Source-linked AI summary
Non-linear estimation is easy
Michel Fliess, Cédric Join, Hebertt Sira-Ramirez
TL;DR
The paper addresses nonlinear state estimation and related identification, diagnosis, and attenuation problems through a methodology based on numerical differentiation. It embeds the necessary system-theoretic definitions in differential algebra and reports online estimations and fault-accommodation simulations. The approach is non-asymptotic and requires no statistical knowledge of corrupting noise, but can be insufficient under very strong noise.
Problem
Nonlinear state estimation and related tasks such as parametric estimation, fault diagnosis, and perturbation attenuation remain largely open despite extensive literature.
Method
The paper uses integration-based numerical differentiators within a differential-algebraic framework that handles system variables and derivatives of arbitrary order.
Results
The paper reports promising results, including excellent online estimation of three inertia moments and efficient fault accommodation in simulations.
Takeaways & Limitations
Nonlinear estimation and related closed-loop tasks can be addressed with simple, non-asymptotic differentiator-based designs that do not require statistical noise knowledge.
Takeaways & Limitations
The techniques may be insufficient in situations involving very strong corrupting noise.
Abstract
from arXiv · showhide
Non-linear state estimation and some related topics, like parametric estimation, fault diagnosis, and perturbation attenuation, are tackled here via a new methodology in numerical differentiation. The corresponding basic system theoretic definitions and properties are presented within the framework of differential algebra, which permits to handle system variables and their derivatives of any order. Several academic examples and their computer simulations, with on-line estimations, are illustrating our viewpoint.
1 Introduction
The paper frames nonlinear estimation and related closed-loop tasks as largely open problems and proposes simple design methods based on numerical differentiation. Its differentiators use integrations, avoid asymptotic estimation and statistical noise models, and support several estimation, diagnosis, and attenuation applications.
- Nonlinear state estimation and related problems remain largely open despite a substantial literature.
- The approach reduces nonlinear estimation to numerical differentiation of derivatives in noisy time signals.
- Derivatives are estimated through integrations, whose iterated time integrals act as low-pass filters for highly fluctuating corrupting noise.
- The differentiators are non-asymptotic and require no statistical knowledge of corrupting noises.
- The framework extends to closed-loop parametric identification, fault diagnosis, fault-tolerant control, and perturbation attenuation.
2 Differential algebra
Differential algebra supplies the formal setting for nonlinear systems whose variables and derivatives may be handled to arbitrary order. The paper uses this framework for observability-related definitions, diagnosis concepts, and illustrative simulations, while noting scope boundaries for non-flat systems and non-algebraic equations.
- Observability, parametric identifiability, detectability, isolability, parity equations, and residuals are formulated within the algebraic setting.
- Differential algebra extends commutative-algebra concepts to differential equations and supports derivatives of arbitrary order.
- 2.1 Basic definitions: The framework introduces differential rings, fields, extensions, differential polynomials, and differential ideals as basic objects.
- The paper illustrates the framework with academic examples and numerical simulations designed to be accessible without the algebraic details.
- The examples are flat, although the estimation techniques are not restricted to flat systems; control of non-flat systems is beyond the article’s scope.
- The module of Kähler differentials links rank and dimension properties to differential transcendence and algebraicity of field extensions.
2.4 Nonlinear systems
The paper models nonlinear input-output systems as finitely generated differential extensions, explicitly separating control, fault, and perturbation variables. Differential algebra supplies the framework for arbitrary-order derivatives and system representations.
- System definition: A nonlinear input-output system is defined as a finitely generated differential extension K/k.The system is represented as K = k⟨S, W, π⟩ over a differential ground field.
- System definition: S contains system variables, including control variables u and output variables y; W denotes faults and π denotes perturbations.These variable classes are introduced separately in the system construction.
- Differential algebra: Differential transcendence bases characterize differential independence by requiring algebraic independence of variables and derivatives of every order.Their common cardinality defines the differential transcendence degree.
- Assumptions: The framework assumes control and fault variables are differentially independent and that control, fault, and perturbation extensions are linearly disjoint.These assumptions formalize the absence of interaction among the variable classes.
- System reductions: Nominal systems ignore perturbation variables, while pure systems additionally ignore fault variables.The corresponding quotient-field constructions preserve differential-algebraic system descriptions.
