Study Of Sparse Representation Of Signal In Fault Feature Extraction And Intelligent Diagnosis Of Rolling Bearing

The rapid development of modern science and technology makes more advanced technology be adopted in the field of machinery fault diagnosis,and has made amazing achievements,which makes today's fault diagnosis rise to a new level.As a new signal processing method,sparse representation technology is quite suitable for mechanical fault diagnosis.Based on the recent research results of sparse representation theory,this paper explores the application potential of sparsity-based methods for feature extraction and fault diagnosis of mechanical fault signals.Rotating machinery in modern industrial production occupies an increasing pro-portion.To make timely warning of the abnormal operation of the rotating machinery not only ensures its safe operation,but also can bring significant economic benefits.The rolling bearing as a key component for rotating machinery to achieve its core func-tions has been widely used.The reliability of its operational status has great influence on the performance of mechanical system.Therefore,to develop novel fault diagnosis method,rolling bearing is a great study object.Considering the above discussions,this paper presents a series of fault feature extraction and fault diagnosis methods based on the sparse representation theory for the diagnosis of rolling bearing.The details are presented as follows:1.The significance of the research on condition monitoring and fault diagnosis of mechanical equipment is expounded from the aspects of historical background,technological development and practical engineering case.Based on the fault mechanism and signal characteristics of the rolling bearing,the existing diag-nostic methods are reviewed and analyzed,including time-domain based meth-ods,frequency-domain based methods and time-frequency domain based method.Besides,we discuss the advantages and disadvantages of each method mentioned above,and the shortcomings of current fault diagnosis are also pointed based on the domestic and international researches.The technology of intelligent diagno-sis is summarized,and introduces the popular neural network and support vector machine in detail.The main applications of sparse representation theory in fault diagnosis,contain signal de-noising,feature extraction and fault identification,are discussed.Finally,the main framework of this paper is presented.2.The basic concepts of sparse representation theory are introduced in detail,and the main problems of sparse coding,namely coefficient solving and dictionary design,are proposed based on the mathematical model of sparse theory.For handling the two problems mentioned above,we provide several widely used algorithms associated with their comparison,which lays the foundation for the discussion of the following chapters.3.Traditional wavelet transform cannot always match the vibration pattern of the fault signal.To solve this problem,a new overcomplete wavelet transform with tunable vibration pattern is proposed based on sparse representation theory.The comparison with traditional time-frequency transform illustrates the proposed method is more suitable for feature extraction,and the SWE features are extracted.The experimental results show that the S WE features are superior to the traditional time-frequency features.4.The structured dictionary is constructed using the structured sparse coding the-ory,which can effectively provide the signals belong to the same category with a unified expression pattern.In the process of solving the coefficients,the opti-mization is performed using the mixed constraint term,so that the features can be selected both on group level and signal structure level,which makes the informa-tion of the signal can be accurately grasped.In order to facilitate the follow-up analysis and diagnosis,the low-dimensional fault feature SS W is further proposed from structured sparse coding.The experiment results show that this feature can effectively suppress signal noise,thus realizing bearing fault diagnosis with high accuracy.5.Combining manifold learning theory with sparse representation,a ManiSC fea-ture extraction framework is proposed to solve the problem of fault diagnosis.In this method,a relationship matrix is established by using the prior knowledge of the data.The basis matrix is found by manifold learning,which maps the data into the sparse domain.Experiments show that the ManiSC feature can effectively represent the geometric characteristic and the intrinsic structure of the original high-dimensional data in a low-dimensional way,and has better robustness and distinguishability than the traditional sparse representation and manifold learning method.6.A novel hierarchical diagnosis network is proposed by combining the sparse rep-resentation with the deep belief network(this method is also called sparse DBN),so that the diagnosis can identify the fault location and fault severity at the same time.Generally,the service life of rolling bearings may be shorten during uses,which bring huge security problems.The diagnosis network we proposed can ef-fectively solve the problem,and the comparison result with other methods shows that the diagnostic network built by the sparse DBN with higher recognition rate for the bearing health condition.The experimental result confirms the hierarchi-cal diagnosis network we proposed has great potential in engineering application.