Data-Driven Federated Fault Detection Of Wind Turbines

As a green,pollution-free and clean energy,wind energy has received increasing attention from countries around the world in recent years.As an important equipment for wind power generation,wind turbines(WTs)are prone to failures due to their harsh operating environment and variable working conditions.Actively carry out research on real-time fault detection technology for WTs is of great significance to reduce the failure rate and operation and maintenance cost of WTs.As a pre-installed monitoring system of WTs,Supervisory Control and Data Acquisition(SCADA)system provides massive and rich high-dimensional and multi-variable operation monitoring data.SCADA data-driven fault detection of WTs has become a current research hotspot.Existing fault detection methods are mainly based on the centralized paradigm,which needs to collect data from different WTs for modeling.However,due to data privacy and security protection among wind power operators or manufacturers,data sharing is often difficult,and there is a typical problem of data islands.In order to solve this problem,the actual scenarios faced by the same and different types of WTs in actual wind farms are analyzed,and a new federated deep learning framework and an efficient federated transfer learning framework are respectively proposed with federated learning and federated transfer learning as the core technologies.Meanwhile,in order to capture the spatial correlations of multi-variable SCADA data and solve the limitation of local WTs computing resources and the communication bandwidth with the cloud,a new multiscale residual attention network and a lightweight multiscale neural network model are designed respectively.The main work of the paper is summarized as follows:(1)Research on federated fault detection method for WTs based on independent and identically distributed data.In view of the problem of data islands faced by the traditional centralized data monitoring paradigm,based on the unique distributed learning advantages of federated learning,combined with the excellent performance of deep learning in nonlinear deep fault features learning,while considering the independent and identically distributed(IID)characteristics of data of the same type of WTs in actual wind farm,a federated deep learning fault detection framework for wind turbines(Deep Fed WT)is proposed,which aims to solve the problem of data islands and achieve accurate detection of WTs faults.Considering the spatial coupling correlations between multiple variables in SCADA data,a multiscale residual attention network(MSRAN)model is proposed,which uses a multiscale residual network and an attention module to extract multiscale important spatial features between different sensor variables,and the proposed MSRAN model and Deep Fed WT framework are validated on two real SCADA datasets.(2)Research on federated fault detection method for WTs based on lightweight model.In view of the limited hardware resources and communication bandwidth faced by the distributed federated paradigm that local WTs in wind farms are mostly located in remote areas and the huge amount of parameters of traditional deep learning model,a new light multiscale neural network(LMSNN)model is designed,which introduces depthwise separable convolution,aiming to reduce the trainable parameters of the local model in the distributed federated framework,and uses Inception v1 to expand the network structures in the width direction and different convolution kernel to enhance the generalization of the model,and maintain strong feature extraction ability while reducing the cost.The effectiveness of the proposed LMSNN model is verified by the real data of direct-drive WTs,and its performance under the distributed framework is compared and analyzed.(3)Research on federated transfer fault detection method for WTs based on non-independent and identically distributed data.In view of the non-independent and identically distributed(Non-IID)characteristics of data of the different types of WTs in actual wind farms,and the domain shift phenomenon caused by the complex and changeable dynamic operating conditions of WTs and the influence of environmental factors in the actual industry,combining federated learning and adversarial learning,a new efficient federated transfer fault detection framework(EFTLWT)is proposed,which aims to bridge the domains gap while protecting data privacy.At the same time,in order to improve the communication efficiency between the local WTs and the cloud server,a partial aggregation strategy is proposed,which only needs to aggregate partial model parameters to reduce the transmission of model weight,thus further reducing the communication bandwidth burden and speeding up the response of the monitoring system.Finally,the performance evaluation of the proposed EFTLWT framework and partial aggregation strategy is conducted through the analysis of the wind farm measured SCADA data.