In the process of deploying target detection models on edge computing de-vices, the contradiction between computing resource constraints and real-time requirements is increasingly prominent. To address this challenge, this paper proposes a lightweight dynamic attention model for real-time fault warning. This model is based on the YOLOv11 architecture and achieves model compression and performance improvement through structural optimization and algorithm innovation. Specifically, three key technologies have been designed: 1) a dynamic sparse linear attention mechanism that focuses on key feature regions through an adaptive window strategy to reduce computational complexity; 2) The parameter free SimAM attention module can model cross channel and spatial feature dependencies without adding trainable parameters; 3) Ghost lightweight convolution reduces feature map redundancy and model parameter count through linear transformation. The experimental results show that the model has high detection accuracy and reference speed on the power fault detection dataset. In the deployment verification of embedded platforms, this method significantly re-duces task latency and energy consumption, providing a feasible solution for high-precision and low-power object detection in edge scenes.