Fractional-order Calculus based Design and Imple ... · Fractional-order Calculus based Design and...
Transcript of Fractional-order Calculus based Design and Imple ... · Fractional-order Calculus based Design and...
Fractional-order Calculus based Design and Imple-
mentation of Robust Industrial Control
Aleksei Tepljakov
October 30, 2013
Talk overview
Aleksei Tepljakov 2 / 86
• Mathematical basis of fractional-order calculus;
• Identification of fractional-order systems:
Model classification, stability, time-domain identification;
Transition from conventional process models to fractional-order ones;
• Tuning of fractional-order controllers:
PIλDµ controllers, robustness characteristics and design specifications;
Transition from conventional controllers to those of fractional-order;
• Analog and digital implementations of fractional-order systems and controllers:
An updated continuous-time approximation method;
A framework for the synthesis of electrical network approximations;
Digital implementation for embedded control applications.
• Applications of fractional-order controllers: A case study.
Introduction: Mathematical Basis ofFractional-order Calculus
Aleksei Tepljakov 3 / 86
Introduction: Historical facts
Aleksei Tepljakov 4 / 86
• The concept of the differentiation operator D = d/dx is awell-known fundamental tool of modern calculus. For a suitablefunction f the n-th derivative is well defined as
Dnf(x) = d nf(x)/dxn, (1)
where n is a positive integer.
• What happens if we extend this concept to a situation, whenthe order of differentiation is arbitrary, for example, fractional?
• That was the very same question L’Hôpital addressed to Leibnizin a letter in 1695. Since then the concept of fractional calculushas drawn the attention of many famous mathematicians,including Euler, Laplace, Fourier, Liouville, Riemann, Abel.
Fractional derivative of a power function:An approach based on intuition
Aleksei Tepljakov 5 / 86
For the power function f(x) = xk the fractional derivative can beshown to be
dαf(x)
dxα=
Γ(k + 1)
Γ(k − α+ 1)xk−α, (2)
where Γ(·) is the Gamma function—generalization of the factorialfunction—defined as
Γ(x) =
∫
∞
0tx−1e−tdt, x > 0. (3)
For example,d1/2(x2)
dx1/2=
Γ(3)
Γ(5/2)x
3/2 =8x3/2
3√π.
Example: fractional-order derivative of afunction f(x) = x
Aleksei Tepljakov 6 / 86
Fractional derivative of a trigonometricfunction: An approach based on intuition
Aleksei Tepljakov 7 / 86
We observe, what happens when we repeatedly differentiate thefunction f(x) = sinx:
d
dxsinx = cosx,
d2
dx2sinx = − sinx,
d3
dx3sinx = − cosx, . . .
The pattern can be deduced: for the nth derivative, the functionsinx is shifted by nπ/2 radians. This is clear when observing agraph of the function. Thus, if we replace n by α ∈ R
+, we have
dα
dxαsinx = sin
(
x+απ
2
)
. (4)
Obviously, a similar equation holds for the cosine function as well.
Example: Half derivative of a sine function
Aleksei Tepljakov 8 / 86
0 1 2 3 4 5 6−1.5
−1
−0.5
0
0.5
1
1.5
t
Am
plitu
de
sin(t)
d0.5/dt0.5 (sin(t))π / 4
The generalized operator
Aleksei Tepljakov 9 / 86
Fractional calculus is a generalization of integration anddifferentiation to non-integer order operator aD
αt , where a and t
denote the limits of the operation and α denotes the fractionalorder such that
aDαt =
dα
dtα ℜ(α) > 0,
1 ℜ(α) = 0,∫ ta (dτ)
−α ℜ(α) < 0,
(5)
where generally it is assumed that α ∈ R, but it may also be acomplex number. We restrict our attention to the former case.
Fractional-order derivative: Definitions
Aleksei Tepljakov 10 / 86
Definition 1. (Riemann-Liouville)
Ra D
αt f(t) =
1
Γ(m− α)
(
d
dt
)m[
∫ t
a
f(τ)
(t− τ)α−m+1dτ
]
, (6)
where m− 1 < α < m, m ∈ N,α ∈ R+.
Definition 2. (Caputo)
C0 D
αt f(t) =
1
Γ(m− α)
∫ t
0
f (m)(τ)
(t− τ)α−m+1dτ, (7)
where m− 1 < α < m, m ∈ N.
Definition 3. (Grünwald-Letnikov)
GLD
αf(t)|t=kh = limh→0
1
hα
k∑
j=0
(−1)j(
α
j
)
f(kh− jh). (8)
Properties of fractional-order differentiation
Aleksei Tepljakov 11 / 86
Fractional-order differentiation has the following properties:
1. If α = n and n ∈ Z+, then the operator 0D
αt can be
understood as the usual operator dn/dtn.
2. Operator of order α = 0 is the identity operator:
0D0t f(t) = f(t).
3. Fractional-order differentiation is linear; if a, b are constants,then
0Dαt
[
af(t) + bg(t)]
= a 0Dαt f(t) + b 0D
αt g(t). (9)
4. If f(t) is an analytic function, then the fractional-orderdifferentiation 0D
αt f(t) is also analytic with respect to t.
