2012 IEEE International Conference on Advanced Communication Control and Computing Technologies (ICACCCT


Centralised Neuro Controller for Two Conical Tank Interacting Level System
I 2 3
Y.R.Ravi, T.Thyagarajan, B.Puviyarasi
Oept.of Electronics&Instru.Engg.,YelammalEngineeringCollege,Chennai.
0ept. ofInstru. Engineering, M.LT Campus, Anna University, Chennai.
0ept. of Electronics & Instru. Engg.,Yelammal Engineering College,Chennai.

Abstract- The implementation of control algorithms for MIMO mapping of the input-output data set. This mapping is
systems is often complicated due to variations in process conformal in nature. Hence the neural network behaves like a
dynamics that occur because of changes in operating point and
"BLACK BOX" which can be used to approximate any non­
the characteristics of non linear dynamic coupling. Such changes
linear, varying process parameters with operating points, like
often render the performance of existing decoupled based
decentralized fixed gain PID and Gain scheduling PID
controllers unsatisfactory. The dynamic uncertainty associated
Madhubala et al. [1] have discussed the performance of
with MIMO systems make existing model based decoupling the genetic algorithm based Fuzzy controller for a single
impractical for real time system. This work presents control of conical tank level process. They examined that the fuzzy
non linear Two Conical Tank Interacting Level System based control is better for compensating the set point and load
(TCTILS) using centralised Neuro controller. The TCTILS is changes than the PI controller. N. S. Bhuvaneswari et al. [2-
considered as two tank benchmark problem used by many 3] proposed a Neuro based Model reference Adaptive Control
researchers. Simulation results show that the centralised Neuro and Optimal Control for single conical tank level process. In
controller realise a good dynamic behaviour of the TCTILS, a
the above works [1]-[3], single conical tank level process is
perfect level tracking with lesser overshoot, lesser settling time,
considered which is a SISO non linear process. V.R.Ravi et
reduced interaction and good rejection of external load
al. developed Gain scheduling Adaptive controller [4] and
decoupled based Decentralized PID controller [5] for
Keywords- Centralised control, Neuro Controller, Conical TCTILS process. Gu Fang et al. [6] proposed time optimal
Tank Interacting Level system feedback control of a non-holonomic vehicle using neural
networks. For this work, the data have been taken as many
I. INTRODUCTION optimal trajectories from the process and trained using neural
networks. If the process parameters vary, the network has to
Conical tanks find wide applications in process industries,
be retrained or otherwise steady state error will occur.
namely hydrometallurgical industries, food processing
In this work,data are collected from an existing industrial
industries, concrete mixing industries, sewage water
treatment plants and wastewater treatment plants. Their shape standard technique such as decoupled based decentralized
contributes to better drainage of solid mixtures, slurries and PID controller and centralised Neuro controller is designed
viscous liquids. In this work, Two Conical Tank Interacting and trained using neural network toolbox of MATLAB.
This paper is organized as follows: In Section II, the
Level System (TCTILS) is taken up for study which is a non­
process of TCTILS is described. In Section III, design of
linear TITO system. Control of TCTILS presents a
Neuro controller is discussed. In Section IV, qualitative and
challenging problem due to its non-linearity and dynamic
quantitative results are discussed. Finally conclusions are
interaction. As the operating point changes, the parameters of
presented in Section V.
