多特征组合集成的风电机组齿轮箱主轴承温度异常预测
Prediction of Main Bearing Temperature Anomaly in Wind Turbine Gearbox with Multi-Feature Combination Integration

风电机组齿轮箱主轴承在高速旋转和重负载工况下工作，其温度异常可能导致轴承的过热或磨损加剧，进而影响设备寿命。本文针对现有预测方法未充分考虑模型输入特征所在设备测点状态的问题，提出一种多特征组合集成的风电机组齿轮箱主轴承温度异常预测方法，抑制输入特征异常对主轴承状态判断的影响，降低模型误报率。首先对数据采集与监视控制系统(SCADA)历史健康数据预处理，根据Spearman相关性选择两组与齿轮箱主轴承相关的测点作为特征组合，然后通过训练、验证和测试建立2个反向传播神经网络(BP)模型用于评估主轴承温度，提取各模型对应的温度残差阈值和平均滑动窗口步长用于实时数据分析，最后通过2个模型的温度残差综合判断主轴承状态是否异常。利用某海上风电场工程实例数据验证模型性能，结果表明主轴承温升故障时，所提方法相对传统SCADA温度报警阈值，提前51.5小时判断主轴承温度异常；发电机绕组温升故障时，准确提示存在异常的特征组合，避免主轴承温度异常误报。
The main bearing of a wind turbine gearbox operates under high-speed rotation and heavy load conditions, and its temperature abnormality may lead to overheating or increased wear of the bearing, which in turn affects the equipment life. Aiming at the problem that the existing prediction methods do not fully consider the state of the equipment measurement points where the input features of the model are located, this paper proposes a multi-feature combination and integration method for predicting the temperature abnormality of the main bearing of wind turbine gearboxes, to inhibit the influence of the input feature abnormality on the judgment of the state of the main bearing, and to reduce the false alarm rate of the model. Firstly, preprocess historical health data from the Data Acquisition and Supervisory Control System (SCADA), select two sets of measurement points related to the main bearing of the gearbox as feature combinations based on Spearman’s correlation, then establish two Back-Propagation Neural Network (BP) models through training, validation, and testing for evaluating the main bearing temperature, extract temperature residual thresholds and average sliding window step sizes from each model for real-time data analysis, and finally, comprehensively assess the main bearing status for anomalies based on the temperature residuals from the two models. The corresponding temperature residual thresholds and average sliding window steps of each model are extracted for real-time data analysis, and finally the temperature residuals of the two models are used to comprehensively determine whether the state of the main bearing is abnormal or not. An offshore wind farm project example data is used to verify the model performance, and the results show that when the main bearing temperature rises, the proposed method can judge the main bearing temperature abnormality 51.5 hours in advance compared with the traditional SCADA temperature alarm threshold; when the generator winding temperature rises, it can accurately prompt the