SDCSN: a hierarchical parallel localization method for pipeline leakage based on vibration signals

Abstract

Abstract In applying deep learning methods to detecting and localising pipeline leaks, improving the fitness of deep learning methods to leak signals is an important task. We propose a novel detection model called Stacked dilated convolutional shrinkage network (SDCSN). This model incorporates a Stacked Dilated Convolution (SDC) module specifically designed for vibration signals, enabling the extraction of rich multi-scale local features. Moreover, implementing the Residual Shrinkage Building Unit (RSBU) module for noise reduction in the network architecture. Building upon this foundation, we introduce a new concept centred around hierarchical leakage discrimination and parallel prediction positioning. This approach enables accurate assessment of leakage levels and precise identification of multiple leakage points. Finally, the performance of the proposed method is verified in real experiments and the optimal settings for the dilated rate are determined. The results demonstrate a maximum classification accuracy rate reaching 98.94%.