【作者】 刘宏伟；

【导师】 冷劲松；

【作者基本信息】 哈尔滨工业大学 ， 力学， 2021， 博士

【摘要】 近年来,风电行业蓬勃发展,风电装机容量大幅上升。风电叶片作为机组的核心构件,其疲劳性能和稳定性严重影响着机组的发电效率和安全。为保障机组的正常运行,众多学者关注于叶片的监测与检测技术,一旦发现叶片故障损伤,便于及时进行维修处理。叶片结构健康监测系统集传感、信号处理和系统集成于一体,实时感知、评估叶片当前状态。随着人工智能技术的不断发展,人们对于监测系统的数据分析利用率和挖掘度也有了更高的要求。全尺寸风电叶片的疲劳涉及叶片出厂前的疲劳试验和机组叶片服役过程。研究叶片疲劳损伤过程中的性能变化与状态预测对于确保机组的长期可靠运行具有重要意义。由于监测系统在长期运行中产生海量数据,一些基于监测数据的风电叶片预测技术应运而生。这些技术往往利用历史运行数据或正常运行数据,建立其运行模式下的基准模型,在评估新状态时,将其与之前状态进行比较,通过评估指标分析其偏差和趋势,从而实现对叶片性能及状态的预测。本文围绕全尺寸风电叶片的疲劳试验和服役过程,针对叶片结构的安全性,进行了基于监测数据的风电叶片状态预测技术研究。搭建了基于光纤光栅(Fiber Bragg Grating,FBG)传感器的风电叶片疲劳试验监测系统,通过实测叶片刚度数据,提出了一种可用于叶片疲劳试验的刚度退化模型;结合历史刚度数据,通过深度学习方法实现了叶片疲劳试验过程中的刚度预测;搭建了基于FBG传感器的风电叶片服役过程监测系统,通过实测应变历程来评估了叶片的当前状态和预测疲劳寿命;针对服役过程中的叶片结冰问题,建立了基于数据采集与监控(Supervisory Control and Data Acquisition,SCADA)数据的机器学习和深度迁移学习模型,实现了源机组和目标机组间的结冰预测。主要工作如下:为建立全尺寸风电叶片疲劳试验过程的刚度衰减模型,对叶片疲劳试验过程进行监测及刚度退化研究。将单点激振下的叶片进行简化并分析讨论;搭建了基于FBG传感器的风电叶片疲劳试验监测系统,与应变片测试结果对比,验证了FBG传感器的稳定性和耐久性;针对复合材料损伤发展的三个阶段,提出了一种基于刚度退化的两参数复合材料疲劳损伤模型,并与现有的刚度退化模型进行对比;将提出的模型应用到叶片疲劳试验中进行验证。为减少叶片疲劳试验的时间成本,鉴于全尺寸风电叶片疲劳试验结果,进行了基于深度学习网络模型的叶片刚度预测研究。将标定下的叶片刚度数据与复合材料刚度退化模型进行拟合,实现了叶片刚度样本量的扩充;将样本进行训练集和测试集划分,采用卷积神经网络(Convolutional Neural Networks,CNN)、长短期记忆(Long Short Term Memory,LSTM)网络、CNN-LSTM网络实现了叶片的刚度预测,并对各个模型网络的预测性能进行对比;对比了不同的输入序列长度和训练集对模型预测结果的影响;为了体现实际叶片疲劳试验过程中的一般性,将刚度数据添加了不同分贝的高斯白噪声后又进行了模型的刚度预测。为建立实时的风电机组叶片服役过程评价手段,对叶片实际服役过程进行监测及寿命预测研究。搭建了基于FBG传感器的风电叶片服役监测系统,实现了机组运行过程中的叶片载荷监测;通过希尔伯特黄氏变换等分析手段,实现了对测点非平稳应变信号的解析;针对风电叶片高周期的应力疲劳问题,基于Goodman应力图和Palmgren-Miner线性损伤累积法则,对叶片进行疲劳损伤计算和寿命评估,实现了风电叶片的寿命预测。为提出基于风电机组SCADA数据的结冰检测方法,基于机器学习和深度迁移学习网络模型对叶片服役过程的结冰检测方法进行了研究。通过GH-Bladed仿真分析,得到了叶片结冰对叶片性能的影响;结合叶片结冰模型和气动模型,对原始SCADA数据进行预处理与特征选择、重构,降低了样本的特征维度;引入统计特征和准确度等评价指标,利用源机组数据和不同分类模型进行训练并对源机组和目标机组预测;针对目标机组预测效果差、模型泛化能力低的问题,通过源机组的预训练模型和少量目标数据样本,实现深度迁移学习,从而提高对该机组结冰检测模型的预测效果。更多还原

