Implementation of Neural Network Based

Control Scheme on the Benchmark Conical Tank

Level System
Sanjay Bhadra Atanu Panda Parijat Bhowmick
Department of Electrical Department of Electronics and School of Electrical and Electronic
Engineering, Communication Engineering, Engineering
UEM, Kolkata IEM, Kolkata, University of Manchester
[email protected] [email protected] Manchester, M13
, 9PI, UK
[email protected]

Shinjinee Goswami Sayan Podder Debayan Dutta

Department of Electronics and Department of Electrical and Department of Electrical and
Communication Engineering, Netaji Subhas Electronics Engineering Electronics Engineering
Engineering College,
, Kolkata IEM, Kolkata IEM, Kolkata
[email protected] [email protected] [email protected]

Zeet Chowdhury
Department of Electrical and
Electronics Engineering
IEM, Kolkata
[email protected]

Abstract—

This paper proposes an elegant controller is a neuron, that is, a processing element accompanied by a
design technique exploiting Neural Network based linear or nonlinear activation function [2].
adaptive control strategy. In this work, an Extended Kalman filter is extensively used for state and parameter
Kalman Filter algorithm has been applied to tune the estimation of discrete time processes which are described by a
-

controller parameters, which are actually the weights of set of linear stochastic difference equations. While the
the neural network. A case study on the water level extended Kalman filter (EKF), well known for parameter
-

estimation for nonlinear processes, utilizes a state estimation

-

control of a conical tank is considered in this paper to

demonstrate the application and usefulness of the technique which only approximates the optimality of Baye’s
proposed control scheme. From the simulation results, it is rule [2][3] .
observed that the proposed neural network adaptive The primary objective of applying an adaptive control law
control scheme ensures satisfactory servo (i.e., set point is to update the controller parameters on real time basis to
-

tracking) and regulatory performance s of the nonlinear better respond to the variation of the dynamic behaviour of a
process. process due to sudden disturbance or plant parameter variation
or model uncertain ties. In this work , neural network is used as
Keywords —

Conical tank, neural network, parameter a major part of the proposed control scheme Note that the
.

estimation, adaptive neural network control, extended Kalman actual output of the process is compared to the desired
filter. response obtained from a reference model, and based on the
calculated error, the controller par ameters are tuned such that
I. INTROD UCTION the process output reaches the desired response within finite
Artificial Neural Network (ANN) is a prominent area of time even in the presence of external disturbances.
research. Over the past three decades, with the revolution in
computation capability of digital computers, AI techniques
such as ANN, which involve chunks of data and require huge II. P ROCESS D ESCRIPTION
mathematical computation s, have been used to design
.
controllers for complex industrial processes [7] Neural
network based control becomes highly effective for the
The dynamics of the system can be described as
practical systems for which exact analytical plant model is not
available.
A Multilayer Perceptron (MLP) is the simplest and
fundamental form of a feed forward ANN which is primarily
- 2 2
w here A(h) = π h tan θ is the area of the tank at different
used to compute an approximated input output model of a
-

height. F in and Fout are respectively the inlet and outlet flows

dh
complex system from a large number of given input and of the system.
output data sets An MLP is comprised of multiple layers of
.

nodes connected via a directed graph in which all the layers

are well integrated . Except the input nodes, every other node
A (h )  Fin  Fout (1)
978-1-7281-0554-3/19/$31.00©2019 IEEE
dt
 A neural network [8][9][10] consists of three layers – input,
PI
output and hidden. Each input in the input layer is assigned a
Controller
FT weight. The different inputs, times their weights are then
assembled in the first neuron layer and a bias is added. The
Fin
output of the ith neuron in the hidden layer or output layer is
Cv
given as

