Data-Driven Input Reconstruction and Experimental Validation

Jicheng Shi, Ying Zhao Lian, Colin N. JonesIEEE Control. Syst. Lett 2022

TL;DR

This paper introduces novel data-driven methods for estimating unknown inputs to a system, using only historical input-output data and Willems' Fundamental Lemma. The proposed "unknown input estimators" (UIEs) can reconstruct inputs asymptotically without needing initial system states or explicit system models, and are validated through simulations and a real-world building occupancy estimation experiment.

What Problem Does This Paper Address?

Many real-world systems have inputs that are crucial for understanding their behavior but are difficult or impossible to measure directly due to cost, privacy concerns, or technical limitations. Examples include the number of occupants in a building (affecting temperature) or the cutting force in machine tools. Traditional methods for estimating these "unknown inputs" often rely on having a precise mathematical model of the system. However, obtaining such models can be challenging. This paper tackles the problem of reconstructing these unknown inputs directly from historical input-output data, bypassing the need for an explicit system model. This is particularly useful in scenarios where online input measurements are not feasible.

Methods

The core of the paper's methodology is built upon Willems' Fundamental Lemma, which provides a way to characterize system responses using historical data. The authors propose two types of data-driven Unknown Input Estimators (UIEs):

  • Open-Loop UIE (op-UIE): This approach directly estimates the unknown input based on historical data and a specific structure derived from Willems' Lemma. The stability of this estimator is analyzed using Lyapunov conditions. The design involves solving an optimization problem, which is then reformulated into a Linear Matrix Inequality (LMI) feasibility problem for practical implementation.
  • Closed-Loop UIE (cl-UIE): This method extends the op-UIE by incorporating a Luenberger observer-like feedback structure. It uses the error between the actual output measurement and a data-driven predicted output to correct the input estimate. This makes the cl-UIE more robust to measurement noise compared to the op-UIE. The stability of the cl-UIE is also analyzed, showing it depends on the stability of a specific matrix related to the system dynamics.

Both methods leverage Hankel matrices constructed from historical input-output data to represent system behavior. The key innovation lies in deriving the estimator's parameters (matrices like $A_{UIE}$ and $B_{UIE}$) directly from these data-driven representations, rather than from a pre-defined system model.

Datasets

The paper uses two types of datasets:

  • Simulated Data: For simulations, a synthetic unstable Linear Time-Invariant (LTI) system is defined with specific matrices (A, B, C, D). Input-output trajectories are generated for 50 time steps, excited by random inputs.
  • Real-World Experimental Data: An experiment was conducted in a whole building (Polydome) at EPFL to estimate occupant numbers. The dataset includes historical indoor temperature, weather conditions, heat pump power, CO2 levels, and manually recorded occupant numbers. Data from five weekdays were used to build the Hankel matrix. This dataset is used to compare the proposed cl-UIE with linear regression (LR) and Gaussian Process Regression (GPR).

Key Insights and Ideas

  • Data-Driven Input Reconstruction without Models: The most significant insight is the successful demonstration that unknown inputs can be reliably estimated using only historical input-output data and Willems' Fundamental Lemma, completely bypassing the need for explicit system models. This opens doors for applications where system identification is difficult or impossible.
  • Asymptotic Convergence: Both proposed UIEs are shown to asymptotically converge to the true unknown input, even without knowing the system's initial conditions. This is a crucial property for practical online estimation.
  • Open-Loop vs. Closed-Loop Trade-offs: The paper clearly distinguishes between op-UIE and cl-UIE, highlighting that while op-UIE is simpler, cl-UIE offers improved robustness to measurement noise due to its observer-like structure.
  • Data-Driven Stability Conditions: The stability of the proposed UIEs is characterized using Lyapunov conditions and LMI formulations, providing a rigorous framework for designing stable estimators directly from data.
  • Experimental Validation in a Real-World Scenario: The application to building occupancy estimation using CO2 levels is a strong validation of the method's practical applicability in a complex, non-linear (though approximated as linear) real-world system.

Key Results

  • Simulation Performance: Simulations show that both the op-UIE and cl-UIE schemes achieve fast convergence in estimating the unknown input, even for an unstable system. The cl-UIE demonstrates better performance in terms of estimation error compared to the op-UIE, especially when considering noise.
  • Experimental Performance: In the building occupancy estimation experiment, the proposed data-driven cl-UIE achieved a Mean Absolute Error (MAE) of 7.88, outperforming linear regression (MAE=8.48) and being slightly worse than Gaussian Process Regression (MAE=7.66). Importantly, the cl-UIE was observed to track the actual occupancy trajectory more smoothly than LR, while GR showed significant fluctuations. The op-UIE did not perform well in this experiment, likely due to measurement noise and system nonlinearities.

Limitations and Future Work

  • Existence of UIE: The paper acknowledges that the existence of a data-driven UIE is not guaranteed for all systems. While model-based existence conditions are known, the conditions for data-driven setups remain an open area for future research.
  • Sensitivity to Noise (op-UIE): The op-UIE is noted to be more sensitive to measurement noise, as its design relies on null space computations which can be affected by noise.
  • System Nonlinearities: The experimental validation was performed under the assumption of a linear approximation for a nonlinear building ventilation system. Extending the data-driven UIE framework to handle significant nonlinearities is a promising future direction.
  • Data Requirements: The performance of data-driven methods is inherently dependent on the quality and quantity of historical data. Further investigation into the data requirements for robust UIE design would be beneficial.
  • Generalization: While the paper demonstrates effectiveness in occupancy estimation, exploring the applicability of these data-driven UIEs to a wider range of systems and unknown input scenarios would be valuable.

Summary updated

Abstract

This paper addresses a data-driven input reconstruction problem based on Willems' Fundamental Lemma in which unknown input estimators (UIEs) are constructed directly from historical I/O data. Given only output measurements, the inputs are estimated by the UIE, which is shown to asymptotically converge to the true input without knowing the initial conditions. Both open-loop and closed-loop UIEs are developed based on Lyapunov conditions and the Luenberger-observer-type feedback, whose convergence properties are studied. An experimental study is presented demonstrating the efficacy of the closed-loop UIE for estimating the occupancy of a building on the EPFL campus via measured carbon dioxide levels.
BibTeX
@article{shi2022data,
  author = {Jicheng Shi and Ying Zhao Lian and Colin N. Jones},
  title = {{Data-Driven Input Reconstruction and Experimental Validation}},
  journal = {IEEE Control. Syst. Lett},
  year = {2022},
}
Figure