Research On The Method Of Local Fault Impulse Signals Extraction In Wheelset Bearings Based On Dictionary Learning

With the rapid development of China high-speed railway,the ever-expanding scale of railway network and the greatly improved vehicle manufacturing technology have made China become the leader of high-speed railway in the world.The high-speed vehicle is a complex system which composed of many electrical and mechanical equipment.The healthy status of some key components in the vehicle has the direct influence on the safety and stability of the vehicle during service process.Wheelset bearing mounted on the wheel is a crucial power transmission element in the vehicle running system.In the process of vehicle operation,excepting the static and dynamic loads generated by vehicle,the rapid and long-term alternating transmission which results from variable speeds and loads in the service condition may lead to the accelerated wear and failure of wheelset bearings,and ultimately endanger the safety of railway service.Therefore,it is significant to monitor the operating condition and fault diagnosis for the wheelset bearings.For a wheelset bearing,when a peeling,scratch,crack and other local defects appear on the surface of its component,a series of discrete vibration impulses will be generated with the rotation of the axle of a wheelset.These fault impulsive components have similar structures.Videlicet,the measured fault vibration signal has shift-invariant characteristics.Sparse representation of signals,as a theory for mining the internal structure and typical representation of the signals,has emerged in recent years.In particular,the shift-invariant dictionary learning algorithm is a special method applied processing signals with shiftinvariant characteristics.This method is able to effectively and efficiently extract those components having shift-invariant characteristic from the given signals.Hence,combined with the local fault mechanism of a wheelset bearing and the characteristics of fault vibration signals,and on the basis of the shift-invariant dictionary learning algorithm,this thesis conducts in-depth research on the extraction of the wheelset bearing local fault impulse signals and fault diagnosis.As a time-domain signal solution,shift-invariant K-means singular value decomposition(SI-K-SVD)dictionary learning algorithm is studied on the local fault impulse signals extraction of a wheelset bearing in this thesis.The improper selection of two key SI-K-SVDrelated parameters,namely,the number of iterations and the pattern lengths,has an adverse influence on the performance of local fault impulse signals extraction.In order to solve key parameters selection problem,the sparsity of the envelope spectrum(SES)and the kurtosis of the envelope spectrum(KES)are applied as indicators for optimizing these two key parameters,respectively.Finally,the optimal parameter SI-K-SVD(OP-SI-K-SVD)based on index guidance is proposed.Meanwhile,for the issue of multi-fault signal extraction,a hierarchical OP-SI-K-SVD(H-OP-SI-DL)is proposed to hierarchically extract those multi-fault impulse signals based on their fault power levels.This method can effectively extract the fault impulse signals generated by different fault,and highlight the fault information both in time-domain and frequency-domain.Due to the lack of specific structure for the dictionary in shift-invariant dictionary learning,incomplete fault impulsive components in the truncated signal are easily lost,and the efficiency of dictionary updating is low.The circular-structure matrix theory is applied for optimizing and limiting the structure of the dictionary,and the dictionary learning with circular-structure matrix(CS-DL)is proposed.The improved method using circular-structure matrix not only realize effective representation for the complete and incomplete fault impulsive components,but also improve dictionary updating efficiency based on property of eigenvalue decomposition.To solve the problem of low computation efficiency and easy to generate local optimal solutions when CS-DL performs sparse coding under the framework of L0 norm,the basic pursuit which under L1 norm framework is applied to replace the original coding method.Because the alternating direction method of multipliers(ADMM)has analytic solution in the stage of sparse coding and can adjust the internal parameters adaptively,ADMM is used to solve the sparse coefficients.Penalty factor,as an important parameter for ADMM sparse coding,affects the sparsity and validity of the extracted fault impulse signals,and SES has also been approved to be suitable for reasonable selection of this parameter.By inheriting the advantages of CS-DL in dictionary updating and combining with ADMM algorithm which carry out sparse coding efficiently,a method called circular-structure ADMM dictionary learning(CS-ADMM-DL)is finally proposed.Good precision and high efficiency computation contribute CS-ADMM-DL more suitable for practical application.The random impact component and the low frequency harmonic component are two common interference components in the experimental signal.By analyzing the influence of the two components on applying CS-ADMM-DL for a wheelset bearing to extract local fault impulse signals from the measured signal,it is found that low frequency harmonic component has an adverse effect on the actual application of this algorithm.In order to eliminate the adverse effect,locally weighted scatterplot smoothing(Lowess)based on signal smoothing in the time domain,is used to preprocess the collected vibration signals.Multiple sets of bench test data and line running test data with different speed levels and faults are applied to verify the ability of combined method of Lowess and CS-ADMM-DL on extracting the wheelset bearing fault impulse signals and fault diagnosis under experimental conditions.