Research On Anomaly Detection And Fault Diagnosis Of Wind Turbine Gearbox Based On Data Driven

In order to achieve the goals of carbon dioxide emission peak before 2030 and carbon neutrality before 2060,the wind power industry still has broad growth space in the future,and the large-scale wind turbine(WT)will become the main model for offshore wind power development and onshore wind power transformation and upgrading.However,the complex structure and relatively harsh operating environment of large WT make its fault rate remain high,which reduces the power generation efficiency and increases the operation and maintenance(O&M)costs.The data-driven based anomaly detection and fault diagnosis technology(DD-ADFDT)is an effective means to solve the above problems.By using the DDADFDT,potential faults can be detected in time and faults can be recognized accurately,which can provide a basis for the maintenance decision-making of O&M personnel.However,the DD-ADFDT still has some problems that need to be solved.For example,it is difficult to obtain a universal anomaly detection threshold under time-varying conditions.It is difficult to accurately extract the features of different faults under multi excitation coupling conditions.The reliability of anomaly detection and the accuracy of fault diagnosis is significantly affected by the quality and quantity of training data.This thesis takes the gearbox of WT as the research object,based on the real data collected by the Supervisory Control and Data Acquisition(SCADA)system and the condition monitoring system of WT,and combined with the data simulated from the transmission chain system simulation platform to study the DD-ADFDT.The main work of this paper is as follows:A gearbox oil temperature anomaly detection method is proposed based on Sparse Bayesian Learning(SBL)and hypothesis testing.The proposed method can be used to solve the problem that it is difficult to find a universal reference threshold to distinguish whether the gearbox oil temperature state is normal or not under the time-varying operating conditions.An SBL based model is constructed to learn the conditional dependence relationship between the state variable related to the gearbox operating state and the gearbox oil temperature based on the measured SCADA data.And the conditional probability density function(PDF)of the timevarying oil temperature can be estimated by the SBL model.Then,the dynamic anomaly detection threshold affected by random operating conditions is obtained by calculating the theoretical distribution interval of the oil temperature under different operation conditions.Besides,an anomaly detection method based on hypothesis testing and sliding window sampling is proposed to discover the abnormal data,which can eliminate the influence of occasional oil temperature out of limit caused by uncertainty disturbance on abnormal detection results,and provides theoretical support for the reliability of detection results from the perspective of mathematical statistics.The case study based on real WT data shows that the proposed method has more advantages than similar methods.Considering the defects in accuracy and universality of many anomaly detection methods,a generalization gearbox anomaly detection method based on the combination of multi probability estimation models is proposed on the basis of the previous study.The method combines the PDF estimated by different probability models such as SBL,kernel density estimation,and quantile regression estimation in the form of weighted combination.Then,the PDF estimated by the combined model can accurately describe the distribution of the target variable in complex situations,which not only improves the sensitivity of the model to anomalies but also enhances the generalization of the model.Experimental verification and comparative analysis based on the SCADA data of real WTs show the proposed method can reliably detect anomalies of gearbox bearing and generator collector ring.And the anomalies are detected earlier than single mode-based methods.Aiming at the problem that the quality and quantity of the historical data of the target are insufficient,which seriously affects the reliability of detection results.An idea that use the SCADA data of WTs with similar operation condition in the same wind farm to improve the reliability of anomaly detection results is proposed.A WT gearbox anomaly detection method based on the combination of self-detection and external detection is first proposed.The proposed method used the iterative self-organizing data analysis techniques algorithm(ISODATA)to qualitatively evaluate the spatial similarity of the operation condition of all WTs in the same wind farm.Then,the SCADA data of multiple WTs with similar operation conditions to the target WT are selected to train a combined probability model to perform anomaly detection for the target WT.Meanwhile,a support vector machine(SVM)based deterministic estimation model is trained by using the historical data of the target WT to perform another anomaly detection.Finally,through the mutual verification of the selfdetection and the external detection to enhance the reliability of the anomaly detection results,which can alleviate the impact of insufficient data of the target WT on the accuracy of the anomaly detection results.Further,a reliable anomaly detection method for gearbox anomaly detection that considers the spatiotemporal similarity of WTs in the same wind farm is proposed.Firstly,the shortcomings of the commonly used state variable similarity comparison methods are analyzed,and a state variable similarity quantification method based on time series piecewise linearization comparison is proposed.Then,a WT operation condition quantification method that comprehensively considers the temporal and spatial similarity of multiple variables is designed.Next,the SCADA data of multiple WTs whose operation conditions are most similar to the target WT are selected as training data,and a stacked long short-term memory network frame is used to construct a state estimation model based on multi-source data training to estimate the target variable.Finally,an improved abnormal data detection method based on the residual effective value comparison and the residual information entropy comparison is proposed to detect the abnormal data.The case study and comparative analysis based on the real data indicate that the proposed method can reliably detect the abnormal state of the gearbox when the training data of the target WT is insufficient or contains abnormal data.As for the weak fault features of gearbox vibration signals are easily covered by irrelevant information under the conditions of multi-excitation coupling and complex working conditions,and the fault feature boundaries of different types fault are difficult to distinguish,which seriously affects the accuracy of fault diagnosis.A hybrid convolution deep residual shrinkage network-based multi-class fault diagnosis method for WT gearbox is proposed.Firstly,a deformable convolutional layer is used in the deep residual shrinkage network to replace the classical convolution layer,which improves the ability to learn irregular features from the timefrequency spectrum and accelerates the learning efficiency of soft thresholds.Further,the improved deep residual shrinkage network is combined with a one-dimensional convolution neural network(ID-CNN).Then,the features extracted from the original vibration signal are fused with the high dimensional features learned from the time-frequency spectrum through the weight distribution mechanism,which makes up for the lack of spatial information loss when the original vibration signal is converted from the time domain to the time-frequency domain by using the wavelet packet transform.It effectively enhances the recognition ability of the model to different fault features.The effectiveness of the proposed method is verified based on the experimental data and the actual vibration monitoring data of WT.The results indicate that the proposed method can accurately recognize different faults of the gearbox under high noise conditions.In order to solve the problem that the gearbox fault diagnosis is inaccurate when the classifier model is trained based on an unbalanced dataset.A data equalization method based on an improved conditional deep convolutional generative adversarial network(C-DCGAN)is proposed.The method uses the data generated by the Gaussian mixture model to replace the one-dimensional random noise as the input condition of the generator of the C-DCGAN,and uses the 1D-CNN to extract the potential features of the real sample as the conditional constraints of the generator of the C-DCGAN.It not only increases the diversity of the samples generated by the generator,but also makes the generated samples contain the features of the real samples.In addition,the Wasserstein distance is used as the generator loss function that replaces the commonly used JS divergence to improve the stability of the model.The Bayesian optimization algorithm is used to optimize the hyperparameters of the network to improve the performance of the model.Case studies and comparative analysis based on the experiment data and the actual vibration signals of the WT gearbox show that the proposed method can generate pseudo-sample data with various forms and feature close to the real samples.It can alleviate the adverse effects of insufficient minority samples and unbalanced training set on the classifier model and improve the accuracy of gearbox fault diagnosis.