Precision Motion Control Laboratory
Department of Mechanical Engineering
Massachusetts Institue of Technology
Address: 77 Mass. Ave., 35-030, Cambridge, MA, 02139
Email: leizhou AT mit DOT edu
[C.V.] [Google Scholar] [LinkedIn]
I am PhD Candidate in the Department of Mechanical Engineering at MIT, under the supervision of Professor David L. Trumper. I am fortunate to have Professor Alexander H. Slocum, Professor Jeffery H. Lang, and Professor James L. Kirtley Jr. on my thesis committee together with my advisor.
My research interest focuses on the design and control for mechatronics systems, with an emphasis on electromagnetic actuators and sensors design, precision machines, control and estimation for motion systems, magnetic levitation, and manufacturing systems.
This project focuses on the design, modeling, and testing of a magnetically suspended reaction sphere driven by an one-axis hysteresis motor (1D-MSRS). The goal of this work is twofold: (a) exploring the possible design of a magnetically suspended reaction sphere driven about three-axes for spacecraft's attitude control; and (b) evaluating the performance of a hysteresis motor for the reaction wheel/sphere application. The novel 1D-MSRS demonstrates a hysteresis motor with a solid steel spherical rotor, which is magnetically suspended in all translational directions by a combination of an electromagnetic actuator and a bearingless motor. Modeling of the sphere’s magnetic suspension is demonstrated, and its motor dynamics is analyzed by an equivalent circuit model. This study also demonstrated that the hysteresis motor has a good potential for being used for high speed, low vibration reaction wheels.
In the induction motor speed sensorless control, the bandwidth of the speed control loop is mainly limited by the convergence rate of the state estimation, especially the speed estimation. In order to improve the motor's speed tracking bandwidth and therefore allowing them being used in applications where the dynamic performance is critical, new state estimation schemes are needed. In this project, we first studied the application of the Moving Horizon Estimation (MHE) for induction motor state estimation. To improve the estimator's robustness with respect to the model parametric uncertainties, we designed a dual-stage adaptive MHE to estimate the parameters and the states simultaneously.We also studied improvements towards the full-order adaptive observer for induction motors. In order to provide a better balance between the fundamental trade-off between bandwidth and noise in the speed estimation, we proposes a new speed estimation method with non-linear adaptation gain and feedforward terms. An iterative tuning method is used to automatically tune the feedforward gains for the speed estimation online. Experiments show that compared with the baseline adaptive estimation approach, the proposed method can demonstrate a 20x bandwidth improvement. Poster: [ Link].
This project targeting at design and further deliver a set of design rule for an advanced thermal energy storage (TES) system directly charged by concentrated solar power (CSP). In the system, the concentrated beam enters directly into the molten salt TES vessel where it is absorbed and stored as sensible heat by the molten salt through the depth; hence the molten salt tank simultaneously acts as a volumetric receiver and a thermal energy storage system. We call this novel concept Concentrated Solar Power on Demand (CSPonD). One major feature of this system is the concept of virtual two-tank: using a loose-fitting horizontal moving raft (divider plate) in the vessel to separate hot and cold salt, and therefore making a more efficient thermocline than other similar designs. My research in this project focuses on the control of the temperature and flow field in the tank, and the design and control for the divider plate.
Flexlab/LevLab is an educational mechatronic system for control and mechatronics teaching in the Department of Mechanical Engineering at MIT. The FlexLab is a system of flexible cantilever beam with permanent magnets attached to it, while the LevLab demonstrates a magnetic suspension system, in which a spherical permanent magnet is magnetically levitated. Linear power amplifiers are used to energize the actuator coils, and Hall effect sensors are used to measure the position of the permanent magnet. Both labs are implemented on one printed circuit board, with actuators, sensors, and power amplifiers arranged on it. This system works together with myRIO from National Instruments, and the real-time controller is implemented in the myRIO. The system is designed to be portable, which allow students to work on experiments outside of the laboratory, and thus making better use of the time in lab. The design file of the Flexlab/LevLab is open source at here:
Here is a list of my teaching/mentoring experiences:
This project is with Dan Rathbone, Jason Fishman, Wilhelmena Figueiredo, Rebecca Kurfess, Kelsey Seto, and under the supervision of Professor Martin Culpepper.
In this project, we work on the modeling, fabrication, measurements and testings for a desktop lathe. There are three main subsystems for the lathe: (1) the spindle, (2) the cross-slide, and (3) the z-axis leadscrew. An additional part that we worked on is the design and fabrication for a metrology frame from the headstock to the lathe tool, which help eliminates errors from some deformation of the machine, and thus improve the precision.
The lathe survived the deadtests (sledgehammer hitting, dropping from desk height). We won the 1st place in the final contests in cutting steel, and the 3rd place in cutting aluminum.
This project studies the effect of structure growth for neural networks during training. The general idea is to start with a simple network and gradually crank up the network size. With the small and simple network, the big features of the data can be captured roughly, and adding some more neurons/layers after training the small net- work allows better capturing the details of the data. Several heuristic network structure update methods are tested, and the experimental results with several standard benchmark datasets are dis- cussed. It is shown that compared with directly training a large (in depth and width) neural net, the growing network demonstrates better accuracy and convergence.
Here is a link to the full project report.
In my spare time, I enjoy drawing and painting, reading fictions, and running. Here are some of my drawing/paintings :-)
| || || || |