基于卷积神经网络自动提取特征的电缆绝缘缺陷识别

Recognition of Cable Insulation Defects Based on Automatic Feature Extraction of Convolutional Neural Network

针对电力电缆局部放电信号人工提取特征严重依赖专业经验,易受主观不确定性影响的问题,提出了一种基于卷积神经网络自动提取特征的电缆局放缺陷识别方法。首先采用随机裁剪方法扩充原始样本数据,在此基础上利用滑动时间窗生成局放信号二维图像信息作为网络模型输入。详细研究了卷积层数、池化方式和激活函数等因素对网络识别性能的影响,生成并优化网络结构。方法能够自动提取电缆局部放电二维图像样本的深层特征,在识别准确率和鲁棒性方面效果突出。试验数据表明,系统对4种典型局放缺陷的总体识别率达到了96%,相比于支持向量机和反向传播神经网络等经典方法,分别提高了3.2%和6.0%,具有良好的应用前景。

Aiming at the problem that the manual extraction of features of partial discharge signals of power cables relies heavily on professional experience and it is susceptible to subjective uncertainty,a cable partial discharge defect identification method based on the automatic extraction of features of the convolutional neural network was proposed.The overall idea is as follows:Firstly,the original sample data is expanded by a random cropping method,and on this basis,a sliding time window is used to generate the two-dimensional image information of the partial discharge signal as the input of the network model.The influence of factors such as the number of convolutional layers,pooling method,activation function,etc.on network recognition performance is studied in detail,and the network structure is generated and optimized.This method can automatically extract the deep features of two-dimensional cable partial discharge image samples,and has outstanding effects in recognition accuracy and robustness.The experimental data shows that the system’s overall recognition rate of the four typical partial discharge defects has reached 96%,which is an increase of 3.2%and6.0%respectively compared with classical methods such as support vector machine and back propagation neural network,therefore,it is of good application prospect.