基于Faster R-CNN算法的变电站 设备识别与缺陷检测技术研究

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，经常发生由于位置偏僻，不支持机器人和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学习算法，精确度低，只适合单个设备类别的缺陷检测，易受到环境影响。基于此，文中出一种改进的识别变电站设备红外缺陷方法。首先，基于Faster R-CNN的目标设备检测，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检测，以实现设备的精确定位；然后，基于稀疏表示分类（SRC）来识别不同的类，因此可以获得输入样本的实际标签；最后，基于温度阈值判别式算法，在设备区域中识别温度异常缺陷。文中算法实现了在红外线图像下的设备识别和检测，使用文中算法对6类设备的红外图像进行检测，准确率达到91.58%，不同类型设备的缺陷识别率为97.63%，缺陷识别准确率达到87.62%。实验结果表明该方法的有效性和准确性。

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

As a transit station for power transportation, substation is an important infrastructure for city operation and people''s life. During the operation of the substation, the problem of untimely detection of the temperature of equipment operation due to the remote location, which does not support the direct detection by robots and drones, often occurs. The traditional substation equipment defect recognition algorithm is based on machine learning algorithms, which has low accuracy, is only suitable for the defect detection of a single equipment category, and is susceptible to environmental influences. Based on this, the article comes out with an improved method for recognizing infrared defects of substation equipment. First, target equipment detection based on Faster R-CNN is performed for six types of substation equipment including bushings, insulators, wires, voltage transformers, lightning rods, and circuit breakers in order to realize the precise location of the equipment; then, different classes are identified based on Sparse Representation Classification (SRC), so that the actual labels of the input samples can be obtained; and finally, based on the Temperature Threshold Discriminant algorithm, temperature anomaly defects are recognized in the device region. The article algorithm realizes the recognition and detection of devices under infrared images, using the article algorithm to detect the infrared images of six classes of devices, the accuracy rate reaches 91.58%, the defect recognition rate of different types of devices is 97.63%, and the defect recognition accuracy rate reaches 87.62%. The experimental results show the effectiveness and accuracy of the method.