This paper presents a method based on convolutional neural network (CNN) for classifying workmanship defects located in 36 kV cross-linked polyethylene (XLPE) cable terminations. The main contributions of the study are to differentiate the poor workmanship defects without a hand-crafted feature extraction process, and to propose a new input type for the partial discharge (PD) recognition algorithm. Experiments are carried out on two different datasets, each of which has five typical cable termination defects. A database comprised of 1200 phase-resolved partial discharge (PRPD) defect patterns are generated, and each PRPD fingerprint recorded for 30 s is converted into an RGB image for inputting them to the CNN. Three case studies are created to increase the robustness of the algorithm by using the two datasets. The algorithm hyperparameters are optimized to improve the performance of CNN. Finally, the proposed method is compared with the state-of-art CNN algorithms used in the literature. The results show that the proposed method is viable for determining the types of potential defects in the cable terminations.

本文提出了一种基于卷积神经网络（CNN）的方法，用于对位于36 kV交联聚乙烯（XLPE）电缆终端中的工艺缺陷进行分类。这项研究的主要贡献在于，无需手工制作的特征提取过程即可分辨出不良的工艺缺陷，并为局部放电（PD）识别算法提出一种新的输入类型。在两个不同的数据集上进行了实验，每个数据集都有五个典型的电缆端接缺陷。生成由1200个相分辨局部放电（PRPD）缺陷图案组成的数据库，并将记录30 s的每个PRPD指纹转换为RGB图像，以将其输入到CNN中。通过使用两个数据集，创建了三个案例研究以提高算法的鲁棒性。优化算法超参数以提高CNN的性能。最后，将提出的方法与文献中使用的最新CNN算法进行比较。结果表明，所提出的方法对于确定电缆终端中潜在缺陷的类型是可行的。