A Multimodal Data-Based Deep Learning Model for Partial Discharge Detection in Solid- and Liquid-Insulated Electrical Equipment

This paper proposes a multimodal data-based deep learning model for partial discharge detection in solid- and liquid-insulated electrical equipment to support early fault diagnosis and electrical accident prevention. Existing approaches often rely on a single data modality, which may limit their ability to capture complementary characteristics of different partial discharge types. The proposed model independently extracts spatial features from phase-resolved partial discharge (PRPD) images and temporal features from time-series signals, and combines them through Late Fusion for final classification. Experiments were conducted using 60,000 multimodal samples acquired from ACSR-OC with solid insulation and an oil-immersed power transformer with liquid insulation, covering normal, noise, void discharge, surface discharge, and corona discharge states. The training, validation, and test sets were constructed without overlapping identical files, and experiments were repeated using five random seeds. Compared with ResNet-18, LSTM-AE, and Cross-Attention Transformer, the proposed model achieved the highest average performance, with an Accuracy of 99.4733 ± 0.3301% and a Macro F1-score of 99.4716 ± 0.3315%. These results show that combining complementary spatial and temporal information can improve partial discharge detection performance and contribute to more reliable electrical equipment condition diagnosis.