Internet of Things-Enabled Degradation Modeling, Inference, and Prognosis
出版項
2020
說明
1 online resource (171 pages)
文字
text
無媒介
computer
成冊
online resource
附註
Source: Dissertations Abstracts International, Volume: 82-02, Section: B
Advisor: Liu, Kaibo
Thesis (Ph.D.)--The University of Wisconsin - Madison, 2020
Includes bibliographical references
Degradation is common in a variety of engineering systems, which can lead to system failures. Enabled by the Internet of Things technology, sensors have been widely used to monitor the degradation process of engineering systems. By analyzing the collected sensor signals, the failure time of an engineering system can be predicted, and appropriate maintenance can be scheduled to avoid unexpected failures. This brings an unprecedented opportunity for developing advanced methodologies that enable and assist (i) the efficient handling of the rich and diverse sensor measurements, (ii) the estimation and inference of the unobserved degradation status, and (iii) the exploitation of the acquired knowledge for more enhanced prognosis of the future dynamics and decision-making for predictive maintenance. This thesis focuses on Internet of Things-enabled degradation modeling, inference, and prognosis to develop data analytics methodologies by effectively combining advanced statistics, machine learning and engineering domain knowledge. The proposed methodologies enable (i) the proper and robust modeling of the degradation process and the inter-relations of the sensors, (ii) an accurate estimation of the unknown degradation status, (iii) an accurate prediction of the future behaviors and failure time, and (iv) the enactment of decisions for predictive maintenance. The first chapter introduces the background and elaborates the challenges in degradation modeling, inference, and prediction enabled by Internet of Things. The objective of this thesis is also highlighted. Chapter 2 focuses on health index methods for degradation modeling and prognostics with multiple sensor signals. While a health index is constructed by combining the multiple sensor signals to characterize the degradation process, existing health index methods are limited to linear fusion function. In this chapter, we propose a novel health index method that extends the linear fusion function to nonlinear functions by incorporating the kernel methods. Chapter 3 focuses on a more fundamental issue regarding the theoretical justification of the health index methods. Existing health index methods are heuristic, and the prognostic performance of the constructed health index cannot be guaranteed. To address this issue, we propose to use indirect supervised learning, where the failure time information is used as an indirect indicator of the underlying degradation status to guide the construction of the health index. In this way, the constructed health index is theoretically guaranteed to characterize the true degradation process. Chapter 4 further proposes a generic framework for multisensor degradation modeling, where a novel concept called failure surface is proposed to define system failure based on multiple sensor signals, and a new method is proposed to estimate the failure surface by transforming the degradation modeling problem into a classification problem. As a result, the proposed method is flexible to explore complicated relations of sensor signals, is capable of handling asynchronous signals, and can automatically screen out non-informative sensors. Chapter 5 proposes a systematic method for degradation modeling and prognosis that can be widely used in different scenarios. After extracting features for each sensor signal, local linear models are adopted to establish the relation between the extracted features and failure time. A goodness-of-fit measure is further proposed to assess the adequacy of the local linear model. If a unit is monitored by multiple sensors, decision-level fusion and feature-level fusion are further used to fuse the information from the sensors. Chapter 6 then summarizes the contributions of the thesis. In summary, this thesis contributes to the Internet of Things-enabled degradation modeling, inference, and prognosis by developing systematic data-driven analytics methodologies. The research possesses a great potential for applications in manufacturing, health care, and energy facilities, etc., where Internet of Things technology has been rapidly adopted
Electronic reproduction. Ann Arbor, Mich. : ProQuest, 2020
