Systems of surveillance of dynamic processes

Lecture 2:

The optimal control at energy performance index

of unsteady and nonlinear systems
Prof. Krzysztof J. Kaliski
Head, Division of Mechatronics

Short introduction
Dynamic system with n degrees of freedom is a system, whose properties are
described by:
M*(nn)

inertia matrix

L*(nn)

damping matrix

K*(nn)

stiffness matrix

f*(n1)

vector of generalised forces

q*(n1)

vector of generalised displacements

Additionally, the system is controlled

Bu (np)

matrix of control commands

u(p1)

vector of control commands, p number of control command

Linear controlled system

Stationary system

&& * + L*q& * + K *q * = f * + B *u u
M *q
Matrices M*, L*, K*, f*, Bu with constant coefficients
Nonstationary system

&&* + L* (t )q& * + K * (t )q* = f * (t ) + B*u (t )u

M * (t )q

Nonlinear controlled system

Prof. Krzysztof J. Kaliski

&& + L(t , q& , q )q& + K (t , q& , q )q = f (t , q& , q ) + B u (t , q& , q )u

M (t , q& , q )q
Description in state coordinates matrix state equations
x& = Ax + Dz + Bu
,

y = Cx + w

where:

x = q& *T

q*T

T
2 n 1

M *1L*
A=
I

vector of state coordinates,

M *1K *
state matrix,

0
2 n2 n

M*1
D=

0 2 n n

matrix of disturbances,

M*1B*u
B=

0 2 n p

input matrix,

Cqn

output matrix,

z f * vector of disturbances,
yq1

vector of outputs, whose components are registered responses of a system,

w q1

vector of measuring disturbances.

Particular case. All components of generalised displacement vector q are registered, that is

to say: y q .

Note. All nonpotential generalised forces, whose action does not influence

the control process (i.e. vector f*), are considered as disturbances.

Description of the controlled system dynamics in domain of state coordinates is universal.

It concerns as well linear (stationary and nonstationary) , as nonlinear systems.

Prof. Krzysztof J. Kaliski

Matrix eigenvalue problem

Uncontrolled system

det (A Is ) = 0

Controlled system

det (A R Is ) = 0

Solution: 2 n poles i = i ji , i = 1, ..., n

For nonstationary system: A = A(t ) , A R = A R (t )
For nonlinear system: A = A(t , q& , q ) , A R = A R (t , q& , q )

Generally, poles have no constant value over the process

Constant poles methods, e.g. PID controllers, are not relevant towards
control of nonstationary and nonlinear systems

Optimal control at energy performance index

Proposition:
On-line optimal control at changing in time energy (algebraic)
performance index with respect to trajectory of the desired motion.

It allows:

formulation of the control problem in time domain and frequency domain, as minimisation
of the whole energy with respect to the trajectory of the desired motion,
effective algorithm of computer simulation,
control of dynamic nonstationary systems,
control of closed loop dynamic nonstationary systems,
control of dynamic nonlinear systems.

Energy performance index, apart from the other functions defining the problem of optimal
control, has its physical interpretation; it expresses an instantaneous energy stored in the
controlled system.

Prof. Krzysztof J. Kaliski

Control of a linear nonstationary system

Control in domain of generalised coordinates

Solution to the first matrix state equation with a use of state transition method

x(t ) = (t , t 0 )x(t 0 ) + (t , )[B( )u( ) + D( )z ( )]d ,

t0

where:
(t ,t 0 ) is solution to homogeneous differential equation: x& = A(t )x,

x(t 0 ) = I .

We define energy performance index, which considers change with respect to time of
kinetic and potential (strain) energy of the system relatively to trajectory of the desired motion
(determined by vector of generalised displacements q and vector of generalised velocities
q& ), as well as instantaneous energy of the control command:

J (t ) =

) (

T
1 * & T
1
1
q& q Q1M * q& * q& + q * q Q 2 K * q * q + u T Ru ,
2
2
2

where:

Q1, Q2 matrices of dimensionless weighing coefficients ,

R

matrix of energy of the control commands.

