ECNet: Network Based on Encoder with Local Convolutional Attention and Cross Layer Information Fusion for Natural Gas Pipeline Leakage Detection

To address the issues of insufficient temporal information feature extraction and detection accuracy in natural gas pipeline leakage detection, we propose ECNet, an intelligent detection network based on an improved encoder [containing local convolutional attention (LCA)] and cross-layer information fusion (CLIF). First, by using the dual-branch cascading method to separately connect two encoders, we can extract features from the two-channel combined leakage time series data in a channel-wise and time-step-wise manner, allowing us to consider both global and local temporal information simultaneously. Second, we design a CLIF method to enhance the model's multilevel expression ability by achieving an interactive fusion of feature information between the intermediate layers and the top layer in the cascaded structure of the dual branches. Finally, we compare ECNet's performance with other deep learning methods by commonly used evaluation metrics. The experimental results show that ECNet has higher detection accuracy, better stability performance, and relatively moderate operating efficiency, which makes it effective and reliable in the field of natural gas pipeline leak detection.