Multisource Multimodal Feature Fusion for Small Leak Detection in Gas Pipelines

Abstract

Improving the ability to detect small leaks to prevent more severe accidents plays an extremely important role in the safe operation of pipelines. To tackle the issue of low diagnostic accuracy associated with single sensors for detecting small leaks, a multisource multimodal feature fusion method for gas pipeline leak detection was proposed. First, the collected data from multiple sensors were transformed into 2-D time-frequency images for input into the feature extraction network. Then, the dual-information fusion (DIF) module was introduced, incorporating the attention mechanism and multiscale feature fusion to enhance the model’s feature expression capability and fully interact with the multimodal features. Second, the channel split multiscale convolution (CSMC) module was designed to accommodate the diversity of input data and improve the model’s generalization capability. The DIF and CSMC modules were cascaded and fused to produce the classification results through the fully connected layer. Finally, the effectiveness of the proposed method was assessed using pipeline leak data collected in the laboratory. The experimental results demonstrate that the proposed multimodal deep learning model can effectively identify the small leak state in pipelines, exhibiting superior diagnostic performance when compared to the current mainstream image classification models.