基于LSTM-BO和SPRT的风电机组故障演化过程分析  

Fault Evolution Analysis of Wind Turbines Based on LSTM-BO and SPRT

作　　者：冯晨龙 刘超[1] 蒋东翔[1] Chen-long Feng;Chao Liu;Dong-xiang Jiang(Department of Energy and Power Engineering,Tsinghua University)

机构地区：[1]清华大学能源与动力工程系

出　　处：《风机技术》2023年第2期76-84,共9页Chinese Journal of Turbomachinery

摘　　要：有效挖掘故障演化的过程有利于分析和总结故障传播规律。本文提出了基于LSTM-BO模型和SPRT算法,利用SCADA历史数据研究风电机组早期故障演化过程的分析方法。首先使用Pearson相关系数法挖掘SCADA数据各监测变量之间的映射关系,获得LSTM-BO网络所需的输入—输出关系对;其次是依赖LSTM网络强大的时序特征提取能力,将预处理后的SCADA数据送入LSTM网络进行训练,从而得到相应的正常行为模型;最后,依赖SPRT在序贯测试方面的优势,将各监测变量模型预测值与实际监测值之间的偏差看作序贯测试的对象,并对其测试结果进行滑动窗口观测,以异常点数目占观测窗口宽度的比值为度量指标,得到SCADA数据部分重要监测变量在故障发生前的演化过程。实例分析结果表明,所提方法能够有效提取故障演化的过程,为后续演化规律的分析提供指导。The effective mining of fault evolution is beneficial to analyze and summarize the fault propagation laws.In this paper,an analysis method based on LSTM-BO model and SPRT algorithm is proposed to study the early fault evolution of wind turbines using SCADA historical data.Firstly,the correlation relationship between each variable of SCADA data and other variables is organized to form a series of input-output relationship pairs.Secondly,the pre-processed SCADA data is fed into the LSTM for training to obtain the normal behavior model corresponding to each variable based on the powerful temporal feature extraction ability of LSTM.Finally,the deviation between the model predicted value and the actual monitored value of each variable is regarded as the item of sequential testing using the SPRT,and these test results are observed by a sliding window to obtain the metric that is defined as a ratio of the number of abnormal points to the width of the observation window.The evolution of some important variables of SCADA data before the occurrence of faults can be obtained using the proposed method and the case analysis verifies its effectiveness for the mining of fault evolution and its guiding significance for the subsequent analysis of the evolution laws.

关 键 词：风电机组 SCADA 故障演化 LSTM-BO SPRT 

分 类 号：TK83[动力工程及工程热物理—流体机械及工程]