Research
Uncertainty quantification and decision making are increasingly demanded with the development of future technology in engineering and transportation systems. Among the uncertainty quantification problems, I am particularly interested in statistical modelling of engineering system responses with considering the high dimensionality and complicated correlation structure, as well as quantifying the uncertainty from a variety of sources simultaneously, such as the inexactness of large-scale computer experiments, process variations, and measurement noises. I am also interested in data-driven decision making that is robust to the uncertainty. Specifically, I deliver methodologies for anomaly detection and system design optimization, which can be applied to manufacturing process monitoring, distracted driving detection, out-of-distribution object identification, vehicle safety design optimization, etc.
Research Interests:
Methodology: out-of-distribution learning, Gaussian process modeling, generative models, anomaly detection
Applications: computer vision, vehicle safety, assistant driving system, smart manufacturing, health care
Out-of-distribution learning:
Deep Neural Networks have been evolving rapidly in the past decade and have demonstrated their promising capacities across many real-world applications. For most of these tasks, one of the fundamental assumptions is that the training and testing data come from the same probability distribution under which DNN models are trained in a closed-world manner. However, this assumption does not necessarily hold in real-world applications where people do not have control over testing samples, making the applications of DNNs unreliable. I am interested in developing methodologies to identify these testing samples under the setup of classification or object detection. My research primarily focuses on the uncertainty quantification and calibration of Deep Neural Network models to detect testing samples lying out of the distribution of training samples.
Design optimization based on computer experiments:
Computer experiments are often used for system design optimization to reduce the physical test cost. To ensure reasonable results, limited experimental physical test results are compared with the simulation results to ensure the accuracy. The differences between the computer simulations and experimental tests are typically caused by three sources: (a) inappropriate specification of unknown simulation parameters; (b) simulation model biases; (c) observation errors from the experimental tests. These three error sources can be combined and generally termed uncertainty. I am interested in developing statistical methodologies to search for optimal constant or adaptive system designs. The technical tools include Gaussian process modeling, robust optimization, and Bayesian optimization.
Generative models of complicated data:
Generative models have been widely investigated to explore target data distributions. The research challenges lie in high data dimension, incomplete dataset, and lack of physical interpretation. I am particularly interested in developing generative models for 3D point clouds and mesh data, with research focus on super resolution, domain adaptation, and physical-informed data generation. Under the support from UMTRI, the generative model is mainly trained on 3D human models obtained from humanshape.org.
Research Projects as PI:
A crash scenario reconstruction framework toward integrative pedestrian safety designs through data fusion and machine learning, Sponsor: Ford Motor Company, Total: $193,281, 2023 -- 2025
Generative Model of High-resolution 3D Shapes via Domain Adaptation, Sponsor: University of Michigan College of Engineering Seeding To Accelerate Research Themes, Total: $60,000, 2023 -- 2024
Adaptive Decision-Making for Airport Operations via Data Fusion, Sponsor: University of Michigan Internal Funds Graham Sustainability, Total: $10,000, 2023 -- 2024
Selected Research Projects as co-PI:
An Automated Crash Simulation Framework Using Parametric Human Model (PI: Hu, J.), Sponsor: Volvo Car Corporation aka Volvo Car Group, Share: $129,473, 2023 -- 2024
NHTSA Field Study Phase 3 L2 Evaluation (PI: Flannagan, C.), Sponsor: Department of National Highway Traffic Safety Administration, Share: $68,154, 2021 -- 2022
Improving Safety Equity between Women and Men in Frontal Crashes through Parametric Human Modeling (PI: Hu, J.), Sponsor: Toyota Technical Center, USA, Inc., Share: $84,656, 2022 -- 2023
Adaptive Restraint Designs through Machine-Learning (PI: Hu, J.), Sponsor: Ford Motor Company, Share: $32,498, 2022 -- 2024
Generating a Synthetic Population of Pedestrian Crash (PI: Flannagan, C.), Sponsor: Ford Motor Company, Share: $21,034, 2021 -- 2022
Data Collection Methods for ADS Operational Safety Monitoring (PI: Flannagan, C.), Sponsor: Department of National Highway Traffic Safety Administration, Share: $76,000, 2022 -- 2024