Research On Wind Turbine Gearbox Fault Warning And Diagnosis Method Based On Multimodal Data Fusion

Data-driven wind turbine gearbox intelligent warning and diagnosis technology is a key technical means to improve the efficiency of wind turbine power generation and reduce operation and maintenance costs.However,most of the existing data-driven wind turbine gearbox fault diagnosis methods are based on a single data modality and only target a single fault form of the gearbox.The correlation between different components of the gearbox and the complementary relationship between different modal data are not fully considered in the fault diagnosis process,which makes it difficult to distinguish the specific fault form of the gearbox and makes the early warning positioning of the faulty components vague and difficult to attribute.In order to make full use of the multimodal data of the gearbox and improve the effect of early warning and diagnosis,the research of wind turbine gearbox fault warning and diagnosis based on multimodal data fusion is carried out according to the research idea of "multimodal data-fault warning-fault location-fault diagnosis-fusion decision".The main research contents are as follows:(1)Analysis of wind turbine gearbox fault formsFirstly,two common fault forms of wind turbine gearboxes are introduced.namely lubrication and cooling faults and gear tooth damage faults;then,the interaction between the two fault forms is analyzed according to the actual operation and maintenance logs of wind farms;finally,the general framework of the multimodal data fusion gearbox early warning and diagnosis method is proposed,which mainly includes two sub-modules of wind turbine gearbox multi-component status early warning and fault diagnosis.The overall framework can provide the basis for the model setting and method research of the subsequent sub-modules.(2)Research on wind turbine gearbox multi-component early warning and location method based on adaptive noise reductionAn adaptive noise reduction based wind turbine gearbox condition warning model is constructed,which can simultaneously monitor the operating status of multiple components in the gearbox,solving the problems of high repetition and low efficiency of the component-by-component modelling method,and enabling early warning and positioning of gearbox lubrication and cooling faults.Firstly,the SCADA data is noise-reduced to remove the uncertainty noise;then,a multi-series input-multi-series output gearbox normal behaviour model is established using the noise reduction data;finally,the prediction residuals of the normal behaviour model are analyzed by combining the sequential probability ratio detection algorithm and the Wasserstein distance to warn and locate the faulty gearbox components.The analysis of the algorithm shows that the proposed adaptive data noise reduction method can improve the prediction accuracy of the normal behaviour model,and the proposed multi-component warning model can achieve early warning and accurate positioning of lubrication and cooling faults,which is more accurate than other single-component and multi-component condition warning models.(3)Research on wind turbine gearbox fault diagnosis method based on multimodal time-frequency diagram fusionAiming at the problems of difficult learning of features and high computational cost of multimodal fusion fault diagnosis model for gearboxes under variable load conditions,a diagnostic method based on image fusion is proposed to achieve diagnostic attribution of gearbox gear tooth damage faults.The method uses multimodal data fusion techniques to improve the accuracy of fault diagnosis;the time-frequency image fusion method converts the fault diagnosis problem into an image classification problem,reducing the amount of model input data and lowering the computational cost.The results show that the fusion of wavelet domain and time-frequency domain data enables the extraction of complementary fault features for gearboxes under variable load conditions,making the proposed diagnostic model more accurate than other single mode and multimodal fusion diagnostic models;the time-frequency image fusion method can reduce the computational cost of the diagnostic model.