Fault Diagnosis And Health Assessment Of Rotating Machinery Based On Kernel Density Estimation And Kullback-leibler Divergence

Rotating machinery has widespread applications in advanced manufacturing and engineering systems, and it plays a very important role in industry fields. The crucial components in rotating machinery, such as bearings, gears, undertake the tasks like transferring torque and supporting loads. These components are oftentimes suffering undesirable stresses and sudden shocks under which initial defects will appear. If maintenance activities cannot be taken properly and timely, the tiny defects will gradually propagate and eventually cause severe damages and unexpected shutdown to the entire systems. Fault diagnosis and health degradation assessment are the two important tools for detecting the operating condition of rotating machinery based on which preventative maintenance can be scheduled in an effective manner. On the other hand, as a great amount of data becomes acquirable and accessible in industry, data-driven methods have become an emerging research area, acting as a complement to the existing fault diagnosis and health assessment methods.Although many data-driven approaches have been developed to identify faults and assess the health status of engineering systems, most of them ignore the statistical relation among samples. Such statistical information may, more or less, contain information related to the fault types and health status of monitored systems. As a matter of fact, samples belonging to the same class may have similar statistical properties. By taking account of the statistical properties, the accurate of fault diagnosis and health assessment can be further improved. Bear this idea in mind, a set of new approaches for fault diagnosis and health assessment of rolling bearing and gears have been proposed in this dessertation. The main contributions are summarized as follows:(1) Development of a hybrid signal process method by combining the Local Mean Decomposition(LMD) and the discrete wavelet transform. The proposed method inherits the merits of both methods, i.e. adaptivity of the LMD and frequency-band clarity of wavelet analysis. Meanwhile, the mode mixture problem of the LMD and the “non-adaptive” weakness of wavelet transform can be solved or weakened, so as to extract a more effective feature for fault diagnosis and health assessment.(2) Development of a feature selection method based on the Kernel Density Estimation(KDE) and the Kullback-Leibler Divergence(KLID) from the statistical perspective. By taking account of the statistical properties of the training samples, a new feature selecting method has been proposed to identify the classes-sensitive features from a set of features. The proposed method cannot only identify the sensitive feature which have similar distribution types but with different mean values, but also identify the sensitive feature possessing the same mean values but with different distribution.(3) Development of a new diagnosis method for rotating machinery based on the KDE and KLID from the statistical perspective. As demonstrated in the fault diagnosis of bevel gears and rolling element bearings, the proposed method has an exceptional performance on the faulty patterns recognition and outperforms the conventional SVM-based and BP network-based fault diagnosis methods. Since the proposed method incorporates the statistical characteristics of the samples within one set, it manifests superior classification accuracy and robust performance.(4) Development of a health assessment method for rotating machinery based on the KDE and KLID. The health assessment can be regarded as a One-Class classifying problem, in which only the normal and abnormal conditions are to be identified. In this work, a health assessment method based on the KDE and KLID has been proposed. A moving window has been used to choose the sample set from the monitoring data, and the corresponding distribution of the sample set can be characterized by the KDE. The health status of the monitored system can thus be assessed by comparing the KLID of the normal condition with that of abnormal condition.