Properties of fractional-order differentiation(continued)
Aleksei Tepljakov 12 / 86
5. For the fractional-order operators with ℜ(α) > 0,ℜ(β) > 0,and under reasonable constraints on the function f(t) (fordetails see [2]) it holds the additive law of exponents:
0Dαt
[
0Dβt f(t)
]
= 0Dβt
[
0Dαt f(t)
]
= 0Dα+βt f(t) (10)
6. The fractional-order derivative commutes with integer-orderderivative
dn
dtn(
aDαt f(t)
)
= aDαt
(
dnf(t)
dtn
)
= aDα+nt f(t), (11)
and if t = a we have f (k)(a) = 0, (k = 0, 1, 2, ..., n− 1).
Fractional-order derivative definitions:Laplace transform
Aleksei Tepljakov 13 / 86
Definition 4. (Riemann-Liouville)
L
[
RD
αf(t)]
= sαF (s)−m−1∑
k=0
sk[
Dα−k−1f(t)
]
t=0. (12)
Definition 5. (Caputo)
L
[
CD
αf(t)]
= sαF (s)−m−1∑
k=0
sα−k−1f (k)(0). (13)
Definition 6. (Grünwald-Letnikov)
L
[
GLD
αf(t)]
= sαF (s). (14)
For the first two definitions we have (m− 1 6 α < m).
On the meaning of the fractional-orderderivative
Aleksei Tepljakov 14 / 86
We shall call F (ft(·), t) a hereditary operator acting on a causeprocess ft(·) to produce a time-shifted effect g(t) which dependson the history of the process ft(τ); τ < t:
g(t) = F[
ft(·); t]
. (15)
We can replace g(t) by the function f(t) or its derivatives, i.e.
df(t)
dt= F
[
ft(·); t]
(16)
and so on.
Some hereditary process examples from physics: Brownian motion;Viscoelasticity.
Part I: Identification of Fractional-orderSystems
Aleksei Tepljakov 15 / 86
Identification of Fractional-order Systems
Aleksei Tepljakov 16 / 86
Author’s contribution:
• Research of time-domain (non)commensurate-order fractional modeloutput error based identification methods;
• Study of fractional-order characteristics of industrial plant models.
Publications:
1. A. Tepljakov, E. Petlenkov, and J. Belikov, “FOMCON: a MATLABtoolbox for fractional-order system identification and control,”International Journal of Microelectronics and Computer Science, vol. 2,no. 2, pp. 51–62, 2011
2. A. Tepljakov, E. Petlenkov, J. Belikov, and M. Halás, “Design andimplementation of fractional-order PID controllers for a fluid tank system,”in Proc. 2013 American Control Conference (ACC), Washington DC, USA,June 2013, pp. 1780–1785
Linear, time invariant fractional-ordersystem classification
Aleksei Tepljakov 17 / 86
Fractional-order transfer functions
Aleksei Tepljakov 18 / 86
A linear, fractional-order continuous-time dynamic system can be expressed bya fractional differential equation of the following form
anDαny(t) + an−1D
αn−1y(t) + · · ·+ a0Dα0y(t) = (17)
bmDβmu(t) + bm−1D
βm−1u(t) + · · ·+ b0Dβ0u(t),
We apply the Laplace transform with zero initial conditions and obtain thefractional-order transfer function:
G(s) =Y (s)
U(s)=
bmsβm + bm−1sβm−1 + · · ·+ b0s
β0
ansαn + an−1sαn−1 + · · ·+ a0sα0
. (18)
In the case of a system with commensurate order q we have
G(s) =
m∑
k=0
bk (sq)k
n∑
k=0
ak (sq)k. (19)
Stability
Aleksei Tepljakov 19 / 86
Theorem 1. (Matignon’s stability theorem) The fractionaltransfer function G(s) = Z(s)/P (s) is stable if and only if thefollowing condition is satisfied in σ-plane:
∣
∣arg(σ)∣
∣ > qπ
2, ∀σ ∈ C, P (σ) = 0, (20)
where σ := sq. When σ = 0 is a single root of P (s), the systemcannot be stable. For q = 1, this is the classical theorem of polelocation in the complex plane: no pole is in the closed right planeof the first Riemann sheet.
Algorithm summary: Find the commensurate order q of P (s), finda1, a2, . . . an in (19) and solve for σ the equation
∑nk=0 akσ
k = 0.If all obtained roots satisfy the condition (1), the system is stable.
Stability regions
Aleksei Tepljakov 20 / 86
Time-domain identification: Methoddescription
Aleksei Tepljakov 21 / 86
Given the transfer function model in (18)
G(s) =bmsβm + bm−1s
βm−1 + · · ·+ b0sβ0
ansαn + an−1sαn−1 + · · ·+ a0sα0
we search for a parameter set θ = [ ap αp bz βz ], such that
ap = [ an an−1 · · · a0 ], αp = [ αn αn−1 · · · α0 ],
bz = [ bm bm−1 · · · b0 ], βz = [ βn βn−1 · · · β0 ],
by employing numerical optimization with an objective functiongiven by an output error norm
∥
∥e (t)∥
∥
2
2, where e(t) = y(t)− y(t) is
obtained by taking the difference of the original model output y(t)and simulated model output y(t).
