Research On Fault Diagnosis Method Of Wind Turbine Drive System Based On Regularized Machine Learning

With the decline of traditional fossil energy and the increasing global warming problem,wind power as a representative of clean green energy has been highly valued and vigorously developed.The capacity of machine assembly machines is increasing,and the proportion of wind power in the power market share.It is also growing.Wind turbines are usually built in harsh environments such as remote mountainous areas,offshore areas,Gobi,etc.The high failure rate of the transmission system is a key component of wind energy conversion.Whether reliable operation determines the safe operation of wind power plays an important role.During the power generation process,the time-varying wind speed and load continue to impact the transmission of the wind turbine,which makes the failure rate of the transmission high.The high maintenance cost and downtime loss not only seriously restrict the development of wind power,but also the large-scale wind power grid connection will greatly affect the stability and safety of the power system.Therefore,effective state detection and fault diagnosis are essential to ensure the safety and reliability of wind turbine operation and reduce maintenance costs.With the in-depth development of intelligent pattern recognition theory,the fault diagnosis method of wind power transmission system based on machine learning method has been developed.In the face of fault diagnosis of wind power transmission system under complex conditions,there are disadvantages such as over-fitting,complex model,poor generalization,poor interpretability,and difficulty in extracting effective features.Therefore,this paper combines regularization theory with machine learning to propose a fault diagnosis method for wind turbine drive system based on regularized machine learning.This paper takes the wind turbine drive system(rolling bearing and gearbox)as the research object,and conducts in-depth research on the single failure of the rolling bearing,the composite failure of the gearbox and the remaining service life of the rolling bearing.The main research contents of the thesis are as follows:(1)Rolling bearings,as key components of wind turbines,play a decisive role in the safe operation of the entire unit.Aiming at the problem of fault diagnosis of rolling bearing of unit,this paper proposes a fault diagnosis method for node-optimized directed acyclic graph large interval distributor(O-DAG-LDM).Combining the advantages of DAG multi-classification expansion performance and LDM two-classifier generalization performance,this paper constructs a DAG structure extended LDM multi-classifier method for rolling bearing fault diagnosis.Inthe framework of DAG-LDM algorithm,the optimization algorithm is used to optimize the DAG nodes to reduce the cumulative error caused by random arrangement and improve the accuracy of LDM fault classification.(2)Gearbox is a device that realizes shifting by meshing of large and small gears.Due to its special structure,its faults are mostly compound faults.It is difficult to extract effective fault features by traditional diagnostic methods,and there is poor anti-interference ability and false positives.High rate issues.In this paper,a fault feature extraction method based on regularized self-encoding neural network is proposed.Combining the feature extraction advantage of self-encoding neural network with the sparse selection feature of Lasso regular term,the penalty can be effectively applied to the self-encoding neural network through Lasso term.Self-encoding neural networks extract key features in composite faults.The extracted features are classified by the support vector machine to achieve the diagnosis of the composite fault of the gearbox.(3)Efficient residual life prediction is essential to ensure the safety and reliability of machine operation and reduce maintenance costs.In order to improve the prediction accuracy of rolling bearing residual life(RUL),a new method of rolling bearing residual life prediction based on elastic net and long and short time memory network is proposed by combining elastic network with long and short time memory network(LSTM).It is called E-LSTM.The LSTM multi-memory cell characteristics are used to fully exploit the spatio-temporal information in the process of bearing degradation,and the elastic network regular term is introduced to control the LSTM structure to overcome the over-fitting problem in the LSTM neural network training process.Experiments show that the method can well characterize the output,better represent the bearing degradation mode,and achieve good residual life prediction accuracy.