基于深度NCA与原型网络的MMC-HVDC故障诊断方法OA
Fault diagnosis method for MMC-HVDC based on Deep NCA and prototypical networks
针对模块化多电平换流器柔性直流输电(MMC-HVDC)系统的故障暂态过程复杂、浅层特征易混叠及在线诊断实时性要求高等问题,提出一种基于深度邻域成分分析(Deep NCA)与原型网络的故障诊断方法.首先,利用一维卷积神经网络从桥臂电流、子模块电容电压和接地电流等多通道暂态序列中自适应提取深层特征;其次,引入 Deep NCA对嵌入空间进行监督度量约束,增强类内紧凑性和类间分离性;最后,基于类别原型实现快速距离判别.仿真结果表明:所提方法在5 类运行状态识别中准确率达到95.20%,高于1D-CNN+Softmax、1D-ResNet和 1D-Transformer等典型深度模型;负荷偏移条件下的平均准确率为91.73%,单类训练样本数减少至50 组时准确率仍为86.93%,表明该方法具有一定的工况适应性和有限训练样本适应能力.
To address the complex fault transient process,feature overlapping in shallow feature spaces,and high real-time requirements for online diagnosis in modular multilevel converter-based flexible high voltage direct current(MMC-HVDC)systems,a fault diagnosis method based on deep neighborhood components analysis(Deep NCA)and prototypical networks is proposed.First,a one-dimensional convolutional neural network is used to adaptively extract deep features from multi-channel transient sequences,including arm current,submodule ca-pacitor voltage,and grounding current.Second,Deep NCA is introduced to impose supervised metric constraints on the embedding space,enhancing intra-class compactness and inter-class separability.Finally,fast distance-based classification is achieved through class prototypes.The simulation results show that the proposed method achieves an accuracy of 95.20%for five operating states,outperforming typical deep learning models including 1D-CNN+Softmax,1D-ResNet and 1D-Transformer.Under load-shift conditions,the average diagnostic accu-racy is 91.73%,and an accuracy of 86.93%is retained when only 50 training samples per class are used.Results demonstrates that the proposed method possesses favorable adaptability to varying operating conditions and the capability to learn from limited training samples.
黄志光;连攀杰;骆玮;刘嘉宸;时艳强;刘伟成
国家电网有限公司 华东分部,上海 200120中国电力科学研究院有限公司,北京 100192国家电网有限公司 华东分部,上海 200120国家电网有限公司 华东分部,上海 200120国家电网有限公司 华东分部,上海 200120国家电网有限公司 华东分部,上海 200120
信息技术与安全科学
柔性直流(HVDC)输电模块化多电平换流器(MMC)故障诊断深度邻域成分分析(Deep NCA)原型网络度量学习
flexible high voltage direct current(HVDC)transmissionmodular multilevel converter(MMC)fault diagno-sisdeep neighborhood components analysis(Deep NCA)prototypical networkmetric learning
《天津工业大学学报》 2026 (4)
62-72,79,12
国家电网有限公司华东分部科技项目(52992425001L)国家自然科学基金项目(52407079)湖南省自然科学基金项目(2025JJ60372)
评论