o  f (Wi * xi  b) (2)
Adaptive neural
network controller LT
where the index i denotes the number of input source, xi
V1 denotes the magnitude of input source and Wi denotes the
Pump
V2 weight. Further, b is a bias term and f() represents the
Fout
activation function. The activation function works as a
squashing function such that the output of a neuron in a neural
network lies within a predefined finite range of values. In the
general form, the equation for a tan-sigmoid function would
Fig 1. Conical tank process appear as
TABLE I ecx  e  cx
PARAMETERS OF CONICAL TANK tanh(cx)  (3)
ecx  e  cx
Parameters Values
where c > 0 is a positive scaling constant.
Maximum tank height (H) 60cm A Multi-layer Perceptron (MLP) network is formed by
Maximum tank diameter (D) 45cm multiple layers of computational units, usually interconnected
Taper angle (θ) 20 in a feed forward way as shown in Figure 2. Each neuron of a
layer has directed path ways to the neurons in the subsequent
III. NEURAL NETWORK AND ITS WORKING PRINCIPLE layer. In practice, sigmoid functions are found to be
A neural network consists of an interconnected group of appropriate candidates for activation function used in such
artificial neurons, and it processes information using a networks [12].
connectionist approach. ANN is an adaptive system that
changes its structure based on external or internal information IV. ADAPTIVE NEURAL NETWORK BASED CONTROL
that flows through the network during the learning phase. An As a conical tank has non-uniformity throughout its
ANN is attributed by the following factors: i) the mode of structure and the process is also nonlinear, an adaptive control
interconnection between different layers of the network, ii) scheme would be more effective than the classical PI
training process for updating the weights of the neurons in controller to meet the satisfactory performance. Further, in
each layer, and iii) the activation function that transforms the contrast to the notion of Robust Control, an adaptive control
weighted input to a neuron into the output [8][9][10]. scheme does not need to know the exact bounds of the
The feed-forward neural (FNN) network emerged as the uncertain/time-varying plant parameters. While the robust
first and elementary type of artificial neural network. In this control asserts that the control law is not required to be
network, the information proceeds in a unidirectional way modified as long as the parameter variation remains within a
from the input to output through the hidden layer. There are specified range, adaptive control makes the control law adapt
no cycles or loops in a FNN [11]. to different operating conditions. Broadly speaking, the aim of
adaptive control is to develop a stabilizing and effective
TABLE II
DESCRIPTION OF THE NEURAL NETWORK controller, the parameters of which is updated subject to the
variation in the plant parameters and/or model uncertainties.
In this scheme, the real-time plant output is compared with the
Type of neural network Feed forward
desired output computed from a reference process model, and
Architecture MLP based on the difference between the actual and desired
outputs, the controller parameters are updated. There are so
Training algorithm EKF
many methods available in literature for training of neural
Number of neurons in input layer 2 network parameter i.e. tuning parameter in our system. In our
Number of neurons in hidden layer 4
proposed work we used EKF for updating the weights of
neural network.
Number of neurons in output layer 1
Activation function (both hidden
Tan-sigmoid
Layer and output layer)
 SP A. Extended Kalman Filter Algorithm and its use for
Reference
mmm Model
learning of parameter:

Prediction step:
1 -
yr
zzz e wˆ k /k  I  w
ˆ k / k 1 (8)
MV + yk
PV
Conical Pk / k 1  Pk 1/ k 1  Q (9)
mmm
Tank PV Measurement covariance:
SetPoint Output T
Sk  Rk  H k * PK / K 1 * H k (10)
layer
Kalman gain:
Hidden T 1
K k  Pk / k 1 * H k * ( S k ) (11)
layer
Kalman Update step:
gain ˆ
w w
ˆ  K * ( yr  yk ) (12)
k /k k / k 1 k
Fig 2. Architecture of feed forward neural network model Pk / k  Pk / k 1  K k * H k * Pk / k 1 (13)
As the EKF theory depends highly on the linearized model,
EXTENDED KALMAN FILTER
V.
the convergence of the filter may be adversely affected if the
In the simplest words, Kalman filter is a type of predictor- initial state estimate is incorrect or the process model is
corrector estimator which is generally applicable to linear inaccurate. Another limitation or drawback of the EKF is that
systems [2][3]. Kalman filter is also considered as an ‘optimal the estimated covariance matrix sometimes underestimates the
filter’ since it minimizes the estimated error covariance. To true covariance matrix which may cause some inconsistency
determine the optimal filter coefficients, the Kalman filter in a statistical sense, particularly, if the ‘stabilizing noise’ is
theory requires a detail information about the probability not added. The EKF can also be perceived as a filter which
distribution of the estimated error. imposes ‘first-order approximation’ to its optimal filter
The mathematics of Kalman filter theory is associated with coefficients. These approximations may produce significant
two sets of equations. The first set represents the time update error in the true posterior mean and the covariance of the
equations which are responsible for projecting the current transformed Gaussian random variable. Because of which, the
states forward in time and produces priori estimates based on EKF may achieve only suboptimal performance and
the error covariance. While the other set contains the sometimes, it may not even converge. The flow chart of EKF
measurement update equations which feed the new measured algorithm is given in Figure 3.
values back into priori estimates to produce more accurate
posteriori estimates. Although, Kalman filtering finds a wide
variety of applications in both engineering and non- VI. SIMULATION RESULTS
engineering problems, it is predominantly used for state and
parameter estimation of linear systems. A. Servo Response of conical Tank System Using ANN
EKF is a nonlinear extension of the basic Kalman filter Controller
which is applicable to nonlinear systems. EKF linearizes a In this control scheme, a sequence of step is given to assess
nonlinear process model around the mean and covariance. The the tracking capability. From Fig 4(a), it can be inferred that
working principle of EKF is guided by a set of non-linear and the NN based adaptive Controller is capable to track the set
stochastic difference equations. The state space model for a point. The controller action was reported in Fig 4(b).
discrete-time stochastic system is of the form
x(k )  f [ x( k -1), u (k -1)]  w(k ) (4) 20
Refernce Model Output
Process Output
where x is the state vector, u is the input vector being known, 18
Setpoint

and w represents the process noise. The output measurement

Level (cm)

16

model is described by 14

y ( k )  h[ x( k )]  v(k ) (5) 12

10
where v stands for the measurement noise.
2000 4000 6000 8000 10000 12000 14000
Time (sec)

