Identification of Cable Partial Discharge Defects Based on a Dual-Attention Mechanism Optimized CNN Model

Partial discharge (PD) is an early indicator of insulation degradation in electrical equipment, and frequent occurrences may lead to severe faults. Due to noise interference, traditional PD detection methods struggle to achieve high accuracy in complex environments. To address the challenge of accurately identifying PD, a novel detection method based on a dual-attention mechanism that integrates Convolutional Neural Networks (CNN) and feature map fusion is proposed. This method is designed to identify typical PD types associated with cable joint defects. First, a PD feature matrix map is constructed, integrating and visualizing features in both time and spatial domains, allowing the model to capture multi-dimensional information of the PD signals comprehensively. Then, a CNN model with a time attention (TA) and Spatial attention(SA) mechanism is developed, effectively enhancing the model's ability to perceive key features and improving detection accuracy and robustness. Experimental results show that the proposed model achieves an overall accuracy of $98.89 \%$ in PD defect recognition. Ablation experiments confirm the effectiveness of the dual-attention mechanism, as removing the time attention mechanism reduces the identification accuracy of surface discharge and air gap discharge to $97.09 \%$ and $91.28 \%$, respectively. Additionally, a comparison with ResNet, DenseNet, and Inception-v3 models further validates the improvement of the proposed model.