Investigation On Bearing Fault Identification And Performance Degradation Assessment Based On KJADE

As an important component of the rotating machinery,the health condition of the rolling bearing is directly related to the stability of the equipment.Thus,the fault diagnosis and performance degradation assessment of bearing is important.And in this paper,the rolling bearing is selected as the research object,the algorithm of kernel joint approximate diagonalization of Eigen-matrices(KJADE)is combined with signal processing methods and machine learning methods for feature extraction,fault identification,performance degradation assessment in bearing fault diagnosis.The theory of the KJADE is first proposed in this thesis,which takes one-dimensional or multidimensional signal as the processing object.The original feature set is extracted from different domains,which could be mapped into high-dimensional feature space through a nonlinear function,so that the linear indivisibility problem in low-dimensional space is transformed into linear separable problem in high-dimensional space.Furthermore,the kernel function is introduced.to replace the inner product for kernel matrix calculation.Then,the nonlinear low-dimensional statistical structure could be estimated through the Eigen-decomposition of the fourth-order cumulative kernel matrix,which eliminates the correlation and redundancy of the original features,and reveals the essential property of the signal.Compared to the traditional JADE algorithm,the KJADE has a better applicability to the nonlinear signal.In the research of the bearing fault identification,the KJADE is used for more effective statistical structure extraction from the original multi-domain feature set that composed by different domains.This thesis mainly developed low-dimensional nonlinear feature fusion technique by using KJADE,the proposed method was implemented in multi-class bearing vibration signals and compared with other typical dimensionality reduction methods,which the results show an effective clustering performance.Moreover,the KJADE-based feature subspace were constructed with SVM for bearing fault identification.Based on the two-class model that could be effective evaluate the difference between the fault samples and healthy samples,the research of the bearing performance degradation assessment is implemented in this thesis.The KJADE is employed to extract the low-dimensional feature from the entire life cycle vibration signal,and obtain the sensitive attribute which can reflect the bearing performance.This thesis explored the bearing performance degradation indicator extraction based on KJADE and two-class model,the proposed method was implemented in the actual signal and obtained a robust and monotonous indicator that is also timely for early fault detection.This thesis still conducted the research of performance degradation trend forecast via ELM,the bearing condition of the next moment could be effectively predicted through ELM prediction model that trained by historical data.Furthermore,the above-mentioned research and analysis were all validated by the experiment data,which are Case Western Reserve University bearing experiment,Cincinnati University bearing entire life cycle experiment and bearing fatigue experiment that designed from our laboratory,respectively.The research in this thesis indicates that the KJADE could be employed to extract the effective and sensitive features that correctly reflect the bearing condition,which is important for fault identification and performance degradation assessment.