International Journal of Computer Applications (0975 – 8887)

National Conference on Computational Intelligence for the Engineering Quality Software (CiQS-2014)

Linear Quadratic Gaussian Control for Two Interacting

Conical Tank Process
S.K.Lakshmanaprabu U. Sabura Banu C.Bharanitharan
Research Scholar Professor PG Student
Electronics and Instrumentation Engineering
BS Abdur Rahman University, Vandalur, Chennai – 48

ABSTRACT The paper is organized as follows. Section II presents the two

In this paper optimal control for two interacting conical tank interacting conical tank process modeling. LQG Controller
process (TICTP) was designed. The optimal control is design described in section III,IV. Simulation results and
obtained by LQG solution with optimal kalman filter. This controller Performance analysis are shown and discussed in
paper describes the theatrical base and practical application of section V. Final conclusions are given in section VI.
an optimal dynamic regulator using model based Linear
Quadratic Gaussian (LQG) control design for nonlinear
2. TWO INTERACTING CONICAL
process. This LQG regulator consists of an optimal state- TANK PROCESS DESCRIPTION
feedback controller and an optimal state estimator. In this
case, a performance criterion is minimized in order to
maintain level of the water in both tanks.

Keywords
Linear quadratic Gaussian controller; Kalman estimator; Two
interacting conical Tank Process; MIMO nonlinear process.

1. INTRODUCTION
Optimal control theory is emerging technique in advanced
process control. In optimal control theory, the calculus
variations and optimization techniques combined used to form
path of controlled variable such that cost function like control
energy and state fluctuation is minimized[1][2][3]. In general
Model predictive control strategy is used mostly on higher Fig 1: Schematic diagram of two interacting conical tank
level , the performance improvement of lower level PID loops
gives improved performance in multivariable systems. But The proposed system consists of two conical tanks which are
Linear quadratic Gaussian control provides optimal and robust in the shape of an inverted cone fabricated from a sheet metal.
control strategy to Multivariable process[4][5]. Linear The height of the process tank is 50cm and the top end,
Quadratic Gaussian design is a optimal control theory which tapering end diameters are 40cm and 14cm. The two tanks are
has many application in control engineering problem. This connected through an interacting pipe with valve (HV1). The
LQG technique widely used in Medical process controllers, in interaction of process can be changed by position of this valve
nuclear power plants and motor control systems. In Linear (HV1). It has a reservoir to store water and this is supplied
Quadratic Regulator(LQR) model, the resulting control law through the pumps to the tanks. Provisions for water inflow
are linear with respect to state variable. The control law is and outflow are provided at the top and bottom of the tank
easy to compute but the disadvantage in LQR is, it fails to respectively. Gate valves, one at the outflow of the tank1 and
perform well when the system is subjected to disturbances [6]. the other at the outflow of the tank2 are connected to maintain
So the special case of controller is designed using separation the level of water in the tanks. Variable Speed pump work as
principle [7]. The concept of Linear Quadratic regulators and actuator and it is used to discharge the water from reservoir
Kalman estimator combined to form LQG control model for tank to process tanks. The speed of pump directly
better servo and regulatory response. Linear Quadratic propositional to the input voltage. It consists of differential
Regulator is optimal robust controller to disturbance rejection pressure transmitter for measuring the bottom pressure created
but for servo problem it may not track set point due to some by water level and it gives height in terms of milliamps.
model mismatch and error in the sensor output. So a Kalman
filter used to minimize the asymptotic covariance of the 2.1. Mathematical Modelling
estimation of error when the process encounter with
disturbances [8][9][10].The state feedback gain matrix uses (1)
dh1 K pp1 U1  1a1 2 gh1 – sign(h1  h2 ) 12 a12 2 g h1  h2
state as input from Kalman estimator not from real process. 
dt h2
The optimal state feedback gain and optimal state estimator  R12 1
H1
gain determined to form optimal LQG controller. The state
feedback LQ-optimal gain is determined by Ricatti equation
[11][12]. The main advantage of LQG control is that closed dh2 K pp 2 U 2   2 a2 2 gh2  sign(h1  h2 ) 12 a12 2 g h1  h2 (2)

loop stability is guaranteed when the system is accurate. dt h2
 R22 2
Cylindrical tanks used in many process industry for discharge H2
of liquid. To control liquid level and flow in process tanks are
still challenging problem in process industries.

