Sie sind auf Seite 1von 6

2019 IEEE 58th Conference on Decision and Control (CDC)

Palais des Congrès et des Expositions Nice Acropolis


Nice, France, December 11-13, 2019

Digital H∞ Robust Control of Mechanical System with Implicit


Observer
Bruno A. Angélico1 , Mateus M. Brugnolli1 , Gabriel P. das Neves1

Abstract— The state-space model of a mechanical system usu- Thus, a static output feedback controller can be designed
ally has in its state vector the angular positions and velocities. conventionally using LMI formulation.
For not using observers, the systems must have position and Following the space-state representation of the system, the
velocity sensors. Considering a digital control formulation, this
work includes the derivative estimates in the system model, and model can be augmented with integrators for output reference
a new representation is built considering the actual angular tracking. With the aid of LMIs, the H∞ control design
positions and their previous values. With this representation, [11] is particularly attractive for the augmented model.
a robust H∞ static state feedback controller is designed, and Each state is obtained through sensors measuring, which are
the robustness of the full system can be guaranteed since the sensitive to exogenous signals, for example, noise signals
state estimator structure is included in the system model. A
servomechanism and a homemade rotary inverted pendulum and disturbances. The H∞ control designs a state feedback
are considered as examples of practical applications. control to reduce the upper bound of the infinity norm of
an exogenous signal of the augmented system. The resulting
I. INTRODUCTION control loop has guaranteed stability and bounded response
Two of the most considered systems in Control Engi- to exogenous signals.
neering laboratories are the servomechanism and the rotary The main advantages of the proposed method are: (i)
inverted pendulum (Furuta pendulum). They are mechanical a discrete state-space model representation with the states
systems that have angular positions and velocities as states obtained with the present and past sensors measurements;
[1]–[3]. To design a state feedback control law, it is desired (ii) flexibility to choose the order of the derivative approxi-
to have both position and velocity measurements available. mation; (iii) as the states are sensitive to exogenous signals,
Assuming full controllability, if all the states are available, a the H∞ control bounds the exogenous infinity norm; (iv) and
proper static robust state feedback controller can be designed. notably, this control structure is simple to implement as any
If only positions sensors are available, one needs to state feedback control. Simulation and practical results with
consider an output feedback scheme. However, the control two classical mechanical systems are provided to validate the
design in these cases is generally a Bilinear Matrix Inequality proposed scheme.
(BMI), which is not a convex problem [4], [5]. II. SYSTEM MODEL WITH IMPLICIT OBSERVER
In the literature, there are methods which require conser-
vatism in the structure of the variables to obtain Linear Ma- Consider an n-th degrees of freedom mechanical sys-
trix Inequalities (LMI) conditions to design output feedback tem1 , where θ1 to θn represent the angular positions of
gains [6]–[10]. For example, researchers apply two-steps [8] each mechanical joint and ω1 to ωn as their respective
or iterative algorithms [10], that increase the feasibility of angular velocities. The states can be arranged as Θ(t) =
θ1 . . . θn > and Ω(t) = ω1 . . . ωn > . Hence, a
   
the control design.
A common feedback control scheme has a derivative linear model can be obtained, such that
state estimator using discrete approximations as a filter [2].
" # " #
Θ̇(t) Θ(t)
However, the sensor measurements are generally noisy and = A + Bu(t) (1)
Ω̇(t) Ω(t)
if the velocity estimator is not designed correctly, the control
system may present a poor performance. with A ∈ R2n×2n and B ∈ R2n×m , with m being the
In this work, the simple Euler backward method is consid- number of inputs.
ered for the velocity estimation. However, instead of consid- After a model discretization, the discrete state-space model
ering it as a separate filter, it is included in the discrete-time can be represented with block-matrices:
state-space representation of the system. The resulting state-  
space model has the current and the past angular positions A11 A 12 " #  1
 d d  Θ[n]
" #
Θ[n + 1] B
as states instead of current angular positions and velocities. = 21 22  +  d  u[n] (2)
Ω[n + 1]  A A  Ω[n]
*This work was supported by Fundação de Amparo à Pesquisa do
d d Bd2
Estado de São Paulo (FAPESP), grant 2017/22130-4 and Coordenação de
xy
∈ Rn×n . Next, the designer may choose a
Aperfeiçoamento de Pessoal de Nı́vel Superior - Brazil (CAPES) - Finance
Code 001. with A
d
1
Dept. of Telecommunications and Control Eng. — Escola difference approximation for the derivative and its order.
Politécnica da USP, São Paulo, Brazil angelico@lac.usp.br,
1
mateus mmb@usp.br, gabriel.pereira.neves This approach is only valid for mechanical systems, where the velocity
@usp.br states can be obtained from position measurements approximations.

