基于改进胶囊网络的电力线巡线异物检测

针对电力线巡线异物检测使用的传统卷积神经网络空间辨识度较差、训练样本需求过多的问题,提出一种改进胶囊网 络模型。 使用数据灰度化、三维块匹配滤波算法预处理巡线数据集。 提出自适应贡献池化降低数据信息丢失量,异物数据深度 信息提取单元提取主要特征来滤除冗余信息、减少数据数量以改善模型性能,改进异物识别主胶囊层和动态路由结构以适应电 力线巡线异物检测的二分类情况。 对自适应贡献池化和最大池化,无池化、传统结构胶囊网络和改进胶囊网络,改进胶囊网络 和 AlexNet、GoogLeNet 分别进行异物识别对比实验和改进胶囊网络的空间辨识度性能进行测试实验。 实验结果表明,在 3 700 张小训练样本条件下,经 20 次训练后,自适应贡献池化比最大池化的改进胶囊网络平均准确率提高 2. 7%,改进胶囊网络比无 池化、传统结构胶囊网络平均准确率提高 3. 6%,改进胶囊网络比 AlexNet、GoogLeNet 的平均准确率分别提高 21. 9%和 12. 6%, 且改进胶囊网络在大小、角度不同的测试数据中仍具有高于 91%的平均准确率。 改进胶囊网络在空间辨识度复杂、少训练样本 情况下仍具有较高的异物识别能力,实现了高效率、高准确率的自动化无人巡线异物检测。

Detection of foreign objects in power line patrol based on improved capsule network

Aiming at the problems of poor spatial recognition and excessive training sample requirements of traditional convolutional neural networks used in power line patrol foreign object detection, an improved capsule network model is proposed. The gray-scale data and three-dimensional block matching filter algorithm are used to preprocess the line survey data set. The adaptive contribution pooling is proposed to reduce the amount of data information loss. The foreign object data depth information extraction unit extracts the main features to filter out redundant information, reduce the number of data to improve model performance, improve the foreign object recognition of main capsule layer and dynamic routing structure to adapt to power line patrol for the second classification of line foreign object detection. For adaptive contribution pooling and maximum pooling, non-pooling, traditional structure capsule network and improved capsule network, improved capsule network and AlexNet, GoogLeNet were respectively compared with foreign object recognition experiment and improved capsule network spatial recognition performance. The experimental results show that under 3 700 small training samples, after 20 trainings, the average accuracy of the improved contribution network of adaptive contribution pooling is greater than the maximum pooling by 2. 7%, and the improved capsule network is better than the non-pooling, traditional structure capsules. The average accuracy of the network is increased by 3. 6%, and the improved accuracy of the improved capsule network is 21. 9% and 12. 6% higher than that of AlexNet and GoogLeNet, respectively, and the improved capsule network still has an average accuracy higher than 91% in test data of different sizes and angles. The improved capsule network has high foreign object recognition ability under the condition of complicated spatial recognition and few training samples, and realizes high-efficiency and high-accuracy automatic unmanned line inspection foreign object detection.