【作者】 刘华新；

【导师】 韩中合；

【作者基本信息】 华北电力大学(北京) ， 动力机械及工程， 2020， 博士

【摘要】 风力发电是当前可再生能源领域技术最成熟、增长速度最快、商业化发展最好的发电方式之一。近年来,我国风电产业实现跨越式发展,装机容量位居世界首位。同时,我国风电机组也面临故障率居高不下等问题,一方面,我国风电产业发展之初风电机组研发制造技术与对装备的迫切需求不匹配;另一方面,针对在运机组状态识别和维修决策的研究还不完善。而且,当前在运机组存在运行数据闲置、利用不充分等问题。因此,本文针对风电机组维修决策问题,利用在运机组的海量运维数据,开展了风电机组部件重要度评判、状态评估、状态预测和状态识别研究,为开展维修决策提供基础。首先,针对重要度评判过程中的不确定性问题,提出了区间取值的灰色模糊重要度评判方法。以特殊模糊数——区间数代替传统模糊隶属度表征评判过程中的模糊性,采用区间灰数表示评判过程的灰色性;为解决区间数排序的问题,采用OWA算子,将模部和灰部进行集结;评判结果有明显的离散性,不易受信息不完全性的影响。综合考虑状态评估模型的精确性和状态评估过程的客观性,提出了基于正态云模型的风电机组整机状态评估方法。通过对SCADA数据进行遴选,在部件重要度评判基础上,构建了包括5个核心部件、14个指标的状态评估指标体系;通过温度指标的变换与修正,消除了输出功率和环境温度对非强制降温部件温度指标的影响,确定了合理的温度指标修正阈值及劣化度;采用熵权权重体现评估指标数据之间的客观差异程度,采用DSmT证据融合准则平衡主观因素和客观数据之间的冲突得到DSmT融合权重;采用云隶属度代替模糊隶属度,提出了一种确定正态云数字特征的新方法,有效解决了重叠性和相对贴近度过大的问题。研究表明,该模型可以跟踪机组的状态劣化趋势,实现早期预警。其次,为实现状态参数预测,提出了基于相空间重构的深度学习状态参数预测方法。利用该方法分别建立了齿轮箱润滑油温度和振动参数的预测模型,并与LSSVM预测结果、BP神经网络预测结果、现场试验数据进行了对比研究。结果表明:该方法在绝对误差和相对误差方面都明显低于其他预测方法50%以上,在状态参数预测中的优越性显著。为实现部件具体状态的识别,提出了改进的卷积神经网络模型,以反映齿轮箱状态的矩阵为学习对象,建立了基于VGGNet结构的改进卷积神经网络模型,通过对状态矩阵进行深度学习,学习到提取齿轮箱状态特征的能力,采用训练后的卷积神经网络模型对其他机组同型号齿轮箱进行状态识别和分析,证明了其准确率可达95.3%,有一定泛化能力。最后,在前述研究内容的基础上,针对部件维修方式决策、定时维修间隔期优化、状态维修时机决策开展研究工作。第一,在部件重要度评判的基础上,结合状态监测手段的完备程度,对部件维修方式进行逻辑决策。结果显示,对于状态监测手段完备的核心部件,如齿轮箱,应采用状态维修;对于重要部件和欠缺状态监测手段的核心部件应采用定时维修。第二,针对定时维修的可修部件,采用不完全维修策略,用役龄修正因子和失效率增长因子对失效率函数进行修正来表征每一次预防性不完全维修的维修效果,根据维修次数的变化动态调整维修间隔。实证研究结果表明,该模型可以兼顾可靠度和经济性。第三,应用比例危险模型进行状态维修时机决策,不仅考虑部件运行时间,更关注状态监测参数对维修时机确定的影响。实证研究中,以平均维修费用最小为目标,求解状态维修阈值,得到状态维修决策曲线,利用实测数据和预测数据进行了测试,计算结果表明该模型可以准确确定维修时机,减少维修的盲目性。更多还原

