Spatio-temporal early warning of rockbursts based on modal fusion

Partial discharge (PD) diagnosis is a critical monitoring method for insulation degradation in gas insulated switchgear (GIS), serving as an early warning of potential faults that enhances the safety and reliability of power systems. In recent years, machine learning has been increasingly applied to PD pattern recognition, significantly improving diagnostic accuracy. However, current methods predominantly rely on unimodal data analysis, either phase-resolved partial discharge (PRPD) spectra or time-domain waveforms, thereby acquiring incomplete information from single modal feature extraction. To address this limitation, this paper proposed a novel cross-modal spatiotemporal fusion architecture called Time-Phase Cross-Attention Network (TPCAN), which discovers feature correspondence between heterogeneous discharge signatures through multi-head attention mechanisms. The architecture integrates DenseNet for hierarchical PRPD spectral feature extraction and a hybrid 1DCNN-Graph Convolutional Network cascade that concurrently captures local waveform transients and global temporal features, then applying cross-attention fusion module enabling dynamic spatial-temporal feature alignment through learnable attention matrices. Comprehensive evaluations on GIS partial discharge data demonstrate TPCAN's superior diagnostic capability, achieving statistically accuracy improvements of 2.37%, 5.01%, 10.27%, and 13.91% compared to CoAtNet, ResNet, 1DCNN, and LSTM, respectively. Further ablation studies found the cross-attention module's ability in enhancing inter-modal PD discriminability and stabilizing model convergence. Parametric analysis of attention head reveals the balance between feature resolution and model complexity that insufficient heads induce incomplete modality interaction, while excessive heads lead to feature over-segmentation. This work advances PD pattern recognition by establishing a new method for multimodal discharge feature fusion, having direct implications for intelligent condition monitoring and fault diagnosis in power equipment.