Vector of generalised displacements q is a solution to matrix algebraic equation:

K *q = f0 ,

where:

Prof. Krzysztof J. Kaliski

f 0 vector of nonpotential generalised forces, which load the system during the
desired motion. In particular case: f 0 f * .
Vector of generalised velocities q& is a first derivative of the q vector, with respect to time.
Matrix Q1 determines influence of kinetic energy (important in case of unsteady
dynamic processes), while matrix Q2 influence of potential energy, whose importance is in
case of stability control of dynamic delayed closed loop systems.

Variation of energy performance index:

J =

)(
)(

)(
)(

1 * & T
q& q Q1M * + M*T Q1T q& * q&
2
.
T
1
1
+ q* q Q 2K * + K *T QT2 q* q + uT R + RT u
2
2

If we apply replacements:

q& * = [I M 0]x = T1x ,

1
424
3
T1

q * = [0 M I ]x = T2 x
1
424
3
T2

we well obtain:

q& * = T1x ,

q * = T2x .

We calculate variation:

t0

t0

x = (t , )B( )u( ) d = (t , )B( )d u .

The task is to control dynamic process; thus the control does not influence the desired
motion. We assume, that:

(q& * q& ) q& * ,

(q * q ) q * ,

and equal to zero variation of the performance index:

Prof. Krzysztof J. Kaliski

t
1 *
T
T
*
*T
&
&
J = q q Q1M + M Q1 T1 (t , )B( )d
t0
2
.
t

T
1 *
1 T
T
T
*
*T
+ q q Q 2 K + K Q 2 T2 (t , )B( )d + u R + R u = 0
2
2
t0

)(

)(

After rearrangements we obtain the optimal control command:

) B ( ) (t, )d {T (M
(K Q + Q K )(q K f )}

u=R+R

T 1

T
1

*T

)[

)]

& *1f + K *1f& +

Q1T + Q1 M * q& * K
0
0

t0

+ T2T

*T

T
2

*1

which minimises the energy performance index.

The optimal control command is a function of generalised displacements q*

and generalised velocities q& * , but its instantaneous values are determined online,
during the controlled process. Also the q& and q vectors are considered in the solution.

In real control process, control commands are restricted.

If we obtain constant value restrictions during online determination of the control
command values, we obtain:
ui (t ) ,
ui (t ) =
U i sgn [ui (t )],

ui (t ) U i
ui (t ) > U i

, i = 1,K, p ,

(4.8)

where:

p number of components of the control commands vector u,

Ui restriction of the absolute value of component ui of the control commands
vector u.

Important advantage. Ability of considering restrictions during on-line solution of the

optimal control at energy performance index.

For a purpose of determination of components of the matrices Q1, Q2 and R the trial and
error approach is recommended.

Prof. Krzysztof J. Kaliski

Control in domain of state coordinates

Control of a continuous system

Energy performance index, which considers change in time of kinetic and potential
energy of the system with respect to trajectory of the desired motion, but now defined by the
vector of state coordinates, is described:

J (t ) =

1
(x x )T Q(x x ) + 1 u T Ru ,
2
2

where:

Q M *
0
Q= 1
,
Q 2K *
0

x = q& T

qT

Variation of the energy performance index:

1
(x x )T Q + Q T (x x ) + 1 u T R + R T u .
2
2

J =

Considering the previous derivations in scope of control in generalised coordinates and

comparing the variation to zero:

1
J = (x x )T Q + Q T
2

) (t , )B( )d + 12 u (R + R )u = 0
t

t0

we obtain the optimal control:

u = R + RT

) B ( ) (t, ) d (Q + Q ) (x x) = K (t ) (x x) .
1

(4.25)

t0

14444444244444443
K R (t )
The latter is a function of matrix of the controller gains KR(t) and of the vector of state
coordinates x, but its on-line determination, apart from the optimal control at quadratic
performance index, does not require investigation of controllability of the system.

