Abstract
Traditional methods in reliability analysis mainly rely on scarce failure data and can no longer fully support modern reliability-related managerial decisions. In contrast, the evolution of sensor and monitoring technologies lead to new opportunities for data-driven methods in enhancing reliability and maintainability of many industrial systems. By employing on a large volume of real-time sensor data regarding to system degradation and characteristics, data-driven methods have been intensely investigated in both literature and practice to address various reliability-related topics, such as fault detection, reliability evaluation, remaining useful life (RUL) estimation, and maintenance optimization. Although quite large literature is presented for data-driven reliability analysis, several open research issues remain to be resolved. The thesis develops and employs advanced data-driven approaches to investigate the RUL estimation, maintenance optimization and reliability evaluation issues for bridging specified research gaps and meeting detailed considerations.Firstly, we study two problems of RUL estimation of degrading systems by taking advantage of the multi-dimensional, timeseries data trajectories:
RUL estimation under unbalanced data trajectory. Along with wide application of sensors, multidimensional timeseries data are commonly available for remaining useful life (RUL) estimation. In RUL estimation, the data trajectories in the training set contain the data up to the units’ failure while the data trajectories in the testing set do not. Although this fact has a significant negative effect on the accuracy of RUL estimation, it is considered by few literature. To deal with this, we propose a joint data-driven approach that adapts two models, AdaBoost regression and Long Short-Term Memory (LSTM), to estimate the RUL based on data trajectory extension. The proposed approach adapts the LSTM to learn the time series dependencies of training data and then extend the trajectories of testing data, aiming at reducing the variance of the lengths of data trajectory between the training and testing sets. Then, the proposed approach adapts the AdaBoost regression to estimate the RUL using the extended time series data. The proposed approach is competitive with state-of-the-art methods by demonstrating on two degradation datasets.
Bayesian degradation inference for RUL estimation with objective prior selection. With multiple sensors are used to assemble degradation signals from the same unit, Bayesian multisensor degradation modeling is often established to predict the RUL. However, one challenging question that remains to be resolved is how to select and construct a prior distribution for the random-effect parameter that captures the unit-to-unit heterogeneity on distinct sensor measurements. Motivated by objective Bayesian inference, we propose to use the Jeffreys prior to construct the prior of random-effect parameters in Bayesian degradation modeling. We present two critical results. (1) The Jeffreys prior for the random-effect parameters in the parameter space is a uniform prior for the degradation path models in the model space. (2) The Jeffreys prior maximizes the mutual information between the random-effect parameters and the observed multisensor degradation data. Result (1) is consistent with the parameterization-invariance property of noninformative priors. This property expresses the prior ignorance about the unit-to-unit heterogeneity as a uniform prior in the model space, and thus allows objective Bayesian inference in degradation modeling. Result (2) enables learning the unit heterogeneity as much as possible from limited degradation data. For the application of the Jeffreys prior in multisensor degradation modeling, we facilitate a supervised classification framework for unit health prognosis.
Afterward, considering that the degradation formulation is usually unknown for a system working in practice, the following work explores the CBM optimization for continuous degrading systems:
CBM optimization under unknown degradation formulation. CBM policy has often been established on degradation models. However, the formulas of the degradation processes are usually unknown and hard to be determined for a system working in practice. To optimize the CBM policy under nonideal conditions regarding to the availability of prior information of degradation formulation, we develop a model-based reinforcement learning approach. The developed approach determines maintenance actions for each degradation state at each inspection time over a finite planning horizon, supposing that the degradation formula is known or unknown. At each inspection time, the developed approach attempts to learn an optimal assessment value for each maintenance action to be performed at each degradation state. The assessment value quantifies the goodness of each state-action pair in terms of minimizing the accumulated maintenance costs over the planning horizon. To optimize the assessment values when a well-defined degradation formula is known, we customize a Q-learning method with model-based acceleration. When the degradation formula is unknown or hard to be determined, we develop a Dyna-Q method with maintenance-oriented improvements, in which an environment model capturing the degradation pattern under different maintenance actions is learned at first; Then, the assessment values are optimized while considering the stochastic behavior of the system degradation. The final maintenance policy is acquired by performing the maintenance actions associated with the highest assessment values. Numerical examples are presented to illustrate the applications.
Finally, we generalize our consideration from single-component systems to multicomponent network systems, for which effective and efficient reliability evaluation is a critical issue in making decisions such as maintenance planning and performance improvement:
System reliability evaluation based on structural similarity. Considering the largescale complex systems that can represent construction of components in a unit, a transportation system, a supply chain, a social network system, and so on, some nodes have similar topological structures and thus play similar roles in the network and system analysis, usually complicating the analysis and resulting in considerable duplicated computations. To ease computation load of evaluating system reliability, we present a graph learning approach to define and identify structural similarity between the nodes in a network or the components in a network system. Based on the structural similarity, we investigate component clustering at various significance levels that represent different extents of similarity. We further specify a spectral-graph-wavelet based graph learning method to measure the structural similarity and present its application in evaluating system survival signature. Numerical examples are investigated to show the insights of structural similarity and effectiveness of the graph learning approach. The proposed structural similarity, component clustering, and graph learning approach are effective in simplifying the complexity of the network systems and reducing the computational cost for complex network analysis.
Through the research of the fours problems related to RUL estimation, maintenance optimization and reliability evaluation, the thesis aims to bridge several gaps that are of interest to researchers and practitioners in effectively and efficiently developing and employing data-driven reliability analysis. The findings in the proposed research will provide scientific insights to support managerial decision making in reliability-related management.
| Date of Award | 3 Aug 2021 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Xiaoyan Zhu (External Supervisor) & Min XIE (Supervisor) |
Cite this
- Standard