SMT-Based Stability Verification of an Industrial Switched PI Control Systems

Stylianos Basagiannis, Ludovico Battista, Anna Becchi, Alessandro Cimatti, Georgios Giantamidis, Sergio Mover, Alberto Tacchella, Stefano Tonetta, Vassilios A. Tsachouridis

TL;DR

This paper demonstrates a scalable approach to formally verifying the stability of industrial-grade aircraft engine controllers. By combining traditional control theory (Lyapunov functions) with symbolic logic (Satisfiability Modulo Theories, or SMT), the authors successfully proved the robustness of a 21-variable hybrid system, bridging the gap between "fast but risky" numerical analysis and "precise but slow" formal methods.

What Problem Does This Paper Address?

In the world of aerospace, "almost stable" isn't good enough. Engineers typically design control systems using differential equations and verify them using numerical simulations. While numerical methods are efficient, they are susceptible to floating-point errors and cannot provide an absolute mathematical guarantee that a system will remain stable under all possible perturbations.

On the other hand, symbolic verification (formal methods) provides a rigorous proof of correctness. However, these techniques often suffer from the "state-space explosion" problem, where the computational complexity grows exponentially with the number of variables. Most symbolic tools struggle with systems larger than a few variables; this paper tackles an industrial model of an aircraft engine with 18 internal states and 3 control inputs, totaling 21 variables—a significant leap in scale for formal verification.

Methods

The authors employ a multi-step pipeline to achieve verification:

  1. System Reformulation: They take a switched Proportional-Integral (PI) controller and reformulate it into an autonomous Piecewise-Affine (PWA) system. This allows them to treat the controller and the engine as a single hybrid dynamical system.
  2. Lyapunov Synthesis: To prove stability, one needs a "Lyapunov function"—essentially a mathematical "energy" function that always decreases as the system moves toward its steady state. The authors used six different methods to generate candidate functions, including solving Lyapunov equations and using Linear Matrix Inequalities (LMIs).
  3. SMT-Based Validation: Once a candidate function is generated numerically, the authors use SMT solvers (like Z3, CVC5, and Mathematica) to symbolically check if the function satisfies the strict requirements of Lyapunov theory. This step ensures that numerical rounding errors didn't result in a "fake" proof.
  4. Robustness Synthesis: They calculate "robust regions" (ellipsoids in the state space). If the engine's state stays within this region, the system is guaranteed to stabilize without triggering a control mode switch, even if the pilot's commands or the engine's environment change slightly.

Datasets

The primary "dataset" is a high-fidelity mathematical model of a dual-spool turbofan engine.

  • Full Model: 18 internal state variables.
  • Benchmarks: To test scalability, the authors used Balanced Truncation to create reduced-order models with 3, 5, 10, and 15 variables.
  • Rounding Tests: They also created "truncated" versions of these matrices (rounding entries to integers) to evaluate how numerical precision affects the verification process.

Key Insights and Ideas

  • Numerical Methods Can Lie: The authors found that some numerically synthesized Lyapunov functions actually failed the symbolic check. This highlights the danger of relying solely on standard optimization toolboxes for safety-critical systems.
  • The "LMIα" Advantage: Among the synthesis methods, the LMIα approach (which adds a convergence speed parameter) proved to be the most robust. It was the only method that could still produce a valid proof even when the numerical data was rounded to just four significant figures.
  • Solver Performance: While SMT solvers like Z3 and CVC5 are powerful, the authors found that an ad-hoc implementation of Sylvester’s Method (checking the determinants of matrix minors) was often the most efficient way to validate positive definiteness in these specific control problems.
  • Scalability is Possible: By using model reduction and smart reformulation, the authors proved that symbolic techniques are finally reaching a point where they can handle industrial-sized problems rather than just academic "toy" examples.

Key Results

  • Successful Verification: The team successfully verified the stability of the full 21-variable system for individual control modes (Thrust control and LPC Spool Speed control).
  • Performance Benchmarks: For a 10-state system, synthesis took less than a second for most methods. For the full 18-state system, synthesis remained fast (under 1 second), but symbolic validation via SMT solvers varied from a few seconds to timeouts, depending on the solver used.
  • Robustness Bounds: The authors successfully calculated the maximum allowable perturbation ($\epsilon$) for reference values, providing a concrete safety margin for engine operation.

Limitations and Future Work

  • The Switching Gap: While the authors proved stability for each individual control mode, they noted that proving stability during the transition (switching) between modes remains difficult. Their attempts to synthesize a "piecewise quadratic Lyapunov function" for the entire switched system failed due to numerical errors in the LMI solvers.
  • Property Scope: The current work focuses on stability and robustness. Future research aims to extend this to more complex "temporal properties"—for example, proving that the system not only stabilizes but does so within a specific timeframe without ever exceeding a temperature limit.
  • Competition: The authors plan to archive this case study as a benchmark for the Competition on Applied Verification for Continuous and Hybrid Systems (ARCH) to encourage other researchers to tackle this industrial challenge.

Summary updated

Abstract

The control of complex systems is typically designed describing the physical system with differential equations. The standard approach to their verification employs numerical analysis, which is suitable to prove stability properties, but is susceptible to numerical errors. On the other side, symbolic techniques give precise analysis results but typically do not scale to industrial size problems. In this paper, we consider the control design of an aircraft engine. The engine model is represented by a linear state space model of 18 internal state variables, 4 outputs, and 3 inputs. The control switches between two PI controllers, one for thrust control and another for low-pressure compressor spool speed control, based on the engine state and pilot commands. After reformulating the PI controllers in terms of differential equations, we obtain a hybrid system with 21 state variables and two modes, for which we want to prove with symbolic techniques the robustness of the stable states to perturbation. We achieved the verification with standard methods to synthesize quadratic Lyapunov functions and SMT techniques to synthesize neighborhoods of the stable states for which we have symbolic proof of stability.
BibTeX
@inproceedings{basagiannis2023smt,
  author = {Stylianos Basagiannis and Ludovico Battista and Anna Becchi and Alessandro Cimatti and Georgios Giantamidis and Sergio Mover and Alberto Tacchella and Stefano Tonetta and Vassilios A. Tsachouridis},
  title = {{SMT-Based Stability Verification of an Industrial Switched PI Control Systems}},
  booktitle = {DSN-W},
  year = {2023},
  url = {https://openreview.net/forum?id=hJo56YK1hp},
}

Loading related content...

Figure