Time-domain identification: Process models
Aleksei Tepljakov 22 / 86
Consider the following generalizations of conventional process models used inindustrial control design.
(FO)FOPDT G(s) = K1+Ts
e−Ls G(s) = K1+Tsα
e−Ls
(FO)IPDT G(s) = Kse−Ls G(s) = K
sαe−Ls
(FO)FOIPDT G(s) = Ks(1+Ts)
e−Ls G(s) = Ks(1+Tsα)
e−Ls
Therefore, due to additional parameters K (gain) and L (delay) we mayupdate the optimized parameter set discussed previously to
θ = [ K L ap αp bz βz ].
Time-domain identification: FOMCONtoolbox
Aleksei Tepljakov 23 / 86
Identification example: Method validation
Aleksei Tepljakov 24 / 86
Suppose a fractional-order system is given by
G2 (s) =1
0.8s2.2 + 0.5s0.9 + 1.
In order to generate basic validation test data, the followingMATLAB commands can be used:
G2 = fotf(’1’,’0.8s^2.2+0.5s^0.9+1’);t = (0:0.01:20)’;u = zeros(length(t),1);u(1:200) = ones(200,1);u(1000:1500) = ones(501,1);y = lsim(G2,u,t)’;iddata1 = fidata(y,u,t);
Identification example: Method validation(continued)
Aleksei Tepljakov 25 / 86
In the time-domain identification tool, generate an initial model
Ginit(s) =1
s3 + s1.5 + 1.
After 82 iterations the following result is obtained with the errornorm enorm = 6.7310 · 10−10:
G(s) =1
0.8s2.2 + 0.5s0.90002 + 1s2.717·10−6 .
It is easy to see, that this model corresponds to the original one.To recover the model completely, choose from the identificationtool menu Model→Truncate and use the default accuracy of 0.001.
Identification example: Original response vs.Identified model (prior to truncation)
Aleksei Tepljakov 26 / 86
0 2 4 6 8 10 12 14 16 18 20−1
−0.5
0
0.5
1
1.5
2
Sys
tem
out
put y
(t)
Initial dataIdentified model
0 2 4 6 8 10 12 14 16 18 200
0.5
1
Sys
tem
inpu
t u(t
)
0 2 4 6 8 10 12 14 16 18 20−6
−4
−2
0
2
4
6
8x 10
−5
Time [s]
Out
put e
rror
Identification example: Identified modelstability check (prior to truncation)
Aleksei Tepljakov 27 / 86
Part II: Tuning of Fractional-order Controllers
Aleksei Tepljakov 28 / 86
Tuning of Fractional-order Controllers
Aleksei Tepljakov 29 / 86
Author’s contribution:
• Development of a flexible, time-domain simulation based PIλDµ controller tuningmethod incorporating robustness specifications derived from open-loopfrequency-domain analysis;
• Research of auto-tuning methods for PIλDµ controllers and implementationthereof in a hardware prototype (results to be published);
• Case studies of applications of PIλDµ controllers to various industrial plantmodels.
Publications:
1. A. Tepljakov, E. Petlenkov, and J. Belikov, “A flexible MATLAB tool for optimalfractional-order PID controller design subject to specifications,” in Proceedings ofthe 31st Chinese Control Conference, W. Li and Q. Zhao, Eds., Hefei, Anhui,China, 2012, pp. 4698–4703
2. A. Tepljakov, E. Petlenkov, and J. Belikov, “Development of analytical tuningmethods for fractional-order controllers,” in Proc. of the Sixth IKTDK Informationand Communication technology Doctoral School Conf., 2012, pp. 93–96
Fractional-order controllers
Aleksei Tepljakov 30 / 86
The fractional PIλDµ controller, where λ and µ denote the orders of theintegral and differential components, respectively, is given by
C(s) = Kp +Ki
sλ+Kd · sµ. (21)
The transfer function, corresponding to the fractional lead-lagcompensator of order α, has the following form:
CL(s) = K
(
1 + bs
1 + as
)α
. (22)
When α > 0 we have the fractional zero and pole frequencies ωz = 1/b,ωh = 1/a and the transfer function in (22) corresponds to a fractionallead compensator. For α < 0, a fractional lag compensator is obtained.
Basics of fractional control: fractionalcontrol actions
Aleksei Tepljakov 31 / 86
Let a basic fractional control action be defined as C(s) = K · sγ .The control actions in the time domain for γ ∈ [−1, 1] with K = 1under different input signals are given below.
0 2 4 6 8 100
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Time [s]
u(t)
γ=0
γ=−0.5
γ=−0.7
γ=−1
Fractional integrator s−γ
0 2 4 6 8 10−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
Time [s]u(
t)
γ=0
γ=0.5
γ=0.7
γ=1
Fractional differentiator sγ
PID controller vs. PI0.5D0.5 controller:frequency-domain characteristics
Aleksei Tepljakov 32 / 86
10−5
10−4
10−3
10−2
10−1
100
101
102
103
104
105
−90
−45
0
45
90
Pha
se (
deg)
Bode Diagram
Frequency (rad/sec)
0
20
40
60
80
100
Mag
nitu
de (
dB)
Classical PID
Fractional PID
Classical PID
Fractional PID
Optimization based PIλDµ tuning
Aleksei Tepljakov 33 / 86
Optimization provides general means of tuning a fractional-orderPID controller given a cost function and suitable optimizationconstraints. There are several aspects to the problem of designinga proper controller using constrained optimization:
• The type of plant to be controlled (integer or noninteger order,nonlinear);
• Optimization criterion (cost function);
• Fractional controller design specifications;
• Specific parameters to optimize;
• Selection of initial controller parameters.
