Abstract:

In response to the dependency of the traditional safety supervision of power plants on manual work and difficulty in tracing back afterwards, a violation identification and control system for power plants was constructed based on large model and edge computing. The system was applied to decouple 31 typical operation scenarios of power plants, develop 55 AI algorithm models, and realize the real-time identification and early warning of violations, equipment abnormalities, and environmental risks. A dynamic task scheduling algorithm based on directed acyclic graph (DAG) was designed to solve the resource allocation problem of edge computing under multi-task concurrence. The AI monitoring mechanism based on dynamic weight was constructed to realize the self-adaptive rotation inspection of the monitoring screen according to the risk level and active focus of high-risk scenes. The pilot results show that the average accuracy of typical violation identification (mAP) @0.5 is 95.7%, 12.3% higher than that of YOLOv8. The alarm delay is less than 200 ms, and the response time of on-site intervention is shortened from 15 min to 4.2 min, effectively improving the intelligent level of power plant security management.

Key words: large model, edge computing, violations, security management, artificial intelligence (AI) monitoring