Research On Fault Diagnosis Methods For Wind Turbines Based On Ordinal Pattern Analysis And Deep Learning

In recent years,global wind industry has experienced a rapid development,and acoordingly,the installed capacity of wind turbines has a remarkable increase both onshore and offshore.However,as wind energy continuouses to rapidly evolve,wind turbines in service are suffering from various failures frequently,which resulted in a series of issues,such as low operation efficiency,short service life,high failure rates and poor reliability.The huge operation and maintenance costs and economical loss caused by long-time downtimes will greatly affect economic benefits of the whole wind farm.Therefore,health monitoring and intelligent fault diagnosis of wind turbines has become one of key urgent techniques to be solved for ensuring a steady,sustained and healthy development fo the wind industry.Along with the spread use and installations of condition montoring systems in wind turbines,a huge volume of condition monitoring data will be generated during the continuous long-time operation of wind turbines.These collected monitoring data usually contain rich information related to the health of the wind turbines and their key components.Therefore,it has become a hot topic in wind turbine condition monitoring domain to how to extract and mine useful fault-related information from the complex monitoring data and then achieve reliable incipient fault detection and intelligent fault diagnosis of wind turbines and their components.To meet the technical demands of wind turbine montoring and diagnostics,the philosophy of “seeing the essence through the appearance” is adoped in this study to analyze the condition monoring data from wind turbines.Our main goal of this study is to discover useful information hidden in original monitoring data based on ordinal pattern analysis and deep learning-based methods,and then to investigate new methods for wind turbine health monitoring and intelligent fault diagnosis.The major research works are listed as follows:Firstly,taking wind turbine bearings as the research object,to capture the subtle change in fault signatures at the early stage,and meanwhile to reveal the changes of nonlinear dynamic characteristics,a new monitoring indicator named ordinal information divergence is proposed based ordinal pattern analysis and information divergence from the perspective of time series similarity analysis.The proposed new indicator can quantitatively describe the ordinal pattern distribution difference of vibration signals in the high-dimensional phase space between the current status and the reference health status.Two experiments,including the damage degree assessment of bearing inner race and the run-to-failure bearing degradation trend analysis,are used to validate the effectiveness of the proposed new indicator.Furthermore,the proposed indicator is compared with traditional time-domain statistical indicators and complexity-based indicators.Secondly,taking wind turbine gearbox as the research object,a stacked denoising autoencoders-based feature learning and diagnosis approach is studied to overcome the limitations of manual feature extraction in traditional fault diagnosis.In the proposed approach,we first design a multi-level-denoising training scheme to enhance the feature learning ability of the existing stacked denoising autoencoder approach under a single noise level,with the aim to capture global and local features hidden in input data.And then a deep leanring network structure,named stacked multilevel-denoising autoencoders(SMLDAEs)is proposed.Furthermore,a SMLDAE-based fault diagnosis system is developed to adaptively extract fault features and achieve the final diangsis tasks.A wind turbine gearbox test rig has been self-designed,and vibration data from different gearbox conditions are collected to validate the effectiveness of the proposed approach in terms of diagnosis accuracy.Thirdly,considering multiscale characteristics inherent in vibration signals from the wind turbine gearbox,and meanwhile to address the limitation in feature extraction ability of traditional CNN architecture,a new multiscale convolutional neural network(MSCNN)architecture is to learn complementary and rich fault pattern features at different time scales from raw vibration signals.Furthermore,an end-to-end fault diagnosis system based on MSCNN is designed,which establishes the complex mapping between raw temporal signals and condition labels directly and enables the adaptive feature learning and fault classification simultaneously.The proposed MSCNN approach is evaluated using vibration data from wind turbine gearbox.Experimental results and comparison analysis with respect to the traditional CNN and several traditional multiscale feature extraction approaches in this field have demonstrated the superiority of the proposed method.Lastly,to deal with the nonlinearity of WTs,uncertainty of disturbances and measurement noise,and temporal dependency in time series data,in this paper,a recently developed unsupervised learning method,denoising autoencoder(DAE),is first introduced and combined with the sliding window technique to propose a new multivariate fault detection model,named sliding window denoising autoencoder(SW-DAE).The proposed approach builds a healthy reference model only relying on the normal data from wind turbines,which can capture the nonlinear correlations among multiple sensor variables and the temporal dependency of each sensor variable simultaneously,which significantly enhanced the fault detection performance.Simulated data from a generic WT benchmark and field SCADA data from a real wind farm are used to evaluate the proposed approach.The results of two case studies demonstrate the effectiveness and advantages of our proposed approach.