Methodologies For Degradation Monitoring And Prognostics Of Rotary Machine Components

In modern industry, various machinery equipment are working under the complex operating conditions of high speed, high temperature, high pressure and heavy load, leading to inevitable machine degradation and failures. These failures, sometimes, can result in huge economic losses and even catastrophic accidents. As the key componets, such as bearings and gears, of rotary machines, their working status will influence the performance of the whole machine. As a result, the development of suitable machine maintenance strategy has received considerable attention in recent years. Current maintenance strategy of rotary machine componets has been progressed from breakdown maintenance and preventive maintenance to condition-based maintenance, and lately towards intelligent prognosis maintenance. Different from traditional diagnostic approaches which only comprise the fault identification and classification after fault occurrence, prognostic approaches have to monitor component degradation during its lifetime and predict the remaining useful life (RUL) of the degradation component. Therefore, prognostic approaches are more beneficial than diagnostic approaches to reduce expensive downtime, maintenance costs and to achieve maximum productivity.For existing prognostic approaches, studies on nonlinear degradation monitoring, real-time online prediction and integration of data-driven model with physics model are still limited. Therefore, based on the phase space reconstruction theory, the degradation monitoring and fault prediction of rotary machine components are investigated in this dissertation. The main research contents can be summarized as follows.(1) The phase space reconstruction theory based on Takens embedding law is intruoduced and the selection methods of two important parameters, time delay and embedding dimension, are investigated. The research shows that phase space reconstruction using mutual information and false nearest neighbors for embedding parameter selection can reconstruct the phase space from a one-dimension measurement series and well preserve the basic structures and dynamic properties of the original system. Hence, mutual information and false nearest neighbors are used to choose time delay and embedding dimension in this study and lay a foundation for the nonlinear feature extraction of mechanical vibration signals.(2) The fault diagnosis and degradation monitoring of rotary machine components based on the modified recurrence quantification analysis (RQA) is investigated. Four nonlinear feature parameters are extracted by the RQA to diagnose rotary component faults. A novel deviation-based recurrence threshold selection method is proposed to modify the traditional RQA for constructing a steady and sensitive recurrence entropy as a nonlinear degradation feature. Combined with the Chebyshev inequality-based health threshold selection, the recurrence entropy is used to track the rotary component degradation and Kalman filter is utilized to predict the initial fault start point at the same time. The experimental studies show that nonlinear features extracted by the RQA can effectively evaluate the damage severity of rolling bearings. The modified RQA can well describe the degradation progress and Kalman filter is able to predict occurrence of the bearing failure in advance, Furthermore, the proposed degradation monitoring algorithm is fast enough to meet the requirement for online tracking and prediction.(3) The enhanced particle filter approach based on the data-driven model is investigated for remaining useful life (RUL) prediction of rotary machine components. It combines an adaptive importance density function selection method with a BP neural network-based resampling smoothing method to reduce the particle degeneracy degree and improve the particle diversity. The experimental studies show that the enhanced particle filter can accurately predict the RUL of bearings and it can achieve better performance than the traditional particle filter-based approach and commonly used support vector regression approach. Furthermore, the proposed RUL prediction algorithm has good real-time performance and can meet the requirement for online prediction.(4) A novel multi-time scale modeling approach integrating the data-driven model with physics model is investigated for failure tracking and RUL prediction of rotary machine components. The phase space warping (PSW)-based data-driven approach is used to extract the dynamic feature of a fast-time dynamic subsystem and the Paris crack growth model-based physics approach is used to predict the physics state evolution of a slow-time subsystem. Furthermore, the PSW approach is enhanced by the multidimensional autoregression (AR) model and the traditional Paris crack growth model is also modified by the time-piecewise strategy. The experimental studies show that the presented approach can well track the bearing failure evolution and accurately predict the RUL of bearings, which outperform the traditional PSW and Paris model, respectively. In addition, the proposed multi-time scale modeling approach verifies the feasibility and effectiveness of combining the data-driven model and physics model for rotary component RUL prediction and provides a new research direction for future studies.