Objective Long-distance oil and gas pipelines increasingly use multiple sensing systems. These include fiber-optic vibration warning, leak monitoring, unmanned aerial vehicle (UAV) inspection, and video surveillance. However, multi-model alerts often suffer from inconsistent time scales, mismatched spatial references, unexplainable evidence conflicts, and heavy manual verification workloads. Pipeline safety risk analysis is moving from single-point alerts to integrated assessment. To support this shift, this study proposes a multi-source sensing risk assessment method for oil and gas pipelines. The method is based on a central cognitive agent.

Methods This study establishes a multi-source sensing data organization framework for pipeline incidents. Outputs from fiber-optic vibration detection, pressure/acoustic monitoring, UAV imagery, and video object detection are mapped into structured risk evidence. Four dedicated sensing models are constructed for fiber-optic early warning, leak monitoring, UAV inspection, and video surveillance. They perform vibration event identification, leak condition evaluation, inspection target recognition, and video object detection. Confidence calibration is uniformly applied across all model outputs. A central cognitive agent is built on the output layer of the dedicated sensing models.The agent treats structured multi-source evidence as event states and incorporates spatio-temporal alignment results. It invokes functional modules for historical case retrieval, operational rule retrieval, and event memory query. It also uses large language models to verify evidence consistency, attribute risks, and generate handling recommendations. For incidents with conflicting evidence or insufficient information, incident states are updated through supplementary retrieval and rule verification. Finally, risk levels, supporting evidence, causal explanations, and handling recommendations are generated. Manually reviewed results are further fed back into incident memory.

Results Experiments are conducted using desensitized playback data from two in-service pipelines. The proposed method achieves a precision of 94.08％, a recall of 91.76％, and an F1-score of 92.91％, with the false alarm rate reduced to 4.90％. Compared with four other risk assessment methods, it obtains a higher F1-score and a lower false alarm rate. In cross-pipeline tests, the method attains F1-scores of 90.42％ for migration from Pipeline 1 to Pipeline 2 and 91.03％ for migration from Pipeline 2 to Pipeline 1, demonstrating favorable migration stability.

Conclusion Without replacing dedicated on-site sensing models, the central cognitive agent enables interpretable, comprehensive reasoning over heterogeneous alerts. It significantly enhances risk identification recall, false alarm suppression capability, and engineering practicability, providing robust technical support for intelligent pipeline inspection and emergency response.