[Objective] Object detection in intelligent surveillance systems at thermal power plants generates frequent false alarms in scenes with ambiguous visual boundaries, such as a partially occluded safety helmet or a small flame that is easily confused with refle cted light. Because every alarm has to be confirmed by hand, these false positives subject control-room operators to chronic alarm fatigue and gradually increase the risk that a genuine hazard is overlooked. This paper aims to solve this problem. 

[Methods] This paper proposes an alarm re-examination method that combines an object detector with a vision -language model (VLM) for semantic reasoning. Unlike a detector, which classifies an isolated bounding box, a VLM can weigh the surrounding scene against the wording of the relevant safety rule, which makes it better suited to the borderline cases that produce most false alarms. In the proposed pipeline, the detector preserves a high recall rate for initial screening so that few real hazards are missed, and the alarms it raises are then passed to the VLM, which performs semantic verification against power industry safety codes before any alarm reaches the operator. The study addresses four typical risk categories in plant operation: safety helmet wearing, work -uniform compliance, smoke, and open flame. For these categories, a progressive prompt scheme was developed in four stages—a keyword baseline, domain-specific definitions, chainof-thought reasoning, and multimodal exemplars—with each stage supplying the model with more task knowledge than the previous one. A domain-specific evaluation set of 1 068 alarm images was collected from the production site of a thermal power plant, and three domestic VLMs were compared on this set so that the influence of both the 2 prompt design and the model could be assessed. 

[Results] Under the optimal prompt configuration, review accuracy increased from a baseline of 46.73% to 81.66%, the F 1 value from 41.54% to 86.36%, and the alarm -filtering efficiency (AFE) from 0.242 to 0.631. At the operator terminal, the total alarm volume requiring handling decreased by 33.6%, and the terminal-side false-alarm ratio fell from 31.9% to 12.5%. The progressive prompt design accounts for most of this improvement, as accuracy rose stage by stag e from the keyword baseline, while the three models behaved consistently under the same prompts. 

[Conclusion] Overall, pairing a high-recall detector with VLM-based semantic verification effectively reduces false alarms in thermal power plant surveillance systems while preserving detection coverage; because it relies only on domestically available models and a modest amount of on -site data, the method offers a practical reference for similar industrial applications.