当前针对数字式仪表检测算法在边缘设备具有实时性差、泛化性差的问题,对此提出一种采用ID-YOLO(instrument detection-you only look once)模型的变电站数字仪表检测识别方法.所提算法以YOLOv5模型为基础,首先设计轻量骨干网络(light weight-YOLO,LW-YOLO)提取图像特征,降低网络参数,提高检测实时性;然后设计了一种双级路由注意力模块(bi-lev-el routing attention moudle,BRAM),提高网络对小数点的检测精度以及网络的鲁棒性和泛化性;最后,引入损失函数α-IoU,通过设定不同的可调节参数α数值得到更准确的真实框与预测框的交并比计算,可以提高模型的检测精度.结果表明:相比于其他基于深度学习的数字仪表检测识别方法,所提方法在不同显示方式的数字仪表识别任务上具有更好的准确性和泛化性,而且可以在检测准确率领先的情况下,将模型在边缘设备上的检测速度从6.87帧/s提升至8.77帧/s,其实时性和检测精度均能够满足实际变电站智能数据采集、检测识别的工程需要.
Digital Instrument Detection Method Based on ID-YOLO
Aiming at the problem that the digital instrument detection algorithm has poor real-time performance and poor generaliza-tion in edge equipment,an instrument detection-you only look once(ID-YOLO)model for substation digital instrument detection and rec-ognition was proposed.The proposed algorithm was based on YOLOv5 model.Firstly,a lightweight light weight-YOLO(LW-YOLO)backbone network was designed to extract image features,reduce network parameters,and improve real-time detection.Then,a bi-level routing attention module(BRAM)was designed to improve the precision of decimal point detection and the robustness and generalization of the network.Finally,the loss function α-IoU was introduced,the detection accuracy of the model was improved by setting different adjustable parameter α to obtain more accurate calculation of the intersection ratio between the real frame and the predicted frame.The results show that compared with other digital instrument detection and recognition methods based on deep learning,the proposed method has better accuracy and generalization on digital instrument recognition tasks with different display modes,and the detection speed of the model on the edge de-vice can be improved from 6.87 frames/s to 8.77 frames/s while the detection accuracy is leading.Its real-time performance and detection accuracy can meet the engineering needs of intelligent data acquisition,detection and identification of actual substation.
digital instrumentyou only look once(YOLO)edge devicetarget detectionlight weight
NETL
NSTL
万方数据
