My research mainly focuses on industrial quality & reliability data modeling and analysis, failure detection and prediction, optimal experimental designs, and prognostics and health management. The research methodologies include statistical-model based approaches by using generalized linear mixed models, Bayesian analysis, simulations and optimization, dimension reduction, and other statistical learning / machine learning techniques.
Optimal experimental designs for product reliability tests
Decisions involved in data collection processes are important especially where not enough resources are allowed such as limited time and budget. We may address such a problem by applying the techniques of optimal experimental designs. In our lab, we study the optimal experimental designs for efficient product reliability tests.
Design of experiments with random effects model
In reliability tests, it is often the case that a completely randomized experimental design cannot be achieved. In this project, we consider effects of multiple test chambers and multiple test unit suppliers to the product failure time for efficient experimental designs. To do this, the iterative search algorithm is developed to determine optimal test conditions, test chamber assignment, and the number of test units allocations for each supplier simultaneously. (INFORMS 2020)
Simulation-based search for optimal DOE
In this project, we suggest “experiments for experiments” approach to create optimal planning of an ALT. We generate simulated data to evaluate the quantity of interest, e.g., 10th percentile of failure time, and apply the response surface methodology (RSM) to find an optimal solution with respect to the design parameters, e.g., test conditions and test unit allocations.
The R package ALTopt has been developed with the aim of creating and evaluating optimal experimental designs of censored accelerated life tests (ALTs). This package takes the generalized linear model approach to ALT planning. Three types of optimality criteria are considered: D-optimality for model parameter estimation, U-optimality for reliability prediction at a single use condition, and I-optimality for reliability prediction over a region of use conditions
System failure detection and prognostic
Early detection of system failure is an important issue for enabling condition-based monitoring and predictive maintenance. In our lab, we develop new algorithms and methods of detecting a system failure beforehand to prevent unexpected, even catastrophic, breakdown of the system.
Early failure detection of advanced manufacturing system
Multiple sensor data collected from a paper manufacturing machine provides useful information on the system’s health condition. Using this data, the goal of this project is to detect system failure beforehand to prevent unexpected breakdown of the system. The dataset is highly imbalanced, multivariate, and high-resolution streaming data. We develop adaptive methods by integrating nearest neighbor-based feature extraction and off-the-shelf machine learning algorithms.
Battery capacity prognostic by functional monitoring data
Conventional approaches to battery prognostic mostly focused on exploring the capacity degradation data. Seldom tried using other performance measures such as temperature, voltage, and current monitoring data that could be useful to predict the capacity. In this project, we aim to predict the capacity of lithium-ion batteries using a data-driven prognostics algorithm by adopting functional principal component analysis (fPCA) applied to battery monitoring data.
Quality & reliability data analysis
Manufacturing system anomaly root cause frequency determination
Even if the root cause of process variability can be identified, substantial cost may be required to remove the root cause, which makes the manufacturer consider the tradeoff between the potential benefit and its cost. In this project, a time series data set representing a key quality characteristic of a medical device component of a manufacturing process is given, and we aim to answer a question of “how frequently abnormal status occur in the manufacturing system” to assist in decision making of financial investment to eliminate the source of the root cause.
Product reliability data analysis with correlated observations
Typical experimental protocols used in accelerated life tests (subsampling and random blocks) should be taken into account in reliability data analysis. In this project, the correlated failure time observations from the step-stress accelerated life tests (SSALT) is modeled by generalized linear mixed model. Two numerical methods for parameter estimation are used, which are adaptive Gaussian quadrature (AGQ) and integrated nested Laplace approximation (INLA).
Health care data analysis
Tracking Alzheimer's disease progression path
To use neuroimaging technique to track Alzheimer's disease (AD) progression, it is desirable to have global indices which can summarize numerous complicated imaging features. In this research, we propose a new global index, derived from non-linear dimension reduction of brain MRI features, to track AD progression over time. We apply locally linear embedding (LLE) to a dataset from ADNI database, which includes 346 brain features derived from the volumes of brain regions and the cortical thickness. High dimensional brain MRI features could be represented well in low dimensional space by using manifold-based dimension reduction techniques.
Nursing activities profile analysis
We investigate an effective monitoring tool for health care providers' work profiles by adapting traditional quality control techniques used in manufacturing process monitoring and control.