基于余弦相似度的 GIS 放电谱图识别精度提高方法

The study of effective feature extraction and optimization algorithms is crucial for improving the accuracy of partial discharge（PD）detection in gas-insulated switchgear（GIS）. To enhance the precision and reliability of GIS partial discharge detection by utilizing the phase information in discharge spectra，this paper proposes a neural network-based PD pattern recognition method that incorporates cosine similarity. By analyzing the spectra of various discharge types，phase features are summarized，and the features of each discharge type are compared with the spectrum of the discharge under evaluation. The resulting phase reference values are incorporated into the network structure for training. The research findings demonstrate that introducing cosine similarity calculations enhances the neural network’s accuracy by 3.9%，and that the improvement is closely related to the spectral phase features—greater phase feature clarity leads to more substantial accuracy gains from the cosine similarity module. This method significantly improves PD fault recognition accuracy，equipping GIS with enhanced warning and fault evaluation capabilities.