TCTILS change due to constantly changing cross sectional
Most of industrial processes are basically MIMO systems. II. PROCESS DESCRIPTION
For such systems, dynamic loop interactions can arise and The Two Conical Tank Interacting Level System
cause difficulties in feedback controller design. MIMO (TCTILS) as shown in Fig. 1 is based on the two-tank
systems possess complicated dynamic coupling. Establishing benchmark problem that has been used by a number of
the accurate dynamic model and decoupling it for designing researchers. The TCTILS is regarded as a setup for
the decoupled based decentralized controller is difficult. investigating, theoretically and experimentally, nonlinear
Also, in the industrial fields, noises and disturbances widely multivariable feedback control schemes. TCTILS consists of
exist. Hence, it is difficult to get accurate mathematical two identical conical tanks (TANKI and TANK2), two
model of the process and decoupler. The traditional independent pumps (PUMPI and PUMP2) that deliver the
decoupled based SISO system control scheme is hard to liquid flows FINI and FINZ to TANKI and TANK2 through the
implement on complicated MIMO systems because of the two control valves CVI and CVz respectively. These two
cumbersome and time consuming design procedure and the tanks are interconnected at the bottom through a manually
larger computational burden. Therefore, a model-free controlled valve, MVI 2 with a valve co-efficient PI 2. FOUTI
intelligent control strategy such as centralized Neuro and FOUTZ are the two output flows from TANKI and TANK2
controller is useful to solve these problems for a process like through manual control valves MYI and MY2 with valve
TCTILS. These problems can be solved by employing coefficients PI and pz respectively. The P12,PI and pz are
universal approximations like neural networks which perform

ISBN No. 978-1-4673-2048-1112/$31.00©20 12 IEE E 343

.l-3h. Schematic diagram of TCTlLS TABLE I OPERA11NO PARAMETERS OF TCTlLs .!:L2 forward neural network. Block diagram of a centralised Neuro Controller for TCTlLS Fig I.· R Top radius of conical tank 19. the input variables of the factors which will improve the speed of the training.. which is assumed as unity. Valve co-efficient of MV.vh. (i = 1. 5 0. is the discharge coefficient. -h. + slgn(h. (2) The outputs of the centralised Neuro controller are FINI and dh .h .. NI to N9 are used as normalization knowledge about the TCTILS.SP2(t) (10) represented in (2) and (3). (4) 344 . Based on prior input .t . The operating parameters of TCTILS are shown in Table 1. de IIloden .· Ou. 12) are the cross sectional area of valves and a.5) 2 Parameter Description Value Input .2. The controlling TCTILS is shown in Fig. D:... 3. In TCTILS. eI (t) = (5) dt may be evaluated using (1)..-. In order centralised Neuro controller are defined individually as to denormalize.W12 vlh. . �I I dA(h'l F'N' -�'vh . the iez(t) = f ez(t) dt (9) [ ] mathematical model for TCTILS process is obtained and SPI2(t) = SPI(t) . level hI in TANKI and level hz in TANK2 are h.output data pair.. iel(t) = f el(t) dt (6) Pi =Siaifii (i=I. FIN' - is identified and is shown in the Fig.00 em' / s dez(t) 5 0.-h. -h 'I-. the optimum neuro IIl:l:i I dA (II. -. Maximum Inflow to TANKI & TANK2 5 00 LPH ez (t) SP2(t) .00 em' / s ez(t) (8) �. It is a 10-8-2 feed dl = �7rR.###BOT_TEXT###quot;2 Fig 2.l�l2vlh ..00 em' / s �12 Valve co-efficient of MV12 35 . 2012 IEEE International Conference on Advanced Communication Control and Computing Technologies (ICACCCT) adjustable coefficients representing the resistance of the del (t) respective valve's opening orifice. ) [ tJ.� (3) network structure for designing centralised Neuro controller dh.I2) (1) Where S. CENTRALISED NEURO CONTROLLER FOR TCTILS manipulated variables FINI and FINz..1 IIldele . It is to be noted that the centralised '" 3 H Neuro Controller is designed in such a way that it can be used to control both hI and hz simultaneously by adjusting the III. 2. Valve co-efficient of MV.p. TCTlLS It. The centralised Neuro input and output data are collected from decoupled based controller consists of multi layered feed forward neural decentralised PID controller [5].. 3 dt FINZ which are the manipulated variables for controlling ] Tt= � 7tR ' �' TCTILS. oun h. In order to normalize the network with nine inputs and two outputs. The valve coefficient p. DNI and DN2 are used as denormalization factors. OUT' h. & FIN.ye..00 em FIN . � . dt Using Mass balance equations around both tanks. 3 H By using trial and error method. The parameters for The block diagram of a centralised Neuro controller for training the neural network are reported in TABLE II. Neural network structure of centralised Neuro controller H Maximum height of TANK I & TANK2 73. -h.uyer. considered as output variables and FINI and FINZ are D::"l considered as respective manipulated variables. -slgn(h .hz(t) (7) �...uyer (10 neurons) (8 neurons) (Z "-'. 2..25 em Fig 3.