Prof. Krzysztof J. Kaliski

In case when not all components of the state vector x are known, this is necessity of
estimation of the rest.
The Luenberger observer:

x& = Ax + Bu + K e (y Cx ) ,
where:
x

estimated vector of the state coordinates,

Ke matrix of gain coefficients of the observer, being as explicit time function Ke=
Ke(t).
The equation above may be rewritten as:
x& = (A K eC )x + Bu + K e y = A 0 x + Bu + K e y ,
14243
A0

A0 state matrix of the system with the observer.

Determination of the Ke=Ke(t) matrix, requires numerous application of timeconsumable procedures. It enlarges time of control in significance.
z

CONTROLLED
PLANT

A
w

OBSERVER
CONTROLLER

A0

) B () (t,)d (Q+Q )

T 1

R+ R

Ke
+

+
+

t0

A scheme of the optimal control at energy performance index in linear continuous

nonstationary system: symbol
denotes signal x x in control system omitting the
observer

Prof. Krzysztof J. Kaliski

Control of a discrete system

Performance of the on-line optimal control at energy performance index in practice

requires sampling of measuring and control signals within finitely small time intervals, called
sampling time t. Its value is determined by properties of the measuring-control apparatus.
Then nonstationary system becomes the discrete one, which differs in significance of
solution the optimal control at energy performance index. State transition is applied for time
interval t, from instant t to t+t.
Then solution obtained by the state transition method:

xt + t = (t + t , t )xt +

t + t

(t + t , )[B( )u( ) + D( )z ( )]d .

t

First component is described as

t k
(t + t , t )x t exp[A t + t t ]x t A
x t = t + t x t ,
k!
k =0
1
42
4 43
4

k
t + t

t + t

but the second one:

t + t

(t + t , )[B( )u( ) + D( )z( )]d

t

1
[Bt + t ut + t + Dt + t z t + t + Bt ut + Dt z t ]t
2

t (k +1) 1
k
A
[Bt + t ut + t + Dt + t z t + t + Bt u t + Dt z t ] =

t + t
k
!
k =0
144244
3 2

e A t + t t

t + t

t + t

1
[Bt + t ut + t + Dt + t z t + t + Bt ut + Dt z t ] .
2

Thus, the state transition in nonstationary discrete system has a form:

xt + t = t + t xt + t + t

1
[Bt + tut + t + Dt + t zt + t + Btut + Dt zt ] .
2

Because, due to the z transformation xt + t = zxt , the state transition equation:

Prof. Krzysztof J. Kaliski

xt = Iz t + t

t + t

1
[Bt + t ut + t + Dt + t z t + t ] 1 + 1 .
2
z

By analogy, we can define the state transition for the observer, but taking into
consideration time-instant earlier with the t:

x t = 0 (t , t t )x t t +

(t , )[B( )u( ) + K ( )y( )]d .

0

t t

Because:

t k
0 (t , t t )x t t exp[A 0,t t ]x t t A 0k ,t
x t t = 0,t x t t ,
k
!
k =0
14243
0 ,t

and:
t

(t , )[B( )u( ) + K ( )y( )]d

0

t t

1
[Btut + K e,t y t + Bt tut t + K e,t t y t t ]t
2
( k +1)

1
k t
A
[Bt u t + K e ,t y t + Bt t ut t + K e,t t y t t ] =

0, t
!3 2
k =0
1
4
424k4

A 0 ,t t

0 ,t

0,t

1
[Btut + K e,t y t + Bt tut t + K e,t t y t t ] ,
2

we shall obtain the state transition equation for the observer:

x t = 0,t x t t + 0,t

1
[B t u t + K e,t y t + B t t u t t + K e,t t y t t ].
2

Because, due to the z transformation x t t =

1
x t , state equation for the observer
z

1
1

1
x t = I 0,t 0,t [B t ut + K e ,t y t ] 1 + .
z
2

Prof. Krzysztof J. Kaliski

In case of the optimal control command:

R t + t + R

(
(

R t + t + R Tt + t

t + t

+ QTt + t (x t xt )

t (k +1)
Q t + t + QTt + t (xt xt ) =
BTt + t A tk+ t
k!
k =0
1
44
42444
3

1
T
t + t

R t + t + R Tt + t

t + t

) B ( ) (t + t , )d (Q
(e ) t (Q + Q ) (x

u t + t = R t + t + R Tt + t

T
t + t

A t + t t T

T
t + t

t + t

tT+ t

xt )

BTt + t tT+ t Q t + t + QTt + t (x t xt ) ,

and finally we get:

u t + t = R t + t + R Tt + t

BTt + t tT+ t Qt + t + QTt + t (xt xt ) .