Optimization based PIλDµ tuning: Costfunction
Aleksei Tepljakov 34 / 86
In case of a linear model we use time-domain simulation of atypical negative unity feedback loop
Gcs(s) =Gc(s)G(s)
1 +Gc(s)G(s). (23)
For the cost function we consider performance indicies:
• integral square error ISE =∫ τ0 e2(t)dt,
• integral absolute error IAE =∫ τ0
∣
∣e(t)∣
∣dt,
• integral time-square error ITSE =∫ τ0 te2(t)dt,
• integral time-absolute error ITAE =∫ τ0 t
∣
∣e(t)∣
∣dt.
Optimization based PIλDµ tuning:Constraints
Aleksei Tepljakov 35 / 86
The design specifications include:
• Gain margin Gm and phase margin ϕm specifications;
• Complementary sensitivity function T (jω) constraint, providingA dB of noise attenuation for frequencies ω > ωt rad/s;
• Sensitivity function S(jω) constraint for output disturbancerejection, providing a sensitivity function of B dB forfrequencies ω < ωs rad/s;
• Robustness to plant gain variations: a flat phase of the systemis desired within a region of the system critical frequency ωcg;
• For practical reasons, a constraint on the control effort u(t)may also be set.
Optimization based PIλDµ tuning: Initialparameter selection
Aleksei Tepljakov 36 / 86
The choice of initial PIλDµ parameters (Kp, Ki, Kd, λ and µ)depends on the selected set of parameters to optimize. There areseveral options available which depend on the particular controllerdesign strategy:
• Optimize all parameters;
• Optimize gains only;
• Optimize orders only.
It is possible to obtain the initial parameters to optimize by usingclassical, integer-order PID design methods, and then enhance theresulting controller by introducing the fractional orders with respectto design specifications.
Optimization based PIλDµ tuning:FOMCON toolbox
Aleksei Tepljakov 37 / 86
Optimization based PIλDµ tuning example
Aleksei Tepljakov 38 / 86
The task is to design a controller for a heating furnace described bya transfer function
G(s) =1
14494s1.31 + 6009.5s0.97 + 1.69.
First we approximate this model by a conventional FOPDT model:
GFOPDT (s) =0.4043
1 + 3440.71se−66.9311s.
Using the Ziegler-Nichols tuning formula the integer-order PIDparameters are obtained such that
Kp = 152.596,Ki = 1.13995,Kd = 5106.72.
Optimization based PIλDµ tuning example(continued)
Aleksei Tepljakov 39 / 86
Next, we take these parameters, fix the gains and optimize only theorders. Model approximation parameters for the Oustaloup refinedfilter are ω = [0.0001, 10000] rad/s, N = 5. The specifications areas follows
• Gain and phase margins Gm ≥ 10 dB, γ ≥ 75;
• Control law saturation ulim = [0; 750].
Using the ITSE performance index for optimization yields thefractional orders λ = 0.76145 and µ = 0.33145. A phase marginϕm = 102.84 is achieved.
The obtained result can be in part explained by relatively largedifferences in the model and the FOPDT approximation.
Optimization based PIλDµ tuning example:Closed-loop transient response comparison
Aleksei Tepljakov 40 / 86
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 100000
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Time [s]
Am
plitu
de
Ziegler−Nichols tuning
Optimized PIλDµ controller
Part III: Analog and Digital Implementationsof Fractional-order Controllers
Aleksei Tepljakov 41 / 86
Analog and Digital Implementations ofFractional-order Controllers
Aleksei Tepljakov 42 / 86
Author’s contribution:
• Study and update of an existing continuous-time method for fractional-orderoperator approximation;
• Research and development of a unified framework for analog approximations offractional-order systems and controllers;
• Development of a method for digital implementation of fractional-order filterapproximation for embedded control applications (to be published).
Publications:
1. A. Tepljakov, E. Petlenkov, and J. Belikov, “Application of Newton’s method toanalog and digital realization of fractional-order controllers,” International Journalof Microelectronics and Computer Science, vol. 2, no. 2, pp. 45–52, 2012
2. A. Tepljakov, E. Petlenkov, and J. Belikov, “Efficient analog implementations offractional-order controllers,” in Proc. of the 14th International Carpathian ControlConference (ICCC), 2013, pp. 377–382
Carlson’s approximation of fractionalcapacitors
Aleksei Tepljakov 43 / 86
This particular approximation of nth roots was considered byCarlson and the method was generalized to obtain rational transferfunction approximations of fractional capacitors in the form(1/s)1/n. The method utilizes the following equation:
Gk+1(s) = Gk(s)(n− 1)
(
Gnk(s)
)
+ (n+ 1)(
H(s))
(n+ 1)(
Gnk(s)
)
+ (n− 1)(
H(s)) , (24)
whereH(s) = 1/s
and the initial estimate is G0(s) = 1. The method works for botheven n = 2m and odd n = 2m+ 1 roots.
