基于改进的Wasserstein生成对抗网络和深度残差网络的电缆中间接头局部放电缺陷识别

随着“碳达峰、碳中和”目标的明确,我国已有较多风光电场使用铝电缆作为直埋输电电缆,来实现成本的降低和清洁能源的利用。由于铝电缆接头PRPD谱图样本数量有限,导致了训练模式识别的网络识别准确率低、泛化能力差等问题,本文通过设计三种电缆接头典型缺陷,搭建局部放电实验平台,运用改进的Wasserstein生成对抗网络训练样本数据,借以生成更多新的图像数据,进而将生成样本和原始样本同时投入深度残差网络训练,该方法识别准确率达97.82%。与数据扩充前后不同层数深度残差网络和普通卷积神经网络训练的准确率进行比较,证明了该方法能够有效提升基于小样本条件下的识别准确率,对实际工程具有一定指导意义。

Based on Improved Wasserstein Generative Adversarial Networks and Deep Residual Networks Partial Discharge Pattern Recognition of Cable Joints

With the clarification of the goal of "carbon peaking and carbon neutrality", many wind and solar farms in my country have used aluminum cables as direct buried transmission cables to reduce costs and utilize clean energy. Due to the limited number of PRPD spectrogram samples of aluminum cable joints, the network recognition accuracy of training pattern recognition is low, and the generalization ability is poor. In this paper, three typical defects of cable joints are designed to build a partial discharge experimental platform. The adversarial network training sample data is generated to generate more new image data, and then the generated samples and the original samples are put into the deep residual network training at the same time. The recognition accuracy of this method is 97.82%. Compared with the training accuracy of different layers of deep residual network and ordinary convolutional neural network before and after data expansion, it is proved that this method can effectively improve the recognition accuracy under the condition of small samples, which has certain guiding significance for practical engineering.