基于堆叠自编码网络的风电机组发电机状态监测与故障诊断

Condition Monitoring and Fault Diagnosis of Wind Turbine Generator Based on Stacked Autoencoder Network

为实现风力发电机的异常检测分析,提出了一种基于风电机组发电机正常状态下数据采集与监控(SCADA)样本数据的堆叠自编码网络深度学习方法。首先将多个自编码网络连接构成深度堆叠自编码网络,选取发电机SCADA状态变量数据作为网络的训练输入,使网络逐层智能提取数据间的分布式规则,从而构建发电机的堆叠自编码学习模型。依据故障状态下发电机SCADA数据内部动态平衡规则被破坏,利用发电机深度学习网络的输入与重构值计算重构误差,并作为整体状态的观测量。通过采用自适应阈值检测重构误差的状态趋势变化,并作为异常预警判定准则,从而实现对发电机故障的判定。当发电机发生异常时,变量的实际值与对应模型的重构值发生较大偏差,表现为状态变量的残差趋势将会偏离原有的动态稳定状态。因此利用状态变量的残差趋势变化对异常变量进行隔离,判定可能的故障原因达到故障诊断的目的。通过对发电机故障前后记录数据进行仿真分析,结果验证了堆叠自编码网络深度学习方法对发电机状态监测与故障诊断的有效性。

In order to realize the abnormal condition detection and analysis o[ the wind turbine generator, a deep learning method of stacked autoencoder （SAE） network is proposed based on condition monitoring data of supervisory control and data acquisition （SCADA） of wind turbine generator. The SAE network is composed of multiple autoencoder networks, and the normal generator SCADA state variable data is selected as the training input of the network. The distributed rules are extracted intelligently from the network layer by layer to build the SAE learning model. Because internal dynamic balance rules of the generator SCADA data are destroyed under the abnormal state, the reconstruction error is calculated using the initial input and reconstruction values as the observation of the whole state. Adaptive threshold is used to detect the trend of the reconstruction error and as a criterion for abnormal early warning to detect generator abnormal condition. When the generator is abnormal, there is a larger deviation between the actual value and its reconstruction value. So the residuals trend of the state variables will deviate from the original dynamic stability state. The trend change of the state variable residual will be used to isolate the abnormal variables to locate the generator fault and achieve the purpose of fault diagnosis. The simulation results show that the SAE network deep learning method is effective for condition monitoring and fault diagnosis of generator.