Data-driven Fault Diangosis Of Wind Turbine Gearbox

The gearbox is the critical transmission device of wind turbines,which consist of main bearing,gearbox,generator,low-speed and high-speed shafts,and etc.It often operates in complex operation regimes and suffers from low-speed,heavy load and varying loads and even strong and shock wind for a long time,thus resulting in high failures and greatly affecting the economic and social benefits of wind power generation.Once one certain component fails,it will induce a chain reaction,cause the whole system down,resulting in huge economic losses and social impact.Therefore,it is of great academic research significance and engineering application value to carry out research on wind turbine drivetrain condition monitoring and fault diagnosis technology for reducing the failure rate and operation and maintenance cost,ensuring the safe operation of the unit and maximizing the economic benefits of the wind farm.The thesis addresses the application requirements and challenges for fault diagnosis technology of wind turbine drivetrain system.Based on the latest theoretical techniques in the field of machine learning and data mining,this thesis aims to study data-driven fault diagnosis methods for wind turbine gearbox,which can provide technical supports to achieve effective monitoring,accurate assessment and fault diagnosis.The main work can be summarized as follows:Firstly,RBSSD often suffers from incomplete decompositions when dealing with low signal-to-ratio signals due to its single dictionary property,and has some limitations especially in practical applications due to the difficulty in parameter selection.To address this issue,a new Multi-Dictionary RBSSD(MD-RBSSD)method is proposed.Instead of using single dictionary,the proposed MD-RBSSD method introduces symlet8 wavelet dictionary and sine dictionary on the basis of the tunable-Q dictionary used in RBSSD.Thus,the low-resonance components obtained using RBSSD is further decomposed to extract fault impulse signatures.Furthermore,the correlated kurtosis is introduced to provide a quantitative evaluation for the decomposition results.Simulations,experiments and engineering case study are used to verify the effectiveness of the proposed method.Secondly,given that vibration signals usually contain multiple temporal structures,this paper proposes a multiscale representation learning(MSRL)framework to learn useful features directly from raw vibration signals,with the aim to capture rich and complementary fault pattern information at different scales.In our proposed approach,a coarse-grained procedure is first employed to obtain multiple scale signals from an original vibration signal.Then,sparse filtering,a newly developed unsupervised learning algorithm,is applied to automatically learn useful features from each scale signal,respectively,and then the learned features at each scale to be concatenated one by one to obtain multiscale representations.Finally,the multiscale representations are fed into a supervised classifier to achieve diagnosis results.Our proposed approach is evaluated using two different case studies: motor bearing and wind turbine gearbox fault diagnosis.Thirdly,to deal with the issue of the interactions and couplings of compound fault features,a new multimodal feature learning and fusion network is designed for compound fault diagnosis,which can learn the complementary characteristics from different modalities,including time domain,frequency domain and time-frequency domain.The multimodal representations from time,frequency and time-frequency domain are first constructed.Then,three individual feature learning networks are designed to adaptively extract fault features from three different modalities,respectively.Furthermore,the extracted multimodal features are fused in a weighted way based on the attention mechanism.The bearing compound fault experiments on a wind turbine drivetrain system was conducted to verify the effectiveness of the proposed method and a comparative study is also performed with traditional unimodal feature learning and simple fusion methods.Lastly,existing studies mainly focus on single vibration signal analysis for fault diagnosis that only contains limited information and it is difficult to achieve accurate description of the status of gearbox.To address this issue,a multi-task learning-based electro-mechanical signal fusion fault diagnosis approach is proposed based on the idea of information fusion,which aims to learn the complementary diagnostic information between vibration signals and current signals.According to the characteristics of vibration and current signals,two individual convolutional feature learning networks are designed to learn useful and complementary features from raw signals.Then,the metric learning is introduced to design an auxiliary task to constrain the samples belonging to the same class to decrease the intra-class distance,thus enhancing the discriminative capacity of the learned features and improving the classification performance from the main task.The feasibility and effectiveness of the proposed method is validated through multi-sensor experiments on the wind turbine gearbox.