【Abstract】 At present,wind power is one of the most mature technologies in renewable energy field,with the fastest growth rate and best commercial development.In recent years,China’s wind power industry has achieved leap-forward development,with installed capacity ranking first in the world.At the same time,China’s wind turbines are also facing problems such as high failure rate.This is partly because that the urgent need for equipment at the beginning of wind power industry led to the design and finalization of wind turbines in a hurry,and technology is not mature.On the other hand,the research on condition assessment and maintenance decision-making of the operating units is seriously lacking.In addition,there are a lot of problems such as idle operation data and inadequate utilization of current operating units.Therefore,aiming at maintenance decision-making of wind turbine,using the massive operation and maintenance data of the units in operation,this paper carried out the research on importance evaluation of components,condition evaluation,condition prediction and condition identification of wind turbine,which provided a basis for maintenance decision-making.First of all,in view of the uncertainty in the process of importance evaluation,a grey fuzzy importance evaluation method of interval value was proposed.The special fuzzy number-interval number-was used to instead the traditional fuzzy membership degree to represent the fuzziness in the evaluation process.The interval grey number was used to represent the gray of the evaluation process,in order to solve the problem that the interval number is not easy to sequence,OWA operator was used to aggregate the module part and the gray part.The evaluation results have obvious discreteness and are not easily affected by incomplete information.Considering the accuracy of condition evaluation model and the objectivity of condition evaluation process,a condition evaluation method of wind turbine based on the normal cloud model was proposed.Based on the selection of SCADA data and the importance evaluation of components,a condition evaluation index system consisting of 5 core components and 14 indicators was constructed.Through the transformation and correction of temperature index,the influence of output power and ambient temperature on the temperature index of non forced cooling parts is eliminated,and reasonable correction threshold and deterioration degree of temperature index were determined.Entropy weight was used to reflect the objective difference between evaluation index data,DSmT evidence fusion criterion was adopted to balance the conflict between subjective factors and objective data to obtain the DSmT fusion weight.Cloud membership degree was used to replace fuzzy membership degree.A method to determine normal cloud digital,which effectively solves the problems of too much overlap and relatively close was proposed.The study shows that the model can track the condition deterioration trend of the unit and realize early warning.Secondly,in order to realize the prediction of condition parameters,a deep learning condition parameter prediction method based on phase space reconstruction was proposed,which was used to establish the prediction models of oil temperature and vibration parameters.The results,which were compared with LSSVM prediction results,neural network prediction results and field test data,showed that the absolute error and relative error were significantly lower than those of other prediction methods by more than 50%,and the superiority of this method in the condition parameter prediction was significant.In order to realize the identification of the specific condition of the components,an improved convolution neural network model was proposed.Taking the matrix reflecting the condition of gearbox as the learning object,an improved convolution neural network(CNN)model based on VGGNet structure was established.The CNN model learned the ability to extract the condition features of gearbox by deep learning the condition matrix.The trained CNN model was used to detect and analyze the condition of the same type gearbox of other unit.It was proved that the accuracy of the model can reach 95.3%,and it had certain generalization ability.Finally,on the basis of the above research.the research work was carried out for the decision-making of component maintenance mode,the optimization of interval of time based maintenance and the decision-making of opportunity of condition based maintenance.First of all,on the basis of component importance evaluation,combined with the application degree of condition monitoring.the logical decision on component maintenance mode was made.The decision results show that condition-based maintenance should be used for the core components with complete condition monitoring means,such as gearbox,and time-based maintenance should be used for the important components and the core components without condition monitoring means.Secondly,the imperfect preventive maintenance strategy should be used for the repairable components with time-based maintenance.The age correction factor and failure rate growth factor were used to modify the failure rate function to characterize the maintenance effect of each preventive imperfect maintenance,and the maintenance interval is dynamically adjusted according to the change of maintenance times.The empirical research results show that the model can take into account the reliability and economy.Finally,the proportional hazard model was applied to the decision-making of condition-based maintenance,which not only considered the running time of components,and more attention was paid to the influence of condition monitoring parameters on the determination of maintenance time.In the empirical study,aiming at the minimum average maintenance cost,the decision curve of condition based maintenance was obtained by solving the threshold value of condition based maintenance,which was tested with the measured data and the predicted data.The calculation results show that the model can accurately determine the maintenance opportunity and reduce the blindness of maintenance.更多还原