Escolar Documentos
Profissional Documentos
Cultura Documentos
Neural Networks for Process Monitoring, Control and Fault Detection: Application to
Tennessee Eastman Plant
Center for Process Simulation, Department of Chemical Engineering, Faculty of Chemical & Natural Resources Engineering,
Universiti Teknologi Malaysia, 80990 Johor Bahru, Malaysia
Abstract
This paper discusses the application of artificial neural networks in the area of process monitoring, process control and
fault detection. Since chemical process plants are getting more complex and complicated, the need of schemes that can improve
process operations is highly demanded. Artificial neural network can provide a generic, non-linear solution, and dynamic
relationship between cause and effect variables for complex and non-linear processes. This paper will describe the application
of neural network for monitoring reactor temperature, estimation and inferential control of a fatty acid composition in a palm
oil fractionation process and detection of reactor sensor failures in the Tennessee Eastman Plant (TEP). The potential for the
application of neural network technology in the process industries is great. Its ability to capture and model process dynamics
and severe process non-linearities makes it powerful tools for process monitoring, control and fault detection.
Keywords: Artificial neural networks; Process monitoring; Process control; Fault detection
1
y= , (1) For a given external input X to the neural network, the
1 + e!A network described above will return an external output
where value Y:
n )
Y = f ( X , N t , wi , i =1, N t ) , (3)
A = the element activation = ! wi xi ,
i =1 where
wi = a weighting term. Ŷ = is the network predicted of Y.
x2
Output layer
Input
. .
x1 w1
Neuron . . Output, y
. .
x2 w2 x(n-1)
. . Output
. . y = 1/(1 + exp(-A)) x(n) Neuron
. wn
xn Adjustable weights, w(i)
Fig. 1: An individual neural network processing unit Fig. 2: An example of a three-layer backpropagation neural
(neuron). network. The input vector is of dimension n and the
output vector is of dimension 1.
The purpose for using the neural network is to obtain a
mapping from a vector X to a vector Y. The size of the The input is the vector being mapped from and the target is
input and output layers are fixed by the number of the vector being mapped to. The output of the network is a
components of X and Y, respectively. The user specifies the predicted output Ŷ. The optimization is set up to minimize
number of neurons Nh in the hidden layer. For a network the difference between the predicted values of Ŷ and the
with a single hidden layer, and a user-specified hidden observed values Y. The finding of the set of weights which
layer size, the number of weighting terms in the network minimize the error between Y and Ŷ is called the training of
is: the network.
ut "1 , ut "2 , ut "3 ,...., et , ! t , t ) , (4) The process system that exhibits non-linear
characteristics can be successfully monitored through the
where
utilization of neural networks. Data usually taken from
f = specified functional relationship,
sensors are used as inputs for the network and outputs are
yt = process output at time t,
the systems that need to be monitored such as temperature,
ut = measured process input at time t,
pressure and product quality. Based on the monitoring, the
et = unmeasured process input at time t, decision is usually made whether the process should
t = discrete time index (integer values only), and continue on, or a control action should be taken to make
θt = set of parameter values at time t.
correction, or faulty condition can be detected. In other
words, process monitoring is the starting step before any
The inputs and outputs can be single (scalar) or
process control or fault detection action is taken.
multivariable (vector) values.
To monitor temperature of a reactor, for example, sets
The identification process for a given set of input-
of data that have an effect to the reactor temperature are
output can be described in three steps:
needed. These data are the inputs for the neural network
and the output is the temperature that needs to be
1. Postulate a structure to the model ( f ).
monitored. The input data contain data of reactor pressure,
2. Calculate the best estimate of θt.
reactor temperature and cooling water temperature with 1
3. Validate the resulting model.
and 2 delayed terms. Elman network is used in this study
with 3 sets of different data needed for training, validation
The identification process is often iterative in nature. For and testing. Fig. 3 shows the predicted of reactor
example, the results of the validation step may suggest a
temperature compared to the actual temperature that
better structure f for describing the process.
monitored by Elman network. The result shows that Elman
Dynamic model identification work performed for
network predict the reactor temperature successfully.
identification is almost always done assuming a simpler )
d
model structure than equation (4). It is usually assumed e 1 .5
output of a neuron in the output layer is equal to zero, the generated when the classifier’s output reached the output
fault represented by that particular neuron is judged to be index 0.8. 0.8 is an assigned value for the residual reactor
absent. A fault-free state is indicated when the values of all pressure. When the residual reactor pressure beyond the
output neurons are equal to zero, and multiple faults are assigned value, it indicated that the actual reactor pressure
indicated when the values of multiple output neurons are has deviated from its normal operating condition. Fig. 8
equal to one. and 9 show fault detection for reactor pressure sensor fault
In this paper, detection of reactor sensor failures in the (F1) and reactor cooling water temperature sensor fault
Tennessee Eastman Plant (TEP) has been studied. Fig. 7 (F3), respectively.
shows the schematic diagram of the process. 1
Fault Classifier
0 .5
1
F
F
I9 0
CW 1 Purg 0 100 200 300 400 500 600 700
T S Compre 0
X
e 1
F I CW ssorLI PI A X
A
1 N B
D I 8 R X
A C
X 2 0 .5
F S F
I
2
C TI Vap/li L X
D
E Y E
LI PI
PI
q X
Z F 0
F F X 0 100 200 300 400 500 600 700
A I C 5
I
Separ E X
G
3 W 1
A 6 1 ator R H
X S
N TI 1
X
A
B A F Strip TI A 3 0 .5
X I CW N X
X
C
L
R per D
X
F
D
Y Rea TI F A
X Z I Stea L E
X
E
X ctor LI m
Conden Y X
F
0
0 100 200 300 400 500 600 700
E
F sate Z X
G
FR F
1 E H
Data Set
C I 4 I Prod
3 R
uct
Fig. 8: Fault detection for reactor pressure sensor fault (F1)
Fig. 7: Schematic diagram of Tennessee Eastman Plant 1
Fault Classifier
(TEP).
0 .5
1
This study focused on the malfunctions of the process F
1
of 4.0% or greater from the normal condition is assumed to
cause malfunction to the process. Similarly, for 3 0 .5
For fault detection scheme, two types of networks are Data Set
References