基于长短期记忆网络的风电机组齿轮箱故障预测

针对风电齿轮箱状态监测数据的多变量动态时空关联性特点,提出了一种基于长短期记忆(long short-term memory,LSTM)网络的齿轮箱故障预测方法,主要包括离线建模和在线监测两个阶段。首先,以齿轮箱油温为目标预测变量,充分考虑其与其它相关输入变量之间在时空维度上的重要关联信息,对历史监测数据进行训练学习,建立齿轮箱正常运行时的油温监测LSTM模型,通过对预测残差进行评估计算设定相应的检测阈值;然后,将训练好的油温监测LSTM模型用于在线测试,通过模型残差分析和阈值比较实现齿轮箱故障状态的检测和预测;最后,通过风电场测试数据对所提出的方法进行验证。结果表明,相比于其它传统方法,该方法表现出更好的预测性能,能够较早预测故障的发生。

Fault Prediction of Wind Turbine Gearbox Based onLong Short-term Memory Network

To capture the characteristics of dynamic temporal-spatial correlation hidden in monitoring data of wind turbine gearbox, a fault prediction method based on long short-term memory (LSTM) network is proposed. The proposed approach mainly consists of two phases: offline training and online detection. First, the oil temperature of the gearbox is taken as the modeling variable and the LSTM-based normal behavior model of wind turbine gearbox oil is built based on historical monitoring data training and learning, which fully takes advantages of the important correlated information between the oil temperature and some relevant input variables. Then, the model residuals are calculated and evaluated to determine the corresponding detection threshold. Furthermore, the well-trained LSTM model is used for online testing.Through model residual analysis and threshold comparison, fault detection and prediction of wind turbine gearbox can be realized.A real motoring data from a wind farm is used to validate the effectiveness of the proposed method. The results show that compared with those traditional methods, the proposed method presents better prediction performance, and can predict the occurrence of gearbox failure earlier.