Research On Condition Monitoring Methods Of Wind Turines Based On Graph Neural Networ

Wind energy,as one clean and renewable energy,has developed in large scale,large speed and large capacity in recent years.As a high-end equipment in the field of wind power,the safe,stable,efficient and reliable operation of wind turbine directly affects the economic benefits of the whole wind farm.However,as wind turbines are usually installed in remote areas,their operating environment is harsh and complex,leading to frequent failures.Therefore,it is of great practical significance and application value to actively carry out research on early fault warning and intelligent diagnosis of wind turbine,timely and accurately grasp the health status of wind turbines,to reduce the fault rate and operation and maintenance cost of wind turbines improve the reliability and utilization rate of wind turbines,and promote the healthy and sustainable development of the wind power industry.Wind turbine is a complex nonlinear strong coupling system with mechanical,electrical and hydraulic characteristics.Its internal structure is extremely complex,including rotor blade,pitch system,transmission system,generation system,yaw system and other subsystems,and different subsystems are closely related,interdependent,and there are many synergies.Supervisory Control and Data Acquisition(SCADA),as the pre-installed monitoring system of wind turbines,can provide state information and data that reflect the health condition of each subsystem of the turbines.It provides favorable data support for condition monitoring and fault diagnosis of wind turbine.However,SCADA data is essentially multivariable time series data with complex spatio-temporal coupling relationships between different sensor variables,and therefore it is difficult to model SCADA data.To this end,in order to fully explore the internal spatio-temporal correlation and dynamic time-variability of SCADA data,the graph neural network is taken as the core method,and the multi-variable spatio-temporal relationship modeling method based on graph neural network is mainly studied,and the wind turbine system health information and complex correlation contained in SCADA data are deeply explored.The condition monitoring model is established to realize the fault early warning and intelligent diagnosis of wind turbine.The main work of this paper is as follows:(1)Based on the analysis of wind turbine system composition and structure,the complex correlation and coupling relationships between and within each subsystem of wind turbine are sorted out.Then,the complex spatio-temporal characteristics and modeling difficulties of SCADA data are summarized from inside and outside.An idea of graph modeling oriented to SCADA data was proposed,in which the multi-variable monitoring data of SCADA was transformed into graph structure data containing sensors(nodes)and topological connection relationships(edges),which could accurately and comprehensively express the inherent spatio-temporal correlation characteristics of the data,laying a foundation for the subsequent research of condition monitoring models based on graph neural networks.(2)Considering the correlation and coupling between different subsystems of wind turbine and the highly spatio-temporal correlation of SCADA data,a Temperature Graph Neural Network(Temp GNN)based fault warning model of wind turbine was proposed.The purpose of this model is to model the temperature variables of each subsystem.Firstly,in order to eliminate the influence of environmental conditions on temperature change,a decoupling model based on fully connected neural network is proposed.Then,the decoupled temperature variable is input into the corresponding low-dimensional embedded representation learned through the gate recurrent unit,and the connection weights between different sensor nodes are adaptive learned through the self-attention mechanism to obtain the adaptive graph structure data of different sensors data.Furthermore,spatio-temporal features were extracted from the improved Chebyshev graph network,and the temperature prediction model of healthy condition was established.Finally,the residual analysis of model prediction was used to realize the early warning of faults.Two fault cases of gearbox and transformer in wind farm SCADA data are used to verify the effectiveness of Temp GNN fault prognosis model proposed,and its superiority is proved by comparing with the mainstream deep learning method.(3)Considering the fact that the target wind turbine lack fault data in practice,it is difficult to establish a reliable fault diagnosis model and the data distribution difference between different wind turbines in the wind farm,a cross-wind turbines fault diagnosis method based on Multitask Spatio-Temporal Graph Convolution Network(MTSTGCN)is proposed.In this method,the data of other wind turbines in the same wind farm is taken as the source domain and the data of the wind turbine to be studied for fault diagnosis is taken as the target domain.Firstly,dynamic graphs are constructed adaptively by the graph learning layer as input.In the feature extractor,Chebyshev convolution is used to extract spatial features and temporal convolution is used to extract temporal features.In view of the shortcomings of insufficient labeling data in engineering applications,domain adaptive transfer learning strategy is adopted to reduce the distribution difference between different WTs,and deep metric learning module is used to expand the distance between classes,so as to improve the accuracy of cross-wind turbines fault diagnosis.The validity of the proposed model is verified by using the SCADA data of two wind farms,and the interpretability of the designed graph learning layer is discussed.