Carlson’s method: drawbacks
Aleksei Tepljakov 44 / 86
Several drawbacks of the approximation method can be outlined:
• The initial estimate for each particular approximation could beproperly chosen;
• The method only allows to obtain approximations for transferfunctions of order 1/n;
• Resulting approximations can be of a very high order;
• The limited frequency range where the approximation is valid.
In fact, the application of this method could be slightly different.That is, it would be beneficial to approximate
G(s) =
(
bs+ 1
as+ 1
)α
. (25)
The revised Carlson approximation method:Algorithm summary
Aleksei Tepljakov 45 / 86
1. Use a suitable decomposition method to obtain fractional powers1/m1, 1/m2, . . . , 1/mk;
2. Use the base transfer function Gbase(s) =bs+1as+1
and for each fractionalpower αk = 1/mk find the initial estimate
G0,k =
∣
∣
∣
∣
jbωm + 1
jaωm + 1
∣
∣
∣
∣
αk
,
where ωm = (ab)−0.5;
3. Obtain the approximation using
Gα(s) =
k∏
j=1
G1/mj
base (s) (26)
applying order reduction, if necessary;
4. Apply the order reduction technique to obtain the final approximation.
The revised Carlson approximation methodexample
Aleksei Tepljakov 46 / 86
Consider a problem of implementing a fractional-order leadcompensator for a position servo described by an integer-ordermodel G(s):
G(s) =1.4
s(0.7s+ 1)e−0.05s.
Design specifications are as follows: phase margin ϕm = 80, gaincrossover frequency ωcg = 2.2 rad/s. The controller C(s) to fulfillthese specifications may be designed such that
C(s) =
(
2.0161s+ 1
0.0015s+ 1
)0.702
.
Let us observe the convergence of the approximation obtainedusing the described method to the ideal frequency response.
The revised Carlson approximation methodexample: initial estimate
Aleksei Tepljakov 47 / 86
0
5
10
15
20
25
30
35
40
45
Mag
nitu
de (
dB)
10−3
10−2
10−1
100
101
102
103
104
105
0
30
60
90
Pha
se (
deg)
Bode Diagram
Frequency (rad/sec)
The revised Carlson approximation methodexample: iteration 1 (2nd order system)
Aleksei Tepljakov 48 / 86
0
5
10
15
20
25
30
35
40
45
Mag
nitu
de (
dB)
10−3
10−2
10−1
100
101
102
103
104
105
0
30
60
90
Pha
se (
deg)
Bode Diagram
Frequency (rad/sec)
The revised Carlson approximation methodexample: iteration 2 (11th order system)
Aleksei Tepljakov 49 / 86
0
5
10
15
20
25
30
35
40
45
Mag
nitu
de (
dB)
10−3
10−2
10−1
100
101
102
103
104
105
0
30
60
90
Pha
se (
deg)
Bode Diagram
Frequency (rad/sec)
The revised Carlson approximation methodexample: iteration 3 (56th order system)
Aleksei Tepljakov 50 / 86
0
5
10
15
20
25
30
35
40
45
Mag
nitu
de (
dB)
10−3
10−2
10−1
100
101
102
103
104
105
0
30
60
90
Pha
se (
deg)
Bode Diagram
Frequency (rad/sec)
Network approximation of a fractional-orderoperator: Fractance
Aleksei Tepljakov 51 / 86
By choosing Z1(s) and Z2(s) in the active filter configuration it ispossible to implement a fractional-order integrator or differentiator.The corresponding network approximations can be computed usinga suitable method. A unified framework adopting multiple methodscan be designed.
General fractance object: FOMCON toolbox
Aleksei Tepljakov 52 / 86
Our goal was to obtain a general enough approach tosystematization of existing network topologies and their generation.In FOMCON, fractance network synthesis and analysis is handledby means of a central component—an object, containing completecircuit information. The object contains references to
• particular network structures (MATLAB functions), howevercomplex, which return computed network transfer functions(impedance values),
• corresponding implementations (also MATLAB functions).
The idea is that a single structure can have several differentimplementations, including, e.g., optimization based ones.
General fractance object relations:FOMCON toolbox
Aleksei Tepljakov 53 / 86
For generality, we also consider the possibility of using inductivecomponents in network structures.
Fractance class properties: FOMCONtoolbox
Aleksei Tepljakov 54 / 86
Class frac_rcl() properties:
• model — a model of the fractional-order system (fotf object);
• structure — network structure (Cauer, Foster, etc.);
• implementation — function that carries out the actual computationof the network component values;
• ω — frequency points used for model validation;
• params — parameters used for implementation and/or in thestructure;
• K — network gain compensation factor(s);
• R,C,L — cell array with component value vectors (size depends onthe number of substructures);
• results — implementation/validation results.
