Feature Fusion Based Fault Diagnosis Of Wind Turbines Rolling Bearing Under Variable Working Conditions

Environmental pollution and energy crisis are two important issues that all mankind must face,wind energy can be utilized to address the above problems due to its advantage of clean,renewable and rich reserves and thus has been widely exploited around the world.The rapid development of the wind power industry requires a reliable and stable condition monitoring and fault diagnosis system to escort its implementation.However,the frequent occurrence of failures for wind turbines in service and the gradual decline in operation efficiency over the years will seriously affect the steady and sustainable development of wind farms.Therefore,the intelligent condition monitoring and diagnosis system for wind turbine is a key technology for the healthy and steady development of wind power industry.As one of the key components of wind power equipment,rolling bearings often encounter situations such as variable speed,variable load,and strong noise pollution during operation,which greatly increases the difficulty of mathematical and physical modeling and limits the application of model-based fault diagnosis methods.At present,the development of wind turbine condition monitoring technology has entered the era of big data.How to obtain useful information from the massive data containing equipment health status and realize real-time early warning and intelligent fault diagnosis has become a research hotspot in the field of wind power equipment monitoring.Deep learning has been evolving to a dominant method in the area of fault diagnosis over the past decades,however,the high demand for training data restricts its implementation,as it is difficult to collect a large number of fault data in the scenario of engineering practice since the occurrence of faults is fleeting,Combining the experimental research and theoretical analysis,starting from the condition monitoring data,this paper performed an in-depth study of a new method of feature extraction and multi-domain features fusion,as well as make the full use of deep learning algorithm to achieve end-to-end fault diagnosis,and proposed cross-device and cross-working condition fault diagnosis transfer learning model.The main work of this paper is as follows:(1)To address the high demanding of deep learning models on training data volumn,a novel rolling bearing fault diagnosis strategy based on multi-channel convolutional neural network(CNN)combined with multi-scale shear fusion(MSCF)data augmentation technology was proposed.The strategy uses multi-scale shear fusion technology to enhance the data of the fault signal,and then transforms the fault signal into a time-spectrum graph by short-time Fourier transform.Multi-sensor data are fused using a multi-channel convolutional neural network(SE-BN-MCNN)introducing channel attention mechanism(Squeeze and Excitation,SE)and Batch Normalization module(BN)to achieve deep feature extraction and fault pattern classification.The effectiveness of MSCF and the robustness of the improved CNN to load changes are proved by comparing the experimental results of the augmented data and the original data at different scales.(2)To solve the big data dependence and high training cost of deep models,an easy-to-implement and cost-effective bearing fault diagnosis method is proposed,where the fractional Fourier transform(FRFT)is used as a feature extractor,the big fault data is segmented into maximum kurtosis based fractional spectrum(MK-FRS)with small sample length,and finally the targeted multi-channel extreme learning machine(MCTELM)is used as a feature fusion and classifier.The proposed approach was verified to achieve a fast and robust bearing fault diagnosis in presence of noise interference and working load variation,it is especially suitable for fault diagnosis in situation that only a limited data sample size and hardware configuration can be available.(3)To cope with the performance degradation of deep models caused by data under variable operating conditions,a new data representation method based on FRFT and Recurrence Plot Transform(RPT)is proposed to make the most of CNN and utilize limited data to achieve bearing fault diagnosis.Among them,FRFT serves as a feature extractor by generating MK-FRS,which is not sensitive to noise interference;RPT serves as a texture feature visualization tool in time domain,frequency domain,and fractional domain,generating different types of recurrent plot datasets;CNN serves as a deep feature extractor and classifier,using a limited size of recurrent plot dataset as model input.Experimental results show that the fusion of MK-FRS,the fractional domain recurrent plot,and time-domain recurrent plot can effectively adapt to the variations of speed and load,and achieve the highest fault diagnosis accuracy.(4)To handle the issue of working condition transfer between different experimental equipment and real world wind turbine equipment,a multi-channel fault diagnosis method based on multi domain fusion feature transfer is proposed.By introducing a deep adaptive CNN model with SE module and combining online sequential extreme learning machine(OS-ELM),the fault diagnosis performance can be further improved.By comparing the diagnostic accuracy,time cost and robustness to the changes in operating conditions,the superiority of proposed strategy over single feature transfer and traditional CNN model transfer was demonstrated.This method is expected to be applied in practical industry scenarios where a large volume of fault data is difficult to obtain.