摘要
由于背景复杂、目标所占像素比例较小,掩膜区域卷积神经网络(Mask R-CNN)模型对输电线路绝缘子缺陷检测能力不足,该文提出一种改进的MaskR-CNN模型。具体地,首先,在特征提取网络中引入卷积注意力模块(CBAM),分别从通道和空间提升小目标特征保持性;其次,使用全局交并比(GIoU)计算目标间的相似度,提升定位准确性;最后,使用Tversky损失计算掩膜分支的损失,以提升不平衡样本下的检测效果。使用某输电运检中心无人机巡检作业所得具有自爆缺陷的绝缘子照片作为数据集对该模型进行验证,实验结果表明,与原始Mask R-CNN模型相比,该方法的平均精确率AP50:90、AP50和AP75分别提升至0.56、0.79和0.72;与三种经典目标检测算法相比,该算法具有较高的检测精度,模型的分割性能有一定提升,且比原始模型具有更好的鲁棒性,可以满足电力巡检中准确性和快速性的要求。
Transmission lines occupy a relatively large proportion in the power system, in order to ensure the safe and stable operation of the power system, it is necessary to regularly inspect the transmission lines, among them, insulators in the transmission line play the role of insulation and support, due to the long-term hanging and working outdoors, prone to self-explosion defects, resulting in short-circuit faults on the line, and even large-scale power outages. With the development of artificial intelligence, the use of unmanned aerial vehicles(UAV) for line inspection, and then based on deep learning target detection methods for insulator defect detection has become an intelligent inspection method with great development potential. Due to the different shooting angles of the UAV,the insulators of the lines obtained by the inspection are different, and the environment in which the insulators are located is different, which results in the occlusion phenomenon of some insulators, to overcome these problems,this paper proposes to make relevant improvements on the basis of the two-stage target detection algorithm Mask region-convolutional neural network(Mask R-CNN), to ensure the detection speed and improve the detection rate of defective insulators by the algorithm.For the insulator defects belonging to the category of small target detection, the convolutional block attention module(CBAM) attention mechanism is introduced in the backbone feature extraction network, so that the network can focus on the defect contour and obtain more interesting high-semantic information in the process of extracting the semantic information of the defect part. Then, in order to improve the limitations of semantic information still possessed by the feature layer, the parallel "bottom-up" path and feature fusion module are added to the original feature fusion network to promote the flow of information and global feature fusion. Once more, with the help of Generalized Intersection over Union(GIoU) to accurately characterize the distance between targets, the positioning performance of the model can be effectively improved when the targets overlap. Conclusively, the part of the original algorithmic loss function is replaced with the Tversky Loss function to alleviate the effect of sample imbalance on model training.Based on the defective insulator dataset obtained by the UAV operation class of a transmission and transportation inspection center of the State Grid, the training of the improved network is carried out, and the model training effect is obtained from the convergence of the loss curve, and the generalization ability of the model is also improved. By using the improved model for defect detection and comparing the visual positioning results,the proposed algorithm avoids the interference of the high likelihood structure around the target to a certain extent,which realizes the effective detection of the insulator defect part, and improves the impact of small targets and sample unevenness on the detection. Compared with the original algorithm, the AP50:95 of the proposed algorithm is increased to 0.56, AP50 to 0.79, and AP75 to 0.72. Finally, the performance of the algorithm is comprehensively compared, and the P-R curves of the improved before and after models under the conditions of AP50 and AP75 are compared, and it can be obtained that the corresponding curves of the improved algorithm are on the outside of the original algorithm curve, which shows the effectiveness of the proposed method, and the performance is better than that of the original algorithm.
作者
苟军年
杜愫愫
刘力
Gou Junnian;Du Susu;Liu Li(School of Automation and Electrical Engineering Lanzhou Jiaotong University,Lanzhou 730070 China)
出处
《电工技术学报》
EI
CSCD
北大核心
2023年第1期47-59,共13页
Transactions of China Electrotechnical Society
基金
国家自然科学基金资助项目(72171106,61863023)。
关键词
绝缘子缺陷检测
掩膜区域卷积神经网络
卷积注意力模块
特征融合
全局交并比
Tversky损失
Insulator defect detection
mask region-convolutional neural network(Mask R-CNN)
convolutional block attention module(CBAM)
feature fusion
generalized intersection over union(GIoU)
Tversky loss
