基于改进Faster R-CNN的变电站设备外部缺陷检测

PDF下载HTML

Title:

External defect detection of transformer substation equipment based on improved Faster R-CNN

作者:

张铭泉, 邢福德, 刘冬

华北电力大学 计算机系, 河北 保定 071003

Author(s):

ZHANG Mingquan, XING Fude, LIU Dong

Computer Department, North China Electric Power University, Baoding 071003, China

关键词:

变电站设备外部缺陷; 深度学习; 目标检测; 卷积神经网络; Faster R-CNN; 特征提取; 特征融合金字塔结构; 损失函数

Keywords:

external defects of transformer substation equipment; deep learning; object detection; convolutional neural network; Faster R-CNN; feature extraction; feature fusion pyra-mid structure; loss function

分类号:

TP391.4

DOI:

10.11992/tis.202207016

文献标志码:

2023-11-14

摘要:

针对变电站设备外部缺陷目标检测任务中目标形状多样，周围环境复杂，当前代表性算法识别准确度低，错检漏检严重的问题，对比了众多目标检测算法在变电站设备缺陷数据集上的检测结果，检测精度较高的是添加了特征融合金字塔结构的Faster R-CNN（faster region-based convolutional network）算法，但其对小目标物体和设备渗漏油的检测精度仍有提升空间，为此设计一种基于Faster R-CNN的改进算法。改进算法通过对输入图像进行数据增强，在网络中添加SPP（spatial pyramid pooling）结构以及改进特征融合方式，对分类以及边界框回归损失函数进行改进的方式来提高缺陷的检测精度。与原Faster R-CNN算法进行对比，改进算法在变电站设备缺陷目标检测数据集的检测结果中AP（average precision）（0.5∶0.95）提高了2.7个百分点，AP(0.5)提高了4.3个百分点，对小目标物体的检测精度也提高了1.8个百分点，试验结果验证了该方法的有效性。

Abstract:

There are challenges in object detection on external defects of transformer substation equipment, such as various target shapes, complex surrounding environment, low recognition accuracy of current representative algorithms, and severe false or missed detection. By comparing the detection results of different object detection algorithms on the transformer substation equipment defect data set, it is revealed that the faster R-CNN algorithm with the feature fusion pyramid structure has higher detection accuracy. However, there are still opportunities to improve the detection accuracy of small target objects and equipment leakage. Thus, in this study, an enhanced, faster R-CNN-based algorithm is developed. It improves the detection accuracy of defects by enhancing the input image data, adding the spatial pyramid pooling structure to the network to improve the feature fusion method, and thereby boosting the classification and bounding box regression loss function. Compared with the original faster R-CNN, the experimental findings demonstrate that the improved algorithm has increased AP (0.5 : 0.95) (average precision) by 2.7% and AP (0.5) by 4.3% in the detection results of the transformer substation equipment with respect to the defect object detection data set and the detection accuracy of small target objects has also been improved by 1.8%. This work confirms the effectiveness of the method proposed here.