Fault Diagnosis For Wind Turbine Gearboxes Based On Densely Connected Convolutional Networks

Wind energy is a green and renewable energy,and the promotion of the development of wind power business has become an important part of China's carbon neutral emissions task.The construction of the wind farms is accelerating,and many wind turbines have come out of the warranty period.Therefore,the demand for the wind turbine operation and maintenance market will be prospering,and efficient operation and maintenance services increasingly become an important factor for the stable and reliable operation of the installed wind turbines.Wind turbines must have a high degree of reliability in order to provide the long-term stable power output.At present,in order to monitor the health status of wind turbines effectively,the numder of monitoring sensors installed in wind turbines are increasingly growing.The volum of monitored status is rich,and the scale of wind power monitoring data is keep growing.The wind turbine operation and maintenance has the characteristics of big data,which is benefitial for comprehensive representing the operating status of wind turbines comprehensively,and improving the efficiency of wind turbine power production,reducing the cost of wind farm operation and maintenance.However,the operating environment of wind turbines is usually harsh,and they are subjected to the influence of extreme weather and irregular load impacts such as sand,rain,snow,wind gusts and strong sunshine for a long time,which makes the performance of their core components degrade continuously.Gearboxes is one of the core components of wind turbines,although its chance of failure is low,but once the failure occurs its downtime is very long and maintenance costs will be very large,which will seriously reduce the economic benefits for the wind farm operation and maintenance.Applying convolutional neural networks to automatically extract key fault feature discriminative information from the vibration data of wind turbine gearboxes,then the health status information of the internal components of the gearboxes can be identified.However,wind turbine gearboxes have characteristics such as weak fault features signals,varying working conditions and few labeled samples,which make it difficult to discriminate the vibration signals of wind turbine gearboxes effectively by directly applying convolutional neural networks.In this thesis,several methods are proposed to address different issues for wind turbine gearbox fault identification with the data information collected by wind farm operation and maintenance.The wavelet packet decomposition and densenly connected neural network work together to address the initial fault diagnosis of wind turbine gearboxes.And the dynamic weighted densely connected neural network,multi-dense block central moment disrepacny based deep domain adaptation network and multi-source deep domain adaptation network are also proposed to solve the wind turbine gearbox fault diagnosis for different challenges.The main research work of this thesis is introduced as follows.(1)A fault diagnosis method for wind turbine gearboxes based on wavelet packet decomposition and densely connected neural network is studied.The proposed method transforms the one dimentional gearbox vibration signal into the initial feature maps by wavelet packet decomposition,then the densely connected neural network will learns the discriminative feature from the initial feature maps,and then the healthy state of the wind turbine gearbox will be identified.This special wind turbine gearbox fault diagnosis procedure provide the baseline for the subsequent study in the next chapters(2)The wind turbine works under varying operating conditions,the speed and load of the wind turbine gearbox fluctuate significantly,and resulting in larger differences in samples within the same category and smaller differences in samples between different categories.A fault diagnosis method based on dynamically weighted densely connected neural network for wind turbine gearboxes is proposed.First,the wavelet coefficients obtained from the wavelet packet decomposition of vibration signals are taken as input,and the a frequency-band-wise weighting moduel is added to the densely connected nueral network to enhance its feature extraction ability by dyanamically adjust the weights of the each arrow feature values in each feature map.The deep learning model automatically learn the weights that should be assigned to each sub feature-band to increase the feature learning ability for gearbox vibration data under variable operating conditions.The proposed method increases the feature learning capability for gearbox vibration data under varying operating conditions.This wind turbine gearbox fault diagnosis method is validated by the real vibration data.(3)It is difficult to collect enough labeled wind turbine gearbox vibration datasets from wind farms,which may leads to the deep learning network model being prone to overfitting and degradation of recognition accuracy.A deep domain adaptation method based on multiple dense block central moment discrepancy is proposed for wind turbine gearbox fault diagnosis under few labeled vibration dataset condition.The method introduces domain adaptation of central moment discrepancy to adapt a large number of labeled samples in the source domain to a small number of labeled samples in the target domain with the aid of adaptive feature extraction by densely connected neural networks.The information contained in the labeled vibration data from the source domain wind turbine is migrated to the target domain wind turbine with only a small number of labeled vibration data for fault diagnosis,reducing the convolutional neural network's need for a large number of labeled data from the target domain.This method reduces the need of large amount of labeled data and improves the generalization ability of the model.(4)When there are labeled vibration datasets from multi-source domains of wind turbine gearboxes,it is difficult to transfer them together to the target domain wind turbine gearboxes to performe the fault diagnosis.A multi-source deep domain adaptation method is proposed to adapt the labeled vibration datasets from the multi-source domain to the target domain for wind turbine gearbox fault diagnosis.During the adaptation process,each pair of source domain and target domain samples will be aligned to reduce the distribution discrepancy among them by the central moment discrepancy method,and the differences between the vibration datasets from multiple source domains will also be minimized by introducing the losses of the multiple source domain-related classifiers.Finally,the model built from multiple classifiers learned by different source domains will be deployed to perform the gearbox fault diagnosis for the target wind turbines.Finally,the research work of the full thesis is summarized and the research directions of the next step are forecasted.