35
 International Journal of Computer Applications (0975 – 8887)
National Conference on Computational Intelligence for the Engineering Quality Software (CiQS-2014)

The nominal values of the parameters and variables are Table 2. Operating conditions and the conventional state
tabulated in table 1. space and transfer function model of the interacting
conical tank process
Table 1. Nominal values of the parameters used
Region Operati State space Transfer function
Parameter Description Value ng matrix G(s)
points
R Top Radius of conical tank 20 cm  6.9 4.3 
I U1s=2.2 A  7.704 s  83.08 49.64 
  2 
h1=0-5 U2s=1.1  6.4 10.8 s 2  17.73 s  47.07 
 s  17.73 s  47.07
H Maximum height of Tank1&2 50 cm h1s=2.5 7.7 0   49.64 11.51s  79.89 
h2=0-4 B   
 0 11.5
 s  17.73 s  47.07 s 2  17.73 s  47.07 
h2s=2.1 2
0-10 V 1 0  0 0
U1, U 2 Input Voltage to pump1 &2 C   ;D   
0 1  0 0 

K pp1 , K pp 2 Pump gain 75 cm3 V .S e c  0.22 0.13 

II U1s =4.5 A   0.49 s  0.17 0.10 
 0.21 0.35  2 
h1=5-15 U2s =2.1 0.49 0 
2
 s  0.568s  0.048 s  0.568 s  0.048 
β1 Valve Co-efficient of MV1 50 cm2 /s h2=4-12 h1s=10.1 B
 0 0.75
 

0.10 0.75s  0.16 

 s  0.568 s  0.0486 s  0.568 s  0.048 
h2s=8.1 1 0  0 0  2 2
C  ; D  0 0 
35cm2 /s
0 1   
β12 Valve co-efficient of Mv12
III U1s=7  0.11s  0.008 0.004 
β2 Valve co-efficient of Mv2 50cm2 /s h1=15- U2s =2
 0.02 0.016 
A
 0.03 0.07 
  2 2
 s  0.105  0.0016 s  0.105  0.0016 


h1s=20.9  0.11 0   0.004 0.26 s  0.007 

25 B
 0 0.26   
𝑎1 , 𝑎12 , 𝑎2 Cross section Area of pipe 1.2272 𝑐𝑚2 h2=12- h2s=13.8 1 0 
C
0 0 
 ; D  0 0 
 s  0.105  0.0016 s  0.105  0.0016 
2 2
 0 1   
20
U1s=7.6  0.01 0.009   0.04 s  0.0006 
IV A 
0.0004
 0.01 0.016  2 05 
h1=25- U2s =5  s  0.02 s  9.91e s 2  0.02 s  9.91e05 
2.2. Open Loop data h1s=35.1 B
0.04 0 
  0.0004 0.048 s  0.0006 
The open loop data was generated in the conical tank system 45  0 0.04  s 2  0.02 s  9.9e05
 s 2  0.028s  9.9e05 
h2s=32.6 1 0  0 0 
by varying the inflow rate F1 in tank 1 and noting down the h2=20- C   ;D   
0 1  0 0 

respective level h1 and h2 and is tabulated in Table 2. The I/O 45

characteristics and linearized region of the two conical tank
process is shown (see Figure 2). 3. MULTI LOOP CONTROLLER
DESIGN
The most attractive advantages of multiloop control system is
simple controller structure and easiness to handle loop failure.
The Variable pairing is selected by Relative gain array
method. The influence between inputs on outputs is calculated
and the high influence input and output pair is chosen for
control. The input1 (voltage applied to pump1) is paired with
output1 (height of tank1) and input2 (voltage applied to
pump1) is paired with output2 (height of tank2).

Fig 2a: U1 Vs h1 Fig.2b: U1 Vs h2

Fig 3: Multiloop Control Scheme

Fig2c: U2 Vs h1 Fig.2d: U2 Vs h2 PI controller transfer function is Gc( s)   k p s  ki 

 
Fig 2: I/O characteristics of the conical tank process  s 

Thus the piecewise Linearization method is used to separate The second order transfer function is reduced into first order
the whole non-linear region into various regions. The plus dead time and diagonal transfer function is tuned using
characteristics are divided into different region. The operating Ziegler Nichols method.
points are found out for each region.

