基于卷积神经网络的直流XLPE电缆局部放电模式识别技术

Pattern Recognition of Partial Discharges in DC XLPE Cables Based on Convolutional Neural Network

针对现有的直流交联聚乙烯(XLPE)电缆局部放电模式识别对强随机性信号的特征提取缺乏一定自适应能力的问题,该文提出基于卷积神经网络(CNN)的模式识别算法,采用卷积神经网络框架CAFFE进行网络训练和识别检测。首先采集四种典型绝缘缺陷电缆的局部放电信号作为样本,再利用自适应的卷积核进行特征提取,池化层进行特征映射,非线性多分类器进行回归分类,最终得到训练完成的CAFFE网络。通过设置不同求解器参数、网络结构和训练样本数量对缺陷识别结果进行对比分析,发现利用改进的Alexnet网络,采用衰减学习率方式的模式识别框架的平均识别正确率最高,达到了91.32%,相比于传统模式识别算法至少提高了8.97%。该方法具有强大的自适应学习能力,为应用于直流电缆故障诊断的模式识别提供了新的思路。

Present partial discharge(PD) pattern recognition of DC cross-linked polyethylene(XLPE) cables has some limitations on the feature extraction of strong random signals. In order to solve this problem, this paper proposes a self-adaptive pattern recognition based on convolutional neural network(CNN). Convolutional architecture for fast feature embedding(CAFFE) was used to train the CNN. First, PD signals of four typical insulation defects were collected as the input samples of C AFFE. Then, the training cycles were iterated by taking self-adaptive convolution kernels to extract features, pooling layers to map features, nonlinear multi-classifiers to classify different types, until the CAFFE network was completely trained. After comparison of different parameters of solver, network structures and numbers of training samples, it is found that pattern recognition framework using the modified Alexnet network and attenuation learning rate method has the highest accuracy of 91.32%. Moreover, it has at least 8.97% improvement compared with traditional methods. The powerfu l self-adaptive learning capabilities of the new method provide a new idea for pattern recognition of DC cable fault diagnosis.