Research On Fault Diagnosis Of Wind Turbine Based On CAE-XGBoost

In order to adhere to the sustainable development path of green energy and implement the set goals of "carbon peaking" and "carbon neutrality",China is accelerating the development and utilization of wind power.However,wind turbines have complex structures and harsh working environments,and once they fault,they will cause incalculable losses.As an important component in the wind turbine,bearings have a high failure rate.Once a fault occurs,it not only affects the normal operation of the wind turbine but also may cause chain reactions of other components.Therefore,the research on fault diagnosis of wind turbine bearings is of great significance.The traditional bearing fault methods mainly based on time-frequency analysis.This article adopts a data-driven approach,integrating signal processing,deep learning,and machine learning methods to identify bearing faults in wind turbines.The main research content is as follows:Firstly,due to the coupled motion of multiple mechanical components within the wind turbine,as well as the changes of transmission load and speed caused by uncertain wind speed,the vibration signal of the bearing has the characteristics of Nonstationarity,nonlinearity and high noise interference.Therefore,it is necessary to preprocess the vibration signal.The Empirical Wavelet Transform(EWT)method is a nonlinear signal processing method that faces problems such as inability to adaptively decompose and mode breakage.This paper proposes an improved IEWT(Improved Empirical Wavelet Transform)method based on the trend of spectral envelope.After decomposing the vibration signal by IEWT,the signal reconstructed by the fault components screened out using the weighted steepness index T,which can reduce noise interference.Compared with EWT and EMD methods,it shows that IEWT can better extract the shock component from the vibration signal.Secondly,extracting the representation features that can characterize vibration signals under different operating conditions can help realize pattern recognition of failures.Since traditional feature extraction methods require a large amount of prior knowledge,this paper adopts a Convolutional Autoencoder(CAE)deep learning network to adaptively learn the deep features of vibration signals.CAE combines supervised learning and unsupervised learning training methods to ensure the effectiveness and discrimination of extracted features.Compared with PCA dimensionality reduction and time-frequency domain statistics,the differentiation of signal feature extracted by the CAE model is better.Training the extreme Gradient Boosting(XGBoost)classification algorithm Optimized by particle swarm optimization by CAE feature extraction and construct a wind turbine fault diagnosis model.Finally,the actual bearing vibration data is used for case analysis.The method in this paper achieved 98.8%accuracy in pattern recognition of 10 bearing states,and has good recognition performance;Compared with PCA&SVM,time-frequency statistical feature&XGBoost,multi-scale permutation entropy&XGBoost and other methods,the superiority of this model has been verified.The robustness experimental analysis shows that the diagnostic method proposed in this paper also has certain anti noise and variable load adaptive capabilities.