. 0.000 data pair used for validation IAE 118. I I " . The learning function used for training and settles to the setpoint faster.. the centralised Neuro controller is able to .1 389.s I.... .regulatory performance of centralised Neuro controller and 345 .. The performances of in the toolbox. less interaction than other algorithms. is LEARNGDM [7] which is the gradient descent with momentum weight and bias learning function. TRAlNLM [7] is a neural network training The observations (qualitative) of the above simulation function which is used in training.ie"e. Servo performance centralised Neuro controller at nominal and at shifted To demonstrate the tracking capability of the centralised operating points.. .. No of input-output 40. .SPI .iel. el . .. 15 . a step disturbance of magnitude 25 LPH is Neuro controller for TCTILS. ..2 207.... l'. ....... �300 ..5 ITAE 87760 168900 IV. . nonlinear differential equations using differential equation B. 2012 IEEE International Conference on Advanced Communication Control and Computing Technologies (ICACCCT) The centralised Neuro controller is trained using neural maintain the tank levels hI and hz at the respective setpoints.. ... FIN I . The output variables are computed by solving the Controller at nominal and at shifted operating points.. e"SP2.. servo with B.01 NEURO CONTROLLER FOR TCTlLS Learning rate 0.:. 1 Performance study for step disturbance in FIN! solver in Matlab R2009a........ .. The number of training epochs is outputs are presented in Fig. ... The servo be inferred that.. 4(a) are introduced.... The process is Simulation studies have been carried out to demonstrate simulated using the nonlinear first principle model «2) and the disturbance rejection capability of the centralised Neuro (3)). 20 SP2 TABLE [[ NEURAL NETWORK PARAMETERS USED FOR TRAINING -h. ..200 . although it does require more memory found to be better.. From the response.. Servo . ..:. .. 80% data are used for training purpose and 20 % . network toolbox of MATLAB R2009a. as there is less overshoot. FINl 1. . fixed at 1000... -112 Parameters Name / type / values o 200 400 600 800 1000 1200 1400 1600 Network type Feed forward back propagation TillIe (second) No of layers Hidden Layer-2.. To demonstrate the disturbance rejection capability of the A...1 No of input-output 10... ...... ... 1. it can then removed at simulation time of 1000tll second... .SPI.----\ No of neurons in 2 output layer Inputs Outputs el. 4(b).. The stopping The servo performance indices of the neuro controller are criterion used in training is 0. Servo response of TCTlLS with centralized Neuro Controller Performance Mean Square Error (a) Process Level outputs (b) Flow control signals Goal 0.regulatory performance regulatory performance and robustness of centralised Neuro controller are carried out on the TCTILS. . that updates weight and study are as follows: bias values according to Levenberg-Marquardt optimization.1 Performance Indices hi h.. process. . Training function TRAlNLM Adaptation learning functio LEARNGDM Activation function Hidden layers: Tan Sigmoid (b) Output layer: Purelin No of epochs 1000 Fig 4.. I. The variations in the controller and the learning rate is 0. . L data for validation purpose. ..25 TABLE [[[ SERVO PERFORMANCE INDICES OF CENTRALISED Gradient 0.. SIMULATlON STUDY Simulation studies such as servo performance. Output layer-I No of neurons in 10 (a) first hidden layer No of neurons in second hidden layer 8 400 .... . and is highly recommended as a first-choice centralized Neuro controller at all the operating points are supervised algorithm.SPI2. . 25 Mean Square Error (MSE ) reported in TABLE III. . In this training 30 c··········· . the setpoint variations as introduced in FINI at simulation time of 400tll second and is shown in Fig.. Centralised Neuro controller is able to maintain the hI and TRAINLM is often the fastest backpropagation algorithm [7] hz at the respective setpoints.000 data pair used for training ISE 292..

