Method for real-time nonlinear system state estimation and control

• US 6,285,971 B1
• Filed: 08/10/1999
• Issued: 09/04/2001
• Est. Priority Date: 08/22/1997
• Status: Expired due to Term

• Current Assignee: Tokyo Electron Limited
• Original Assignee: Voyan Technology Incorporated (HP Inc.)
• Inventors: Shah, Sunil C., Pandey, Pradeep
• Primary Examiner(s): Teska, Kevin J.
• Assistant Examiner(s): PHAN, THAI Q
• Application Number:
  US09/372,396
• Time in Patent Office: 756 Days
• Field of Search: 703/2, 703/7, 700/29, 700/33, 700/44, 700/38, 706/14, 706/33, 708/300
• US Class Current: 703/2
• CPC Class Codes: G05B 13/04 involving the use of models or simulators

Abstract

A method for the estimation of the state variables of nonlinear systems with exogenous inputs is based on improved extended Kalman filtering (EKF) type techniques. The method uses a discrete-time model, based on a set of nonlinear differential equations describing the system, that is linearized about the current operating point. The time update for the state estimates is performed using integration methods. Integration, which is accomplished through the use of matrix exponential techniques, avoids the inaccuracies of approximate numerical integration techniques. The updated state estimates and corresponding covariance estimates use a common time-varying system model for ensuring stability of both estimates. Other improvements include the use of QR factorization for both time and measurement updating of square-root covariance and Kalman gain matrices and the use of simulated annealing for ensuring that globally optimal estimates are produced.

Claims

1. A method for estimation of state variables of a physical nonlinear system with physical exogenous inputs that can be modeled by a set of continuous-time nonlinear differential equations, including input signals representing the physical exogenous inputs to the system, and measurement signals representing accessible measurements that are indicative of the state variables to be estimated, the method comprising:

a) creating a nonlinear differential equation model of a nonlinear system using initial estimated state variables;

b) obtaining the nonlinear system measurement signals in response to the nonlinear system input signals;

c) creating an updated nonlinear differential equation model using the nonlinear differential equation model, the nonlinear system input and measurement signals, and a state variable estimation method for refining the initial estimated state variables, the state variable estimation method for producing updated estimated state variables using integration methods for state variable estimation of the nonlinear system with exogenous inputs;

d) using a discrete-time covariance matrix update; and

e) using a common linear perturbation signal model for estimating the updated state variables and a corresponding updated covariance matrix, wherein the linear perturbation model is a discrete-time time-varying linear system model.

2. The method of claim 1 wherein the square-root of the covariance matrix is propagated.

3. The method of claim 1 further comprising a step for simulated annealing by running multiple estimators and choosing the best estimate based on an error metric.

4. The method of claim 3 wherein an increased process noise covariance in the state variable estimation method is used, the increased process covariance generated in accordance with a prescribed schedule that begins by amplifying the initial process noise covariance.

5. The method of claim 1 wherein the state variable estimation method is based on an extended Kalman filter, modified for state variable time propagation by integration.

6. The method of claim 1 wherein the nonlinear differential equation model is given by {dot over (x)}=f(x, u)+w and z=h(x, u)+v where x is the state {dot over (x)} is a time derivative of the state variable vector x, f(x,u) is a nonlinear function of x and input vector u, z is a measurement vector which is function, h(x,u) of both x and u, and w and v are mutually independent noise vectors and wherein a state variable estimate, {circumflex over (x)}_k, for time t=kT is propagated to a state variable estimate, {circumflex over (x)}_k+1, for time t=(k+1)T by integrating the nonlinear differential equation model over a time interval of T as follows:

${\hat{x}}_{k+1}={\hat{x}}_k+{\int <br><br>}_{s=0}^T<br><br>e^{A_k<br><br>\left(T-s\right)}<br><br>f_k<br><br>\text{\hspace{1em}<br><br>}<br><br><br><br>s,$

where f_k=f({circumflex over (x)}_k|k, u_k) evaluated at time t=(kT), and A_k is a discrete state feedback matrix, evaluated at time t=kT, from a linearized approximation to f(x,u).

7. The method of claim 6 wherein integrating is performed by evaluating a matrix exponential as follows:

$e^{\left[\begin{array}{cc}{A}_k& f_k\\ 0& 0\end{array}\right]<br><br>T}=\left[\begin{array}{cc}{e}^{A_k<br><br>T}& {\int <br><br>}_0^T<br><br>e^{A_k<br><br>\left(T-s\right)}<br><br>f_k<br><br>\text{\hspace{1em}<br><br>}<br><br><br><br>s\\ 0& I\end{array}\right]=\left[\begin{array}{cc}{F}_k& G_k\\ 0& I\end{array}\right],$

so that a time propagated state vector is obtained as follows,

8. The method of claim 7 further comprising a step for simulated annealing by running multiple estimators and choosing a best estimate based on an error metric.

