Research

Process Data Analytics of High-Dimensional Processes

The availability of massive operation data and the advancement of machine learning and AI tools provide extensive resources to gain insights into complex systems and also importantly, better monitor the process faults. Early detection of process faults is critical to prevent disastrous consequence of large-scale systems in the process industry when equipment failures occur. Often these data are dynamically correlated and of high dimensions due to the large amount of sensors installed across multiple units of the process. Exemplary processes include petroleum refining, chemical production, food processing, etc. Our research in this direction is set to combine statistical process control, deep learning, system identification, and data analytics to extract knowledge from high-dimensional dynamical data, and establish sensitive & reliable algorithms for effectively detecting and diagnosing system faults. Timely detecting and accommodating the potential failures in large-scale plants can prevent catastrophic accidents from happening. This is critical for ensuring personnel and equipment safety and reducing economic loss.

Modeling, Control and Monitoring of Complex Dynamic Systems

Dynamic systems are ubiquitous across almost every aspect of our life such as vehicles, power grids, buildings, aircrafts, and stocks. Taming the dynamics of a complex system with automatic controllers is of significance especially for manufacturing to improve the efficiency & product quality, reduce operation cost, and ensure constraint satisfaction. Complex dynamic systems, e.g., HVAC systems, energy storage for power grids, pulp & paper-making, often contain a large number of variables with strong correlations, making it difficult to understand or control them. Our research in this direction aims at developing first-principle or data-driven models (e.g., with system identification techniques) for such systems, upon which advanced model-based controls (e.g., linear and nonlinear model-predictive control) can be enabled for controlling these systems. In addition, system dynamics often drift at different operating points which cause controller degradation over time. We combine machine learning techniques with control theory into a learning-and-control framework where the system dynamics, disturbance, and control performance can be learnt from the operation data to enable the adaption of controllers to various operation situations.

Data-driven Model Reduction and Control Tuning

Many dynamical systems are infinite-dimensional described by PDE models such as the lithium-ion battery (DFN model, single particle model), fluid dynamics (Navier-Stokes), etc. Simulating and control of such systems are challenging due to the complexity in addressing nonlinear PDE or high-dimensional ODE models. Model reduction has been popular to develop low-order models as a surrogate for simplifying the problem. Traditional model reduction techniques, such as proper orthogonal decomposition, balanced truncation, require the knowledge of the full-scale model which may not be accessible in practice. Our research in this direction focuses on using data-driven approaches such as dynamic mode decomposition/autoencoders to directly obtain low-order models from data, in case where first-principle model is absent. The data are often spatio-temporal and can be in the form of snapshots of velocity/temperature fields or image sequence. In addition, we develop computationally efficient algorithms with data-driven optimizations such as Bayesian optimization for tuning large-scale advanced controllers (e.g. MPC) for complex systems.

Page updated

Google Sites

Report abuse