-------.. --. ---- 28 ---.::a � V §2G J--i--------f'�--�----�-- -..--. �_ . 30 .--- � Distu.. - 30 ---. ---. 1.1. of TCTILS with centralized Neuro controller for step disturbance in F1NZ is presented in Fig.---�P. second. The servo .18 .. .. 2012 IEEE International Conference on Advanced Communication Control and Computing Technologies (ICACCCT) its corresponding variations in the control signal are shown in Fig.--------------.. .-----. ' ..ielll l"I='--'=""i_===_'_i='_'_l 100 . if\ . 00 ::.regulatory response the respective set points. ..L _DI�:�. .--. ----.----------.100 500 GOO 700 800 900 1000110 Tilne (second) Ti ..------.=:::1-- 600 700 800 Til1l:ae (second) 900 1000 1100 .a step change in centralized Neuro controller is reported in TABLE V. ----..-bal ce i5 200 . The following observations are drawn from simulation studies: t • In the simulation period from 400 ll second to 4501h B. - -"22 � applie I -------- f. . The the set points hI and hz are introduced at simulation time of observations drawn from this simulation study are similar to 5001h second and 750tl1 second respectively..��:�c�� ---- �2"'" te:'" �·(siu·I�DQ·li . r--.------.4 463 the controller is satisfactory. 400 .PP!!�!!-.----.500 -.1000 . The servo performance indices are computed and reported in TABLE IV.. .----..----. ------------. 5(a). ----...4 ITAE 91130 161600 346 .-... The robustness performance of the centralized Neuro ISE 309. ----.. . 5. .-----. --'--C --==-'l � 3 00 -.-----------.r= ----'-'.-Dbt'tirban-ce--- ::1 = :::::: I:::: :::::::: 0 .---cc--+----'-'. ..5 471. -... ..regulatory response of TCTlLS with centralized Neuro (b) controller in the presence of setpoint change and load change in FIN2 Fig 5.regulatory response of TCTlLS with centralized Neuro (a) Process Level outputs (b) Flow control signals controller in the presence of setpoint change and load change in FINI (a) Process Level outputs (b) Flow control signals TABLE V SERVO -REGULATORY PERFORMANCE INDICES FOR CENTRALISED NEURO CONTROLLER IN THE PRESENCE OF • At simulation time of 10001h second.1 -' -----------.---'--C----'-'.if--l. C.:-----....-. .:.--SP2 ---.-------------. "00 .-.------ If ------------../-------- �2" ---.--.-RPp-l. 20 ----. FINI at simulation time of 450tl1 second) are removed simultaneously and it can be inferred that the performance of ISE 5 5 8. I-ClllOye -I noo 110C: (b) Fig 6_ Servo .-----. The change in PI Z physically represents the change in interaction between two Performance Indices hI h2 tanks.. This clearly demonstrates that the 400tll second and is maintained upto simulation time of 999tl1 controller is able to reject the disturbance at the nominal second.-.. 5(a)) as well as the step disturbance of 25 LPH (which was introduced in the input flow to TANKl Performance Indices hI h2 i. ---. --- I( : . step change in SETPOINT CHANGE AND LOAD CHANGE IN FIN2 the set point hI (refer Fig.-----..e. _ �..- ---. The servo with regulatory performance indices for • With the disturbance being persistent.hr----�----�--v_�-----+----� � 1' :' 32 28 . 2 Performance study for step disturbance in FIN2 second of Fig. ..: -11.---- . t--..' .-------------... 6.-.REGULATORY PERFORMANCE INDICES FOR Neuro controller is tested by varying the valve coefficient PI Z CENTRALISED NEURO CONTROLLER IN THE PRESENCE OF SETPOINT CHANGE AND LOAD CHANGE IN F INI of MVI Z from its nominal value (35cm%). This part of the simulation IAE 306_7 232_9 demonstrates that the centralized Neuro controller is able to ITAE 15 7900 192300 reject the disturbance quicker as well as maintain the process variables at the respective setpoints..�---+---4---f.----.--.__ f--- 20 . �--.. Later it is removed at simulation time of 10001h operating point.00 600 700 800 900 . '- ----+:O�-+----___i --------- SP2 i?26r+-----i�--� ----��_t ..___ L===...'--"2'----'-.-_- -'-S- P '..1 .... r ------"-. it can be inferred that the centralized Neuro Controller is able to reject the disturbance quickly and Similarly a step change in the input flow to TANK2 (FINz) bring the levels hI and hz back to the nominal value of the of magnitude 25 LPH is introduced at simulation time of respective setpoints. Servo....-------'I=='------+--------f'�--___l ..----.-----. The robustness of centralized TABLE IV SERVO . Robustness performance Simulation studies are carried out to demonstrate the robustness of the centralized Neuro controller at nominal and at shifted TCTILS parameter. --------------...1 ..-----. --..::.--------. :------1----. e (second) (a) (a) 500 ! 500 "---- .. B.---j" ----'-'. It can be noted that of the performance study for step disturbance in FINI as that the controller is able to maintain the levels hI and hz at discussed in section IV.1 IAE 122..8 198.-.h J r···· r=== � 22 Distul!ba �ce_.����::-� -�------ ROO 90n .

