基于改进实例学习算法的风电机组齿轮箱状态监测

Wind Turbine Gearbox Condition Monitoring Based onImproved Instance Learning Algorithm

风电机组齿轮箱的运行过程是复杂的非线性过程,采用实例学习(IBL)算法建立模型可有效对其进行状态监测。针对实例学习模型对训练数据质量敏感的特点,提出综合考虑多种性质的两步主动学习样本选择方法。首先提出一种基于拉丁超立方体抽样思想的网格划分初始样本选取方法,并基于z-score方法剔除其中的离群点。然后第一步基于信息性和代表性的综合得分选出候选样本来避免离群点影响,第二步基于多样性使第一步的候选样本稀疏化,从而避免冗余点影响。最后,基于指数加权移动平均控制图对实例学习回归模型输出的残差进行分析,并根据故障率对风电机组齿轮箱实现状态监测。利用某风电机组实际故障数据进行验证。结果表明:所提出的方法能选出优质样本,模型精度在验证集上较未改进前有所提升,且运算效率提升约50%,可实现齿轮箱异常的早期预警。

The operation of a wind turbine(WT)gearbox is a complex nonlinear process.Utilizing instance-based learning(IBL)algorithm to establish models can effectively monitor conditions of the gearbox.In view of the fact that IBL algorithm is sensitive to the quality of training data,a two-step active learning(AL)sample selection method considering multiple properties was proposed.Firstly,a grid partitioning initial sample selection method based on Latin hypercube sampling was proposed,and outliers were removed using z-score.Then,the first step was to select candidate samples based on the comprehensive scores of informativeness and representativeness to avoid the influence of outliers.The second step was to sparsize candidate samples based on diversity to avoid the impact of redundant points.Finally,the residual of instance-based learning regression(IBLR)model was analyzed based on exponential weighted moving average control chart,and the status of WT gearbox was monitored based on the failure rate.The model was verified using actual fault data of a certain WT.Results show that the proposed method can select high-quality samples,improve the model accuracy compared to the unmodified model on the validation set,and increase the computational efficiency by about 50%,which can effectively achieve early warning of the gearbox faults.