Early Fault Warning for Gearboxes in Real-world Wind Turbines Using Spatio-Temporal Causal Reasoning Network

The gearbox of a wind turbine is essential to ensure the safe operation of wind power generation. However, existing fault warning methods often struggle to effectively and timely detect potential faults, resulting in insufficient reliability. In this paper, we proposed a Spatio-Temporal Causal Reasoning Network (STCRN) for early fault warning of gearboxes. First, multi-level feature extraction was performed on the raw vibration signals of the wind turbine. A causal graph was constructed by inferring dependencies among sensors, informed by correlation analysis and physical principles. Additionally, the health indicator was developed. Second, a Spatio-Temporal analysis module was designed to analyze the stability of time series and the topological characteristics of the spatial causal network, as well as to capture abnormal propagation patterns in the Spatio-Temporal dimensions through dual analysis. Third, the detection results were enhanced by confidence based on causal evidence, improving the accuracy of fault localization. This module focuses on root cause identification and detection confidence enhancement. Finally, intelligent fault detection, precise localization, and graded warnings based on causal-enhanced probability were used to provide actionable guidance for operational and maintenance decisions. The proposed method was tested on a real-world wind turbine in China during 2021. The experimental results showed that the method not only accurately locates faulty components but also achieves reliable fault warnings, with an Anomaly Rate and Fault Probability of 1.000 and 0.950, respectively. Its warning performance outperforms existing methods such as Isolation Forest, One-Class SVM, Local Outlier Factor, and XGBoost.