数字化智能配电站房边缘图像检测技术

Edge image detection technology for digital intelligent power distribution station

城市中配电站房分布广、基数大,为了便于对大规模的配电站房进行实时监控并避免配电站房的信息泄露,提出了一种基于YOLOv5s模型的图像识别边缘计算方法。YOLOv5s模型被简化以移植到微控制器单元。改进模型采用ShuffleNetV2网络替换原模型的CSPNet骨干网络;去除Focus层,避免多次切片操作;摘除ShuffleNetV2骨干网络的1024卷积和7×7池化层;对YOLOv5的颈部网络进行了剪枝操作。通过实验验证,在不同场景下,所提出的图像识别方法的网络参数约为95 ms/帧(优于YOLOv5s的480 ms/帧),能够有效识别和准确定位火情,探测精度可达95.5%。YOLOv5计算所得的结果将被传输至边缘节点进行整合并发送至云平台进行处理。

The power distribution stations in cities is widely distributed and has a large base.In order to conveniently monitor large-scale power distribution station in real-time and avoid information leakage,an image recognition edge computing method based on YOLOv5s model is proposed.The YOLOv5s model is simplified to be ported to microcontroller units(MCU).The improved model uses ShuffleNetV2 network to replace the CSPNet backbone network of the original model.The Focus layer is removed to avoid multiple slicing operations.1024 convolution and 7×7 pooling layers of ShuffleNetV2 backbone network is removed.The neck network of YOLOv5 is pruned.Through experimental verification,the network parameters of the proposed image recognition method in different scenarios are about 95 ms/frame(better than 480 ms/frame of YOLOv5s),which can effectively identify and accurately locate fires,with a detecting precision of 95.5%.The results calculated by YOLOv5 will be transmitted to edge nodes for integration and sent to the cloud platform for processing.