Research On Wind Turbine Gearbox Fault Diagnosis Method Based On Deep Learning

With the rapid development of China’s wind power industry,the monitoring and fault diagnosis of wind turbine operation status can effectively reduce the operation and maintenance costs caused by fault downtime.As an important component of wind turbines,gearboxes can bring serious economic losses if they fail.Sparse effective features exist for the vibration signals of wind turbine gearboxes.The traditional non-probabilistic diagnosis model has the problem of poor fault feature extraction and identification when monitoring massive data,and the diagnostic process is highly deterministic.Therefore,in this paper,a Bayesian Parallel Deep Learning(Bayesian PDL)model with dynamic framework probabilistic based on deep learning theory is created for wind turbine gearboxes for fault diagnosis.The proposed model achieves adaptive matching of the network structure with the fault data,which allows the fault feature recognition to be improved.Meanwhile,the network hyperparameters are probability-making the output diagnostic results with high confidence.The main research of this paper is as follows.(1)The research background and significance of fault diagnosis methods for wind turbine gearboxes are introduced.The current research status of intelligent fault diagnosis methods at home and abroad is summarized.Planetary gearboxes of wind turbines as the object of study,the composition structure of the bearings and gears in the planetary wheel system and the reasons for the occurrence of fault signals are analyzed.To solve the problem that the acquired signal has high noise and low fault features,a differential continuous wavelet transform(DCWT)fault signal preprocessing method is proposed.Reducing irregular signal fluctuations using differential transform.The resolution of time-frequency features is improved by continuous wavelet transform.Through experiments,it is verified that the DCWT method can make the pre-processed signal have the advantages of both high time and frequency resolution.(2)The traditional intelligent fault diagnosis model of wind turbine gearbox has the characteristics of fixed network structure,which leads to the problem of poor ability of digging fault features.Thus,a Parallel Deep Learning(PDL)framework for fault diagnosis model is proposed.The frame structure of the proposed model can be changed dynamically depending on the input data features,which achieves adaptive matching of data features with the frame structure.Meanwhile,an attentional feature fusion mechanism is proposed within the network framework.This mechanism focuses the network attention on the recurring fault features.The fusion of fault features with high attention values can improve the fault feature recognition capability of PDL.Through experiments,the proposed model can better match data and enhance fault features,and its diagnostic accuracy can reach 99.91% ± 0.0054,which has stronger feature recognition ability and higher diagnostic performance than traditional intelligent diagnostic models.(3)The traditional intelligent diagnosis model has a high degree of determinism,which makes the fault diagnosis results of wind turbines have low confidence.Based on the dynamic PDL framework,a probabilistic Bayesian PDL fault diagnosis model is proposed.The proposed model converts the network hyperparameters from traditional point estimates into a probability distribution expression,quantifying the uncertainty in the network during fault diagnosis.The problem of highly determined hyperparameters of the network model leading to low confidence output of the results is solved.The experiments showed that the fault diagnosis accuracy for wind turbine gearbox bearings and gears was 99.44% ± 0.0101 and 99.10% ± 0.0034,respectively,while the uncertainty distribution of the diagnosis accuracy was the highest.The Bayesian PDL fault diagnosis model has better data fitting and feature identification ability than the traditional non-probabilistic model,and higher uncertainty and robustness than the probabilistic diagnosis model.