978-1-7281-1398-2/19/$31.00 ©2019 IEEE 1171

Authorized licensed use limited to: University of Waterloo. Downloaded on May 31,2020 at 19:54:55 UTC from IEEE Xplore. Restrictions apply.
 
Considering the First-Order Euler backward approximation 2 12 22 In
K2 = − 2Ad + 2Ad + ,
for Ω[n], it follows that 3 + 2Ts 2
1  12 
A11 A12 Ad + A22
    
Θ[n + 1] Θ[n] K3 =
 Θ[n + 1] − Θ[n]  =  d d  Θ[n] − Θ[n − 1]  + 3 + 2Ts d ,
21
Ad Ad 22
Ts Ts 2Ts  1 
K4 = Bd + Bd2 .
3 + 2Ts
B1
 
 d  u[n]. (3) A. Servomechanism
Bd2 The servomechanism considered is shown in Figure 1.
The following two equations are obtained: There is a DC motor with a tachometer speed sensor coupled
12 ! to the same axis2 . There are two stages of belt pulleys, each
A A12
d Θ[n − 1] + B 1 u[n] one with a reduction of n = 1/3, which couples the motor
Θ[n + 1] = A11 d + d Θ[n] − d
Ts Ts shaft to the axis of the potentiometer. For position control,
(4) the potentiometer axis is considered, where there is also a
and disc with a graduated scale to indicate the position.
A22
!
Θ[n + 1] − Θ[n] 21 d
= Ad + Θ[n] −
Ts Ts
A22
d Θ[n − 1] + B 2 u[n]
d (5)
Ts
Adding up (4) and (5), it turns out that
Θ[n + 1] =
1  
Ts A11
d + A 12 + T A21 + A22 + I Θ[n] −
d s d d n
1 + Ts
1  12 
A + A22 d Θ[n − 1] +
1 + Ts d
Ts  1 
Bd + Bd2 u[n], (6)
1 + Ts
resulting in a new state-space representation:
" # " #" # " #
Θ[n + 1] K1 K2 Θ[n] K3 Fig. 1. Picture of the servomechanism.
= + u[n], (7)
Θ[n] 1 0 Θ[n − 1] 0
The armature voltage is represented by VM . The voltages
with
at the outputs of the tachometer and potentiometer are
1  
K1 = Ts A11
d + A 12 + T A21 + A22 + I ,
d s d d n respectively represented by VT and VP . The speed at the
1 + Ts motor axis is given by ΩM , while the velocity and position
−1 
22 , K = Ts
  
on the potentiometer axis are given by ΩP and ΘP . The
K2 = A12 + A 3 B 1 + B2 .
1 + Ts d d 1 + Ts d d relation between armature voltage and velocity in the motor
This procedure can be considered for any order mechan- axis is, approximately, given by a first-order system with gain
ical system, provided that the positions and velocities are K and time constant T . The tachometer and potentiometer
grouped separately. If a second difference approximation is gains are given by KT [V / (rad / s)] and KP [V / rad]. The
considered for the derivative, such that, simplified linear diagram of the system is shown in Figure
2.
3Θ[n] − 4Θ[n − 1] + Θ[n − 2]
Ω[n] = , (8)
2Ts
then the following representation can be obtained
      
Θ[n + 1] K1 K2 K3 Θ[n] K4
 Θ[n]  =  1 0 0  Θ[n − 1]+ 0  u[n],
      

Θ[n − 1] 0 1 0 Θ[n − 2] 0 Fig. 2. Schematic of the servomechanism.


(9)
where The nominal parameter values are: KP = 1.7, KT =
3A12 3A22 0.017, K = 50, T = 0.28.
!
2 11 d 21 d
K1 = Ts Ad + + Ts Ad + + In ,
3 + 2Ts 2 2 2
The speed sensor is not used in this work.

