Reliable multi-point temperature monitoring and fault localization for wind turbine stator windings through feature-decoupled multi-task LSTM

Monitoring the distributed thermal state of generator stator windings is critical for preventing catastrophic insulation failures in wind turbines. As wind farms scale up, multi-task learning is increasingly adopted as an alternative to independent models, significantly reducing the model-maintenance burden of monitoring wind farm clusters. However, traditional multi-task monitoring models are vulnerable to “lateral contamination”, which leads to ambiguous fault localization when a specific winding overheats. This paper develops the feature-decoupled multi-task long short-term memory (FD-MTL-LSTM) network, an intelligent monitoring framework that employs a feature-decoupled data flow to isolate cross-sensor interference. By strictly separating shared environmental features from private historical temperature sequences, the model ensures that anomalies remain encapsulated within the corresponding prediction task. Experimental results based on high-capacity offshore wind turbine data show that the model not only maintains high precision under normal conditions but also exhibits exceptional diagnostic robustness. Notably, for a documented thermal runaway event, the proposed model identified the specific faulty winding with a lead time of over 35 hours, whereas traditional multi-task approaches failed to pinpoint the location accurately and generated cascading false alarms. The results demonstrate that the FD-MTL-LSTM framework effectively enhances the reliability of thermal condition monitoring, offering significant potential for reducing unplanned downtime and maintenance costs.