基于时空自编码网络的风电齿轮箱状态监测

在风电机组状态监测问题中，常规自编码网络通常仅使用截面SCADA(supervisory control and data acquisition)数据，使得网络对数据时间特征的学习不足。因此，提出一种基于时空自编码网络的风电齿轮箱状态监测方法：使用1维卷积网络(1DCNN)级联双向长短时记忆网络(Bi-LSTM)作为编码层，序贯提取面板数据的空间及时间特征，以输入的重构误差作为预警指标实现在线状态监测。使用河北省某风电场实际数据验证，结果表明：相比故障记录时刻，时空自编码网络能提前20 d发出报警信号，且故障检出率和误报警次数均优于常规方法；通过分析重构误差各分量的贡献率，可知该齿轮箱故障中主要异常参数为油路压力和油池温度。

Condition Monitoring of Wind Turbine Gearbox Based on Spatial-temporal Autoencoder Network

Conventional autoencoder networks only use cross-section supervisory control and data acquisition data when monitoring wind turbine conditions, providing insufficient data to the network to learn about the temporal data characteristics. Therefore, a method of monitoring wind turbine gearbox conditions is proposed using a spatiotemporal autoencoder network. First, we use a one-dimensional convolutional neural network cascade bidirectional-long short-term memory network as the encoder layer to abstract the spatiotemporal characteristics of panel data sequentially. Second, input reconstruction errors are used as the warning index to realize online state monitoring. Finally, the results are verified using the actual data of a wind farm in Hebei province. The results demonstrate that, compared with the fault recording time, the spatiotemporal autoencoder network can send the alarm signals 20 days earlier, and the fault detection rate and false alarm times are better than the conventional methods. By analyzing the contribution rate of each component of the reconstruction error, it is observed that the main abnormal parameters of the gearbox fault are oil pressure and oil pool temperature.