Heterogeneous data fusion model for gas leakage detection

Abstract

This paper proposes a gas identification model based on multi-source data fusion, termed SCGA (Spatio-temporal Cross-attention Gas identification Algorithm). The method initially employs a pre-trained ShuffleNetV2 network to process thermal images, extracting spatial-level feature representations. Subsequently, a hybrid architecture integrating 1D Convolutional Neural Networks (1DCNN) and Bidirectional Gated Recurrent Units (BiGRU) is designed to comprehensively capture temporal dynamics in gas sensor data. To effectively fuse features from these heterogeneous sources, a cross-attention mechanism is introduced to model spatio-temporal dependencies, thereby generating discriminative fused features. Finally, the fused features are fed into the classification module via residual connections to perform gas identification. Experimental results demonstrate that the proposed SCGA model achieves a recognition accuracy of 99.22 % on real-world datasets while maintaining stable loss values within the range [0, 0.04]. Compared to models using only gas sensor data (SVM, 1DCNN, BiGRU), SCGA improves accuracy by at least 4.12 percentage points. Against models utilizing solely thermal image data (MobileNetV3, ShuffleNetV2, ResNet18), it exhibits an improvement of at least 1.14 percentage points. Furthermore, SCGA outperforms the multi-source fusion baseline with direct feature concatenation (SCG model) by 1 %. These results validate the high precision and robustness of the SCGA framework for gas recognition tasks.