The existing multi-source heterogeneous fire detection faced with issues of high privacy leak risk for data, low small-flame detection accuracy, and high communication overhead, so this paper proposed a federated fire detection model based on integrating YOLO11n-P2 detection model and Fire-Yolopa detection with partial aggregation. This framework is also the first to use YOLO11 series and federated learning together for fire detection tasks. Using the federated learning will allow for each client to participate in cooperation and share their knowledge with each other without having to share their real data, solving the issue of closed data silo in fire detection. Through experimental results, it can be seen that the optimal client model has an mAP50 value of 80.2%，only different from the centralized baseline model by 2.68%, and the small flame recall rate improves from 50.00% to 54.64% This result shows the improvements for Small flame detection in the terms of accuracy as well as the overall detected. This work will give a reusable technical framework which can help the community to collaboratively improve on small objects and privacy as well as deployment edge.

In large-scale power dispatching cloud environments, massive traffic dispatching suffers from low efficiency and insufficient observability. This paper proposes a seed-end-cloud collaborative intelligent dispatching and observability optimization scheme. A three-tier pipeline architecture based on a lightweight agent, a stateless processing cluster, and distributed storage is designed. Source traffic shaping is achieved through real-time data aggregation at the agent end and Zstandard compression technology. A server-led dynamic rebalancing mechanism and CNI network plugin traffic shaping strategy are used to address instantaneous traffic surges and uneven load distribution. Storage layer performance is optimized through a self-developed high-availability connection pool and a dual-trigger batch write mechanism. In a simulated 10,000-node power dispatching scenario, the system throughput reaches 503.7 thousand records per second, with end-to-end latency controlled within 115.6 ms, providing feasible support for high-concurrency, high-reliability power services.