A Weighted Multi-Scale Global-Local Feature Fusion Framework for Cross-Turbine Diagnosis of Wind Turbine Faults

Data-driven approaches for diagnosing wind turbine faults have gained significant popularity, particularly due to the promising diagnostic performance of deep learning models. However, training fault diagnosis models for every wind turbine is both time-intensive and resource-demanding. Therefore, for wind turbines with limited data or applications with limited computational resources, transfer learning becomes crucial. To tackle these challenges, we propose a weighted multi-scale global-local feature fusion network (WMGLFFN) framework, which is comprised of two parts. The first part utilizes multi-scale wavelets and a spatio-temporal convolutional neural network to learn global and local features. The second part is a multi-task learning module, which learns invariant features for different turbine domains using a dynamically weighted domain adversarial network. Meanwhile, in order to enhance the feature discrimination ability, we introduce deep metric learning. We used four existing networks to compare with the proposed method and validated it with SCADA data from two wind turbines. Our WMGLFFN significantly outperformed traditional deep learning methods. The cross-validation experiments performed on two turbines prove the effectiveness of the proposed method.