A Study on the Identification of Partial Discharges in Typical Defects of Cross-Linked Polyethylene Cables Using the CNN-SE-RAP Algorithm

To address the reliance on manual interpretation and the limited discrimination capability of conventional methods for highly similar phase-resolved partial discharge (PRPD) patterns in cross-linked polyethylene (XLPE) cables, a convolutional neural network incorporating squeeze-and-excitation and region average pooling (CNN-SE-RAP) is proposed for defect recognition. A partial discharge test platform based on the pulse current method was established, and four typical cable defect models, including void, needle, fracture, and external-damage defects, were constructed. Together with the normal operating condition, a five-class PRPD dataset was acquired at different voltage levels. Grayscale conversion, local contrast enhancement, and local binary pattern extraction were applied to enhance local texture characteristics. The SE module was employed to adaptively emphasize discriminative channel features, while RAP preserved discharge distribution information across different phase regions. Experimental results show that the proposed CNN-SE-RAP model achieved an overall accuracy of 95.25%, outperforming support vector machine (SVM), k-nearest neighbors (KNN), backpropagation neural network (BPNN), and conventional CNN models. The results demonstrate that the proposed method can effectively extract spatial distribution, local texture, and phase-dependent regional features from PRPD patterns, providing an effective approach for intelligent diagnosis of typical cable-body defects.