36
 International Journal of Computer Applications (0975 – 8887)
National Conference on Computational Intelligence for the Engineering Quality Software (CiQS-2014)

Table.3 CDM based PI and ZN based PI values for law of LQR controller is “u=-kx” that reduce the state
TICTP fluctuation. Normally this gain is called state feedback gain or
Conical Transfer Function ZN based PI LQ-optimal gain. Control action is based on the state of
FOPDT Loop1 Loop2 process, if there is error in process state measurement due to
model mismatch and noise, then controller will take action
 5.227e 1s 2.576 e 1s  Kp=8.827 Kp=4.711 only for error measurement. To avoid this issue, kalman
 51.22 S  1 66.29 S  1 estimator is used to estimate the state of process.To form the
 1s Ki=2.758 Ki=1.535
 2.576 e 4.727 e 1s  LQG regulator, simply connect the Kalman filter and LQ-
 66.29 S  1 24.65 S  1  optimal gain K as shown (see Figure 5).

4. LINEAR QUADRATIC GAUSSIAN

CONTROLLER DESIGN
Linear-quadratic-Gaussian (LQG) control is a modern state-
space technique which uses a state-space model of the plant.
The model of process will not be a perfect replica of real
process. The real time system is corrupted by noise and
disturbance. The state space model of system which is
affected by process noise „w‟ and measurement noise „v‟ is
given below,

Fig 5: Linear quadratic Gaussian controller

State cannot be fully measurable in real time system, so the

states are estimated based on the output. This regulator has
state-space equation is
̂  A x
x ˆ  Bu (6)
Estate estimated error e  x  x ˆ ,
Fig4 : Block diagram of LQG Regulator
Dynamic equation of error is
e  A x  A x ˆ  Ae (7)
x  Ax  Bu  G w (3)
The error goes to zero asymptotically when the State matrix A
y  C x  Du  H w  v (4) is stable. If the state matrix A is unstable then error become
uncontrollable and estimated states x̂ grows further apart
Assume w and v are white noises.LQG regulator consist of from process state x. So correction term L is introduced to
kalman state estimator and optimal feedback gain. The rank of provide direct corrective action to minimize the model error.
controllability and observability matrix of system is same as

ˆ  Ax
x ˆ  B u  L( y  y
ˆ) (8)
number of state of process. So the system is fully observable
and controllable. Estimator can be used for estimate all the ˆ (0)  0
where x
states. ˆ Cx
y ˆ (9)
The performance of LQG control is measured by performance But in real time process, states are generally corrupted by „w‟
index, is process noise and „v‟ is measurement noise. State space
model of noise corrupted system is model in equation 6,7.


 x Q x  u T R udt
Now error dynamic of observer changes due to corrupted
J (u )  T (5)
white noise.
e  ( A x  B u  w)  ( A x  B u  L(C x  v  C xˆ )
0
Minimum values of J(u) represents minimum effect of (10)
controller energy „u‟ and minimization of state variable e  ( A  LC)e  w  L v
changes in process. The selection of Q and R matrix is logical The error estimation will not go to zero because of process
and meaningful. The Q and R matrixes selected for minimum noise „w‟ and measurement noise „v‟. So LQG technique, the
value of performance index j(u) which means that the output-feedback problem reduced by the estimated state.
minimization of square of manipulated input and Assume the measurement noise v and process noise w are
minimization of square of state variables fluctuations. In uncorrelated zero mean Gaussian white noise. Kalman filter is
MIMO process, Influences between each input and output robust filter and very good optimal estimator when the system
variables varies. The weightage for each manipulated variable affected by Gaussian white noise. Specifically, it minimizes
varies based on the impact of output variables. The weighting the asymptotic covariance of the estimation error x  xˆ .
lim E ( x  x ˆ )T 
matrices Q,R are fixed for minimizing the J(u). Q,R are the
ˆ) ( x  x
controller design parameters , large Q penalizes transients of x t  (11)
, large R penalizes usage of control action „u‟. The optimal Kalman gain L minimizes E e(t ) 2 is
The state feedback gain is fixed for the minimum value of L  PC *V 1 . where P is the unique positive- semi definite
J(u). This minimization of gain matrix „K‟ find by solving solution of the ARE
Riccati equation[11][12]. The regulator response of this linear
P A*  A P  P C * V 1 C P  W  0 (12)
quadratic regulator will be never affected by disturbance. But
the setpoint tracking is not possible in LQR controller. The

37
 International Journal of Computer Applications (0975 – 8887)
National Conference on Computational Intelligence for the Engineering Quality Software (CiQS-2014)