- Representations: The resulting state and output representations are implicit, and derivatives of control variables generally cannot be removed.The nominal and pure constructions use differential-polynomial relations and associated operator modules.
2.5 Variational system19
The variational system linearizes the nonlinear differential-algebraic model, while nominal and pure systems remove perturbation and fault variables to expose reduced dynamics and transfer descriptions.
- Linearization: The variational system is the linearized system associated with K/k, with analogous nominal and pure systems.These systems are defined through the corresponding modules of Kähler differentials.
- Full system: System states and outputs satisfy implicit polynomial differential relations whose coefficients depend on controls, faults, perturbations, and finitely many derivatives.The relations take the form Ai(ẋi, x) = 0 and Bj(yj, x) = 0.
- Flatness: A flat system admits a finite set of pure flat outputs whose number equals the number of independent control variables.These outputs provide a linearizing description of the dynamics.
- Nominal system: Nominal relations remove perturbation variables from their coefficients, leaving dependence on nominal controls, faults, and their derivatives.The nominal system is obtained from the full representation by ignoring perturbations.
- Pure system: The pure system is obtained by ignoring fault variables and has reduced state dimension npure ≤ nnom.Its pure transfer matrix is represented as A^-1B over a skew quotient field.
2.7 Observability and identifiability
Observability, identifiability, and fault detectability are expressed through algebraic relations among controls, outputs, states, parameters, and faults. Fault detectability has a transfer-matrix characterization.
- Observability: A nonlinear system is observable when every system variable is a differential function of the control and output variables.Equivalently, the pure system field is algebraic over the field generated by pure controls and outputs.
- Fault detection: A fault is detectable when it influences the output through the nominal system.The definition is formulated using the relevant field extension and fault variables.
- Identifiability: A parameter is algebraically identifiable when it satisfies an algebraic equation over k0⟨u, y⟩, and rationally identifiable when it is a differential rational function of controls and outputs.Both definitions allow derivatives up to finite order.
- Fault detection: The fault variable wι is detectable if and only if the ιth column of the fault transfer matrix TW is non-zero.This gives a direct transfer-matrix test for detectability.
- Fault isolation: A subset of faults is differentially algebraically isolable when each component satisfies a parity differential equation over controls and outputs.Algebraic isolability is the order-0 special case, while rational isolability makes the parity equation linear algebraic.
3.2 Analytic time signals
For analytic signals, numerical differentiation approximates a truncated Taylor expansion through operational calculus and yields online estimates of derivatives. Fault-isolability notions form a hierarchy with an output-count bound.
- Analytic signals: An analytic signal is approximated near zero by a truncated Taylor expansion of order N.The approximation applies on an interval (0, ε), with 0 < ε ≤ ρ.
- Numerical differentiation: Operational calculus transforms the truncated expansion into an operational series used to estimate derivatives at the initial time.Replacing the truncated operational expression by the operational signal defines the numerical estimates.
- Fault isolation: Rational isolability implies algebraic isolability, which implies differentially algebraic isolability.The paper uses differential algebraic isolability as the default meaning of “isolable.”
- Fault isolation: If isolable faults are represented by W′, then card(W′) ≤ card(y).The bound is derived by comparing differential transcendence degrees of fault-output extensions.
- Approximation: The approximation error is O(t^(N+1)) and becomes negligible as t approaches zero or N tends to infinity.This provides the stated analytic justification for computer implementations.
3 Derivatives of a noisy signal
The paper estimates derivatives and initial values of noisy signals using operational-calculus manipulations and iterated integrations. The resulting estimates are presented as denoised signals, with an approach that is independent of probabilistic noise models.
- Noise attenuation: Iterated time integrals attenuate highly fluctuating additive noises by acting as low-pass filters.The integrations arise after multiplying by s^-N̄, with N̄ chosen sufficiently large.
- Interpretation: The estimated value of x(0) should be viewed as a denoising of the corresponding signal.
- Scope: The approach is independent of probabilistic assumptions and can also handle multiplicative noises.The paper defines highly fluctuating functions through infinitesimal finite-interval integrals and cites simulations and laboratory experiments.
