Abstract: Objective Long-distance oil and gas pipelines have been increasingly equipped with sensing models such as fiber-optic vibration early warning, leakage monitoring, UAV inspection, and video surveillance. However, multi-model alarms still suffer from inconsistent temporal scales, non-unified spatial references, difficulty in explaining conflicting evidence, and heavy manual review workload. To support the transition from isolated alarms to comprehensive risk judgement, a multi-source risk assessment method based on a central cognitive agent is proposed. Methods A pipeline-segment-event-oriented data organization mode is established, and fiber-optic vibration signals, pressure-acoustic signals, UAV images and video detection results are uniformly mapped into structured risk evidence. Four specialized sensing models are constructed for fiber-optic warning, leakage monitoring, UAV inspection and video surveillance to perform local event recognition and confidence calibration. A central cognitive agent is built on top of these models, taking structured multi-source evidence as the event state. Based on the current evidence state, the agent retrieves historical cases, business rules and event memory, and employs a large language model to perform evidence consistency analysis, risk attribution and disposal suggestion generation. For events with conflicting or insufficient evidence, the event state is further updated through supplementary retrieval and rule checking before a risk conclusion or manual-review request is generated.Results Experiments were conducted using desensitized replay data from two in-service pipelines. The proposed method achieved a precision of 94.08%, a recall of 91.76%, and an F1-score of 92.91%, while reducing the false alarm rate to 4.90%. Compared with the attention-based fusion model, the F1-score increased by 3.61 percentage points and the false alarm rate decreased by 2.50 percentage points. In cross-segment tests, the F1-scores reached 90.42% and 91.03% in two migration directions, respectively. Conclusion Without replacing field-specific models, the central cognitive agent can provide explainable comprehensive judgement of heterogeneous alarms and support intelligent patrol and emergency response of long-distance pipelines. (6 Figures, 4 Tables, 30 References)