A Weakly Supervised Deep Learning Method For Incipient Fault Diagnosis Of Wind Turbine Gearboxes

To meet China's energy development needs in the new era,the State Council issued a white paper on “China's Energy Development in New Era.” Among them,wind energy,as a clean energy source,has received vital attention for its large reserves,wide distribution,renewable,and other advantages,and has been vigorously developed in recent years.However,with the increase in service time,a large number of wind turbines began to gradually out of the warranty period.The early detection of wind turbine failure and accurate diagnosis is the inevitable trend of the wind power industry to take a new high-quality development path.Wind turbine gearboxes are the primary source of wind turbine failures because they are subjected to high dynamic loads for a long time,and their failures cause the most prolonged downtime and huge losses.Therefore,it is crucial to study wind turbine gearbox early diagnosis methods,early detection,and precise location of faults to reduce wind turbine operation and maintenance costs.Compared with signal analysis and traditional machine learning mechanical fault diagnosis methods,deep learning enhances wind turbines' nonlinear fault feature extraction capability through a deeper network structure.It reduces human intervention by extracting fault features without combining rich engineering practical experience,thus realizing the shift from an experience-driven artificial feature paradigm to a data-driven representation learning paradigm.However,due to the complex structure,intense noise,and significant differences in operating conditions of wind turbine gearboxes,especially in the early stage of wind turbine gearbox fault occurrence,the fault features are weak,and the state is in continuous development.The mapping relationship between the fault mode and the fault representation is fuzzy,making the incipient fault identification difficult and thus causes the problems of lack of incipient fault samples and sample mislabeling of wind turbine gearboxes.It is necessary and urgent to study fault diagnosis methods of wind turbines with weakly supervised deep learning,to improve the adaptive feature mining capability of deep neural networks for weak fault features under intense noise and variable operating conditions,and to meet the needs of wind turbine gearboxes' fault diagnosis with less labeled samples and noisy labeled samples.The research of wind turbine fault diagnosis methods with weak supervised deep learning is essential and urgent.It is of great practical significance to improve the stability and reliability of the safe operation of wind power equipment in China.This paper proposes a weakly supervised deep learning wind turbine gearbox incipient fault diagnosis method to address the above wind turbine gearbox incipient fault diagnosis problem,combined with its nonlinear adaptive feature extraction advantages.The proposed methods include weak feature hybrid attention enhanced residual network,multi-association layer deep semi-supervised network,and adaptive loss weighted meta-residual network based on the study and expansion of existing deep neural networks.The main research work of the Dissertation is as follows.(1)A wind turbine gearbox fault diagnosis method based on the hybrid attention deep residual network is proposed for weak wind turbine gearbox incipient fault characteristics and intense background noise.The technique effectively presents the time-frequency information of vibration signals through wavelet packet decomposition.It improves the frequency attention mechanism of the deep residual model to highlight the weak fault feature frequency bands in the wavelet coefficients,enhancing the feature extraction ability of the model in the intense noise background.Meanwhile,the error backpropagation of the channel attention mechanism automatically gives different attention weights to the channels,which improves the features learned by different network channels.Finally,the proposed method's superior performance is verified through experiments and actual fault data of wind farms.(2)A multi-layer deep semi-supervised fault diagnosis method is proposed to address the lack of labeled samples for incipient faults in wind turbine gearboxes.The method combines the advantages of traditional semi-supervised learning and deep learning.It uses wavelet packet transform to analyze the local details in the time-frequency domain and highlight the impact signal features by partially labeled and large sets of unlabeled fault samples used to train the proposed model.Experimental results confirm that the proposed method eliminates the reliance on manual feature extraction and many accurately labeled examples and can achieve better results than traditional strongly supervised deep learning models under less labeled conditions.(3)To address the problem of noise-labeled wind gearbox incipient faults,an adaptive loss-weighted meta-residual network model is proposed for the diagnosis of noise-labeled wind gearbox faults.The method establishes a mapping of weighting functions through a weighting network and a meta-network cloned from the original residual network.It adaptively learns the weights from the data with clean labels to dynamically weigh the original residual network's loss function to improve the robustness of the fault diagnosis model to noise tags.And finally,the proposed method's effectiveness for fault diagnosis under a large number of noise tags is verified through simulation experiments and wind farm historical data.(4)Based on the wind turbine drive train network monitoring and diagnosis system developed by the group and the fault diagnosis method proposed in the previous chapters,the proposed algorithm is implemented by Python(3.62)programming language and deep learning frameworks such as Tensorflow(1.15.4)and Keras(2.3.1)according to the wind gearbox fault diagnosis requirements.A weakly-supervised deep learning wind turbine gearbox fault diagnosis system is implemented and integrated with the previously developed networked monitoring and diagnosis system for wind turbine drive trains.The wind farm's actual measurement data verify each functional module of the system.Besides,this chapter presents the application of the model proposed in the previous chapter,which is deployed in a wind turbine diagnosis and early warning system of a domestic company after training and tuning by historical wind field data.Finally,after the above experimental research exploration,this paper summarizes and outlooks the problems and development trends faced by weakly supervised deep learning in the application of wind turbine gearbox incipient fault diagnosis.