Gas Detection and Classification Using Multimodal Data Based on Federated Learning.

Area of Science:

• Environmental Science
• Computer Science
• Sensor Technology

Background:

• Gas leak detection is vital in industries and homes to prevent environmental and human harm.
• Low-cost sensors have limitations in reliability and long-distance detection.
• Multimodal data fusion offers a robust solution for enhanced gas identification.

Purpose of the Study:

• To introduce a novel multimodal dataset for gas detection, combining sensor and thermal imaging data.
• To develop and evaluate AI models for effective gas classification using this dataset.
• To explore federated learning for privacy-preserving gas leakage classification.

Main Methods:

• A multimodal dataset of 6400 samples was created using gas sensors and a thermal imaging camera.
• Convolutional Neural Networks (CNNs) with variants like Bi-LSTM and dense LSTM were trained on thermal and sensor data.
• Federated learning was implemented for privacy-preserving classification.

Main Results:

• AI models demonstrated effective classification of various gas types (smoke, perfume) and neutral environments.
• Fusion of sensor and thermal data improved classification accuracy.
• Federated learning achieved accuracy comparable to traditional deep learning methods.

Conclusions:

• The multimodal dataset and developed AI models offer a valuable resource for gas leakage detection research.
• Combining thermal imaging with sensor data enhances detection capabilities.
• Federated learning provides a viable, privacy-preserving approach for AI-driven gas classification.