Fault Diagnosis Of Wind Turbine GearBoxes Based On Deep Residual Learning

The wind turbine gearbox is one of the core components of a wind turbine.In the event of failure,due to its complicated structure and large volume,the downtime caused by the failure is relatively long,which leads to economic losses,so the early fault diagnosis of wind turbine gearbox is an important research task.At present,the vibration signals which contain much health condition information of wind turbine gearboxes are often used to conduct fault diagnosis of wind turbine gearboxes.Due to the complex vibration transfer path of the wind turbine gearboxes,continuous changes in operating conditions,and environmental noise,the frequency components of the vibration signal are extremely complex and there is a large amount of redundant information.The traditional method based on vibration signal processing is difficult to detect the fault frequency,and the statistical features commonly used in shallow machine learning methods are often unable to accurately characterize the health status of wind turbine gearboxes,and ultimately lead to misjudgment and missed judgment of wind turbine gearbox failure.Wind turbine gearbox fault diagnosis methods based on deep learning have been studied in recent years because deep learning methods,especially convolutional neural networks,can automatically learn discriminative features from high-dimensional data to replace traditional statistical features.However,due to the relatively deep layers of deep learning algorithms and large amounts of parameters,traditional deep learning methods often encounter the problem of training difficulties.Deep residual learning is a new type of convolutional neural network,also known as deep residual network.It reduces the training difficulty of deep neural networks by introducing a number of identity mappings.Therefore,this paper introduces deep residual learning into the fault identification of wind turbine gearboxes.Aiming at the characteristics of wind power gearboxes,such as varying working conditions,strong noise interference and multi-sensor monitoring,the research was conducted from the following aspects:(1)A fault diagnosis method based on wavelet packet decomposition and deep residual learning is investigated for wind turbine gearboxes.The developed method performs wavelet packet decomposition on the vibration signals in the first step,and lets the deep residual network to automatically learn discriminative features from the wavelet packet coefficients obtained from the wavelet packet decomposition,which can take advantages of the outstanding parameter training ability of deep residual networks,in order to replace the artificially configured traditional feature set for fault diagnosis of wind turbine gearboxes.Experimental results demonstrated the effectiveness of the “wavelet packet decomposition → deep residual networks” method,and provided a baseline for the subsequent chapters.(2)Since the inter-class variances can be enlarged and intra-class variances can be relatively decreased under the circumstance of various operating conditions,a deep residual network with dynamically weighted wavelet coefficients is proposed for the fault diagnosis of wind turbine gearboxes.The developed method takes the wavelet packet coefficients matrix obtained from the wavelet packet decomposition as the input of the deep learning method,adopts a frequency-band-wise weighting way to enlarge the contributions of the frequency bands which have small inter-class variances and decrease the contributions of the frequency bands which have large inter-class variances,and lets the deep learning method to automatically learn the weights that should be assigned to these frequency bands,in order to improve the feature learning ability of the deep learning method on the vibration data under various operating conditions.(3)Aiming at the problem that the fault recognition performance of conventional deep learning methods often decrease under strong background noise,a deep residual network with trainable wavelet-associated thresholding is developed for fault diagnosis.The developed method takes the wavelet coefficients matrix obtained from the wavelet packet decomposition as the input,inserts a differentiable soft thresholding function into each residual building unit as an individual nonlinear transformation layer,optimizes the involved thresholds as trainable parameters along with the other weights and biases,performs a batch normalization before each differentiable soft thresholding function,and achieves the unification of “signal denoising,feature learning,and fault recognition” in the architecture of deep learning algorithm.(4)Aiming at the problems that the vibration signal collected form a single sensor often cannot fully represent the health state of the whole wind turbine and the vibration signals from various sensors unavoidably contain much redundancy,a fault diagnosis method called multiple sensor multiple wavelet coefficients fused deep residual network is developed.The developed method takes the wavelet coefficients matrix as the input of deep learning algorithm,uses an element-wise maximum output layer to perform the feature-level fusion for the information from various sensors and various wavelet packet transforms,in order to achieve the adaptive reduction of redundant information in the architecture of deep learning algorithm and learn a more discriminative feature set.Finally,the research work of the full thesis is summarized and the research directions of the next step are forecasted.