基于SMOTETomek过采样方法与领域自适应迁移学习的风电机组故障诊断  

FAULT DIAGNOSIS OF WIND TURBINES BASED ON SMOTETOMEK OVERSAMPLING METHOD AND DOMAIN ADAPTIVE TRANSFER LEARNING

作　　者：张伊杰 刘宝良 王承民[1] 杨镜非[1] 谢宁[1] Zhang Yijie;Liu Baoiang;Wang Chengmin;Yang Jingfei;Xie Ning(School of Electronic Information and Electrical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China;School of Electric Power,Shenyang Institute of Engineering,Shenyang 110136,China)

机构地区：[1]上海交通大学电子信息工程与电气工程学院,上海200240 [2]沈阳工程学院电力学院,沈阳110136

出　　处：《太阳能学报》2024年第10期635-644,共10页Acta Energiae Solaris Sinica

基　　金：国家自然科学基金(51777121);2021年辽宁省揭榜挂帅科技攻关专项(2021JH/10400009)。

摘　　要：为在不平衡数据上得到准确分类的故障诊断模型,提出将SMOTETomek过采样方法与领域自适应迁移学习相结合的故障诊断算法框架。首先利用滑动窗口采样技术将数据采样成二维时空窗口数据,然后执行SMOTETomek过采样操作,可保留并丰富完整的时序故障特征。针对过采样算法引入噪声信息的问题,引入领域自适应迁移学习算法在原始数据与过采样后的数据之间提取不变特征,使得过采样算法的引入的噪声信息可被过滤掉。在中国某实际风电场的实验结果显示,所提方法可在高度不平衡的数据上完成模型训练,准确识别各类型故障并精确辨识故障过程对应的时间窗口,诊断性能显著优于基于先前用于应对数据不平衡所普遍使用的过采样方法得到的模型。The installed capacity of wind power has grown significantly in recent years,and wind power accounts for an increasing proportion of the total generation capacity,while its fault process can pose a greater threat to the safety and stability of grid operation,so it is important to accurately diagnose and predict the faults occurring on wind turbines.SCADA data-driven fault diagnosis algorithms have been widely researched and applied,however,the high imbalance in the distribution of the number of SCADA normal and fault data poses a major challenge for establishing a high-performance fault diagnosis model.In order to obtain a fault diagnosis model that can accurately give fault categories on unbalanced data,this paper proposes a fault diagnosis algorithm framework that combines SMOTETomek oversampling method with domain adaptive migration learning.The data is first sampled into two-dimensional temporal window data using sliding window sampling technique,and then SMOTETomek oversampling operation is executed on this basis to retain and enrich the complete temporal fault features.To address the problem of noise information introduced by the oversampling algorithm,this paper introduces a domain adaptive migration learning algorithm to extract invariant features between the original data and the oversampled data,so that the noise information introduced by the oversampling algorithm can be filtered out.Experimental results in a real wind farm in China show that the proposed method can complete model training on highly unbalanced data,accurately identify each type of fault and accurately give the time window of the fault process,and the diagnostic performance is significantly better than that of the model obtained based on the previously commonly used oversampling method.

关 键 词：风电机组 故障诊断 监督控制和数据采集系统 深度学习 SMOTE过采样方法 领域自适应 

分 类 号：TK83[动力工程及工程热物理—流体机械及工程]