Fractance class methods: FOMCON toolbox
Aleksei Tepljakov 55 / 86
The following particular methods are implemented:
• tf(), zpk() — return the impedance Z(s) in transfer function orzero-pole-gain format, corresponding to the fractance circuit, fornetwork analysis. This can be used to automate, e.g. frequencyresponse analysis.
• prefnum() — locates closest component values according to thepreferred series and replaces network components accordingly. It isalso possible to provide custom values of components so that thealgorithm will choose the closest matches among a custom set ofvalues. This can be necessary to analyze the changes of the systemfrequency characteristics due to variation of component values.
• zscale() — implementation of impedance scaling used to shift thevalues of discrete electronic components into the feasible domain.
Approximation of fractional operators: TheOustaloup filter
Aleksei Tepljakov 56 / 86
The Oustaloup recursive filter gives a very good approximation offractional operators in a specified frequency range and is widelyused in fractional calculus. For a frequency range (ωb, ω) and oforder N the filter for an operator sγ , 0 < γ < 1, is given by
sγ ≈ K
N∏
k=−N
s+ ω′
k
s+ ωk, K = ωγ
h, ωr =ωh
ωb, (27)
ω′
k = ωb(ωr)k+N+1
2 (1−γ)
2N+1 , ωk = ωb(ωr)k+N+1
2 (1+γ)
2N+1 .
The resulting model order is 2N + 1.
A modified Oustaloup filter has been proposed in literature [3].
Oustaloup filter approximation example
Aleksei Tepljakov 57 / 86
Recall the fractional-order transfer function
G(s) =1
14994s1.31 + 6009.5s0.97 + 1.69,
and approximation parameters ω = [10−4; 104], N = 5.
0 0.5 1 1.5 2 2.5 3 3.5
x 104
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Step Response
Time (sec)
Am
plitu
de
Grunwald−Letnikov
Oustaloup filter
Refined Oustaloup filter
10−5
100
105
−135
−90
−45
0
Pha
se (
deg)
Bode Diagram
Frequency (rad/sec)
−250
−200
−150
−100
−50
0
Mag
nitu
de (
dB)
Oustaloup filter
Refined Oustaloup filter
Original plant
Oustaloup filter
Discrete-time approximation: The zero-polematching equivalents method
Aleksei Tepljakov 58 / 86
Continuous zeros and poles, obtained using the Oustaloup recursivefilter, are directly mapped to their discrete-time counterparts bymeans of the relation
z = esTs , (28)
where Ts is the desired sampling interval. The gain of the resultingdiscrete-time system H(z) must be corrected by a proper factor.
This implementation method has been successfully used in ourprevious work. We remark, that for the synthesis of continuouszeros and poles in (27) with the intent to obtain a discrete-timeapproximation the transitional frequency ωh may be chosen suchthat
ωh 62
Ts. (29)
Discrete-time approximation offractional-order controllers
Aleksei Tepljakov 59 / 86
After acquiring a set of discrete-time zeros and poles by means of(28), the fractional-order controller may be implemented in form ofa IIR filter represented by a discrete-time transfer function H(z−1).In general, one has two choices:
1. Implement each fractional-order component approximation ofthe controller in (21) separately as Hλ(z−1) and Hµ(z−1); thismethod offers greater flexibility, since the components may bereused in the digital signal processing chain, but requires morememory and is generally more computationally expensive;
2. Compute a single LTI object approximating the wholecontroller; this method is suitable when there is a need for astatic description of a fractional-order controller.
Digital controller implementation: IIR filters
Aleksei Tepljakov 60 / 86
In this particular work we choose the second option, that is we seeka description of the controller in the form
H(z−1) = Kb0 + b1z
−1 + b2z−2 + · · ·+ bmz−m
a0 + a1z−1 + a2z−2 + · · ·+ anz−n. (30)
For practical reasons, the equivalent IIR filter should be comprisedof second-order sections. This allows to improve computationalstability when the target signal digital processing hardware haslimited DSP capabilities. Thus, the discrete-time controller mustbe transformed to yield
H(z−1) = b0
N∏
k=1
1 + b1kz−1 + b2kz
−2
1 + a1kz−1 + a2kz−2. (31)
Part IV: Applications of Factional-orderControl
Aleksei Tepljakov 61 / 86
Applications of Fractional-order Control
Aleksei Tepljakov 62 / 86
Author’s contribution:
• Tuning and digital implementation of controllers for simulated and physicallaboratory models;
• Implementation of an analog controller for a simulated model.
Publications:
1. A. Tepljakov, E. Petlenkov, J. Belikov, and M. Halás, “Design and implementationof fractional-order PID controllers for a fluid tank system,” in Proc. 2013American Control Conference (ACC), Washington DC, USA, June 2013, pp.1780–1785
2. A. Tepljakov, E. Petlenkov, J. Belikov, and J. Finajev, “Fractional-order controllerdesign and digital implementation using FOMCON toolbox for MATLAB,” inProc. of the 2013 IEEE Multi-Conference on Systems and Control conference,2013, pp. 340–345
3. A. Tepljakov, E. Petlenkov, J. Belikov, and S. Astapov, “Digital fractional-ordercontrol of a position servo,” in Proc. 20th Int Mixed Design of Integrated Circuitsand Systems (MIXDES) Conference, 2013, pp. 462–467
Case study (1): Fractional-order control ofthe coupled tank system
Aleksei Tepljakov 63 / 86
The system is modeled in continuous time inthe following way:
x1 =1
Au1 − d12 − w1c1
√x1, (32)
x2 =1
Au2 + d12 − w2c2
√x2,
where x1 and x2 are levels of fluid, A is thecross section of both tanks; c1, c2, and c12 areflow coefficients, u1 and u2 are pump powers;valves are denoted by wi : wi ∈ 0, 1 and
d12 = w12 · c12·sign(x1 − x2)√
|x1 − x2|.
