Multi-Dimensional Leakage State Identification for Natural Gas Pipelines Using MKTCN and KARN

Natural gas pipeline leaks are characterized by sudden occurrences, complex operating conditions, and wide-ranging hazards, requiring high-precision, multi-task state determination methods. Traditional approaches can only output single-dimensional tasks and cannot simultaneously capture multi-dimensional leakage state information. In this study, we propose a multi-task recognition model for pipeline leakage that integrates Multi-scale Temporal Convolution Network (MKTCN) and Gated Attention Residual Network (KARN). MKTCN captures long- and short-term temporal features through varied convolutions and dilation rates, which are then fused using 1×1 convolutional layers. The KARN module, with stacked 1D convolutions and a channel attention mechanism, generates deep temporal representations that enhance feature relevance and suppress noise interference. These features are fed into a shared fully connected backbone for further fusion, followed by separate task heads for leakage distance, pressure, noise, and sensor recognition. Weighted cross-entropy loss and gradient clipping balance task learning rates and importance, enabling end-to-end joint training and recognition. Experimental results show that the model achieves accuracies of 0.9911, 0.9317, 1.0, and 1.0 for the four tasks, with Task Gap, Conflict Norm, and Synergy metrics of 0.0668, 0.0055, and -0.0491, respectively, demonstrating its stability and multi-task synergy.