Kalman filter is equivalent to designing an LQR controller on 1 0 ; Equal weightage given to state h1 and h2 of
Qn  
the dual system (A*,C*,B*,D*) with Q = W, R = V. The 0 1

system is fully observable and controllable. So, this estimated the tank. And 1 0 fixed for minimizing the
Rn  
0 1

state given to the kalman gain L.
controller energy and state variation.
5. DESIGN OF LQG CONTROLLER
1 0  ; 1 0 
FOR TICTP Qn    Rn   
(16)
Steps for Designing Linear Quadratic Gaussian Regulatory, 0 1  0 1 
i. Check for controllability and observability of the The row size of QN specifies the length of w and NN is set
system. to 0 when omitted. By means of kalman filter, the estimated
ii. Choose Q and R such that Q = MTM, with (A,M) states are found
detectable, and R = RT > 0   0.119 0.0005 
A 
iii. This equation is the matrix algebraic Riccati  0.0208 0.275 
equation (MARE), whose solution P is needed to   0.089 0.0167 
B  
compute the optimal feedback gain K.  0.0167 0.1994 

iv. Solve the Riccati equation PA + ATP + Q – PBR- (17)
 1 0  0 0
1 T
B P = 0, And compute K = R-1BTP,Simulate the C   ;D  
. 0 1
 0 0

initial response of x  ( A  BF ) x for different
initial conditions. Kalman returns the estimator gain L and the steady-state error
v. If the transient response specifications and/or the covariance P (solution of the associated Riccati equation).
magnitude constraints are not met, and again step 1
is considered to re-choose the value of Q and R. L  PC T R 1
(18)
0  AP  PAT  PC T R 1CP  Q, P  0,
5.1. Checking of Controllability and
Q  E ( wwT ), R  E (vvT )
Observability From the above equation A,B,C,D,R,Q can be substituted and
The controllability of matrix depends on A and B matrix and
can find the value of P, With the help of all those values, the
observability of system depends on matrix A and C matrix.
state feedback „L‟ can be easily found and can be used for
The system is said to be fully observable if all the states of
designing an LQG controller for TICTP.
system can be externally measured.
 0.0891 0.0167 
A system is controllable if condition W is satisfied P
 0.0167 0.1994 

W: rank [B AB …….An-1 B] = n (13)
The state feedback gain will be
Therefore the value of  0.089 0.017 
L
0.114 0 0.003 0.004 
W   0.017 0.199 

 0 0.26 0.004 0.020 
6. ANALYSIS
And rank (W) = 2, Finally the system satisfies the condition Two interacting conical tank system was analyzed by
W, therefore the system is controllable. developing simulation model. The interaction of this process
A system is observable if condition M is satisfied is highly nonlinear. So piecewise linearization method is used
 C  to separate the whole non-linear region into various linear
M: rank  CA  regions. The characteristics are divided into different region.
 n The operating points are found out for each region. Linear
 .  (14)
 n 1  state space format is obtained for four different regions by
CA  using jacobian matrix method.
 1 0 
Therefore the value of   The combination of state estimator and state feedback LQ
0 1
M   gain will be LQG controller for TICTP. Thus the calculated
 0.0298 0.0162 
  values are used for obtaining the proper LQG controller which
 0.0375 0.0760  helps to reject disturbance and to track set point accordingly.
According to the process interaction, Q and R values are
And rank (M) = 2; The system satisfies the condition M, chosen. Since in TICTP, the interaction is equal Q and R
therefore the system is observable. matrix chosen as [1 0; 0 1] and respective gain value is used
Here the system is both controllable and observable. in state feedback. For third linear region Linear Quadratic
Therefore the design of LQG controller is possible. Since the Gaussian regulatory was designed.
third region of TICTP is considered for the design of LQG
controller. Similarly it will be designed for rest of the region. The response of disturbance rejection at steady sate is shown
The state space for third region is (see Figure 6) and performance validation between LQG and
PI is tabulated in table.4. The disturbance is applied to study
 0.0298 0.0162  0.1140 0  the regulatory performance. At 300sec and 400sec positive
A ;B   0
 0.0375 0.0760   0.2629  disturbance is applied to both tanks. The disturbance is
(15) rejected quickly by LQG controller.
1 0  0 0
C  ; D  0 0 
0 1   

38
 International Journal of Computer Applications (0975 – 8887)
National Conference on Computational Intelligence for the Engineering Quality Software (CiQS-2014)

