基于无人机图像与迁移学习的线路绝缘子状态评价方法

Insulator state evaluation method based on UAV image and migration learning

针对目前绝缘子运维过程存在着规程过于繁杂,过于依赖运维人员的人工识别等问题,文中提出了一种绝缘子状态评价方法,该方法采用历史绝缘子缺陷图像作为训练样本,通过迁移学习在小样本数据处理的优异性能实现基于深度卷积神经网络绝缘子的缺陷识别模型训练,并借助卷积神经网络的特征提取能力实现绝缘子缺陷量化评分,结合历史样本与专家经验实现考虑运行年限、外界环境等因素实现绝缘子综合状态评价。通过实例分析表明文中迁移学习模型训练后绝缘子缺陷识别准确率可达到90%以上,而采用全新学习在同样的样本条件下识别准确率仅为70%,且文中建立的评价模型在日常运维中能够更为灵敏地体现绝缘子的缺陷状态,说明文中评价方法具有相当可靠性,可为运维人员的日常维护安排提供经验。

In view of the problems existing in the process of insulator operation and maintenance,such as too complicated regulations and too dependent on manual identification of operators,this paper presents an insulator condition evaluation method,which uses historical insulator defect images as training samples and realized the basis of excellent performance of small sample data processing through migration learning.Based on the training of defect recognition model of deep convolution neural network and the feature extraction ability of convolution neural network,the quantification score of insulator defect can be achieved,and the comprchensive state evaluation of insulator can be realized by considering the operation life and external environment with the help of historical samples and expert expericnce.An example shows that the recognition accuracy of the proposed transfer learning model can reach more than 90% after training,while the recognition accuracy of the new learning model is only 70% under the same sample conditions,and the evaluation model established in this paper can more sensitively reflect the defect status of insulators in daily operation and maintenance.It shows that the evaluation method established in this paper is quite reliable and can provide experience for the daily maintenance arrangement of operation and maintenance personnel.