SUBJECT: Ph.D. Dissertation Defense
BY: Yanfei Lu
TIME: Thursday, July 25, 2019, 11:00 a.m.
PLACE: MARC Building, 201
COMMITTEE: Dr. Steven Liang, Chair (ME)
Dr. Thomas Kurfess (ME)
Dr. J. Kurt Jacobus (MSE)
Dr. Shreyes Melkote (ME)
Dr. Thomas Habetler (ECE)


As the technology in mechanical engineering advances, increasing complex mechanical systems are developed to achieve better quality and accuracy with prolonged useful life. However, with the added components and processes included in the mechanical system, the relationship between the input process parameter and output process parameter is difficult to define precisely. The system exhibits a high dimensional nonlinear relationship from the input parameter to the output parameter. Additionally, with more components included in the system, the probabilities of a machinery breakdown increase. Although the quality of the components improves, it does not offset the complication of the added complexity. The addition of machine learning tools helps solving the complex problems without the requirement of complete understandings of the sophisticated systems or analytical models. The major two categories machine learning are implemented to understand mechanical systems are regression and classification. In machinery diagnosis, the machine learning (ML) tools are mainly used to adjust model parameters to gain enhanced performance of noise reduction and fault location recognition. In machinery prognosis, the ML tool are mainly used to predict defect size and the RUL. The contribution of this study is to implement intelligent algorithms with adaptation and various machine learning tools to assist the diagnosis of incipient fault signature of rotating machineries and propose a possible route for diagnose machinery remaining life. More specifically, various signal processing techniques including wavelet decomposition, principal component analysis, envelope analysis, spectral kurtosis, dictionary learning, deep learning are implemented to enhance the fault signature covered in background noise. The optimization algorithms contain recursive least square, LASSO, unscented Kalman filter to update the diagnostic model parameters both linearly and nonlinearly. As the machine degradation occurs, the parameters of the diagnostic model need to be updated to accurately capture the degradation trend and fault location. The need for adaptation of the model parameters cannot be overemphasized. The development of the new approaches involves two major parts. The first part of this study focuses on various signal processing techniques to obtain critical features in the acquired signal from the mechanical system during the early degradation stage. Case studies using the vibrational signal of rolling element bearings will be demonstrated to prove the validity and applicability of the developed algorithm. The second part of this study concentrates on prognosis of the fault size and location using the deep learning algorithm.