A Bio-Inspired Spiking Neural Network with Few-Shot Class-Incremental Learning for Gas Recognition

Area of Science:

• Artificial Intelligence
• Sensor Technology
• Chemical Sensing

Background:

• Gas sensor performance degrades due to drifting, aging, and environmental factors like temperature and humidity.
• Maintaining gas recognition accuracy requires effective methods to counteract performance decline.
• Incremental online learning offers a solution for adapting sensor networks to changing conditions.

Purpose of the Study:

• To develop a bio-inspired spiking neural network (SNN) for accurate recognition of nine flammable and toxic gases.
• To enable few-shot class-incremental learning for rapid retraining of the gas recognition system.
• To validate the SNN's robustness and effectiveness compared to traditional algorithms in real-life fire scenarios.

Main Methods:

• Development of a bio-inspired spiking neural network (SNN) architecture.
• Implementation of few-shot class-incremental learning for online retraining capabilities.
• Comparative analysis against Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Principal Component Analysis (PCA) + SVM, PCA + KNN, and Artificial Neural Network (ANN).

Main Results:

• The proposed SNN achieved the highest accuracy of 98.75% in five-fold cross-validation for identifying nine gases across five concentrations.
• The SNN demonstrated a 5.09% higher accuracy compared to other evaluated gas recognition algorithms.
• The network's ability for rapid retraining with new gas types at a low accuracy cost was confirmed.

Conclusions:

• The bio-inspired SNN offers a robust and effective solution for gas recognition, overcoming limitations of traditional methods.
• The SNN's few-shot class-incremental learning capability ensures sustained performance in dynamic environments.
• The developed SNN is highly suitable for real-life fire scenarios requiring reliable flammable and toxic gas detection.