Research On Methods In Prognostic And Health Management For Gearbox Of Wind Turbine

The wind turbines,as the key renewable energy conversion equipment,have played an important role in easing the energy crisis and the environmental pollution problem.Risk reduction,ensuring safety of wind turbine operation for keeping and improving the economics have become the important research directions.The gearbox in a wind turbine is the key mechanical parts in the transmission system,the reliability of which will influence the security and stability of the wind turbine operation.Thus,it is necessary to design prognostic and health management methods according to the gearbox.Ensuring the reliability of working condition and stabilizing the safety of wind turbine operation will improve the economics of operation and maintenance of wind farms.In addition,it is urgently needed to develop the reliable and efficient PHM methods for monitoring and managing the health of WT mechanical components in such a complex environment.For this purpose,this paper develops and studies the key techniques in PHM methods including the methods of fault diagnosis and remaining useful life estimation based on signal processing and machine learning.The features extraction is the key process in both fault diagnosis and remaining useful life prediction.This paper has proposed two adaptive algorithms respectively for fault diagnosis and prognosis based on variational mode decomposition and non-dominated sorting genetic algorithm-II with considering the differences of the purposes.Considering the defect that expert experience-based feature extraction relies too much on prior knowledge,Deep Learning(DL)is introduced to realize automatic feature extraction,and an end-to-end fault diagnosis model is established,and the influence of data fusion mode on the performance of the diagnosis model is studied in this paper.In addition,considering the limitations of supervised learning in the application of residual life prediction,a performance degradation index construction method based on unsupervised learning was proposed,and the reliability of DL model for RUL prediction was studied.The main research contents and conclusions are as follows:1.To extract reliable features from wind turbine bearing vibration signals that are nonlinear,non-stationary,and easily masked by mechanical/environmental noise.Considering the difference in purpose of feature extraction for fault diagnosis and life prediction,based on the variational modal decomposition algorithm,based on NSGA-II,a fitness function is established through information entropy and mutual information to suppress the over-decomposition phenomenon of VMD filtering and prevent it from occurring.A multi-objective optimization model is established,considering that the fault diagnosis feature should have a significant impact component;and the life predictionoriented feature should have monotonicity with the degradation time,so the filter is constructed with kurtosis and monotonicity to filter the Pareto solution set to form the Diagnosis Oriented Adaptive Variational Mode Decomposition(DOA-VMD)and Prognosis Oriented Adaptive Variational Mode Decomposition(POA-VMD)algorithms.The reliability and superiority of the proposed method are verified by bearing inner ring damage data and 2MW wind turbine high-speed shaft bearing degradation data set.2.To improve the generalization of the fault diagnosis model developed based on convolutional neural network in noisy environment.Combining the robustness of VMD algorithm in filtering and noise reduction with the superiority of CNN in feature extraction,a VMD-CNN fault diagnosis model is proposed.Study the influence of VMD parameter changes on the noise resistance of VMD-CNN model;through the visualization of VMD-CNN filtering features and try to explain the reason and mechanism of VMDCNN anti-noise performance changes;consider the impact of feature fusion and decision fusion on model performance,proposed a weighted majority voting method and constructed VMD-CNN-II.The study found that an appropriate decision fusion method can improve the generalization of the model when diagnosing faults in noisy environments.3.To improve the "end-to-end" fault diagnosis model based on DL,it lacks adaptability when dealing with the vibration signal of wind turbine gearbox due to the single feature input.A composite multi-scale(HMS)feature extraction module is established based on multi-scale coarse-grained and dilation convolution to extract time multi-scale and frequency-domain multi-scale information directly from vibration signals.Considering the contribution of different scale features to fault pattern recognition,an attention mechanism is introduced to perform weighted fusion of multi-scale advanced features in the frequency domain.The loss function solves the weights by gradient descent to avoid model bias.The reliability and superiority of the proposed method are verified by the experimental data of gearbox failure.4.For different data types,the model-based and data-based methods are used to carry out research on the estimation method of the remaining life of wind turbine bearings.Establishing a reliable degradation index can improve the accuracy of the residual life estimation of the exponential model.Therefore,POA-VMD is used to filter the vibration signal of the 2MW wind turbine bearing and build a degradation index.The residual life of the wind turbine bearing is estimated by the exponential degradation model,which is like the baseline method.The comparison proves the superiority of this method5.To avoid the dependence of the degradation index built on the data-driven model on the prior knowledge,a multi-scale resolution autoencoder is established,and the distance of the encoded feature is used as the degradation index,and the distance calculation method is used to study the degradation feature.Sensitivity and reliability effects.The influence of four kinds of deep neural networks on the reliability of establishing degradation index and remaining life label mapping was studied,and a bearing remaining life prediction method without any prior knowledge was established.The reliability of the proposed method is studied through the whole life data set,and the results show that this method can accurately estimate the remaining life of bearings under the conditions of autologous/alienated and variable loads.