On the equivalence of model-based and data-driven approaches to the design of unknown-input observers

Basic Information

• Paper ID: 2311.00673
• Title: On the equivalence of model-based and data-driven approaches to the design of unknown-input observers
• Authors: Giorgia Disarò, Maria Elena Valcher (University of Padova, Italy)
• Classification: math.DS (Dynamical Systems)
• Publication Date: November 2023 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2311.00673

Abstract

This paper investigates data-driven methods for the design of unknown-input observers (UIO). Specifically, the authors provide necessary and sufficient conditions for the existence of unknown-input observers for discrete-time linear time-invariant (LTI) systems based on finite-time window data. The paper demonstrates that under weak assumptions on the collected data, the solvability conditions derived from the data-driven approach are equivalent to those obtained from model-based methods. In other words, the data-driven conditions do not impose additional constraints compared to classical model-based conditions expressed in terms of the original system matrices.

Research Background and Motivation

1. Problem Description: In many control engineering applications, knowledge of the internal system state is essential for solving fundamental problems such as state feedback stabilization and fault detection. However, system states are typically not directly accessible and require the design of appropriate observers to estimate the state vector.
2. Challenges: Real system dynamics are often affected by disturbances, measurement errors, or other unknown signals that cannot be used for identifying state evolution. Traditional observer design assumes the system model is known, but in some cases, the system cannot be uniquely identified from available data alone.
3. Limitations of Existing Approaches:
   • Model-based methods require known system matrices
   • Existing data-driven methods (e.g., 18) provide only sufficient conditions; when the obtained system is not asymptotically stable, it remains unclear whether a UIO can be designed
4. Research Motivation: Leveraging abundant available data to develop data-driven techniques for solving control engineering problems, particularly state estimation when precise system models cannot be obtained.

Core Contributions

1. Necessary and Sufficient Conditions: Provides verifiable necessary and sufficient conditions for validating UIO existence from data
2. Complete Parameterization: Provides complete parameterization of all candidate UIOs
3. Equivalence Proof: Demonstrates that under specific data assumptions, the solvability conditions of the data-driven method are completely equivalent to those of model-based methods
4. Theoretical Breakthrough: Establishes a bijective correspondence between data-driven and model-based approaches

Methodology Details

Task Definition

Consider a discrete-time LTI system Σ:

x(t+1) = Ax(t) + Bu(t) + Ed(t)  (1)
y(t) = Cx(t)                    (2)

Where:

• x(t) ∈ ℝⁿ: state vector
• u(t) ∈ ℝᵐ: known control input
• y(t) ∈ ℝᵖ: output vector
• d(t) ∈ ℝʳ: unknown input (e.g., disturbances)

UIO Definition: An LTI system Σ̂ of the form:

z(t+1) = A_UIO z(t) + B^u_UIO u(t) + B^y_UIO y(t)  (3)
x̂(t) = z(t) + D_UIO y(t)                           (4)

Such that the estimation error e(t) = x(t) - x̂(t) asymptotically converges to zero, independent of initial conditions and unknown input dynamics.

Model-Based Approach

Necessary and Sufficient Conditions (Theorem 2): A UIO exists if and only if one of the following conditions holds:

1. There exist matrices (A_UIO, B^u_UIO, B^y_UIO, D_UIO) satisfying:
   • A_UIO is Schur stable
   • D_UIO CE = E
   • B^u_UIO = (I - D_UIO C)B
   • A_UIO(I - D_UIO C) + B^y_UIO C = (I - D_UIO C)A
2. Matrix conditions:
   • rank(CE) = rank(E) = r
   • rankzIₙ - A -E; C 0 = n + r, ∀z ∈ ℂ, |z| ≥ 1
3. The triple (A,E,C) is strongly *-detectable

Data-Driven Approach

Data Organization: Historical data is organized into matrix form:

• Uₚ = uₐ(0) ... uₐ(T-2) ∈ ℝᵐˣ⁽ᵀ⁻¹⁾
• Xₚ = xₐ(0) ... xₐ(T-2) ∈ ℝⁿˣ⁽ᵀ⁻¹⁾
• Xf = xₐ(1) ... xₐ(T-1) ∈ ℝⁿˣ⁽ᵀ⁻¹⁾
• Yₚ = yₐ(0) ... yₐ(T-2) ∈ ℝᵖˣ⁽ᵀ⁻¹⁾
• Yf = yₐ(1) ... yₐ(T-1) ∈ ℝᵖˣ⁽ᵀ⁻¹⁾

Core Assumption: The matrix Uₚᵀ Dₚᵀ Xₚᵀᵀ has full row rank m+r+n.