1172

Authorized licensed use limited to: University of Waterloo. Downloaded on May 31,2020 at 19:54:55 UTC from IEEE Xplore. Restrictions apply.
 
Considering the state vector as x(t) = ΘP (t) ΩP (t) ,
the input u(t) = vM (t), and the output y(t) = ΘP (t), the
following state-space representation is obtained:
" # " #" # " #
Θ̇P (t) 0 1 ΘP (t) 0
= + u(t). (10)
Ω̇P (t) 0 −1/T ΩP (t) Kn2 /T
After substituting the nominal values:
" # " #" # " #
Θ̇P (t) 0 1 ΘP (t) 0
= + u(t).
Ω̇P (t) 0 −3.5714 ΩP (t) 19.8413
(11)
With Ts = 0.1 s, the ZOH discrete-time equivalent of (11)
is:
" # " #" #
ΘP [n + 1] 1 0.0841 ΘP [n]
= +
ΩP [n + 1] 0 0.6997 ΩP [n]
" #
0 Fig. 3. Built rotational inverted pendulum.
u[n]. (12)
19.8413
According to Equation (7), a new discrete-time represen-
tation can be found such that
" # " #" #
ΘP [n + 1] 1.8761 −0.8761 ΘP [n]
= +
ΘP [n] 1 0 ΘP [n − 1]
" #
0.0884
u[n]. (13)
1.6685
B. Rotary Inverted Pendulum
Figure 3 presents the built system. The base and the tower
are made of wood. On the top of the tower, there is support
with a gear system with reduction ratio n = 4 connecting
the DC motor to the pendulum arm. The motor itself has an
incremental encoder used to obtain the angular displacement
of the arm. In rotation axis of the pendulum, there is another
encoder to acquire its angular displacement. The control
hardware is the Teensy 3.2, which has an ARM Cortex-
M4 microcontroller with 256 Kbytes of programming flash
Memory and 64 Kbytes of RAM. The motor driver is an
H-bridge, model VNH5019.
A schematic of the system can be found in Figure 4. The
Fig. 4. Illustration of the pendulum.
system dynamics equations for the rotational inverted pen-
dulum were defined using the Euler-Lagrange formulation.
This method requires the definition of a Lagrangian, which potential energy of the pendulum. However, the movement
is the sum of the kinetic energy of each body minus the of the rotating arm is parallel to the horizontal axis, thus
potential energy, as in Equation (14). P0 is always zero. Then, the potential energy P1 can be
L = K0 + K1 − P (14) described using the displacement of the center of mass of
the pendulum (CoM), as in Equation (16).
The Euler-Lagrange method consists in the dynamic equa-
tions of the system using (15). The elements θi are known P = m1 g`1 cos θ1 (16)
as generalized variables of the system and Qi are known as
generalized forces. The kinetic energy of the rotating arm K0 is composed
! only by the rotation. The inertia I0 is calculated considering
d ∂L ∂L that this joint can be modeled as a bar around one of its
− = Qi (15)
dt ∂ θi ˙ ∂θi sides. Hence,
2
The potential energy of the system can be divided into I θ̇ m (2` )2
P0 , the potential energy of the rotating arm, and P1 , the K0 = 0 0 ; I0 = 0 0 + m0 d2 (17)
2 12

1173

Authorized licensed use limited to: University of Waterloo. Downloaded on May 31,2020 at 19:54:55 UTC from IEEE Xplore. Restrictions apply.
The kinetic energy of the pendulum K1 can be described the direction of the motor rotation, the following linear model
using its rotation, the velocity of the CoM in the x-direction is obtained around x(t) = [0 0 0 0]> :
and in y-direction, as in Equation (18),
ẋ(t) = Ax(t) + Bu(t)
2
I θ̇ m 
K1 = 1 1 + 1 `21 θ̇12 + r2 θ̇02 + `21 θ̇02 sin2 (θ1 )+ y(t) = Cx(t), (20)
2 2
 m (2` )2 where
2r`1 θ̇0 θ̇1 cos(θ1 ) ; I1 = 1 1 (18)
12 
0 0 1 0

