【作者】 胡志伟；

【导师】 秦毅；

【作者基本信息】 重庆大学 ， 工程（机械工程）（专业学位）， 2021， 硕士

【摘要】 风机通常布置在高山、荒漠和海上等风力资源丰富的偏远地区,恶劣的环境、复杂多变的工况使得风机故障频发,在各种风机故障中传动系统的齿轮箱故障造成的损失最大,因此对风电齿轮箱的运行状态进行监测是一个具有重要意义的课题。目前,风场基本配备数据采集与监测系统(Supervisory Control and Data Acquisition,SCADA)和状态监测系统(Condition Monitoring System,CMS)进行风机监控,但CMS包含多个部件的耦合振动信号,振动信号受工况影响大、信息单一且采样间隔长,很难通过建立阈值进行有效的状态监测,为了通过更多的变量和阈值来进行实时状态监测,本文使用采样间隔小且变量更多的SCADA系统进行齿轮箱状态监测。传统的监测方法主要通过使用信号处理和机器学习等方法分析并监测SCADA系统提供的运行数据来实现,但随着SCADA系统结合大数据技术给齿轮箱状态监测带来新的问题,一方面数据质量参差不齐,原始数据中存在大量的劣质数据和缺失数据,难以直接挖掘出有价值的信息甚至会造成模型的监测能力变差;另一方面齿轮箱SCADA数据数量庞大的同时不同变量之间存在复杂的耦合关系,传统监测方法很难对大量数据进行充分的学习。针对风电齿轮箱的数据质量和状态监测方法的问题,本文使用风机机理结合深度学习方法进行风机SCADA数据预处理,然后构建一种记忆增强自编码的深度学习模型结合SCADA数据进行风电齿轮箱的状态监测。本文进行了下列研究。首先,针对SCADA原始数据质量较差的问题,提出风机运行机理结合bin算法的数据清洗方法、基于邻比机组的门控循环网络(Gate Recurrent Unit,GRU)数据修复方法和数据预处理策略设计。通过风机运行机理结合bin算法去除齿轮箱劣质SCADA数据,再通过GRU修复模型修复缺失数据,并使用实测SCADA数据集验证预处理方法的效果,结果表明该预处理方法能极大地提高SCADA数据质量。其次,针对传统齿轮箱状态监测方法学习能力不足、监测效果不佳的问题,提出一种记忆增强自编码网络(Memory-augmented Deep Autoencoder,MAE)融合SCADA数据的风电齿轮箱状态监测模型,并设计风机齿轮箱状态监测流程。先通过清洗后的数据样本计算皮尔森相关系数选取模型所需的SCADA变量,然后构建记忆增强自编码监测模型学习数据特征,利用SCADA数据重构误差作为状态监测的指标,再结合指数加权移动平均值(Exponentially Weighted Moving-Average,EWMA)控制图计算阈值进行监测。使用风场实测SCADA数据进行实验,证明本模型比传统监测模型具有更好的监测效果。最后,开发风电齿轮箱监测功能模块并部署在某风场的风电机组状态监测微服务平台上,使用本文提出模型构建相关服务,设计监测服务的功能、接口、数据库和数据模型,并通过微服务平台的可视化前端验证风电齿轮箱监测功能模块的有效性。更多还原

【Abstract】 Wind turbines are usually located in remote areas with rich wind resources,such as mountains,deserts and sea.The bad environment and complex and changeable working conditions make wind turbine failures occur frequently.Among all kinds of wind turbine failures,the gearbox failure of transmission system causes the greatest loss.Therefore,it is of great significance to monitor the operation status of wind turbine gearbox.At present,the wind farm is basically equipped with supervisory control and data acquisition(SCADA)and condition monitoring system(CMS)to monitor wind turbines,however,CMS contains coupling vibration signals of multiple components.The vibration signals are greatly affected by the working conditions,the information is single and the sampling interval is long,so it is difficult to establish the threshold for effective condition monitoring.In order to carry out real-time condition monitoring through more variables and threshold,this paper uses the SCADA system with small sampling interval and more variables for gearbox condition monitoring.Traditional monitoring methods are mainly realized by using signal processing and machine learning methods to analyze and monitor the operation data provided by SCADA system.However,with the combination of SCADA system and big data technology,new problems are brought to gearbox condition monitoring.On the one hand,the data quality is uneven,and there are a lot of poor quality data and missing data in the original data,so it is difficult to directly mine valuable data On the other hand,there is a large number of gearbox SCADA data,and there is a complex coupling relationship between different variables,so it is difficult for traditional monitoring methods to fully learn a large number of data.Aiming at the problems of data quality and condition monitoring method of wind turbine gearbox,this paper uses the wind turbine mechanism combined with deep learning method to preprocess the wind turbine SCADA data,and then constructs a memory enhanced self coding deep learning model to monitor the condition of wind turbine gearbox combined with SCADA data.In this paper,the following studies are carried out.Firstly,aiming at the problem of poor quality of SCADA raw data,the data cleaning method of wind turbine operation mechanism combined with bin algorithm,the data repair method of gate recurrent unit(GRU)based on adjacent ratio unit and the design of data preprocessing strategy are proposed.The bad SCADA data of gearbox is removed by f wind turbine operation mechanism combined with bin algorithm,and then the missing data is repaired by GRU repair model.The effect of pretreatment method is verified by actual SCADA data set.The results show that the pretreatment method can greatly improve the quality of SCADA data.Secondly,aiming at the problems of poor learning ability and poor monitoring effect of traditional gearbox condition monitoring methods,a wind power gearbox condition monitoring model based on memory enhanced deep autoencoder(MAE)and SCADA data is proposed.According to the monitoring process of wind turbine gearbox,Pearson correlation coefficient is calculated through the cleaned data samples,and SCADA variables required by the model are selected.Then memory enhanced self coding monitoring model is constructed to learn data features.SCADA data reconstruction error is used as the indicator of condition monitoring,and combined with exponentially weighted moving average Moving average(EWMA)control chart.The experimental results show that this model has better monitoring effect than the traditional monitoring model.Finally,the wind turbine gearbox monitoring function module is developed and deployed on the wind turbine condition monitoring micro service platform of a wind farm.The model proposed in this paper is used to build related services,and the function,interface,database and data model of the monitoring service are designed.The effectiveness of the wind turbine gearbox monitoring function module is verified through the visual front-end of the micro service platform.更多还原