摘要
油色谱数据的缺乏和不均衡会导致训练过拟合、模型缺乏代表性、测试集效果不理想等问题,从而难以对变压器的状态进行准确评价。针对该问题,将强化学习中的策略梯度算法引入生成式对抗网络GAN(Generative Adversarial Networks),提出了一种基于策略梯度和GAN的变压器油色谱案例生成方法。仿真结果表明,与传统的样本扩充算法相比,利用所提方法合成的样本质量较高。对包含9种故障状态共700组样本的变压器油色谱数据利用所提方法进行油色谱故障样本扩充,利用基于BP神经网络模型的变压器故障分类模型对将扩充后样本作为训练集训练得到的神经网络模型和仅用真实数据作为训练集训练得到的神经网络模型进行了对比,结果表明利用扩充的样本后,变压器故障分类准确率得到了提高。变压器故障诊断实例表明利用所提方法得到的结果与实际情况相符。
The lack and imbalance of oil chromatogram data lead to problems such as training over fitting,lack of representativeness of the model,and unsatisfactory results of train sets,which makes it hard to accurately evaluate the state of transformers.Aiming at this problem,the policy gradient algorithm in reinforcement learning is introduced into GAN(Generative Adversarial Networks),and an oil chromatogram case generation method of transformer based on policy gradient and GAN is proposed.The simulative results verify that the samples synthesized by the proposed method are of higher quality than those synthesized by the traditional sample expansion algorithm.The transformer oil chromatogram data containing 700 groups of samples of 9 fault states is expanded by the proposed method.The neural network model trained by only the real data as the training set is compared with the neural network model trained by the expanded data by using the transformer fault classification model based on BP neural network model.The results show that the accuracy of transformer fault classification is improved by using the expanded data.The actual transformer fault diagnosis case show that the results gained by the proposed method are consistent with the actual situation.
作者
李雅欣
侯慧娟
胥明凯
李善武
盛戈皞
江秀臣
LI Yaxin;HOU Huijuan;XU Mingkai;LI Shanwu;SHENG Gehao;JIANG Xiuchen(Department of Electrical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China;State Grid Shandong Power Supply Company,Jinan 250002,China)
出处
《电力自动化设备》
EI
CSCD
北大核心
2020年第12期211-217,共7页
Electric Power Automation Equipment
基金
国家自然科学基金资助项目(51477100)
上海交通大学新进青年教师启动计划基金(基于人工智能的电力设备故障诊断)。
关键词
变压器
油色谱
样本扩充
生成式对抗网络
强化学习
策略梯度
power transformers
oil chromatogram
sample expansion
generative adversarial networks
reinforcement learning
policy gradient