0 0 0 1 
The torque of the DC motor of 12 Volts is proportional to A= ,
 
the duty-cycle of the Pulse Width Modulation (P W M ) sig-
0 −9.1047 −0.6066 0.0132 
nal applied at the armature terminals, subject to mechanical 0 86.2615 1.0110 −0.1249
losses. The torque can be expressed as in Equation (19),
−52.7483 > ,
 
K   B= 0 0 31.6490
τ = t 12 P W M − Ke θ̇0 , (19)
R and
where P W M ∈ [−1, 1] stands for the control signal (per- 
C= 1 0 0

0 .
centage of PWM and signal of the motor rotation). Lastly,
the generalized forces Qi of Equation (15) are defined as After ZOH discretization with Ts = 0.025 s, it resulted in
external forces and reaction forces with relation to each
generalized variable θi . For the rotating arm, there is a torque x[n + 1] = Ad x[n] + Bd u(t)
applied by the DC motor and there is a reaction torque due y[n] = Cx[n], (21)
to viscous damping of motor shaft and gearbox (b0 ). For the
pendulum, there is a reaction torque due to viscous damping where x[n] = θ0 [n] θ1 [n] ω0 [n] ω1 [n] > , and3
 
of pendulum bearing and encoder coupling (b1 ).
1.0000 −0.0028 0.0248 −0.0000
 
• Q0 = τ − b0 θ̇0 , b0 = 0.0001 (N · m)/(rad/s);
• Q1 = −b1 θ̇1 , b1 = 0.000001 (N · m)/(rad/s).  0 1.0270 0.0003 0.0252 
Ad =  ,
 
The solution of Equation (15) for each generalized vari-  0 −0.2276 0.9849 −0.0025
ables θ0 and θ1 are the expressions of θ̈0 and θ̈1 . The values 0 2.1697 0.0253 1.0239
of the parameters for the built rotational inverted pendulum
are shown in Table I. The CoM of the rotating arm was and
determined using SolidWorks software. The radius d was
−1.3185 > .
 
calculated from the fixed axis with relation to (w.r.t.) the Bd = 0.0098 −0.0165 0.7863
CoM of the rotating arm. For the inverted pendulum, it was more
convenient to consider a second difference
TABLE I
approximation for the derivative, resulting in a sixth-
PARAMETERS OF THE BUILT PENDULUM . order discrete model with states vector x[n] =
 >
Parameter Value θ0 [n] θ1 [n] θ0 [n − 1] θ1 [n − 1] θ0 [n − 2] θ1 [n − 2] ,
m0 Rotating Arm Mass [Kg] 0.393 and
m1 Pendulum Mass [Kg] 0.068 2.3211 −0.0063 −1.6521 0.0033 0.3311 −0.0008
2`0 Rotating Arm Length [m] 0.3650
2`1 Pendulum Length [m] 0.207 0.0252 2.3958 −0.0336 −1.7037 0.0084 0.3440 
r Radius from Fixed Axis to Pendulum [m] 0.210 
1.0000

d CoM of the Arm w.r.t. Fixed Axis [m] 0.022 0 0 0 0 0 
Ad= ,
 
g Gravity Acceleration [m/s2 ] 9.81  0 1.0000 0 0 0 0 
 
 0 0 1.0000 0 0 0 
The parameters of the motor DC are shown in Table II. 0 0 0 1.0000 0 0

TABLE II and
Bd = 0.0131 −0.0219 0 0 0 0 > .
 
PARAMETERS OF THE MOTOR .

Parameter Value
Kt Motor Torque Constant [N m/A] 0.02 The proposed representation with implicit observer is
Ke Back EMF Constant [V s/rad] 0.08 simple for implementation and its states are a sequence of
R Armature Resistance [Ω] 2.4 output signals. In the pendulum example, the current output
measurement and two past measurements are considered as
After properly defining the state vector as x(t) = system states.
[θ0 (t) θ1 (t) θ̇0 (t) θ̇1 (t)]> , and u(t) = P W M (t) ∈
3
[−1, 1] representing the duty-cycle of the PWM signal and For the ease of presentation, only four decimal digits are shown.

1174

Authorized licensed use limited to: University of Waterloo. Downloaded on May 31,2020 at 19:54:55 UTC from IEEE Xplore. Restrictions apply.
It was also considered that Bw = 1 0 0 > . Applying
 
III. CONTROL DESIGN
In both cases the servo mode was considered and the the Theorem (3.1) the following controller is obtained
controllers were designed considering an integrator at the K 0 = 6.8083 −4.2531 −0.5272 .
 
