基于Faster R-CNN算法的变电站设备识别与缺陷检测技术研究

Research on substation equipment identification and defect detection technology based on Faster R-CNN algorithm

变电站作为电力运输的中转站,是城市运转、人民生活的重要基础设施。变电站在运行过程中,经常发生因位置偏僻,不支持机器人或无人机直接进行探测而造成的设备运作温度检测不及时的问题。传统的变电站设备缺陷识别算法是基于机器的学习算法,精确度较低,只适合单个设备类别的缺陷检测,易受环境的影响。基于此,文中提出一种识别变电站设备红外缺陷的方法。首先,基于Faster R-CNN算法的设备识别,对6种类型的变电站设备包括套管、绝缘体、电线、电压互感器、避雷针和断路器进行目标识别,以实现设备的精确定位;然后,基于稀疏表示分类(SRC)的算法获得输入样本的实际标签;最后,基于温度阈值判别式算法,在设备区域中识别设备温度的异常缺陷。文中的方法实现了在红外线图像下的设备识别和缺陷检测,运用文中设计的方法对6类设备的红外图像进行检测,准确率达到91.58%,不同类型设备缺陷的平均识别准确率为91.62%,整体缺陷图像的识别准确率达到87.62%。实验结果表明了该方法的有效性和准确性。

As a transit station for power transportation,substations are an important infrastructure for city operation and life of people.During the operation of the substation,the problem of untimely detection of the temperature of the equipment operation due to the remote location,which does not support direct detection by robots or drones,often occurs.Traditional defect recognition algorithms for substation equipment are based on machine learning algorithms,which have low accuracy,only suitable for defect detection of individual equipment categories,as well as susceptible to environmental influences.On this basis,a method to recognize infrared defects of substation equipment is proposed in this paper.Firstly,equipment identification based on Faster R-CNN algorithm is used to identify the target of six types of substation equipment including bushings,insulators,wires,voltage transformers,lightning rods,and circuit breakers so as to realize the precise location of the equipment;then,an algorithm based on sparse representation classification(SRC)is used to obtain the actual labels of the input samples;finally,the region of equipment is used to identifies the abnormal defects of the device temperature based on the temperature threshold discriminative algorithm.The method in this paper realizes equipment recognition and defect detection under infrared images,and the accuracy of detecting infrared images of six types of equipment using the method designed in this paper reaches 91.58%,and the average recognition accuracy of defects of different types of equipment is 91.62%,and the recognition accuracy of the overall defect image reaches 87.62%.The experimental results demonstrate the effectiveness and accuracy of the proposed method.