基于深度学习的配电柜指针仪表示值识读研究

Research on Reading Indication Value of Pointer Meter in Power Distribution Cabinet Based on Deep Learning

摘要:

为准确读取配电柜指针式仪表的示值，保证巡检机器人作出相应决策，提出一种结合改进YOLOv5和PSPNet模型的指针仪表检测及示值识读方法。首先利用主干网络替换为轻量化网络MobileNetv3的YOLOv5算法检测定位表盘区域；然后采用特征提取网络替换为MobileNetv2的PSPNet算法对表盘的刻度线区域和指针进行分割，并通过最小二乘法圆拟合和霍夫直线检测法得到指针回转中心及指针的偏转角度；最后结合指针偏转角度和相邻主刻度线与回转中心连线的偏转角度，通过公式法求取仪表示值。实验结果表明，该算法能够准确提取配电柜上的指针仪表表盘，并对表盘中的刻度和指针进行精准分割，在误差允许的范围内指针仪表示值识读相对误差最大为6.5%，满足实际工程应用的需求。

Abstract:

To accurately read the indicated value of the pointer meter in the distribution cabinet and ensure that the patrol robot makes the corresponding decision, a pointer meter detection and indicated value reading method combining improved YOLOv5 and PSPNet (Pyramid Scene Parsing Network) model are proposed. Firstly, the YOLOv5 algorithm was used to detect and locate the dial area by replacing the backbone network with the lightweight network MobileNetv3. Secondly, the PSPNet algorithm was used to segment the dial scale area and the pointer by replacing the feature extraction network with MobileNetv2, and the center of rotation of the pointer and the deflection angle of the pointer was obtained by Least Squares and Hough Line Detection. Finally, the deflection angle of the pointer and the deflection angle of the line connecting the adjacent main scale line and the rotation center was combined to obtain the instrument representation value by the formula method. The experimental results show that the algorithm can accurately extract the dial of the pointer meter on the distribution cabinet and accurately segment the scale and the pointer in the dial, and the relative error is small within the allowed error range, which meets the requirements of practical applications.