1. NONLINEAR AND ADAPTIVE SIGNAL ESTIMATIONBy
Dr. P.A. Ramamoorthy
Nithya Gopalarathinam

2. CONTENT ORGANIZATION Kalman Filter Estimation
Concepts leading to proposed estimation technique (dynamics corresponding to proper Electrical Circuits)
Proposed Signal Estimation
Results
Conclusion

3. Kalman Estimator Used to estimate signals observed under noise.
Typically derived by minimizing the squared error between the observed and the estimated signals.
Leads to a closed form optimal solution under Gaussian noise.
The Kalman gain is updated based on known noise statistics alone and not on the actual error.
Estimation could deviate from the actual if the noise statistics is different or time-varying, and could become unstable as well.

4. Kalman Estimator (Contd..)
Known Plant model:
Task is to get an optimal estimate of the state vector given the noisy output vector.

5. Kalman Estimator (Contd..)
The standard procedure is to define an estimate and a squared error given by:
Where, and is the error in the state estimate.

6. Kalman Estimator (Contd..)
Kalman gain matrix K(t) in the above estimate is found by minimizing the squared error for all time t leading to the solution:
Kalman gain is updated using the system parameters and the noise statistics which in turn is used in the signal estimation.

7. Kalman Estimator (Contd..)
Involves a feedforward architecture
Error in the output estimate which is readily available is not used in Kalman gain update

8. Adaptive Signal Estimator
Has a bottom-up approach
Uses NLTV electrical circuits
Dynamics in the form of I/O mappings of various elements
Electrical Circuits and Elements:
One-port and multi-port NLTV passive resistors (static element or one with no memory)
Multi-port NLTV gyrators and transformers (lossless, memory-less elements )
One-port NLTV reactive elements (inductors and capacitors), which are lossless elements with memory or energy storage capacity
Multiport Mutual inductors

9. Classic top down approach and the new bottom up approach

10. Adaptive Signal Estimator
Introduces a closed loop between the error estimation and the gain adaptation.
Leads to an inherently robust estimator.
Involves making the combined estimator and the gain-update dynamics to resemble the dynamics of a nonlinear time varying (NLTV) electrical circuit having the required properties

11. Adaptive Signal Estimator

12. Adaptive Signal Estimator (Contd..)
The estimation and gain update dynamics can be
The Error dynamics can be written as:
Where,

13. Adaptive Signal Estimator (Contd..)
The combined error dynamics (in the absence of any noise in the plant) should have the origin as the globally uniformly asymptotically stable (UAS).
From the circuit perspective, the dynamics should point to a NLTV circuit made of NLTV reactive elements, static multi-port lossless elements, and globally passive resistors.
All the reactive elements must have their relaxation points at and only at the origin

14. Adaptive Signal Estimator (Contd..)
Obtain corresponding dynamics in terms of the I/O mappings of the various elements.
can have complex error measures and obtain superior estimation.
The dynamics lead to direct update of the gain which makes it possible to design the estimator easily regardless of whether the plant is LTI or NLTV.
We can introduce any kind of complex error measure.

15. 2nd Order System -Example
Consider a 2nd order system
Signal Estimation and Gain dynamics for this plant can be written from circuit considerations as
can be written as

16. 2nd Order System -Example
To obtain a particular solution, we will fix as non-symmetric PD and solve for
comes from the weighted error function used in the estimation dynamics .
is obtained by solving

17. Adaptive Signal Estimator

18. Adaptive Signal Estimator
While comparing with the circuit dynamics and solving for the various quantities, we fix the impedance matrix of the static passive network and solve for the inductance matrix, which should turn out to be PD .
Similar to testing for stability of LTI systems using Kalman-Yakabovich lemma where for a given system matrix A, we assume arbitrary PD matrix Q and solve for the matrix P from

19. Results and Simulations
Picked various quantities in the dynamics almost randomly (subject to the given constraints), used a 16 point average of given as
We can have another error Criterion as :

20. Results and Simulations (Contd..)

21. Results and Simulations (Contd..)

22. Results and Simulations (Contd..)

23. Results and Simulations (Contd..)

24. Conclusion
The error becomes almost zero for the new estimation algorithm .
The new estimator is also robust to variations in the signal statistics and the chosen estimator parameters .
The new gain update uses a highly coupled NLTV dynamics that responds in a proper way to eliminate the error in the estimation

25. QUESTIONS?