多任务元学习网络的气体绝缘组合电器局部放电同时诊断与定位

Research on Gas-Insulated Switchgear Partial Discharge Diagnosis andLocation Method Based on Multi-Task Meta-Learning

针对气体绝缘组合电器(GIS)局部放电诊断和定位作为两个独立的任务来完成而忽略了两个任务间的联系、且难以部署和泛化到现场小样本场景下的问题,提出了多任务元学习网络以同时实现现场小样本场景下的GIS局部放电诊断与定位。首先,为了充分挖掘两个任务间的关联关系并保留每一任务的差异特征,构建了多任务网络,在特定任务层引入注意力机制,为每个任务从浅到深选择重要特征,保证每一任务特征差异化提取的质量。其次,为了将所开发的多任务网络部署应用到现场小样本场景下,采用元训练的方法对模型进行训练。再次,在元测试阶段使用目标任务的少量数据进行微调,实现了小样本下GIS局部放电诊断和定位。最后,在现场样本上对模型性能进行了验证。实验结果表明,提出的多任务元学习网络在GIS局部放电诊断上的准确率达到94.53%,定位平均误差和均方根误差分别为10.78 cm和12.97 cm,与单任务网络和其他模型相比性能提升明显,为GIS局部放电诊断和定位提供了新颖的解决思路。

To address the issue where gas-insulated switchgear(GIS)partial discharge diagnosis and localization are treated as separate tasks without considering their connection and in response to challenging deployment and generalization to on-site small sample scenarios,this paper introduces a multi-task meta-learning network to simultaneously achieve GIS partial discharge diagnosis and localization in on-site small sample scenarios.Initially,a multi-task network is developed to fully explore the correlation between the two tasks while preserving their differential characteristics.An attention mechanism is introduced at specific task layers to select important features from shallow to deep levels for each task,ensuring the quality of differentiated extraction of features for each task.Subsequently,a meta-training method is adopted for model training to deploy the developed multi-task network in on-site small sample scenarios.Fine-tuning is performed using a small amount of data from the target task in the meta-testing stage,and GIS partial discharge diagnosis and location with small samples are enabled.Finally,the model performance is verified using field samples.Experimental results demonstrate that the multi-task meta-learning network proposed in this paper shows an accuracy of 94.53%in GIS partial discharge diagnosis,with an average error in the location of 10.78 cm and a root mean square error of 12.97 cm.Exhibiting superior performance relative to the single-task network and other models,the proposed multi-task meta-learning network presents a novel solution for GIS partial discharge diagnosis and location.