# Gas Detection and Identification Using Multimodal Artificial Intelligence Based Sensor Fusion

**Authors:** Parag Narkhede, Rahee Walambe, Shruti Mandaokar, Pulkit Chandel, Ketan Kotecha, George Ghinea  
**arXiv:** [2106.06180](https://arxiv.org/abs/2106.06180)  
**Date:** 2021-06-11  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2106.06180v1/

## Key points

- Industrialization and technological advancements require innovative engineering technologies that are cost-effective, efficient, and easy to implement.
- Gas leaks in settings like coal mines, chemical industries, and home appliances are causing increasing accidents.
- A novel approach using multimodal AI fusion techniques for detecting and identifying gaseous emissions has been proposed to address this challenge.
- Traditional gas detection methods face limitations due to the colorless, odorless, and tasteless nature of many gases.
- Sensor fusion is crucial for reliable detection in real-world applications as single-sensor systems may lack accuracy.
- The study collected 6400 gas samples using a 7-semiconductor gas sensors array and a thermal camera with an early fusion method of multimodal AI applied to fuse data from both sensors.
- Testing revealed an impressive accuracy rate of 96% for the fused model compared to individual model accuracies of 82% (Gas Sensor data using LSTM) and 93% (thermal images data using CNN model).
- The paper discussed various methodologies for multimodal data fusion including early fusion, late fusion, and hybrid fusion techniques along with the importance of Convolutional Neural Networks (CNN) for processing thermal image data.
- Future research directions include collecting datasets comprising multiple gases to enhance training robustness in deep learning frameworks and utilizing dedicated hardware processing modules for efficient operation.

## Summary

In the realm of industrialization and technological advancements, the need for innovative engineering technologies that are cost-effective, efficient, and easy to implement is paramount. One pressing issue is the increasing number of accidents caused by gas leaks in various settings such as coal mines, chemical industries, and home appliances. To address this challenge, a novel approach utilizing multimodal AI fusion techniques for detecting and identifying gaseous emissions has been proposed. Traditional gas detection methods often fall short due to the colorless, odorless, and tasteless nature of many gases. Single-sensor systems may lack accuracy, making sensor fusion crucial for reliable detection in real-world applications. In this study, 6400 gas samples were manually collected using two specific sensors: a 7-semiconductor gas sensors array and a thermal camera. The early fusion method of multimodal AI was applied to fuse data from both sensors. The network architecture featured a feature extraction module for each modality which were then fused through a merged layer followed by a dense layer to provide a single output for gas identification. Testing revealed an impressive accuracy rate of 96% for the fused model compared to individual model accuracies of 82% (Gas Sensor data using LSTM) and 93% (thermal images data using CNN model). These results underscored the superiority of fusing multiple sensors and modalities over relying on a single sensor. Further elaboration on system architecture was provided in Section 3 of the paper while Section 4 delved into detailed discussions on obtained results. The paper concluded in Section 5 by outlining future research directions. The theoretical background highlighted various methodologies for multimodal data fusion including early fusion, late fusion, and hybrid fusion techniques. Convolutional Neural Networks (CNN) were also discussed as essential tools for processing thermal image data. Moving forward, the focus will be on collecting datasets comprising multiple gases to enhance training robustness in deep learning frameworks. Additionally, dedicated hardware processing modules will be essential for ensuring efficient operation. The proposed architecture holds promise for high-risk applications like leak detection in chemical plants or identification of explosives where reliability is critical. By leveraging advanced AI technologies and robust training methods with diverse datasets, this research paves the way towards more effective gas detection systems in various industries.