改进BiLSTM-GRU-DMD的风机叶片结冰故障检测与状态评估模型  

Detection and Status Assessment of Wind Turbine Blade Icing Using Improved BiLSTM-GRU-DMD

作　　者：陈闻鹤 程龙生[1] 常志朋 文卜玉[4] 陈宗祥 CHEN Wenhe;CHENG Longsheng;CHANG Zhipeng;WEN Buyu;CHEN Zongxiang(School of Economics&Management,Nanjng University of Science&.Technology,Nanjing,Jiangsu 210000,China;School of Business,Anhui University of Technology,Manshan,Anhui 243000,China;Key Laboratory of Multidisciplinary Management and Control of Complex Systems of Anhui Higher Education Institutes,Anhui University of Technology,Manshan,Anhui 243000,China;School of Information Engineering,Eastern Liaoning University,Dandong,Liaoning 118000,China;School of Electrical and Information Engineering,Anhui University of Technology,Maanshan,Anhui 243000,China)

机构地区：[1]南京理工大学经济管理学院,江苏南京210000 [2]安徽工业大学商学院,安徽马鞍山243000 [3]安徽工业大学复杂系统多学科管理与控制安徽普通高校重点实验室,安徽马鞍山243000 [4]辽东学院信息工程学院,辽宁丹东118000 [5]安徽工业大学电气与信息工程学院,安徽马鞍山243000

出　　处：《工业工程与管理》2023年第5期108-118,共11页Industrial Engineering and Management

基　　金：国家自然科学基金资助项目(71673001);国家留学基金资助项目(202206840062);辽宁省自然科学基金博士科研启动基金计划(2022-BS-287);江苏省研究生科研与实践创新计划项目(KYCX21_0356);安徽省高校人文社会科学基金重大项目(SK2021ZD0034);安徽省普通高校重点实验室开放基金重点项目(CS2020-ZD02)。

摘　　要：针对风电机组运行监测数据的不平衡性与时序性,提出一种新的风机叶片结冰故障诊断与状态评估方法。首先,利用自适应过采样方法均衡风机结冰样本数据集的不平衡性;然后,改进堆叠双向长短时记忆网络和门控循环单元检测风机叶片结冰故障,利用焦点损失函数作为损失函数侧重于难分类样本优化模型,并结合改进非洲秃鹫优化算法优化超参数,从而提升检测准确性;最后,测度结冰样本与非结冰样本的动态马氏距离,并转换为结冰指数评估叶片结冰状态。真实风场数据验证表明:提出的风机结冰故障检测与状态评估方法,在结冰检测方面,其F1分数、精确率和召回率分别达到0.9678、0.9607和0.9751,优于其他基线模型和优化算法,有效地减少了错报率和漏报率。同时,在定量化评估风机叶片不同阶段的结冰状态方面具有优势,可以为风电设备视情维修提供支持。A novel method was proposed for icing fault diagnosis and condition assessment of wind turbine blades. It could address the unbalanced and temporal data in wind turbine operational monitoring. First,the adaptive oversampling method was used to equalize the imbalance data in training set. Then,a stacked model was built to detect turbine blade icing. The stacked model was composed of bi-directional long and short-term memory network and gated recurrent unit. The focal loss function was applied as loss function focusing on samples with difficult classification. An improved African vulture optimization algorithm was also used to optimize the hyperparameters for improving the accuracy. Finally,the dynamic Mahalanobis distance was proposed to measure the icing and nonicing samples and converted into the icing indexes to assess the icing status of blades. The real wind turbine data validated that the proposed method can achieve F1 score(0.967 8),accuracy(0.960 7) and recall(0.975 1) in icing detection. The results are better than other baseline models and optimization algorithms. It effectively reduces the rates of false alarm and missing alarm. At the same time,it can obtain advantages in quantitatively assessing the icing status at different stages. This model can provide support for emergency management and maintenance decisions for wind turbine equipment.

关 键 词：风机结冰检测 动态马氏距离 非平衡数据 堆叠循环神经网络 改进非洲秃鹫优化算法 

分 类 号：TM315[电气工程—电机]