• Transformer encoder detects gearbox faults using SCADA data only • Normal behavior learned from one turbine, then adapted to others • Kendall’s Tau used for SCADA variable selection related to the gearbox • Fault detection achieved months in advance with no labeled fault data • Validated on real-world wind farm data with high accuracy Gearbox failures remain one of the most critical and costly issues in wind turbine operation, often leading to extended downtime and increased maintenance expenses. This paper presents a novel methodology for the early detection of gearbox faults using Supervisory Control and Data Acquisition (SCADA) data and a fine-tuned Transformer encoder model. Unlike traditional condition monitoring systems, which rely on high-frequency sensors or require extensive labeling, the proposed approach uses the existing low-frequency SCADA infrastructure and employs a semi-supervised learning strategy. A normal behavior model is first pre-trained on healthy data from a single wind turbine and subsequently fine-tuned to adapt to the operational characteristics of additional turbines in the same wind farm. Key contributions include an optimized variable selection process based on Kendall’s Tau correlation, data imputation and smoothing techniques to address missing values and noise, and a fault prognosis indicator derived from model residuals. Results obtained with real SCADA datasets demonstrate the effectiveness of the method in detecting incipient gearbox faults under realistic operating conditions.
España et al. (Sun,) studied this question.