PSEW-VLM: An Interactive Distributed Fiber Optic Sensing Model for Threat Identification and Decision-Making in Oil and Gas Pipelines

Abstract

Oil and gas pipelines are core infrastructure for energy transportation. They present critical challenges in safety monitoring and threat detection to ensure energy supply and public safety. Conventional AI-based recognition models for early warning in distributed fiber-optic sensing still exhibit significant limitations. They struggle to establish cross-modal semantic associations and provide explainable results. This article proposes a pipeline safety early warning model based on the visual language model (PSEW-VLM). The model constructs a multimodal expert system that processes fiber-optic sensing energy distribution maps and domain text knowledge. It integrates contrastive language-image pre-training (CLIP) cross-modal contrastive learning with low-rank adaptation (LoRA). This design addresses limitations in cross-modal reasoning, human–computer interaction, and explainable text generation. Experimental results demonstrate that PSEW-VLM precisely recognizes various threat and interference events in complex scenarios, with an accuracy of 96.59 %. It also supports multimodal natural language interaction, enhancing the practicality of human–machine collaboration. The innovative value of this research lies in integrating large-scale model technology with industrial sensing data. This work introduces a practical multimodal inference framework that bridges distributed fiber optic sensor (DFOS) measurements with domain textual reasoning.