Development of high-performance and lightweight deep learning architectures for multimodal gas leak detection and classification

Accurate and rapid classification of gas leaks is of critical importance for ensuring the safety of life and property in industrial facilities and living spaces. Traditional unimodal detection systems often produce false alarms or fail to detect low-concentration leaks due to their susceptibility to environmental changes and cross-sensitivity issues. This thesis aims to overcome these limitations and optimize the balance between accuracy and computational cost in gas detection through multimodal deep learning approaches that combine thermal imaging and chemical sensor data. Within the scope of the study, two different architectural approaches have been developed and systematically analyzed. The first approach is the "High-Capacity Hybrid Fusion" model, designed for safety-critical scenarios with high computational power availability. This architecture utilizes a hybrid Convolutional Neural Network-Transformer structure to capture temporal and global relationships in sensor data, and a ResNet50 backbone for spatial features in thermal images. Combining autoencoder-based early fusion and a weighted voting mechanism (late fusion), this model achieved an accuracy rate of 99.92% on the MultimodalGasData dataset with approximately 25 million parameters, demonstrating one of the highest performances in literature. The second approach is the "Lightweight Early Fusion" model, optimized for resource-constrained edge devices and distributed sensor networks. In this architecture, a Multilayer Perceptron for sensor data and a ShuffleNetV2 backbone for thermal images were preferred to minimize computational load. Containing only 338,160 parameters, this compact architecture reduced the model size by approximately 71 times compared to the hybrid model, providing a 98.6% saving in the number of parameters. Despite this radical reduction, the model proved its efficiency success by reaching a competitive accuracy rate of 98.91%. The experimental results obtained indicate that the proposed hybrid architecture can be effectively used in high-risk zones requiring error-free detection, while the lightweight architecture is suitable for real-time and low-cost monitoring applications. This thesis expands the design space in the literature by offering both performance-oriented and efficiency-oriented solutions to the multimodal gas classification problem.