Case study (1): Fractional-order control ofthe coupled tank system (continued)
Aleksei Tepljakov 64 / 86
Our task is to control the level in the first tank. We identify thereal plant from a step experiment with w1 = w12 = 1, w2 = 0 in(32). The resulting fractional-order model is described by a transferfunction
G2 =2.442
18.0674s0.9455 + 1e−0.1s. (33)
We notice, that this model does not tend to exhibit integer-orderdynamics. Due to the value of the delay term the basic tuningformulae for integer-order PID tuning do not provide feasibleresults. It is possible to select some starting point manually andrun optimization several times. However, it is important to choosethe correct frequency domain specifications to ensure controlsystem stability.
Case study (1): Experiments with controllerimplementation: Hardware platform
Aleksei Tepljakov 65 / 86
Case study (1): Fractional-order control ofthe coupled tank system (continued)
Aleksei Tepljakov 66 / 86
In our case the goal is to minimize the impact of disturbance, soconstraints on the sensitivity functions could be imposed. Ourchoice is such that
∣
∣T (jω)∣
∣ ≤ −20 dB, ∀ω ≥ 10 rad/s and∣
∣S(jω)∣
∣ ≤ −20 dB, ∀ω ≤ 0.1 rad/s. The gain and phase marginsare set to Gm = 10dB and ϕ = 60, respectively. Additionally, inorder to limit the overshoot, the upper bound of control signalsaturation was lowered from 100% to 60%. Using the IAEperformance metric we finally arrive at the following PIλDµ
controller parameters by optimizing the response of the nonlinearsystem in Simulink:
Kp = 6.9514, Ki = 0.13522, Kd = −0.99874,
λ = 0.93187, µ = 0.29915. (34)
Case study (1): Fractional-order control ofthe coupled tank system (continued)
Aleksei Tepljakov 67 / 86
0 10 20 30 40 50 60 70 80 90 100 110 1200
0.05
0.1
0.15
0.2
0.25
Leve
l [m
]
0 10 20 30 40 50 60 70 80 90 100 110 1200
0.2
0.4
0.6
0.8
1
Con
trol
law
u(t
) [%
]
0 10 20 30 40 50 60 70 80 90 100 110 1200
1
t [s]
w2
Real plant, controller in SimulinkDiscrete−time simulationReal controller, model in Simulink
Real plant
Real plant
Case study (2): Fractional-order control ofa servo system
Aleksei Tepljakov 68 / 86
Case study (2): Fractional-order control ofa servo system (continued)
Aleksei Tepljakov 69 / 86
The following transfer function is identified:
G(s) =192.1638
s(1.001s+ 1).
The generic PD controller parameters provided by INTECO areKp = 0.1,Kd = 0.01. We shall use these parameters as the initialones for the optimization.
The results of optimization are such, that after 100 iterations thegains of the PD controller have been found as Kp = 0.055979 andKd = 0.025189.
After fixing the gains and manually perturbing the value of µ to0.5, the optimized PDµ controller is obtained with µ = 0.88717.Phase margin of the open loop control system is ϕm = 65.3.
Case study (2): Fractional-order control ofa servo system (continued)
Aleksei Tepljakov 70 / 86
Experimental setup for evaluating the digital implementation of thefractional-order PID controller:
Case study (2): Fractional-order control ofa servo system (continued)
Aleksei Tepljakov 71 / 86
0 10 20 30 40 50 60−50
0
50
100
150
200
Ang
le [r
ad]
0 10 20 30 40 50 60−1
−0.5
0
0.5
1
Con
trol
law
u(t
)
0 10 20 30 40 50 600
0.2
0.4
Time [s]
Dis
turb
ance
System with integer−order PD controller
System with PDµ controller (in Simulink)
System with PDµ controller (external)Reference angle
Case study (3): Network approximation of aFO lead compensator
Aleksei Tepljakov 72 / 86
Recall the example, where our goal was to obtain an analogimplementation a fractional controller for a model of a positionservo
G(s) =1.4
s(0.7s+ 1)e−0.05s.
We now provide the results of approximating the controller
C(s) =
(
2.0161s+ 1
0.0015s+ 1
)0.7020
by an electrical network by using a deterministic method,implemented as part of the unified network generation frameworkin FOMCON, for obtaining the parameters of the network.
Case study (3): Electrical networkapproximations
Aleksei Tepljakov 73 / 86
In order to implement it, the following steps are carried out:
• We choose R1 = 200kΩ and C1 = 1µF due to the timeconstant τ .
• The basic structure is the Foster II form RC network and theimplementation is done by means of the mentioned algorithm.
