Abstract: Traditional machine learning methods were used in fault early warning of wind turbine gearboxes, the models were usually designed only by studying the relationship between data and faults, and the selection of parameters and model structure were lack of physical basis, resulting in poor interpretability and weak generalization capabilities of the models. The structure and actual operation control mode of the wind turbine gearbox were studied, the relationship between the operation mechanism and the data of corresponding supervisory control and data acquisition system was analyzed, and the operation data change trend was given qualitatively followed by deterioration of the typical gearbox faults. Finally, a series of one-class support vector machine(OCSVM) based models were constructed according to change law of the data distribution to realize the early fault warning of the wind turbines gearbox systems. Experimentsal results show that all of the proposed models may locate the fault positions of the wind turbine gearbox systems, which has clear physical significance. 

Key words: wind turbine, failure analysis, fault diagnosis, fault detection, operating mechanism analysis

摘要： 传统基于机器学习的风电齿轮箱故障预警模型往往仅从数据着手分析数据与故障的映射关系，在参数和模型结构选择上缺少物理依据，导致模型的可解释性和泛化能力不强。从风电齿轮箱的结构和实际运行控制方式出发，分析了运行机理与对应的数据采集与监视控制系统数据的关系，定性地给出了齿轮箱典型故障发生时运行数据的变化趋势，然后根据数据分布变化规律选择参数和模型，建立了一系列基于单分类支持向量机的风电齿轮箱系统故障预警模型。实验结果显示各模型能够准确定位风电齿轮箱系统故障，具有清晰的物理意义。

关键词: 风力发电机, 故障分析, 故障诊断, 故障检测, 运行机理分析