Académique Documents
Professionnel Documents
Culture Documents
M.S. Ahmed
1
Introduction
Controller parameters are needed to be modified
against the change in the operating point
Response of a nonlinear plant generally cannot be
shaped to a desired pattern using a linear controller
One of the main difficulties in designing the nonlinear
controller is the lack of a general structure for it
2
Researches on Adaptive Neural Control of Nonlinear Dynamical
Systems
Narendra and Parthasarathy (1990): dynamic back-propagation for
identification and control employing MFNN
Chen and Khalil (1992, 1995): use of MFNN in adaptive control of feedback
linearizable minimum phase plants represented by an input-output model
(local convergence)
Jagannathan and Lewis (1996a,b): use of MFNN in adaptive control of
feedback linearizable plants with all states accessible (convergence to a
stable solution through the Lyapunov approach)
Sanner and Stoline (1992): use of Gaussian RBF for the adaptive control of
feedback linearizable continuous time plants (globally stable under mild
assumptions)
Rovithakis and Christodoulou (1994):adaptive control through two step
procedure employing a dynamic neural netwrok (convergence to zero
error)
Polycarpuo (1996): a more general class of feedback linearizable plants
Ahmed (1994, 1995), Ahmed and Anjum (1997): adaptive control of
nonlinear plants of unknown structure (local convergence)
Wang, Liu, Harris and Brown (1995): indirect adaptive control of similar
plants employing neural networks
3
Adaptive Control of a SISO Plant
4
The Control Scheme
5
Artificial Neural Networks
6
Regulation
8
State Tracking
9
Algorithm Analysis
Closed loop system:
10
Adaptive Control of a MIMO Plant
Extension of the above development to MIMO
nonlinear plants depends on the existence of a
pseudo-linear plant model that warrants
existence of a suitable Lyapunov function.
11
Adaptive Control of a MIMO Plant
12
Regulation
13
Algorithm Analysis
Closed loop system:
14
State Tracking
15
Algorithm Analysis
16
Remarks
Theorems 1-4 do not imply that the controller
parameters in will converge to those in *.
In the state tracking it is also possible to take the
regressor vector as [eT rT]T instead of [xT rT]T.
It was assumed that there exists a parameter set for the
nonlinear controller that can drive the control system to
zero error. When this assumption does not hold only
bounded error can be secured.
17
Simulation Studies
Example 1: SISO plant
18
Simulation Studies
MLP based adaptive control:
19
Simulation Studies
Example 2: MIMO plant
20
Simulation Studies
21
Simulation Studies
22
Simulation Studies
23
Conclusion
A globally stable neural net MRAC for a class of nonlinear systems
Nonlinear canonical state variable description
Time varying pseudo-linear state feedback control gain generated
from the output of a ANN set
Plant need not be feedback linearizable
Regulation and tracking schemes are proposed
Adaptation of the controller parameters based on Lyapunov
function to ensure global convergence
Extension to MIMO plants
Simulation studies
Drawback of the proposed controller:
In some cases a large number of adjustable controller parameters
may be needed
24