• To obtain the differentiator, we use the propertyZd(s) = 1/Zi(s), where Zd(s) and Zi(s) correspond toimpedances of a differentiator and an integrator, respectively.
• This is done by setting the impedances in the active filtercircuit such that Z1(s) = Zi(s) and Z2(s) = Rk, where Rk
serves as a gain correction resistor.
Case study (3): Electrical networkapproximations (continued)
Aleksei Tepljakov 74 / 86
b = 2.0161; wz = 1/b;alpha = 0.702;Gc = fotf(’s’)^alpha / wz^alpha;
params = struct; params.R1 = 200e3;params.C1 = 1e-6; params.N = 4;params.varphi = 0.01;
imp2 = frac_rcl(1/Gc, ...’frac_struct_rc_foster2’, ...’frac_imp_rc_foster2_abgen’, ...logspace(-2,3,1000), ...params);
Case study (3): Electrical networkapproximations (continued)
Aleksei Tepljakov 75 / 86
The controller is obtained from the object using
C = 1/zpk(imp2);
Now we set the resistor values to the preferred series with 5% tolerance,and the capacitor values are substituted for closest components out ofthe 10%-series:
imp2 = imp2.prefnum(’5%’,’10%’,[],’5%’);
Finally, the bill of materials can be generated using engnum():
[vals, str] = engnum(imp2.R);
The variable str will contain the following:
’360 k’ ’200 k’ ’75 k’ ’27 k’ ’9.1 k’
The gain setting resistor Rk has the preferred value of 390kΩ.
Case study (3): Electrical networkapproximations (continued)
Aleksei Tepljakov 76 / 86
Case study (3): Electrical networkapproximations (continued)
Aleksei Tepljakov 77 / 86
−40
−30
−20
−10
0
10
20
30
40
50
60
Mag
nitu
de (
dB)
10−3
10−2
10−1
100
101
102
103
0
30
60
90
Pha
se (
deg)
Bode Diagram
Frequency (rad/s)
Fractional differentiator
Ideal response of lead compensator
Network approximation
Case study (3): Electrical networkapproximations (continued)
Aleksei Tepljakov 78 / 86
Bode Diagram
Frequency (rad/s)
−80
−70
−60
−50
−40
−30
−20
−10
0
10
20
System: CompensatedGain Margin (dB): 24.9At frequency (rad/s): 23.3Closed loop stable? Yes
System: UncompensatedGain Margin (dB): 23.2At frequency (rad/s): 5.28Closed loop stable? Yes
Mag
nitu
de (
dB)
100
101
102
−270
−225
−180
−135
−90
−45
0 System: CompensatedPhase Margin (deg): 85.5Delay Margin (sec): 0.71At frequency (rad/s): 2.1Closed loop stable? Yes
System: UncompensatedPhase Margin (deg): 49.1Delay Margin (sec): 0.774At frequency (rad/s): 1.11Closed loop stable? Yes
Pha
se (
deg)
UncompensatedCompensated
Case study (3): Electrical networkapproximations (continued)
Aleksei Tepljakov 79 / 86
Case study (3): Frequency response aroundωcg = 2.2 rad/s
Aleksei Tepljakov 80 / 86
Case study (3): Frequency response aroundωcg = 2.2 rad/s
Aleksei Tepljakov 81 / 86
Bode Diagram
Frequency (rad/s)100
101
0
30
60
90
Pha
se (
deg)
0
2
4
6
8
10
12
14
16
18
20
Mag
nitu
de (
dB)
Fractional lead compensatorElectrical network approximation (real)
Case study (3): Electrical networkapproximations: Results
Aleksei Tepljakov 82 / 86
0 20 40 60 80 100 120−10
−5
0
5
10
15
Am
plitu
de
0 20 40 60 80 100 120−5
0
5
Time [s]
Con
trol
law
u(t
)
Response with simulated model and real controllerSimulated responseSet point
Analog controllerSimulated controller
FOMCON project: Fractional-orderModeling and Control
Aleksei Tepljakov 83 / 86
• Official website: http://fomcon.net/
• Toolbox for MATLAB available;
• An interdisciplinary project supported by the Estonian DoctoralSchool in ICT and Estonian Science Foundation grant nr. 8738.
Acknowledgments
Aleksei Tepljakov 84 / 86
References
Aleksei Tepljakov 85 / 86
[1] K.S. Miller, B. Ross, An Introduction to the Fractional Calculus andFractional Differential Equations, Wiley, New York, 1993.
[2] I. Podlubny, Fractional Differential Equations, Academic Press, SanDiego, CA, 1999.
[3] C.A. Monje, Y.Q. Chen, B.M. Vinagre, D. Xue, V. Feliu,Fractional-order Systems and Controls Fundamentals andApplications, Springer-Verlag, London, 2010.
[4] R. Hilfer, Applications of Fractional Calculus in Physics, WorldScientific, Singapore, 2000.
[5] I. Podlubny, Geometric and Physical Interpretation of FractionalIntegration and Fractional Differentiation, Fractional Calculus andApplied Analysis, vol. 5, no. 4, pp. 367-386, 2002.
Discussion
Aleksei Tepljakov 86 / 86
Thank you for listening!