Predictive Modeling for Prognostics and Health Management

預診斷和健康管理框架下的模型預測方法

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date14 Sep 2018

Abstract

Predictive modeling methods utilize modeling, identification techniques, and learning algorithms to predict future outcomes, which provides promising support for diagnosis, prognosis, and health management for both human beings and engineering products or systems. There are remaining many challenges and difficulties in improving the prognostic and health management with predictive modeling methods, including the following:

• The historical data are incomplete, and the sources of the incomplete data are very complicated;

• Many risk factors are health relate; it is difficult to identify critical factors and quantify the effects due to unknown causes;

• Noise may inevitably interfere with measurement data and is difficult to remove;

• The degradation process from a healthy state to an unhealthy state is non- stationary, nonlinear, and non-Gaussian; and the mechanism principles are unclear in many cases.

The corresponding methods for solving the preceding difficulties are presented in this thesis. The main contributions are highlighted as follows:

1. Different imputation methods such as Local Linear Interpolation (LLI) and Global Statistic Approximation (GSA) are proposed and incorporated to deal with complicated types of incomplete data in practice.

2. Risk factors are ranked according to the importance with respect to both linearity and nonlinearity. The wrapper method for feature selection is implemented with importance metrics to avoid model overfitting, which also increases the interpretability of our prediction models.

3. A personalized modeling method is proposed to capture the properties of each individual and common characteristics across individuals. The time series analysis methods is utilized to establish the individual model, and the general regression modeling methods are employed to model common properties across the entire dataset.

4. The Bayesian inference methods, such as Kalman filter, extended Kalman filter, and particle filter, are utilized to update the model online with the noise data;

5. Some novel empirical models are established to capture the health degradation characteristics as much as possible.

6. A dual particle filter framework for identifying a highly nonlinear dynamic model is proposed and successfully applied to the remaining useful life prediction of lithium-ion batteries.

Throughout the thesis, two main case studies are conducted to address and validate the proposed methods. The first case study addresses the diagnosis and prognosis of scoliosis, and the second addresses the remaining useful life prediction for lithium-ion batteries. The validation experiments demonstrate that our methods are effective and efficient, and have the potential to be implemented in real practice.

    Research areas

  • Prognostics and health management, data driven model, data preprocessing, remaining useful life prediction, particle filter