DeepFedWT: A federated deep learning framework for fault detection of wind turbines

Abstract

Data-driven fault detection of wind turbines has gained increasingly attention. Currently, most existing methods require sufficient labeled data to train a reliable model in a centralized way. However, it is difficult to collect enough labeled data due to data privacy and strict confidentiality of wind farm owners. To this end, we propose a federated deep learning framework (DeepFedWT), which allows multiple decentralized WTs to collaboratively build a fault detection model using their local private data. Specifically, we designed a multi-scale residual attention network (MSRAN) model to extract informative features from raw multivariate sensor data, which first integrates a multiscale residual learning block to extract spatial features among different sensor variables at multiple scales and adopts a feature attention block to highlight important features highly associated with faults, and finally enables an enhanced fault detection. Experimental results on two real WT datasets demonstrate the effectiveness of our proposed DeepFedWT framework.