【Abstract】 In recent years,with the vigorous development of wind power industry,the installed capacity of wind power has increased considerably.As the core component of wind turbines,the fatigue performance and the stability of blades seriously affect the power generation efficiency and the safety of wind turbines.In order to ensure that the wind turbine blades are in normal working condition,many scholars pay attention to the monitoring and detection technology of wind turbine blades.Once the fault and the damage are found,it is convenient for timely maintenance.The structural health monitoring system of wind turbine blades integrates sensing,signal processing and system integration to sense and evaluate the current state of blade s in real-time.With the continuous development of artificial intelligence technology,people have higher requirements for the data analysis utilization and the mining of monitoring systems.The fatigue of full-scale wind turbine blades involves the fatigue test of blades before delivery and the service process of blades.The research on performance change and state prediction during blade fatigue damage is of great significance to ensure the long-term reliable operation of wind turbines.Since the monitoring system generates massive data in long-term operation,wind turbine blade prediction technologies based on monitoring data came into being.These technologies often use historical operation data or normal operation data to establish the benchmark model under its operation mode.When the new state is to be evaluated,it is often compared with the normal state.The deviation and the tendency are analyzed through the evaluation indexes,so as to realize the monitoring and the prediction of blade state.This paper based on the fatigue test and service process of full-scale wind turbine blades,aiming at the safety of blade structure,carries out the research of wind turbine blade state prediction technology based on monitoring data.A fatigue test monitoring system for wind turbine blades is established based on fiber Bragg grating(FBG)sensors.A stiffness degradation model for blade fatigue testing is proposed through the measured blade stiffness data.Combined with the historical stiffness data,the stiffness prediction of the blade during the fatigue test is realized by deep learning methods.A monitoring system for the service process of wind turbine blades based on FBG sensors is built to evaluate the current state and predict the fatigue life of the blade by strain data.To solve the problem of blade icing in service,machine learning and deep transfer learning model s based on data acquisition and monitoring(SCADA)data are established to realize icing detection in the source wind turbine and the target wind turbine.The main work is listed as follows:To establish the stiffness attenuation model of full-scale wind turbine blade fatigue test process,the fatigue test process is monitored and the stiffness degradation is studied.The blade under single-point excitation is simplified and analyzed;A fatigue test system of a wind turbine blade based on FBG sensors is built,and the stability and durability of the sensors are verified by comparing with the test results of the strain gauges.Aiming at the three stages of composite damage development,a two-parameter composite fatigue damage model based on stiffness degradation is proposed and compared with the existing stiffness degradation models.The proposed model is applicable to the blade fatigue test for verification.To reduce the time cost of blade fatigue test,in view of the fatigue test results of full-scale wind turbine blades,a method of blade stiffness prediction based on deep learning network models is studied.The stiffness data under calibration is fitted with the stiffness degradation model of composite materials to expand the sample size of blade stiffness.The samples are split into a training set and a test set.Convolution neural network(CNN),long-term short memory(LSTM)network,and CNN-LSTM network are used to predict the blade stiffness,and the prediction performance of each model network is compared.The effects of different input sequence lengths and training sets on the model prediction are compared separately.In order to reflect the generality of the actual blade testing process,the stiffness data is added with different decibels of white Gaussian noise,and then the stiffness prediction of the model is carried out.To establish a real-time evaluation method for the service process of wind turbine blades,the service process monitoring and life prediction of wind turbine blades are studied.To realize the blade load monitoring during the operation of a wind turbine,a monitoring system of wind turbine blades based on FBG sensors is built.Through Hilbert Huang transform and correlation analysis,the non-stationary strain signal of the measuring point is analyzed.Aiming at the high cycle stress fatigue problem of wind turbine blades,based on Goodman stress diagram and Palmgren-Miner linear damage accumulation rule,the fatigue damage calculation and life assessment of blades are carried out,and the life prediction of wind turbine blades is realized.To propose an ice detection method based on SCADA data of wind turbine s,a blade icing detection method for the service process through machine learning and deep transfer learning methods is studied.Through GH-bladed simulation analysis,the influence of blade icing on blade performance is obtained.Combined with the blade icing model and aerodynamic model,the original SCADA data is preprocessed,feature selected,and reconstructed so that the feature dimension of the sample is greatly reduced.Statistical characteristics and evaluation indexes such as Accuracy are introduced to train and predict the source and the target wind turbine using single source turbine data and different classification models.For the problem of poor prediction effect and low generalization ability of the target turbine,the deep transfer learning models are realized through the pre-trained models of the source turbine and a small number of target data samples,so as to improve the prediction effect of the icing detection model of the turbine.更多还原