Automatic system identification and controller synthesis for embedded systems

1. A method, comprising:

performing a system identification experiment;

performing a system identifying processing by fitting a model to data from the system identification experiment;

performing model reduction to generate a model numerically suitable for controller synthesis by removing inconsequential states, wherein the inconsequential states cause controller optimization methods to fail;

performing control synthesis using the generated model or reduced models, including disturbance spectrum estimates, to generate a candidate controller design to be used during system operation; and

checking controller robustness using the identified model to ensure stability of the system while maximizing closed-loop bandwidth and performance.

2. The method of claim 1 , wherein the generated model is a parametric dynamic model from measurable inputs to measurable outputs and unknown disturbance inputs to measurable outputs.

3. The method of claim 1 , wherein the generated model is a discrete-time transfer function or matrix of discrete-time transfer functions or an equivalent state space model, where orders of a transfer function numerator and denominator polynomials or a number of states are specified prior to estimating the parameters of the generated model.

4. The method of claim 1 , wherein the generated model is a locally or globally optimal solution to a system identification problem of minimizing energy of prediction errors given a set of measured input data, measured output data, and a specified model structure.

5. The method of claim 1 , wherein the generated model comprises estimated covariances of the estimated parameters of the generated model that are subsequently used to compute uncertainties of estimated frequency responses of the generated model, and

the uncertainties of the estimated frequency response are key enablers for automating the automated system identification and control synthesis and validation processes.

6. The method of claim 1 , wherein the reduced model is parametric dynamic model from measurable inputs to measurable outputs and unknown disturbance inputs to measurable outputs.

7. The method of claim 1 , wherein the reduced model is a discrete-time transfer function or matrix of discrete-time transfer functions or an equivalent state space model, where orders of a transfer function numerator and denominator polynomials or a number of states are specified prior to estimating the parameters of the model.

8. The method of claim 7 , wherein the reduced model is a smaller model in terms of transfer function polynomial orders or number of state space model states, produces larger prediction errors than a generated model, and is suitable for controller synthesis.

9. The method of claim 1 , wherein the disturbance spectrum estimates is computed from an estimated disturbance model,

the estimated disturbance model is part of a generated model.

10. The method of claim 1 , wherein the performing of the system identifying processing comprises estimating uncertainty, the uncertainty comprising frequency-dependent uncertainties.

11. A method for automated system identification and controller synthesis, comprising:

continuously performing, by a computing system, a system identification experiment and system identification processing until a frequency-dependent relative uncertainty is less than a predefined threshold across one or more frequencies under consideration;

performing, by the computing system, a model reduction; and

continuously performing, by the computing system, controller synthesis until a closed-loop system is robustly stable given estimated uncertainty in a model and closed-system performance metrics are maximized.

12. The method of claim 11 , wherein the performing of the system identification experiment comprises generating and injecting an excitation signal, and recording a total input and output response.

13. The method of claim 11 , wherein the performing of the system identification processing comprises fitting a model, and estimating the frequency-dependent relative uncertainty with the fitted model.

14. The method of claim 13 , further comprising:

evaluating a quality of the fitted model using estimated frequency-dependent uncertainties.

15. The method of claim 14 , further comprising:

generating, from the fit model, an estimated disturbance model; and

computing an estimated disturbance spectrum from the estimated disturbance model.

16. The method of claim 11 , wherein the predefined threshold is based on a user-defined threshold based on a confidence level.

17. The method of claim 13 , wherein the predefined threshold is a constant value against which a weighted relative frequency-dependent uncertainty is compared.

18. The method of claim 11 , wherein the performing of the model reduction comprises

casting the fitted model into a controller or observer canonical state-space form, wherein coefficients of the fitted model are directly used as elements in select rows or columns of a state-space model matrices, and

performing the model reduction on a canonical state-space model.

19. The method of claim 18 , wherein the performing of the model reduction comprises

applying the model reduction prior to any use or transformation of the canonical state-space model.

20. The method of claim 11 , wherein the reduced model is a parametric dynamic model from measurable inputs to measurable outputs and unknown disturbance inputs to measurable outputs.

21. The method of claim 11 , wherein the reduced model is a discrete-time transfer function, matrix of discrete-time transfer functions, or an equivalent state space model, where orders of a transfer function numerator and denominator polynomials or a number of states are specified prior to estimating the parameters of the model.

22. The method of claim 11 , wherein the performing of the control synthesis comprises

creating a controller; and

performing a robustness check.

23. The method of claim 22 , wherein the creating of the controller comprises creating an auto-tuned proportional-integral-derivative (PID) controller, creating a frequency-shaped linear quadratic Gaussian (LQG) controller, or synthesizing an H-infinity controller.

24. The method of claim 23 , wherein the LQG controller requires multiple iterations of checking the controller robustness and resynthesizing the controller to reduce the bandwidth by adjusting the frequency-shaped weighting function until robustness is achieved.

25. The method of claim 23 , wherein the H-infinity control synthesis produces a robust controller on a first attempt, though the controller synthesis process itself may internally use an iterative process.

26. The method of claim 22 , wherein the performing of the robustness check comprises performing a first stability check to ensure that the closed-loop system model is stable when a generated controller is applied to an identified model.

27. The method of claim 26 , wherein the performing of the first stability check comprises computing eigenvalues of the closed-loop system model to ensure the eigenvalues have magnitudes less than one.

28. The method of claim 27 , further comprising:

when the identified model is uncertain, performing a stability check utilizing a reduced model.

29. The method of claim 28 , further comprising:

incorporating the uncertainty model and a model reduction error is incorporated as additive complex unstructured uncertainties in a robust stability analysis based on mu-analysis theory.

30. The method of claim 29 , further comprising:

checking the robustness using mu-analysis when bounds on an additive complex unstructured uncertainty is computed from the model uncertainty.

31. The method of claim 30 , further comprising:

checking, using mu-analysis, a maximum singular value of a perturbed closed-loop system at each frequency in a dense grid of frequencies.

32. The method of claim 31 , further comprising:

iteratively reducing performance goals until robust stability is ensured, when robust stability of the system is unensured.