Research On Fault Simulation And Diagnosis Of Wind Turbine Gearbox

As one of the typical examples of renewable energy,wind energy can alleviate the shortage of oil resources in China and the adverse impact of other traditional power generation methods on the environment,therefore,wind power generation has received high attention and rapid development.Due to the harsh working environment,complex and variable loads,and unstable quality of wind turbines(WT),WT are extremely prone to various types of failures.Gearbox failure is one of the faults with a high incidence,and the resulting WT downtime and high maintenance costs have a serious impact on economic benefits of wind production.This paper applies machine learning algorithms for fault diagnosis of common gear faults in gear boxes that are difficult to diagnose early.By optimizing the parameters of the algorithm to optimize the diagnostic model,an effective algorithm for early diagnosis of gear faults is presented based on a multifaceted comparison of the classification effects of the diagnostic models.The main work contents are as follows:(1)Based on SIMPACK,a rigid-flexible coupled WT model with simulation capabilities such as WT aerodynamic loads,wind power generation control,and gear meshing was built to simulate five gear states,namely,normal,low-speed level sun gear corrosive pitting and wear,and high-speed level pinion corrosive pitting and wear.The conventional WT operation parameters and shaft vibration signals are collected under each state as multi-source data.Multiple eigenvalues are extracted from the data using sliding windows in the time and frequency domains,then the extracted high-dimensional eigenvalues are filtered using recursive feature elimination.Finally,10 eigenvalues are retained as the input feature set for the fault diagnosis machine learning algorithm.(2)Using the improved whale optimization algorithm(IMWOA),grid search,and the combination of the two methods,the parameters of various commonly used machine learning algorithms and ensemble learning algorithms are optimized.The classification effects of various machine learning algorithms are compared using the F1 value of each state in the classification result and the weighted F1 value based on the proportion of sample category numbers as evaluation criteria.Based on the evaluation results,it is found that SVM and CatBoost classification algorithms perform better than other algorithms.At the same time,the impact of multi-scale feature extraction and filtering on the performance of classification algorithms is summarized.(3)Aiming at the problem that the number of normal state samples for gearbox gears is far greater than that of fault state samples in actual fault diagnosis,SVM and CatBoost are respectively combined with oversampling,undersampling,and mixed sampling strategies.Among them,the SMOTEENN data enhancement strategy effectively retains and enhances the information of the small number class samples,reduces the interference of unrelated samples in most classes to the model.Combined with the strong classification model CatBoost,the classification effect gets best.Finally,the effectiveness and practicality of the SMOTEENN-CatBoost algorithm for early fault diagnosis of gears are further verified using experimental data from a gearbox fault simulation test rig.