How can event-triggered controllers for LPV systems be learned from data?

Controllers learned from data, not models

The paper targets a practical bottleneck: nonlinear engineering systems can be approximated by LPV representations, in which a linear system is scheduled by a measurable time-varying signal, but deriving those models by first-principles analysis or system identification is computationally heavy and offers limited theoretical support for control synthesis. The authors instead ask whether a stabilizing, event-triggered LPV controller can be computed from recorded trajectories alone.

Their answer is a design pipeline with three linked outputs: a condition on how much data must be collected, a set of semidefinite programs whose feasible solutions give the gain-scheduling controller, and a triggering rule that decides when the controller-to-actuator channel actually transmits. Stability is certified in a robust-control sense, so the guarantee is stated for perturbed LPV systems rather than idealized noise-free ones.

The data must be rich enough to represent the system

The paper's first technical step is a condition called θ-persistence of excitation of order L, which requires the minimum singular value of a Hankel matrix built from the recorded signal to be at least θ. In plain terms, the collected trajectory must excite the system richly enough that the data matrices span the behavior the controller needs.

For LPV systems the authors derive a specific version of this condition. With an excitation order of (1+ℓ)n+1 and a data length T at least n(1+ℓ)(1+m(1+ℓ))−1, they show that the stacked data matrix Θ=Col(U, UP, X, XP) has rank (1+ℓ)(n+m), provided the excitation level exceeds √Lo·cω·δ/ρ, where δ bounds the perturbation and ρ is a system-dependent constant. This is the paper's answer to how long and how rich a data record must be before direct data-driven LPV synthesis is meaningful; it is a prerequisite, not a performance claim.

Stability certificates become solvable semidefinite programs

With the data representation in place, the authors derive stability certificates for the event-triggered LPV system using Petersen's lemma and the full-block S-procedure. These tools convert a matrix inequality that depends on the scheduling signal into a finite set of constraints checked at the vertices of the polytopic scheduling set, which is the same strategy used in model-based LPV approaches.

The practical consequence is that the controller gains and triggering parameters come from convex semidefinite programs rather than from an infinite family of inequalities. The paper is explicit that this convexification is also a source of conservatism, alongside data uncertainty from bounded disturbances and finite data length. Readers should therefore treat feasibility as a certificate under the paper's assumptions, not as a claim of optimality.

The triggering rule cannot fire at every instant

The event-triggered scheme transmits at instants k̂i chosen by a threshold: the next event occurs when νkᵀΨ1νk ≥ xkᵀΨ2xk + v, where νk aggregates the sampling-induced error and the scheduling-dependent mismatch, Ψ1 and Ψ2 are positive definite matrices to be designed, and v is a positive constant. Because νk depends on the scheduling signal as well as the state error, the rule is not a direct copy of event-triggered designs for linear time-invariant systems.

The authors prove that the rule inherently prevents triggering at every instant. If an event fired at each step, the sampling error would be zero and νk would vanish, making the left side zero while the right side stays strictly positive because v > 0 — a contradiction. This is a structural property of the design, not an empirical observation from the simulations.

Tracking needs an added integral compensator

Stabilization alone does not cover reference tracking, so the paper introduces an auxiliary integral compensator to build an augmented LPV system and derives corresponding data-based stability certificates for robust output tracking. The tracking guarantee is again stated as robust stability of the augmented closed loop, with the compensator carrying the integral action needed to drive output error.

This extension matters because the authors note that existing data-driven LPV control work cannot directly handle robust output tracking. The tracking results inherit the same data and excitation prerequisites as the stabilization results, so the data-length condition must be rechecked for the augmented system rather than assumed from the stabilization case.

A wheeled-vehicle example cuts updates by over 98.98%

The numerical validation includes a wheeled-vehicle lateral tracking example with scheduling signal pk = (2vk − (vmax+vmin))/(vmax−vmin) constrained to [−1, 1], a simulation horizon of 45 s, and a bounded random disturbance with ‖ω̂k‖ ≤ 0.1. The required data length was set to T̂ = 17, yielding rank Θ̂ = 8 as required for the augmented system.

Against a periodic-sampling baseline with Ntr = N = 2250 triggering instants and average inter-event interval Δk = 1, the data-driven event-triggered controller triggered fewer than 25 times in every tested parameter setting. The reported communication enhancement ratio η = 1 − Ntr/N was at least 98.98%, with average inter-event intervals ranging from about 79.8 to 144.7 sampling instants across the five cases. These are simulation results under the stated disturbance bound and parameter settings, not hardware measurements.

The guarantee is practical exponential ISS, not exact stability

The stability notion used throughout is practical exponential input-to-state stability, which allows a nonzero ultimate bound ν rather than requiring the state to converge exactly to zero. In the event-triggered case the bound includes a term c0·v tied to the triggering offset v, so the triggering parameter directly shapes how tight the ultimate bound can be.

Two scope conditions qualify the results. The design assumes the full state xk is measurable; the authors note that if it is not, an observer-based estimate could be used, but the estimation error would then enter both the closed-loop dynamics and the triggering condition and would need to be incorporated into the stability analysis. The framework also assumes bounded perturbations and does not explicitly model measurement noise during data collection, though the authors suggest a possible extension. The reported numbers come from numerical examples, so the 98.98% figure should be read as a simulation outcome under those settings.

Data richness is the binding constraint to check first

The paper's own ordering makes the priority clear: before solving any semidefinite program, a practitioner must confirm that the collected trajectory meets the θ-persistence condition at the required order and length, because the rank guarantee for the data representation depends on it. Only after that prerequisite holds do the controller gains and triggering matrices become meaningful outputs of the convex programs.

The comparison that most changes how a reader should evaluate this work is between the two communication regimes: periodic sampling updates at every one of 2250 instants, while the learned event-triggered scheme updates fewer than 25 times in all five tested cases. The design question is therefore not whether event triggering saves communication — the example shows it does — but whether the data record is rich enough, and the perturbation bound tight enough, for the certificate to hold in the first place.

Sources used in this answer