Study On Intelligent Monitoring Methods For Operating Condition Of Main Bearings Of Large-scale Direct-driven Wind Turbines Based On SCADA Data

As a clean,low-carbon,green and renewable energy source,wind energy has attracted the attention of many countries and the energy industry.In recent years,more and more large wind farms and large-capacity wind turbines have been put into use one after another.However,due to the harsh operating environment and complex operating conditions of wind turbines,some core components of wind turbines frequently fail,and the O&M(Operation and Maintenance)costs of wind farms have increased sharply.This seriously restricts the development of wind power technology and the wind power industry.Therefore,to ensure the stable and reliable operation of wind turbines,the research on the operation condition monitoring method of the core components of large-scale wind turbines not only has important theoretical significance but also has important engineering application value.Based on the wind turbine operation data collected by SCADA(Supervisory Control and Data Acquisition,SCADA)system,this paper takes the main bearing of large-scale direct-driven wind turbines as the research object and carries out the research on the construction of intelligent monitoring model for wind turbine operating condition,operating condition parameter prediction,operating health condition identification,and modeling and realization of online intelligent monitoring,etc.,aiming to provide theoretical basis and technical support for the operating condition monitoring and warning of large-scale direct-driven wind turbines.The main research of this paper is as follows:(1)Aiming at the problems that the physical mechanism of wind turbine operating condition parameters is complex,and the complex nonlinear and non-stationary characteristics between parameters are difficult to model,a new method for constructing an intelligent monitoring model of wind turbine operating condition is proposed,which not only breaks through the learning capability limitations of shallow machine learning models but also solves the problems of network over-fitting and model parameter adjustment difficulties when the model depth increases.First,the specific component to be studied is determined,and combined with its physical mechanism,the relevant SCADA parameter data related to the condition change of the specific component is extracted,and its condition feature vector is initially constructed.Secondly,the basic intelligent building blocks are selected according to the characteristics of the condition parameter data,the framework of the intelligent model and the corresponding training algorithm are designed and constructed,and the deep features inside the input data are learned through training samples.Thirdly,the model is trained and evaluated according to the evaluation index,and the structure and hyper-parameter of the model are determined.Finally,a Stacked Sparse Autoencoder with Multi-Layer Perceptron(SSAE-MLP)intelligent model is designed,and the feasibility of this kind of intelligent model construction method is verified by comparative analysis of wind turbine vibration data.It has superior performance that other shallow machine learning models do not have.(2)Aiming at the problems such as inaccurate temperature prediction of wind turbine main bearings,false alarms and missing alarms,etc.,a temperature prediction method(Stacked Long-Short Term Memory with Multi-Layer Perceptron,SLSTM-MLP)consisting of multiple long-term short-term memory units and a multi-layer perceptual regression layer is proposed,which solves the problems of insufficient mining of temperature time series features and time series features between parameters,difficulty in determining the model structure and hyper-parameters,and poor reproducibility of the model.First,multiple condition parameters are combined into a feature matrix to construct a multivariate time series dataset.Secondly,with a single LSTM unit as the basic component,multiple LSTM units are stacked to build a deep model with a perceptual regression layer to mine the nonlinear and non-stationary dynamic characteristics between the main bearing temperature itself and its related condition parameters.Thirdly,the performance of the proposed model is evaluated by analysis from three aspects of different sample sizes,different sampling times and different sampling frequencies.Finally,by simulating faults,it is verified that the proposed model has better performance in predicting the main bearing temperature of large-scale wind turbines.(3)Aiming at the problems such as complex operating conditions,numerous influencing parameters,and strong nonlinear parameters that make it difficult to identify the operating health condition of the main bearing,a method for identifying the operating health condition of the wind turbine main bearing based on dual attention mechanism and bidirectional long short-term memory model(Dual Attention-Based Bi-LSTM,DA-Bi-LSTM)is proposed,which solves the situation that different operating conditions have different control strategies,different operating condition parameters have different contribution degrees,and condition parameters have different influence degrees at different historical moments during the identification process of the main bearing operating health condition.First,two attention calculation modules are designed to extract important features of different input parameters and important features of input parameter time series respectively.Secondly,the two extracted feature information are fused,and Bi-LSTM building blocks are used to perform forward and backward feature extraction on the fused information.Thirdly,the extracted features are used to reconstruct the input data.Finally,compared with the Bi-LSTM model without adding the attention module and the Bi-LSTM method considering only a single attention mechanism,it is verified that the proposed method has better performance and stability,and during the monitoring process,the proposed model has clearer and better interpretability.(4)Aiming at the problem that direct application of the existing intelligent model will lead to low prediction and recognition accuracy due to the difference in wind resource characteristics of the target wind turbine and the lack of data samples in the process of offline intelligent modeling and online condition monitoring of wind farms,this paper applies transfer learning theory,combines offline modeling knowledge with target domain learning,proposes a condition parameter temperature prediction model and an operating health condition identification model,and online intelligent monitoring software for the main bearing condition of large-scale direct-driven wind turbine is designed and developed.First,the principles of parameter prediction and condition recognition based on transfer learning are given,and their models are designed.Secondly,the overall structure of the software is given,and each functional module of the software are designed in detail.Finally,the design and implementation of interfaces are given.The software is developed under the Windows platform using technologies such as Python language,Py Qt5,Echart,Matplotlib,Pyside2,Pyodbc and Mysql database.The design and development of online intelligent monitoring software have realized the intelligent modeling method of vibration monitoring,main bearing temperature prediction and early warning,main bearing operating condition health identification and early warning,and further promote the research work for engineering applications.