9. The method of claim 8 wherein each estimator is initialized by a different random initial state vector.

10. The method of claim 8 wherein the error metric is a scalar value, c_k, computed as c_k=trace(S_k|kS_k|k^T).

11. The method of claim 8 wherein an increased process noise covariance is used in the state variable estimation method, the increased process noise covariance generated in accordance with a prescribed schedule that begins by amplifying the initial process noise covariance.

12. The method of claim 11 wherein the prescribed schedule comprises:

a) computing a temperature schedule, ξ

_k, that decreases as a function of time index k;

b) computing a gain multiplier as (1+ξ

_k); and

c) multiplying the initial process noise covariance, Q₀, by the gain multiplier for producing a process noise covariance at time index k as Q_k=Q₀(1+ξ

_k).

13. The method of claim 12 wherein ξ

_k is computed as ξ

_k={square root over (2+L c /ln(2+L +kT))}, where c is a prescribed constant, and ln(.) is a natural logarithm function.

14. The method of claim 1 further comprising:

estimating an updated covariance of the updated estimated state variables using a common state transition matrix for estimating the updated estimated state variables and the updated covariance for improved estimation quality; and

computing a controller which uses the set of current-step system state variable estimates produce a set of system input control signals for controlling the physical inputs to the system, the controller is obtained by use of a performance index that specifies an objective performance criterion for the system.

15. A state vector estimation method for a real-time nonlinear system with exogenous inputs, the system represented by a set of continuous-time nonlinear differential state equations that include a state vector representing the system state vector that are to be estimated, and a nonlinear system matrix that defines the functional relationship between the input vector, the state vector, and the time-derivative of the state vector, the state vector being estimated from a set of accessible measurements that are representative of the state variables, the set of measurements, labeled as a measurement vector, taken at prescribed times and represented by a measurement vector, both the state vector and the measurement vector respectively contaminated by a state noise vector and a measurement noise vector, the method comprising the following steps:

(a) constructing a set of state equations representing the nonlinear system with exogenous inputs;

(b) establishing a set of state variable values for a current-time state vector, and a set of matrix element values for one each covariance matrix for a current-time state vector, for a current-time state noise vector, for a current-time input noise vector, and for a current-time output measurement noise vector;

(c) acquiring a current-time measurement vector;

(d) updating the state covariance matrices using the current-time measurement noise covariance vector;

(e) estimating an updated current-time state vector from the current time measurement vector by use of a state vector estimating filter that operates on the current-time state vector using the current-time measurement vector, the covariance matrices for the current time state vector, for the current time state, and for the current-time measurement noise vector, the estimated updated state vector representing the state vector at the measurement current-time;

(f) projecting the estimated updated state vector by integrating the system state equations over the prescribed time interval between the prescribed measurement current-time and the next prescribed measurement current-time for obtaining a current state vector using a discrete-time covariance matrix update;

(g) projecting the updated state vector covariance matrix using the results of the system equation integration of step (f) using a linear perturbation signal model for estimating the updated state variables and a corresponding updated covariance matrix, wherein the linear perturbation model is a discrete-time time-varying linear system model; and

(h) at the next measurement time, iterating steps (c)-(g).

16. The method of claim 15 wherein the square-root of the state covariance matrices are propagated.

17. The method of claim 15 wherein the updating of the state covariance matrix is performed by a matrix transformation that produces a triangular square-root covariance matrix.

18. The method of claim 17 wherein the matrix transformation is a QR factorization.

19. The method of claim 15 wherein the system equation is linearized at the initial value of the state vector, the linearization creating a locally linearized state feedback gain matrix that defines the state vector contribution to the time derivative of the state vector, given the current time input vector and the current-time state estimate.

20. The method of claim 19 wherein the linearization also creates a locally linearized input gain matrix for relating the input vector contribution to the time derivative of the state vector, given the current state and the current input vector.

21. The method of claim 19 wherein the step of updating the state vector by integrating the augmented locally linear state feedback gain matrix is performed by a matrix exponential.

22. The method of claim 15 wherein the steps updating the current-time state vector covariance matrix and the step of estimating an updated current-time state vector is not executed because the current-time measurement is not usable or available, and the current-time projected estimated updated state vector, together with the current-time projected updated state vector covariance matrix, are used as the estimated updated current time state vector and updated current time covariance matrix, respectively.

23. A method for estimation of state variables of a physical nonlinear system, the method being applicable to systems that can be modeled by a set of continuous-time nonlinear differential equations and to systems that can be modeled by discrete-time nonlinear difference equations, the method using measurement signals representing accessible measurements that are indicative of the state variables to be estimated, the method comprising:

a) creating an extended Kalman filter (EKF) for estimating system state variables from the measurement signals in a dynamic system; and

b) applying simulated annealing by running multiple EKFs and choosing the best estimate based on an error metric, wherein simulated annealing is initiated by a random perturbation.

24. The method of claim 23 wherein an increased process noise covariance is used in the EKF, the increased process noise covariance generated in accordance with a prescribed schedule that begins by amplifying the initial process noise covariance.

25. The method of claim 23 wherein an increased state vector covariance is used in the EKF, the increased state vector covariance generated in accordance with a prescribed schedule that begins by amplifying the initial state vector covariance.