Research On Fault Early Warning Method Of Wind Turbine Gearbox System Based On Graph Convolutional Neural Network

The speed increase gearbox is one of the most vital systems of doubly-fed asynchronous wind turbines,once a failure occurs,it will lead to a decrease in the power generation efficiency of the unit,interruption of power transmission and even system shutdown,resulting in huge economic losses.Although the units currently in operation are basically equipped with SCADA systems,which can provide alarms for some faults,their alarm timeliness is not strong,and the general fault is already very serious when the alarm is issued.In order to be able to detect the failure of the gearbox system in time,it is necessary to carry out intelligent fault warning technology research.At this stage,the gearbox fault diagnosis method based on SCADA data mainly focuses on the prediction of a single variable of the gearbox oil pool temperature,and lacks the description of the whole gearbox system,this paper aims to establish the overall graph structure of the gearbox system,describe the flow relationship of energy,matter and information between the measurement points in the gearbox system,and realize the reasonable prediction of parameters under the coupling of multiple components.Major research contents of this thesis are as fellows:(1)Establish a gearbox system fault tree.The operation control method of wind turbine,especially gearbox cooling and lubrication system,is analyzed,the typical faults of gearbox systems and fault characteristics under typical faults are discussed based on the actual data on site,and the fault tree of gearbox system is established by using the fault tree analysis method,which provides a basis for subsequent model fault identification and location.(2)Improve the data preprocessing and indicator calculation process.The characteristics of the theoretical wind speed-power curve under different wind speeds were discussed,and the bad data were screened out according to this law,and the quartile method was used to screen out the abandoned wind curtailment data.Component condition metrics based on fault characterization are also defined.(3)Gearbox system fault warning based on self-attention mechanism.The short-time series samples at each moment are constructed by data sliding window processing,the time information in the short-time series samples is learned by the self-attention mechanism,the value of the target measurement point at the next moment is predicted by aggregating the working condition parameters,the model is built for all key parameters respectively,and the prediction results of multiple key parameters are compared with the measured values to calculate the Euclidean distance between the two to obtain the health status of the components and realize fault warning.(4)Gearbox system fault warning based on graph convolution.Considering the complex coupling of each measurement point in the gearbox system,this paper proposes to use the graph abstract information transfer relationship between the measurement points,and the graph structure is improved by calculating the physical structure of the gearbox and the importance of datadriven parameters,and the spatial position information of the working conditions and related measurement points is extracted by combining the graph convolutional neural network to optimize the prediction effect of the measurement points.(5)Gearbox system fault warning with spatiotemporal feature fusion.A fault early warning model framework based on spatiotemporal feature fusion is proposed,and the parameter prediction effect of normal operation stage and fault operation stage under multi-component coupling is optimized by using temporal self-attention and graph convolution to aggregate temporal information and spatial position information,respectively.