The state space representation of a feed-forward Neural Fig 4(a). Servo response using number of hidden layer neuron as four
Network (FNN) is as follows
wk / k  wk / k 1 (6)
u ( k )  NN (Wk , yk , SPk ) (7)
 Start 20

Reference Model Output

18
Process Output
Setpoint

Level (cm)
16
Initialize W(0/0), P(0/0),Q ,R
14

12
One Step ahead estimation of weight vector &
computation of priori covariance matrix 10

0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4
Time (sec) x 10
4
Residual computation using measured output &
one step ahead predicted output Fig 5(a). Regulatory response using number of hidden layer neuron as four

5
Computation of innovation covariance matrix & Manipulated Variable
cross covariance between weight and NN output 4

Flow (cm3/sec)
Computation of Kalman gain 3

Update the weight vector and covariance matrix 2

Stop 1
0.5 1 1.5 2 2.5
Time (sec) x 10
4

Fig 3. Flow chart of EKF Fig 5(b). Controller output at regulatory response for same number of hidden
layer neuron
5
Manipulated Variable
4.5
0.28
4
0.26
Flow (cm3/sec)

3.5
Weights form Hidden Layer to Output

3 0.24
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
2.5 1 x 10
4

2 0.5
1.5
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
1 0.65 4
2000 4000 6000 8000 10000 12000 14000 x 10
Time (sec)
0.6
Fig 4(b). Controller output at servo response for same no. of hidden layer
0.55
neuron 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0.6 x 10
4

0.55

B. Regulatory Response of conical Tank System Using 0.5

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
ANN Controller Times (sec) x 10
4

Fig 5(c). Neural network weights are updating at same regulatory response
In this work, disturbances are introduced at different stage
to assess the tracking performance of the proposed control CONCLUSION VII.
scheme. From Fig 5(a), it can be inferred that the NN based From the simulation study, it can be concluded that the
adaptive controller is able to track the set point in presence of neural network based adaptive control scheme designed for
external disturbance. The control action is shown in Fig 5(b). the water level control of a conical tank process achieves
Fig 5(c) shows evolution of weights with respect to time satisfactory performance. The future work extends validation
when a disturbance is introduced. of this control scheme on the hardware set up.

REFERENCES
[1] Stephen C. Stubberud and Mark Owen, “Artificial Neural Network
Feedback Loop with Online Training”, IEEE International Symposium
on Intelligent Control,Vol. 25, pp. 1281–1288, 1996.
[2] Daniel Oliveira Cajueiro and Elder Moreira Hemerly, “A Chemical
Reactor Benchmark for Parallel Adaptive Control Using Feed forward
 Neural Networks”, IEEE Transactions on Neural Networks, Vol. 29 ,pp.
335-348, 2000.
[3] P. Gilts, J. Henriques, P. Canalhot, H. Duarte-RamosS and A. Dourado,
“Adaptive Neural Model-Based Predictive Control of a Solar Power
Plant”, IEEE Control Systems Magazine, Volume 12, pp. 227-238,
2002.
[4] Shengquan Li, Juan Li, Jinhao Qiu, Hongli Ji and KongJun Zhu,
“Control design for arbitrary complex nonlinear discrete-time systems
based on direct NNMRAC strategy”, Journal of Process Control, Vol.
30, pp. 44–48, 2010.
[5] Vincent A. Akpan and George Hassapis, “Adaptive Predictive Control
Using Recurrent Neural Network Identification”, IEEE Trans. Syst.,
Man, Cybern.,Vol. 15, pp. 1281–1288, 2009.
[6] H. D. Patiño and Derong Liu, “Neural Network-Based Model Reference
Adaptive Control System,”IEEE Control. System Magazine, Vol. 30,
pp. 44–48, 2000.
[7] C.W. Dawson1 and R.L. Wilby2, “Hydrological modelling using
artificial neural networks”. Progress in Physical Geography 25,1 (2001)
pp. 80–108
[8] Kröse B.; Smagt P. (1996). “An Introduction to Neural Networks”, The
University of Amsterdam, Amsterdam.
[9] Gurney, K. (1997). “An Introduction to Neural Networks”, Routledge,
ISBN 1-85728-673-1, London.
[10] Rojas R. (1996). “Neural Networks: A Systematic Introduction”,
Springer, ISBN 3-540-60505-3, Germany.
[11] Lawrence K.Saul, Michael I. Jordan, “Attractor Dynamics in
Feedforward Neural Networks”, Neural Computation 12, 1313–1335
(2000) , @ 2000 Massachusetts Institute of Technology.
[12] Atanu Panda and Rames C. Panda, “Adaptive Non-linear Model Based
Control Scheme Implemented on the Non-linear Processes”, Nonlinear
Dynamics, vol. 91, no. 4, pp. 2735-2753.
13] Sreepradha Sivaramkumar, Rames C. Panda, bhuvaneswari N. S., Atanu
Panda, “Operational control of an integrated drum boiler of a coal fired
thermal power plant”, Energy, vol. 159, no. 9, pp. 977-987.