Wind turbine fault diagnosis and early warning are important to reduce wind farm operation and maintenance costs and improve power generation efficiency. In this paper, we take the Supervisory Control and Data Acquisition (SCADA) data as the research object and research wind turbine health data purification, fault diagnosis model building, and unit operation status monitoring from a completely data-driven perspective. Firstly, for the problem that Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm cannot identify high-density anomalous data. An anomaly data processing scheme combining a density clustering algorithm and normal power interval estimation is proposed. The accuracy of extracting health data from wind turbines is improved. Secondly, to address the problem that the eXtreme Gradient Boosting (XGBoost) algorithm has more hyperparameters, we propose an optimization scheme based on the Bayesian Optimization Algorithm (BOA) and tree model for feature weight measurement, which improves the efficiency and accuracy of intuitive mapping from SCADA system monitoring data to fault features. Finally, a wind turbine condition monitoring scheme based on the information fusion of multi-characteristic monitoring parameters is designed. The wind turbine condition monitoring scheme proposed in this paper can warn generator system failure 3.67 hours, gearbox system failure 5.17 hours in advance, and hydraulic system failure 2.33 hours in advance.

In order to resolve the contradiction between the rapid growth of wind turbines installed capacity and the lagging operation and maintenance technology, this article uses supervisory control and data acquisition (SCADA) system data to realize wind turbines fault diagnosis and early‐warning research. First, this article proposes a method for cleaning abnormal data in SCADA systems. It combines density‐based spatial clustering of applications with Noise (DBSCAN) and median absolute deviation (MAD) to estimate the normal power range, effectively removing outliers from the wind‐speed‐power curve and providing high‐quality health data samples for further research. Then, for wind turbines fault diagnosis, this article designs a lightweight gradient boosting machine (LightGBM) model based on a genetic algorithm (GA). This model achieves an overall fault diagnosis accuracy of 97%, effectively diagnosing different types of faults in wind turbines. Finally, this article proposes a multivariate state estimation technique (MSET) normal model based on the DBSCAN algorithm for fault warning in wind turbines. This article designs two threshold‐setting methods: fixed threshold and adaptive‐dynamic threshold. For gearbox faults, the fixed threshold method provides warnings 55 min in advance, while the adaptive dynamic threshold method advances the warning time to 68 min. For generator faults, the warning times are 134 and 148 min, respectively. Fault warning is conducted by calculating the residuals between predicted and actual values.

In order to resolve the contradiction between the rapid growth of wind turbines installed capacity and the lagging operation and maintenance technology. This article uses Supervisory Control And Data Acquisition (SCADA) system data to realize wind turbines fault diagnosis and early warning research. First, this article proposes a method for cleaning abnormal data in SCADA systems. It combines Density-Based Spatial Clustering of Applications with Noise (DBSCAN) and Median Absolute Deviation (MAD) to estimate the normal power range. It provides high-quality health data samples for further research, based on the results obtained. Then, for wind turbines fault diagnosis, this article designs a Lightweight Gradient Boosting Machine (LightGBM) model based on a Genetic Algorithm (GA). This model can effectively diagnose different types of faults in wind turbines. Finally, this article proposes a Multivariate State Estimation Technique (MSET) normal model based on the DBSCAN algorithm for fault warning in wind turbines. This article designs two threshold setting methods: fixed threshold and adaptive dynamic threshold. Fault warning is conducted by calculating the residuals between predicted and actual values.