My doctoral research in electrical engineering is on data-driven modeling and control of nonlinear systems: how to bring a recurrent network identified from data into a framework where robust-control tools apply.
1Research axes
Continuing to build the bridge between deep learning models and certified control.
1.1Recurrent networks in a form control tools can use
Using GRU models in the LPV-LFT framework
Problem.
Neural networks are increasingly used in control problems, but they connect poorly to the classical tools of automatic control.
Approach.
Rewrite a modified Gated Recurrent Unit (GRU) as a Linear Parameter-Varying / Linear Fractional Transformation (LPV-LFT) model, so that a network identified from data can be handled with dedicated control tools.
Contribution.
An LPV-LFT formulation of a modified GRU.
Checked.
Performance tests on constrained and unconstrained GRUs measure the losses caused by the assumptions of the formulation.
1.2Gain-scheduled H∞ control of identified nonlinear systems
LPV-LFT-based data-driven control of nonlinear systems with GRU models
Problem.
A prediction model is not yet a controller. The question is how to go from a recurrent model identified on a nonlinear plant to a controller that meets performance requirements.
Approach.
A strictly equivalent LPV-LFT formulation of the modified GRU, H∞ gain-scheduled LPV controllers, and an Unscented Kalman Filter (UKF) that estimates the GRU state so the varying parameters can be computed in real time.
Contribution.
A methodology connecting data-driven identification and robust LPV control.
Checked.
The methodology is evaluated on a nonlinear benchmark.
Certifying stability when linearization is inconclusive.
Work on exact certification with sum-of-squares (SOS) programs. It is under review and has not been peer reviewed yet.
Invariant-Manifold Jets for Explicit Facial Reduction in SOS Lyapunov Programs
Problem.
At a non-hyperbolic equilibrium, linearization can be inconclusive. The SOS programs that search for a Lyapunov function become degenerate there: the slow dynamics on the invariant manifold impose exact cancellations, which floating-point solutions may violate near the equilibrium.
Approach.
Compute these constraints before optimization, from a finite jet of the invariant manifold, then remove them from the program by an explicit facial reduction.
Contribution.
Two theorems: kernels forced on the Gram matrices of the decrease blocks, in any manifold dimension, and explicit exposing certificates obtained from rational moments.
Checked.
On a scalar benchmark of 80 systems, reduced central-point optimization and rational reconstruction certify every instance, with exact checks.
This teaching example applies the method to a mass-spring-damper system. It is separate from the benchmark in the papers.
3.1From the mass to the car
Adjust the parameter of the mass of the documentation, and watch the output, the control input and the range stated by the synthesis. Then see the same principle on the active suspension of a car, against a fixed controller.
Body acceleration [m/s²]50−5Fig. 1. A still of the simulation, on the broken road with the LPV controller. Below, the body acceleration over 4 seconds, with the controller (solid) and without actuator (dashed): 21% less in root mean square.
3.2H∞ LPV-LFT control, step by step
From the physical model to the synthesis formulation, with every equation and two interactive modules: the energy of a signal and the worst-case input.