- Initial-value estimation: Initial values x(ν)(0), for ν = 0, 1, ..., N, are obtained from a triangular system of linear equations.The system follows from rewriting polynomial signals in operational-calculus notation.
- Operational calculus: Multiplication by s^-N̄ removes terms involving derivatives of the operational variable when N̄ > N.
4 Feedback and state reconstructor
A flexible-joint manipulator is controlled using its flat output while a numerical differentiator reconstructs the unmeasured motor state. Simulations test tracking, estimation, and robustness under additive output noise.
- System description: The manipulator couples a DC motor to an inverted pendulum through a torsional spring.
- State reconstruction: The unmeasured motor angle θm(t) is estimated from available signal information.The pendulum angle θl is identified as the flat output y.
- Simulation: The simulations use the physical parameters reported from Fan and Arcak and test robustness with additive white Gaussian noise N(0; 0.01) on the output.The paper reports that offline estimation of ¨y and θm, with a small delay allowed, is better than online estimation of ¨y.
- Feedback control: Feedback is designed so the output y = θl tracks a smooth reference trajectory.
5 Parametric identification
The paper applies its numerical-differentiation methodology to parameter estimation and feedback stabilization for a fully actuated rigid body. It reports online inertia estimation under noise and improved stabilization when estimated parameters replace false values.
- Rigid-body model: The rigid-body example uses measured angular velocities and applied torques, while the moments of inertia are poorly known.
- Feedback stabilization: The feedback law stabilizes the rigid body around the origin using estimated inertia values.
- Parametric identification: The method estimates I1, I2, and I3 by replacing angular velocities and their derivatives with numerical estimates.
- Numerical simulations: The simulations report excellent online estimation of all three moments of inertia under additive Gaussian white noise N(0; 0.005).
- Numerical simulations: Stabilization with the estimated values is quite better than stabilization obtained with false inertia values.
6 Fault diagnosis and accommodation
The section develops algebraic estimation and feedback methods for actuator faults and unknown perturbations, then tests fault accommodation in closed-loop simulation. The controller uses on-line estimates to restore tracking after a fault, with robustness checked under measurement noise.
- Problem setting: The cascade tank example assumes only y = x2 is measured, while the actuator failure and perturbation are constant unknown quantities.
- Estimation: The method estimates the constant perturbation before the actuator failure and then algebraically isolates the failure signal.The failure starts at an unknown time tI that is not small.
- Control design: The failure-accommodating controller combines estimated failure and perturbation signals with a robustifying integral action.
- Simulation: At t = 2.5s, fault-tolerant control becomes effective after a fault w = 0.7 occurs at tI = 1.5s.
- Simulation: Comparison with the no-accommodation case confirms efficient fault accommodation under additive Gaussian white noise N(0; 0.01).
7 Perturbation attenuation
The perturbation-attenuation section uses estimated output derivatives and an estimated perturbation in a GPI feedback regulator. Simulations compare tracking with and without perturbation estimation and report strong performance under measurement noise.
- Controller: The closed-loop characteristic polynomial is selected with roots whose imaginary parts are strictly negative.
- Robustness: The controller is robust with respect to un-modeled piecewise constant errors, similarly to proportional-integral-derivative regulators.
- Simulation: The perturbation experiment introduces a permanent value C = 1.25 at tI = 4 while tracking y⋆(t) = sin ωt with ω = 2.5[rad/s].
- Simulation: Simulations compare output tracking without and with estimating ze(t), adding Gaussian white noise N(0; 0.025) to the measurement.The reported results are “quite remarkable.”
8 Conclusion
The conclusion presents numerical differentiation as a non-asymptotic approach to nonlinear estimation that does not require statistical knowledge of corrupting noise. It also identifies fault-tolerant control as an application of the perturbation-estimation technique and points to further developments.
- Conclusion: The perturbation-estimation technique extends to fault-tolerant linear control by treating the perturbation z(t) as a fault variable.
- Conclusion: The paper proposes a non-asymptotic nonlinear-estimation approach that requires no statistical knowledge of corrupting noises.
- Future work: The authors state that further numerical improvements will be investigated and cite additional theoretical advances and concrete case studies as future work.
- Non-linear extension: The nonlinear extension replaces the term y(t) in the system with the product y(t) ˙y(t).