At the latest, the output equation has a form:

y t = Ct x t + w t .

zt+t
CONTROLLED
PLANT
Dt+t
wt

Bt+t

1
Iz t + t
2

xt

1
t + t 1 +
z

CONTROLLER

ut+t

R t + t + R

T
t + t

Ke, t

T
t + t

Q t + t + Q

T
t + t

x t
1
z

yt

OBSERVER

xt
1
T
t + t

Ct

1 1

1
I 0,t 0 ,t 1 +
2
z

ut

+
+

Bt

A scheme of the optimal control at energy performance index in discrete linear nonstationary
system: symbol
denotes signal x x in control system omitting the observer

Prof. Krzysztof J. Kaliski

Control of nonlinear discrete system

In case of nonlinear continuous systems, matrices in state equations and in the observer
equation are not only functions of time t, but also functions of state coordinates x. Thus:
A = A(x, t ), B = B(x, t ), C = C(x, t ), D = D(x, t ), K e = K e (x, t ), z = z (x, t ) .

Because looking for the exact solution to dynamic equation of nonlinear continuous
system makes a lot of difficulties, the optimal control in the relevant discrete system has
only a practical importance, and can be performed on a basis of knowledge of matrices A,
B, C, D, Ke and z for time t, but not as in case of the linear nonstationary one for time
t+t.

Then we will obtain the state transition matrix:

t + t = A tk
k =0

t k
t (k +1)
and after discrete integration t + t = A tk
k!
k!
k =0

Following equations shall have respective forms:

xt = Iz t + t

t + t

1
[Bt ut + Dt z t ] 1 + 1 .
2
z

1
1

1
x t = I 0,t 0,t [B t ut + K e ,t y t ] 1 + .
z
2

u t + t = R t + R Tt

)
BTt tT+ t Qt + QTt (xt xt xt ) ,

y t = Ct x t + w t .
)
Vector x t is a vector of correction values of the state coordinates vector in the desired

motion xt . Necessity of including the latter yields from solution of strongly nonlinear
problem, with a use of the linear approximation. Vector of disturbances z is determined for

Prof. Krzysztof J. Kaliski

time t, at which vector of state coordinates x is known. Vector of control command u, as

result of the state transition, is determined for time t+t.
)
There is no general method for determination of the xt vector components. However it

is possible to define it for particular cases of the optimal control, for example motion of twowheeled mobile platform.

zt

CONTROLLED
PLANT

Dt
wt
+

Bt

1
Iz t + t
2

1
t + t 1 +
z

xt

OBSERVER
Ke, t

xt

ut+t

BTt tT+ t Q t + Q Tt

)
xt
1
z

yt

Ct

CONTROLLER

R t + R Tt

x t

1
1

1
I 0 ,t 0 ,t 1 +
2
z
z

ut

+
+

Bt

A scheme of optimal control at energy performance index in discrete nonlinear system:

symbol
denotes signal x t x t in control system omitting the observer

References
1. Kaliski K. J., Galewski M. A.: A Modified Method of Vibration Surveillance with a Use
of the Optimal Control at Energy Performance Index. Mechanical Systems and Signal
Processing 2015, 58-59, 41-52.
2. Kaliski K. J., Buchholz C.: Mechatronic Design of Strongly Nonlinear Systems on a
Basis of Three Wheeled Mobile Platform. Mechanical Systems and Signal Processing
2015, 52-53, 700-721.

Prof. Krzysztof J. Kaliski