A Reproducible Benchmark for Gas Sensor Array Classification: From FE-ELM to ROCKET and TS2I-CNNs | JoVE Visualize

Area of Science:

Machine Learning
Sensor Technology
Data Science

Background:

Classifying low-concentration Gas Sensor Array (GSA) data presents significant challenges due to low signal-to-noise ratio (SNR), sensor heterogeneity, drift, and limited sample sizes.
Existing methods often struggle to achieve robust performance under these adverse conditions, necessitating novel approaches and comprehensive benchmarking.

Purpose of the Study:

To benchmark various time-series classification methods, including time-series-to-image (TS2I) Convolutional Neural Networks (CNNs), against established baselines for low-concentration GSA data.
To provide a reproducible and fair comparison of different classification strategies to guide model selection in electronic nose (e-nose) applications.

Main Methods:

Reproduced a strong FE-ELM baseline and compared it with vector baselines, traditional time-series classifiers (TCN, MiniROCKET), and TS2I-CNNs.
Utilized the GSA-LC and GSA-FM datasets with a rigorous 20x5 repeated stratified cross-validation (n=100) for robust evaluation.
Assessed performance using accuracy and Macro-F1 scores, employing paired t-tests with Holm-Bonferroni correction for statistical significance.

Main Results:

The ROCKET time-series classifier achieved the highest accuracy on both GSA-FM (0.9721 ± 0.0480) and GSA-LC (0.9578 ± 0.0433) datasets, significantly outperforming TCN and MiniROCKET (p < 0.05).
Among image-based models, CNN-RP demonstrated the most robustness, while CNN-GASF showed limitations, particularly on the GSA-LC dataset.
RGB fusion strategies and transfer learning with ResNet-18 did not offer consistent advantages, indicating dataset-dependency and limited generalizability.

Conclusions:

ROCKET is the top-performing model for classifying low-concentration GSA data, offering superior accuracy and reliability.
CNN-RP serves as the most viable TS2I alternative for these challenging conditions, providing a robust image-based approach.
The study establishes a reproducible benchmark, offering practical guidance for model selection in e-nose systems and clarifying the capabilities of TS2I methods.