Data analysis/selection/optimization for estimation:
Sponsor: LG Chem Inc.
Estimating the internal states of a physical plant is an important task of the control system. For example, the major function of the battery management system (BMS) is to monitor the critical battery states regarding performance, including state of charge (SOC), state of health (SOH), and power capability among others. Estimation is typically performed by an algorithm, such as a Kalman filter, which produces estimates based on a plant model and the measured input/output data. Traditionally, the research on estimation is dominated by studies on algorithm and model with data somehow neglected. We are particularly interested in studying the role of data in estimation and finding answers to fundamental questions such as 1) do we have enough data for estimating a certain system variable? 2) what are the best data we should use? 3) what is the systematic way of finding them? We are exploring both the (physical) model-based and data-driven approaches to seek the answer. The former uses models to interpret data and is particularly suitable for the cases where limited data but adequate physical knowledge of the system is available. The latter, on the other hand, could achieve high-fidelity fitting of the input and output data using methods such as machine learning. It requires massive data for learning but no physical knowledge about the system. We are looking for innovative methods to combine these two approaches to advance the frontier in data analysis.
Combination of Model-based and Data-driven Approaches for Estimation
System-level estimation and control under sparse sensing and uncertainty:
Systems with large numbers of single cell units are commonly seen in practice especially in energy applications. For example, the battery packs of electric vehicles often consist of hundreds or even thousands of single cells, and so are the fuel cell systems and solar arrays among others. It is often of great interest to monitor and control the behavior of every single cell. Traditional approaches are usually built upon one assumption: the available sensors deployed in the system could provide sufficient information for estimation and control, or, in control terminology, the system is observable. Unfortunately, this ideal condition is rarely satisfied in reality due to the constraints on cost, packaging, and system complexity. As a result, system uncertainty becomes inevitable, which could lead to significant estimation and control errors. We are interested in exploring answers to the following 3 fundamental questions: 1) what is the worst-case error under all possible combinations of uncertainty? 2) how to design robust observer/controller to minimize the worst-cases error? 3) where are the best locations to deploy the limited sensors? By addressing these issues, we are looking forward to establishing a framework for performance evaluation, control design, and sensor deployment of large complex systems. Current applications for this research thread is the development of system-level solutions for battery management.
System-level Estimation and Control of Battery Packs
Integration of control with design and optimization:
Collaborator: Professor Yi (Max) Ren, Arizona State University
Traditionally, the controller development for a mechatronic system is decoupled from the design and optimization of the system itself. As an example, when designing a robotic arm or an intelligent structure, the first step is usually formulating the material and topology to achieve good mechanical performance, such as compliance and thermal/electrical conductivity. This step is then followed by sensor/actuator deployment and controller/observer design to satisfy the control and estimation requirements. A major drawback of this practice is that the control design will often be limited by the fundamental constraint imposed by the system structure because the first design step did not consider control-related metrics such as controllability and observability. Our approach to address this issue is to formulate a multi-objective optimization framework for integrated system (structure) and control design. We will incorporate control metrics into the original design problem and find the optimal solution satisfying multiple design goals by using the latest development in the science of optimization. In this way, we look to significantly improve the control performance with little or no sacrifice of the desirable system properties. Current applications include battery pack geometry design/sensor deployment and co-optimization of compliance and load observability of cantilever beam through topology design.
Multi-physics Modeling and Simulation:
Understanding and modeling the dynamics of a physical system is the basis for controlling its behavior. We are interested in performing multi-physics measurement, modeling, and simulation of dynamic systems. Current applications are primarily battery and energy systems. For example, Lithium ion battery is a complicated multi-physical system featuring coupled electrical (e.g. voltage, current, and state of charge), thermal (e.g. temperature, heat generation, and heat transfer), and electrochemical (e.g. charge transfer kinetics, lithium ion diffusion, and side reaction) dynamics. We have been working on observing and quantifying these dynamics by using traditional and frontier in-situ and ex-situ measurement technologies, such as cycling test, electrochemical impedance spectroscopy (EIS), and neutron imaging among others. Based on the measurement data, we have developed models to capture and simulate the evolution of battery internal states under a wide range of operating conditions. These models could enable physics-based design, estimation, control, and diagnostics of battery system.
Multi-physics Modeling of Lithium Ion Battery