(28)
desired tracking position input, as presented in Figure 5,
resulting in an augmented system by one state, such that, B. Rotary Inverted Pendulum
x0 [n] = x[n] v[n] > . In the Furuta pendulum, it is desired
 
Considering the Furuta pendulum, the following matrices
to track the arm position, resulting in a seventh order system. were defined
For the servomechanism, the resulting order is equal to three. " #
Czp
Cz = , (29)
0[14×7]
h i>
Dz1 = 0[1×7] Dz1 p 0[1×7] , (30)
h i>
Dz2 = 0[1×7] 0[1×7] Dz2 p , (31)

Fig. 5. Discrete time state feedback with integrator. where


√ √ √ √ 
Czp = diag 2 2 2 2 2 2 0.1 ,
The augmented system matrices are represented as
3 3 3 3 3 3 3 >,
 
" # " # Dz1 p =
0 Ad 0[n×1] 0 Bd i>
Ad = , Bd = , h
−C 1 0 Dz2 p = 10−3 10−3 0 0 0 0 0 ,

and Cd0 = C 0 , while the control gain matrix is given by


 
and Bw = 1 1 0 0 0 0 0 > . Applying Theorem
 
K 0 = KC −KI .
 
(3.1) the controller is designed, such that
For the Furuta pendulum, it is desired to robust stabilize
K 0 = KC −KI
 
the states and to track the arm position (θ0 ). Consider the
following system representation = [ −35.9280 −70.8605 45.0465 ...
x0 [n + 1] = (A0d − Bd0 K 0 )x0 [n] + Bw w[n] (22) ... 80.5902 −11.0908 −19.3796 0.0477 ]
0 0 (32)
z[n] = (Cz − Dz1 K )x [n] + Dz2 w[n],
For the ease of presentation, only four decimal digits were
where w[n] is an exogenous input and z[n] the controlled shown, but double precision was considered in the simulation
output. The controller can be designed by the Theorem 3.1. and practical experiments.
Theorem 3.1 (Bounded Real Lemma [11]): If, and only
if, exists a matrix W = W > and Z such that IV. SIMULATION AND PRACTICAL RESULTS
min µ (23) For the simulation, some prior experiments were con-
W =W > >0 ducted to obtain a gross estimation of the measurement noise
variance. For the rotary inverted pendulum, it was observed a
subject to
noise variance 10−6 for the arm encoder sensor and 10−7
W A0d W + Bd0 Z
 
0 Bw for the pendulum encoder sensor. For the servomechanism,
the observed noise of the potentiometer sensor presented a
W Cz> + Z > Dz>
 
? W 0 
 1  > 0 (24) variance of 10−4 . Based on these values, Gaussian noise
? ? I Dz2 
 
sources were included in sensor measurements.
? ? ? µI The practical results are presented together with the sim-
then the state-feedback controller u[n] = −K 0 x0 [n], with ulation ones for the sake of comparison. Note that the states
K 0 = −ZW −1 , stabilizes the system (22) and minimizes are a sequence of the system output measurements. A video
√ demonstrating the performance of the proposed controller in
the upper bound kHk∞ < γ, where γ = µ.
The matrices Cz , Dz1 and Dz2 are tuned to achieve the real system is submitted as an attachment.
desired system performance for an exogenous input w[n].
A. Servomechanism
A. Servomechanism Figures 6 and 7 present the results considering the tracking
For the servomechanism, it was assumed that mode. As can be seen, the controller was able to stabilize
 √  the system and make it track the reference signal. The
Cz = diag 1 0.2 0.1 , (25) simulation result presented an overshoot somehow larger
 √ √ √ > when compared to the real system. The actuator has a
Dz1 = 1.5 0.25 1.5 0.25 1.5 0.5 , (26)
saturation in 5V, but in both simulation and practical results,
Dz2 = 1 0 0 > .
 
(27) the control signal had not exceeded the limits.

