN	Hybrid	LSTM + attention + 1D CNN
LSTM-FCN	Hybrid	LSTM + 1D CNN
InceptionTime	CNN	Residual inception modules for multi-scale temporal pattern detection
FCN	CNN	Lightweight 3-layer 1D CNN with global pooling
ResNet	CNN	Deep residual CNN architecture
Bosch AI Studio	Proprietary	Baseline classifier (black-box)

---

🚀 Installation

pip install -r requirements.txt

Requirements include: torch, tensorflow, scikit-learn, matplotlib, numpy, and pandas.

---

⚙️ Usage

Each model is executed via its corresponding script. Use either dataset depending on the case study.

Common CLI Parameters

--data_path          Path to CSV file (e.g., resistance_values.csv)
--output_dir         Directory for saving outputs
--device             'cuda' or 'cpu' (default: 'cuda')
--sequence_len       Length of input sequences (default: 10)
--search_trials      Number of random search configurations (default: 10)
--seed               Random seed (default: 42)

---

🧪 Training Pipeline

1. Load and preprocess dataset
2. Execute random hyperparameter search
3. Train best configuration
4. Evaluate on test set
5. Save model, metrics, confusion matrix, and plots

All models use the same preprocessing logic to ensure fair benchmarking.

---

🔍 Hyperparameter Spaces

ConvTran

HSPACE = {
    'emb_size': [32, 64, 128],
    'num_heads': [2, 4, 8],
    'dim_ff': [64, 128, 256],
    'epochs': [200, 500, 1000],
    'lr': [1e-3, 1e-4, 1e-5],
    'batch_size': [16, 32, 64],
    'dropout': [0.0, 0.1, 0.2]
}

See scripts for other model-specific hyperparameter spaces.

---

📈 Results Summary

Case Study 1: CoffeePow-4 (Balanced)

Model	Accuracy	F1 Score	FPR	Params	Latency (ms)
FCN	97.21%	97.05%	0.92%	92,100	0.41
ConvTran	96.66%	96.42%	1.12%	24,298	0.92
InceptionTime	96.94%	96.82%	1.04%	477,700	63.0
ALSTM-FCN	95.54%	95.32%	1.50%	266,440	63.3
LSTM-FCN	93.87%	93.46%	2.05%	292,612	62.4
ResNet	93.02%	92.63%	2.35%	2,012,552	65.3
Bosch Model	87.81%	86.98%	4.03%	N/A	N/A

Case Study 2: Aroma-7 (Imbalanced)

Model	Accuracy	F1 Score	FPR
ConvTran	95.16%	96.03%	0.85%
ResNet	93.68%	94.87%	1.11%
InceptionTime	93.47%	94.76%	1.16%
ALSTM-FCN	90.11%	92.37%	1.77%
FCN	77.47%	85.12%	3.96%
LSTM-FCN	69.26%	78.82%	5.44%

🔍 ConvTran preserves performance under real-world imbalance, outperforming convolutional and recurrent baselines on minority cream classes.

---

🧾 Metrics

• Accuracy: Overall correct predictions
• Macro F1 Score: Harmonic mean of per-class precision and recall
• False Positive Rate (FPR): Rate of incorrect class activations
• Confusion Matrix: Insightful view of class confusions

---

📚 Citation

---

👥 Credits

---

🧠 Insights from the Paper

• ConvTran fits within 115 KB and uses only 24k parameters, making it ideal for microcontrollers
• Handles long-range dependencies via self-attention
• Maintains performance as the number of odor classes increases
• Clearly superior to proprietary black-box models in both accuracy and interpretability

---

📂 Repository Structure

---

📌 License

Released under the MIT License. All datasets and model configurations are made available for reproducibility and future research.