A Novel Visual Transformer for Long-Distance Pipeline Pattern Recognition in Complex Environment

The effective warning of dangerous events along long-distance pipelines is critical to ensure the safety of oil and gas transportation. Distributed optical fiber sensing (DOFS) technology can assist operators to identify threat vibration signals. However, due to the complex and changeable environmental background noise along long-distance pipeline, most of the existing methods only extract one-dimensional features of the signal, making it difficult to distinguish various types of environmental noise, strong interference, and dangerous events. Besides, the samples of different classes in the actual scene are unbalanced. The sample size of dangerous events is often smaller than others. To address these problems, we use image encoding to transform the time-series signals collected by the DOFS system into image data, and fully extract the time dependence and the correlation between different elements in the signal. Moreover, a visual transformer model PipelineADWinT is proposed. The self-attention mechanism of diagonal-axial window designed in this model can perfectly combine image encoding features, and obtain local to global multi-scale features through hierarchical structure. By optimizing the loss function, the model’s ability to handle the class imbalance problem is enhanced. The experimental results show that PipelineADWinT has more comprehensive classification performance and fewer false alarms than all the baseline models, which proves the effectiveness and superiority of the model.