S.S..." Applied soft computing.. the authors have suggested a simple and 347 ... V..a TCTILS was of centralized Neuro controller are computed and reported in designed and its mathematical model was identified for TABLE VI. In order to evaluate the servo control and 50 % changes in PIZ are shown in the Fig. 36.­ ······-sr2 F. 7. 2011.. .. Proceeding of IEEE 2011 3'd (s. Vol.. Further. 1-9 == •...R..5 167.. -I Optimal Control of a Non-Linear process using Intelligent � ------. INDIA for the project titled "Implementation Of Advance Controllers For Interacting Non Linear System" vide File no : sr. "Neuro based �:::::::� �� Model Reference Adaptive control of a conical tank level process.1.Ravi acknowledges the financial support of AICTE. Mathworks .1 87600 388. 2004. " controller is capable of controlling TCTILS in entire ICARCV -2002 .2 200. Uma & T.. it is concluded that the centralised Neuro controller is a robust controller. ..== :::=�.. pp. .K. "Development and tuning of filzzy controller for a conical level system.. Rangaswamy. Tiruchirappalli. disturbance [7] Neural Network Toolbox User guide R2012a. 30% dynamic interaction.8 298700 ACKNOWLEDGEMENT 50% 490..2002 operating region with better tracking capability.. an optimum neural network was is increased up to 50% from its nominal value. servo with regulatory control performance and robustness performance indices such as ISE . 40% 378... 9.. From the Fig.. The performance.. Control. Thyagarajan T.7 197. � International conference on Electronics Computer Technology (ICECT 2011)... MATLAB R2009a to use it as centralised Neuro controller for controlling TCTILS... Rangaswamy.·.. it is concluded that the centralised Neuro feedback control of a non-holonomic vehicle using neural networks. ... CONCLUSION In this paper... 0 [3] N. "Time-optimal performance... vol. From the performance indices values. 386-392 (b) [5 ] Ravi V R.. c . IH. 297-302 From the servo performance and servo with regulatory [6] Gu Fang..5 314500 V.339/ 2007-2008. ------ . " 40 U -······ � . 1. Proceeding of IEEE 2011 International (a) Process Level outputs (b) Flow control signals Conference on Emerging trends in Electrical and Computer Technology. which exhibits significant nonlinear dynamics and corresponding variations in the control signal for 0 %.. 2008.. This proves identified and trained using neural network toolbox of that centralized Neuro controller is a robust controller. .. Boopathy.3 289. [2] N... 7. ." Proceedings of International Conference Intelligent Sensing and Information Processing.../o I i Controllers.4 165300 centralised Neuro controller at all the operating points are found to be good..R.1 35 7.2 �OO/'" [I] Madhubala. pp. ...3 175800 5 29. -h.. it is concluded that the numerical study..K. .. less interaction and settles to 30% 343... 1 2009...7 240. A De-centralised PID Controller for Fig 7... From the robustness performance.5 132000 487.. pp. .. Bhuvaneswari. M.cond) 26� Ti .0 216..0 274. Radhakrishnan.6 180100 response with less overshoot. Control and Intelligent Systems. No... pp.9 106000 463. The numerical simulation results Increase hI h2 show that the centralised Neuro controller is able to control of TCTILS effectively. 45 0-45 5 . Application of Adaptive Control technique for Interacting Non linear Systems. 2011..��' �§���� (a) India.. -200------- 1---- FIN2- �O . -----.l-�C:.. Thyagarajan T. - ....OA.• 2 O'Yo REFERENCES -h2-3c:. Sarat Chandra Babu and T.4 201..R..2 176700 centralised Neuro controller scheme helps to produce 20% 326..... Secondly..3 136..5 223100 605. J..2 225200 the setpoint faster in the entire operating region. No.4 116.2012 rejection capability and lesser interaction effect. pp. Bhuvaneswari. it Parameter ISE IAE ITAE ISE IAE ITAE is observed that the robustness and control performance of /l1'=35 0% 291.\. it can be concluded that the 10% 320. G.. Uma & T.. Robotics and Vision. 2012 IEEE International Conference on Advanced Communication Control and Computing Technologies (ICACCCT) controller is tested by increasing the PI2 value from 0% (no straightforward procedure for designing centralized Neuro change in PIZ) to 50% in steps of 10% change. -F1N2- 3 0 oA.. Automation..6 290. input . . Firstly. The robustness controller for the Two Conical Tank Interacting Level System performance of centralized Neuro controller and their (TCTILS).. "Adaptive and � 3 U- -FLN2-O"/.. G.:/. N : 8023/ RID / BOR/ MOD ..output data are collected centralized Neuro controller is able to produce good servo from decoupled based decentralised PID controller with tracking performance even though the process parameter PIZ TCTILS.. ...4 293900 65 1....182-190 [4] Ravi V R. T.. IAE and ITAE robustness of the centralized Neuro controller."'A. Robustness response of TCTlLS with centralised Neuro controller Interacting Non linear Systems.