Main Result (Theorem 9): The following conditions are equivalent:

1. There exists a UIO of the form (3)-(4) such that T_Σ ⊆ T_Σ̂
2. There exist matrices T₁,T₂,T₃,T₄ such that Xf = [T₁ T₂ T₃ T₄]Uₚ; Yₚ; Yf; Xₚ and T₄ is Schur stable
3. Data conditions: ker(Xf) ⊇ ker(Uₚ; Yₚ; Yf; Xₚ) and rankzXₚ-Xf; Uₚ; Yₚ = n+m+r, ∀z∈ℂ,|z|≥1
4. The triple (A,E,C) is strongly *-detectable

Technical Innovations

1. Equivalence Establishment: First rigorous proof of complete equivalence between data-driven and model-based conditions
2. Bijective Correspondence: Establishes bijective relationship between UIO parameter matrices and data-driven parameters (Corollary 10)
3. Practical Algorithm: Provides simplified computational methods, transforming the problem into a more tractable form

Experimental Setup

Numerical Example

Using the example from reference 4:

A = [-1 -1  0; -1  0  0; 0 -1 -1]
C = [1  0  0; 0  0  1]
E = [-1; 0; 0]

Experimental Parameters

• Time window: T = 20
• Unknown input: uniformly randomly generated in the interval (-2,2)
• Initial conditions: randomly selected
• Disturbances: randomly generated in the interval (-10,10)

Comparison Methods

Comparison with the method proposed in reference 18

Experimental Results

Main Results

1. Model Verification: Verifies the conditions rank(CE) = rank(E) = r and strong *-detectability
2. Data-Driven Design: Successfully reconstructs matrix C from data and verifies the conditions of Theorem 9
3. Performance Comparison:
   • The A_UIO matrix obtained by the proposed method is nilpotent
   • The method in reference 18 yields eigenvalues {0, 0, -0.2}
   • Both methods work, but the proposed method allows selection of convergence speed

Case Analysis

Figure 1 shows the dynamics of state estimation error:

• Solid line (black): proposed design method
• Dashed line (red): method from reference 18
• Both methods achieve asymptotic convergence, but the proposed method provides greater design freedom

Experimental Findings

1. Data-driven conditions correctly identify the existence of UIO
2. Complete parameterization allows selection of desired convergence characteristics
3. When data satisfies the assumptions, the data-driven method introduces no additional constraints

Related Work

Model-Based UIO Design

• Luenberger observer theory 12,13
• Unknown input decoupling methods 9,11,21
• Geometric methods 2 and generalized inverse methods 15
• Singular value decomposition techniques 6

Data-Driven Methods

• Two-step approach: system identification followed by observer design
• Single-step approach: direct use of data avoiding identification phase 18
• Data-driven simulation and control 14

Positioning of This Work

Compared to existing work, this paper is the first to establish theoretical equivalence between the two methods and provides a complete parameterized solution.

Conclusions and Discussion

Main Conclusions

1. Under weak data assumptions, data-driven UIO design conditions are completely equivalent to model-based conditions
2. Data-driven methods introduce no additional constraints
3. Provides complete parameterization of all possible UIOs
4. Establishes a theoretical bridge unifying the two design paradigms

Limitations

1. Requires historical state measurement data (feasible in laboratory environments)
2. Requires knowledge of unknown input dimension
3. Depends on full row rank assumption of data
4. Applicable only to linear time-invariant systems

Future Directions

1. Extension to nonlinear systems
2. Consideration of time-varying systems
3. Investigation of robustness to noisy data
4. Development of algorithmic implementations for practical applications

In-Depth Evaluation

Strengths

1. Outstanding Theoretical Contribution: First rigorous proof of equivalence between data-driven and model-based methods
2. Mathematical Rigor: Provides complete necessary and sufficient conditions with rigorous mathematical proofs
3. Practical Value: Provides verifiable data conditions and practical algorithms
4. Completeness: Provides parameterization of all UIOs, not just a single solution
5. Clear Presentation: Logical structure is clear with sufficient technical detail

Weaknesses

1. Restrictive Assumptions: The requirement for state measurement data is quite stringent in practice
2. Linear Limitation: Applicable only to linear systems, limiting the scope of applications
3. Limited Numerical Validation: Provides only one numerical example
4. Missing Robustness Analysis: Does not consider the effects of data noise and modeling errors

Impact

1. Theoretical Significance: Provides important theoretical foundation for data-driven control theory
2. Practical Value: Provides feasible design methods for engineering applications
3. Research Inspiration: Offers new research perspectives for other observer design problems
4. Reproducibility: Provides clear algorithmic steps facilitating reproduction

Applicable Scenarios

1. Scenarios where system models are unknown but historical state data is available
2. State estimation problems requiring handling of unknown disturbances
3. Observer design for discrete-time linear control systems
4. Fault detection and diagnosis applications

References

The paper cites 24 relevant references covering important works in observer theory, data-driven control, and system identification, providing a solid theoretical foundation for the research.

Overall Assessment: This is a high-quality theoretical paper making important contributions to the field of data-driven observer design. While there are some limitations for practical applications, its theoretical value and contribution to advancing the field are noteworthy.