Research On Fault Diagnosis Methods For Large Wind Turbine Generators

With the gradual advancement of industrialization,the development of mechanical equipment is gradually moving towards large-scale,high-speed,and refinement.Wind turbines are experiencing rapid development as the main machinery and equipment for converting wind energy into electricity.However,this progress has also led to an increase in unit failure problems.Improper maintenance and failure problems can affect the direct economic benefits of wind farms and even threaten maintenance personnel’s lives.As the main mechanism for energy conversion in a wind turbine,the wind turbine generator plays an irreplaceable role.Its health status is crucial for the normal and stable operation of the unit.Therefore,a fault diagnosis study is conducted for large doubly-fed wind turbine generators as follows:(1)Large wind turbines were explained to be wind turbines with an installed capacity of 1.5MW and above 1.5MW.Secondly,wind turbines were classified based on different wind turbine drive structures.Then,the working principle of doubly-fed generators was introduced.Finally,several types of mechanical faults and electrical faults common to generators were introduced through the perspectives of fault mechanism and vibration characteristics,respectively,to lay the groundwork for wind turbine generator fault diagnosis research.(2)A wind turbine rolling bearing fault diagnosis method based on S-transform,Convolutional Neural Networks(CNN),and Bidirectional Gated Recurrent Units(Bi GRU)was proposed to address the problem of nonlinearity and non-smoothness in the vibration signal of wind turbine rolling bearings,which makes it difficult to extract fault features.Since the vibration signal is time series data and contains fault information at t moments and t±1 moments,the S-transform was used for time-frequency analysis of the vibration signal.This transforms the one-dimensional vibration signal into a two-dimensional time-frequency map,solving the problem of poor single-domain analysis.The time-frequency map was then input into the constructed Convolutional Bidirectional Gated Recurrent Units(CBi GRU)network model,including CNN and Bi GRU layers.The model sequentially extracts spatial features and positive and negative temporal features from the time-frequency diagram to enrich the extracted fault features.Finally,the experimental results are classified using the Softmax function.Comparing the results with traditional deep learning methods,the diagnostic methods used in this chapter achieve the highest accuracy rate.Different time-frequency transformation methods or different network models are used,and the accuracy rate of the test set reaches 93.17%.(3)In actual wind turbine diagnosis,the signal background noise is often strong,and the weak fault features were frequently submerged in the complex and variable background noise.Additionally,the use of fixed convolutional kernel sizes in the convolutional operation may result in the loss of local features.To address these issues,a wind turbine rolling bearing fault diagnosis method is proposed,which combines Singular Value Decomposition(SVD),S-transform,and Multiscale Convolutional Neural Network(MSCNN).Firstly,the vibration signal is reconstructed by applying singular value decomposition and selecting singular value curvature spectrum to reduce signal noise.Secondly,the time-frequency analysis of the reconstructed signal for each state is performed to construct a time-frequency diagram dataset for diagnosing wind turbine generator rolling bearing faults.Then,the training set is fed into the multi-scale convolutional neural network,and the local and global fault features of the time-frequency map were extracted in a parallel multi-scale manner using convolutional kernels of different sizes,enabling more comprehensive feature extraction for fault diagnosis.Finally,a comparative experimental study is conducted by selecting different time-frequency analysis methods and various network models.The visualization and accuracy diagrams are used to illustrate that the diagnosis method used in this chapter is superior to traditional fault diagnosis methods.(4)The feasibility of diagnosing electrical faults of generators using vibration signals was demonstrated in this chapter through both theoretical analysis and practical verification.When electrical faults occur,the vibration characteristics of the corresponding time and frequency domain spectrum were analyzed.Subsequently,a multi-scale convolutional neural network based on one-dimensional signals was proposed to diagnose several types of typical generator faults using vibration signals.The comparison test results show that different types of faults can still be effectively classified using vibration signals,with a diagnosis accuracy of 92.7%,which is higher than traditional diagnosis methods.The superiority of the proposed method in this chapter is reflected through a comparative study conducted using visualization methods.