Hierarchical-Attention-Based Knowledge Distillation Network with Category Structure for Natural Gas Pipeline Leakage Detection Under Imbalanced Data

Accurate detection of natural gas pipeline leakage (NGPLD) under imbalanced data presents a significant challenge due to the rarity of leakage incidents. To this end, this paper proposes a novel hierarchical-attention-based knowledge distillation network with category structure (HACKDN), which simultaneously accounts for multi-level information and category structure relations. Concretely, three one-dimensional (1-D) convolutions with varying kernels and a spatial attention mechanism (SAM) are exploited to capture critical temporal features across multiple scales, enhancing the model's feature representation capability. Subsequently, an attention-split (A-S) network is devised, integrating a hierarchical-split (H-S) block and a channel attention mechanism (CAM) to extract robust and discriminative features with key channels at various levels. This network is embedded into a multi-teacher knowledge distillation (KD) network, improving the performance of shallow network in recognizing minority categories. Finally, intra- and inter-category losses are introduced to optimize the category structure relations in the embedding space, promoting the compactness and separability of category representations. Experimental results demonstrate that the proposed HACKDN surpasses several state-of-the-art methods, significantly improves the accuracy of NGPLD under imbalanced data, providing an efficient and reliable tool for real-world pipeline monitoring.