7. CONCLUSION
A Linear Quadratic Gaussian (LQG) control design for two
interacting conical tank system has been investigated to
achieve robust properties. The Linear Quadratic Regulator
(LQR) designed by minimizing a performance criterion based
on comfort and controller energy considerations; and than a
Kalman filter (KF) is used to estimate the state of the system.
However, the objective of this paper has been to examine the
use of optimal state-feedback controllers in TICTP. The
integrator is added for servo tracking, integral constant is
based on trial and error method. There is no standard method
Fig 6: Regulatory response of LQG controller available for selection of Q and R. Here, based on the impact
of states on process output, the Q and R matrix selected. The
Table.4 Performance Indices for Regulatory Response of LQG controller performance is compared with ZN tuned PI
LQG and PI controller controller. This paper has shown the dynamic behaviour of
Control IAE ISE ITAE nonlinear multivariable system. The simulation results for
Strategy variety of input processes shows that the method adopted in
PI Loop1 237.1 1816 9772 this research work gives better controller performance.
Loop2 113.4 1024 3349
LQG Loop1 146.37 1173 2683 8. REFERENCES
Loop2 81.21 877.5 1395 [1] R.W.H.Sargent, “optimal control”, journal of
computational and applied mathematics, volume 124 ,
The servo and regulatory response of LQG and PI controller is issue1-2, pp.361-371, 2000.
shown (see Figure 4). The small overshoot can be eliminated [2] Erding cong , Minghui hu , Shandong Tu, Huihe shao ,
by proper selection of integral gain. For this simulation “A New optimal control system design for chemical
response integral gains chosen by trial and error method. The process”., Chinese journal of chemical engineering ,
LQG controller response is very smooth and servo tracking of vol.2, issue 12 , pp.1341-1346, 2013.
process variable is good. Initially the set point for h1 and h2
[3] Bryson Jr. AE , “Optimal control 1950 to 1985”, IEEE
are about 24cm and 14cm respectively. At the time of 500s,
control system magazine , vol.16 , issue 3 , pp.26-
the height of the tank1 and tank2 are 20cm and 18cm changed
33,1996.
respectively, the controller keeps on tracking the changed
setpoint, and disturbances applied at 400sec, 800 sec to check [4] Wei Zhang , Jianghai hu, Jianming lian., “Quadratic
the behavior of the controller. The LQG controller tracks the optimal control of switched linear stochastic systems”,
setpoint and rejects the disturbance effectively. The systems and control letters , vol.59, issue 11, pp.736-744,
performance of LQG controller is compared with PI controller 2010.
and tabulated in table.5. The disturbance rejection of LQG is [5] N.kumaresan, P.Balasubramaniam., “Optimal control for
guaranteed. stochastic linear singular system using neural networks”,
Journal of process control, vol.19,issue3 , pp.482-488,
2009.
[6] M. Guay , R.Dier, J.Hahn, P.J. Mulellan., “Effect of
process nonlinearity on linear quadratic regulator
performance” , vol.15 , issue 1, pp. 113-124, 2005.
[7] V.Sundarapandian, “A separation theorem for robust
pole placement of discrete-time linear control systems
wih full order observers”, Mathematical and computer
modeling, vol.43, issue1-2, pp.42-48, 2006.
[8] Bak, D., Michalik, M. & Szafran, J. 2003, "Application
of Kalman filter technique to stationary and non
stationary state observer design", Power Tech
Conference Proceedings, IEEE Bologna IEEE, , pp. 6.
Fig 7 : Servo and Regulatory response of LQG controller Vol. 3, 2003.
[9] E. Hendricks, O. Jannerup and P. H. Sørensen, "Optimal
Table.5 Performance Indices for Servo and Regulatory observers: Kalman filters," in Linear Systems Control
Response of LQG and PI controller
Anonymous Springer, pp. 431-491, 2008.
Control IAE ISE ITAE
Strategy [10] G. Bishop and G. Welch, "An introduction to the kalman
PI Loop1 336.2 3398 28680 filter," Proc. of SIGGRAPH, Course, 8, 2001.
Loop2 110.2 752.9 13030 [11] R.S. Bucy , “Global theory of the Riccati equation” ,
LQG Loop1 256 1318 18920 Journal of computer and system sciences, vol.1, issue.4,
Loop2 76.49 494.8 9912 pp.349-361, 1967.
[12] A.C.M. Ran, R.Vreugdenhil , “Existence and comparison
theorems for algebraic Ricatti equations for continous-
and discrete systems”, Linear Algebra and its
Application,vol.99 , pp. 63-83 ,1988.

39