GDW-YOLO-based internal discharge spectral waveform recognition in GIS

GIS (Gas-insulated switchgear) is a critical component of modern power grids, and accurate awareness of its internal insulation condition is essential to power system safety. PD (Partial Discharge) is a key indicator of insulation degradation, and the ultraviolet spectra it emits contain rich diagnostic information about incipient faults. However, ultraviolet spectral waveforms are characterized by highly variable feature scales, complex background noise, and diverse morphological patterns, all of which make automated identification particularly challenging. To address this issue, this paper proposes an intelligent ultraviolet spectral waveform recognition algorithm based on an improved YOLO(You Only Look Once)v8 framework. First, to overcome the difficulty of jointly capturing multi-scale characteristics such as narrow-band spikes and broadband envelopes in spectral waveforms, a multi-scale dilated attention (MSDA) module is introduced at the end of the backbone network. By combining parallel dilated convolutions with different dilation rates and an efficient coordinate attention mechanism, the proposed module establishes a feature enhancement strategy that can simultaneously emphasize local details and global contextual information. Second, to better suit edge-computing scenarios, a lightweight heterogeneous cross-stage network CSPHet (Cross Stage Partial with 2 convolutions and Fast) is designed to replace the original C2f (Cross Stage Partial with 2 convolutions and Fast) module. This structure incorporates depthwise separable convolutions and heterogeneous branch design, significantly reducing the number of parameters and computational complexity while preserving efficient gradient flow. Finally, to improve localization accuracy under ambiguous waveform boundaries, the loss function is refined by adopting WIoU (Wise Intersection over Union) v3 as the bounding-box regression loss. In this study, an ultraviolet spectral dataset was constructed using a self-developed GIS defect simulation platform. After data augmentation, the dataset expanded to 5,142 samples. Experimental results on this dataset show that the proposed model achieves an mAP(mean Average Precision)@0.5 of 94.7%, while its parameter count and computational cost are only 78% and 75% of those of the original YOLOv8n model, respectively. In addition, the inference speed satisfies real-time requirements. These results demonstrate that the proposed method provides an efficient and reliable solution for online monitoring and early fault diagnosis of GIS equipment.