不平衡样本下基于迁移学习-AlexNet的输电线路故障辨识方法  

Fault Identification Method for Transmission Line Based on Transfer Learning-AlexNet with Imbalanced Samples

作　　者：王建[1] 吴昊[1] 张博[1] 南东亮 欧阳金鑫[1] 熊小伏[1] WANG Jian;WU Hao;ZHANG Bo;NAN Dongliang;OUYANG Jinxin;XIONG Xiaofu(State Key Laboratory of Power Transmission Equipment&System Security and New Technology,Chongqing University,Chongqing,400044,China;Electric Power Research Institute of State Grid Xinjiang Electric Power Co.,Ltd.,Urumqi,Xinjiang,830011,China)

机构地区：[1]输配电装备及系统安全与新技术国家重点实验室(重庆大学),重庆市400044 [2]国网新疆电力有限公司电力科学研究院,新疆维吾尔自治区乌鲁木齐市830011

出　　处：《电力系统自动化》2022年第22期182-191,共10页Automation of Electric Power Systems

基　　金：国家自然科学基金资助项目(51707018);重庆市出站留(来)渝博士后择优资助项目(2020LY23)。

摘　　要：输电线路不同故障类型和故障原因的故障样本集具有类不平衡性,为基于图像深度学习的故障分类辨识带来挑战。文中提出类不平衡样本下基于迁移学习-AlexNet神经网络的输电线路故障辨识方法。首先,统计分析了输电线路故障的特征与概率分布,使用MATLAB/Simulink仿真产生了符合实际情况的不平衡故障样本集。然后,以故障暂态波形图像为输入集,采用迁移学习-AlexNet神经网络构建故障分类器,降低了故障特征提取的复杂性。算例测试结果表明,现有按类平衡故障样本集开展故障辨识的方法,分类准确率偏于乐观,即使采用抽样法也无法准确识别类不平衡样本中的小样本故障类型;而所提方法可以很好地应对类不平衡故障样本集,相比于经典的卷积神经网络,对故障类型与故障原因的辨识准确率也更高,训练模型用于类似线路的真实故障录波数据也能很好地辨识出故障类型。The fault sample sets of transmission lines with different fault types and fault causes have a characteristic of class imbalance, which brings challenges to fault classification and identification based on image deep learning. A fault identification method for transmission lines based on transfer learning-ALexNet neural network is proposed for class imbalanced samples.Firstly, the characteristics and probability distribution of transmission line faults are statistically analyzed, and an imbalanced fault sample set that conforms to the actual situation is generated using MATLAB/Simulink simulation. Then, using the fault transient waveform image as the input set, the fault classifier is constructed by transfer learning-AlexNet neural network, which reduces the complexity of fault feature extraction. The case test results show that the existing fault identification methods with class balanced fault sample sets have a speciously high classification accuracy. Even if the sampling method is adopted, the fault type of a small sample in the class imbalanced sample cannot be accurately identified. However, the proposed method can give a better solution to the class imbalanced fault sample set. Compared with the classical convolutional neural network, the identification accuracy of the fault type and fault cause with the proposed method is also better. The training model used for the real fault recording data of similar lines can also identify the fault types well.

关 键 词：输电线路 故障辨识 迁移学习 AlexNet神经网络 图像学习 不平衡样本 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术] TM75[电气工程—电力系统及自动化]