Wind Turbine Gearbox Intelligent Fault Diagnosis Based On Electromechanical Multi-view Learning

Gearbox is a key component of wind turbine drivetrain system,and its operating state directly affects the operating efficiency of the whole system.Gearbox has complex internal structure and often operates in harsh environments where the load is constantly changing,and as a result,it is subject to various failures.Once the failure occurs,it will lead to shutdowns and therefore result in high maintenance costs and loss of power generation.Therefore,timely and accurate fault diagnosis of gearbox has important practical significance and practical application value.However,most of the existing fault diagnosis methods for wind turbine gearboxes mainly rely on a single signal source,such as vibration or current.Due to the limited information contained in a single signal,it is difficult to accurately and comprehensively describe the health state of the gearbox,which leads to the low reliability and accuracy of fault diagnosis.Taking the wind turbine gearbox as the research object and considering the electromechanical coupling characteristics of wind turbine drivetrain system,this thesis focuses on studying intelligent fault diagnosis methods based on electromechanical multi-view learning for wind turbine gearboxes.Instead of using a single signal,the proposed methods regard the gearbox vibration signal and generator current signal as two different and related monitoring views from the prospective of multi-view learning,and aim to capture correlated and complementary diagnosis information between vibration and current signals to improve the fault diagnosis performance of the single signal,and finally realize high precision classification diagnosis of gearbox.The main research contents are as follows:Firstly,the common fault types and causes of gears and bearings in a wind turbine gearbox were first analyzed and summarized.Then the principle of gearbox fault generation and the basic principle of gearbox fault detection based on current signals and vibration signals were analyzed in detail.Finally,the experiments on gears and bearings are designed and performed and two gearbox fault datasets containing multi-view electromechanical monitoring information under different health conditions of gears and bearings are built.Secondly,considering the correlated and coupling characteristics between gearbox vibration signals and generator current signals,a multi-view canonical correlation feature learning and fusion method is proposed to learn the correlated features between current and vibration signals.The original current and vibration signals are first decomposed with the multilevel wavelet packet transform(WPT)and then statistical features of each wavelet coefficient at different decomposed levels are calculated and used as input of the canonical correlation analysis(CCA)network to learn the maximum correlation between vibration and current features.Finally,the output enhanced features are used to realize fault identification and classification.The effectiveness and reliability of the proposed method are verified by bearing and gear fault diagnosis experiments,and compared with traditional diagnosis methods relying on the single vibration and current signals.Thirdly,to dealing with the problem of weak fault information and low diagnostic accuracy of single current signal and meanwhile to capture the complementarity between current signal and vibration signals,an enhanced fault diagnosis method based on multi-view multi-task collaborative feature learning was proposed,with the aim to enhancing the diagnostic ability of current signals with vibration signals as an auxiliary view.The proposed method adopts deep canonical correlation autoencoder(DCCAE)to perform collaborative feature learning and multi-objective optimization and realizes collaborative extraction of the consistent and complementary fault features between vibration and current signals,and thus obtains the enhanced current fault features.The fault diagnosis and identification can be accomplished only through the current signals in the on-line test stage.The effectiveness and superiority of the proposed method are verified through the gearbox fault diagnosis experiment with comparative studies with the traditional multi-view feature learning methods.