Pattern Recognition Method for GIS Insulation Defects Based on Multi-source Sensing Data Fusion and Lightweight Neural Networks

Abstract

To address the problems of incomplete single-source sensing information, complex field noise, scarcity of real defect samples, and high deployment cost of deep models in online identification of GIS insulation defects, this paper proposes a pattern recognition method for GIS insulation defects based on multisource sensing data fusion and lightweight neural networks. First, a multi-source partial discharge dataset composed of ultra-high-frequency, acoustic emission, optical, and pulse current signals is constructed, and multimodal representations such as PRPD, PRPS, waveform, and statistical features are generated. Second, a dual-stage fusion strategy is designed, in which homogeneous sensing information is adaptively weighted by multi-head attention and dominant heterogeneous modality features are selected through a Maxout mechanism. Furthermore, a lightweight multi-task neural network is developed to simultaneously implement insulation defect classification and discharge severity assessment, and MAML meta-learning is introduced to improve rapid adaptation in few-shot scenarios. Experimental results show that the proposed method achieves an accuracy of 97.36% and an F1-score of 97.01% on the regular test set, an accuracy of 84.96% in the 4-way 1-shot task, with only 2.41 M parameters and a single-sample inference time of 6.93 ms. The results verify the effectiveness of the proposed method in terms of recognition accuracy, robustness, and edge deployment. In the proposed training strategy, MAML is applied to the shared lightweight encoder and taskspecific classification/severity branches during meta-training, while only the fusion projection and task heads are fine-tuned with a few labeled field samples during edge adaptation.