1175

Authorized licensed use limited to: University of Waterloo. Downloaded on May 31,2020 at 19:54:55 UTC from IEEE Xplore. Restrictions apply.
2 0.2
2 10

8 1 0.1
1.5
6
0 0
1
4
0.5 -1 -0.1
2

0 0 -2 -0.2
0 10 20 30 40 50 0 10 20 30 40 50

-0.5 -2 5 1

-4 0.5
-1
-6 0
-1.5 0
-8 -0.5

-2 -1
0 10 20 30 40 50 -10
0 10 20 30 40 50
-5 -1.5
0 10 20 30 40 50 0 10 20 30 40 50

Fig. 6. Servomechanism controlled states. Fig. 8. Pendulum controlled states.

4 1

2 0.5

0 0

-2 -0.5

-4 -1
0 10 20 30 40 50 0 10 20 30 40 50

Fig. 7. Servomechanism control signal. Fig. 9. Pendulum control signal.

B. Rotary Inverted Pendulum R EFERENCES


Results concerning the Furuta pendulum are shown in [1] K. Ogata, Modern Control Engineering, 5th ed. Upper Saddle River,
Figures 8 and 9. As can be observed, the controller presented NJ: Prentice Hall, 2010.
[2] G. Franklin, J. Powell, and A. Emami-Naeini, Feedback Control of
a good performance in stabilizing the states and making the Dynamic Systems, 6th ed. Upper Saddle River, NJ: Pearson, 2010.
arm position to track a step reference signal. The control ef- [3] N. Nise, Control Systems Engineering, 6th ed. Hoboken, NJ: John
fort has reached 50% of the PWM limit during the reference Wiley & Sons, 2011.
[4] S. Kanev, C. Scherer, M. Verhaegen, and B. D. Schutter, “Robust
changes. output-feedback controller design via local bmi optimization,” Auto-
We stress that, since the rotary pendulum arm is not matica, vol. 40, no. 7, pp. 1115 – 1127, 2004.
precisely aligned with the horizontal plane, the pendulum [5] Q. Tran Dinh, W. Michiels, S. Gros, and M. Diehl, “An inner convex
approximation algorithm for bmi optimization and applications in
angle is stabilized in a position slightly apart from the zero control,” in 2012 IEEE 51st IEEE Conference on Decision and Control
encoder measurement. (CDC), Dec 2012, pp. 3576–3581.
[6] J. C. Geromel, C. C. de Souza, and R. E. Skelton, “Lmi numerical
V. CONCLUSIONS solution for output feedback stabilization,” in Proceedings of 1994
American Control Conference - ACC ’94, vol. 1, June 1994, pp. 40–
A digital H∞ robust control procedure was proposed for 44 vol.1.
mechanical systems, where the derivative estimations are [7] X. Du and G.-H.Yang, “Lmi conditions for h∞static output feedback
included in the plant model, in such a way that the control control of discrete-time systems,” in 2008 47th IEEE Conference on
Decision and Control, Dec 2008, pp. 5450–5455.
policy is designed considering the structure of the observer. [8] C. M. Agulhari, R. C. L. F. Oliveira, and P. L. D. Peres, “Robust
The proposed scheme was validated in two practical systems, h-inf static output-feedback design for time-invariant discrete-time
a servomechanism, and a rotary inverted pendulum. As polytopic systems from parameter-dependent state-feedback gains,” in
Proceedings of the 2010 American Control Conference, June 2010,
observed, the proposed approach resulted in good perfor- pp. 4677–4682.
mances for both systems, representing a feasible approach [9] Y. Kurotaki, T. Nagashio, and T. Kida, “An optimal design of h-inf
for designing digital output feedback robust controllers for static output feedback controller using lmi for collocated gyroscopic
system,” in 2012 IEEE 51st IEEE Conference on Decision and Control
mechanical plants. (CDC), Dec 2012, pp. 7776–7780.
[10] S. Bhattacharyya and S. Patra, “Static output-feedback stabilization
for mimo lti positive systems using lmi-based iterative algorithms,”
IEEE Control Systems Letters, vol. 2, no. 2, pp. 242–247, April 2018.
[11] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear matrix
inequalities in system and control theory. SIAM, 1994, vol. 15.

1176

Authorized licensed use limited to: University of Waterloo. Downloaded on May 31,2020 at 19:54:55 UTC from IEEE Xplore. Restrictions apply.

Das könnte Ihnen auch gefallen