Session 2793

A Neural Network Lab Experiment

Robert Lynn Mueller

The Pennsylvania State University
New Kensington Campus

Abstract

Neural networks are becoming widely used in complex control problems. Many academic exercises
approach neural network applications using only software simulations; however, simulations alone
do not give students a full appreciation of the power and complexity of neural network-based
controls. This paper describes a laboratory experiment that uses a temperature and airflow process
simulator to demonstrate neural network control applications. The simulator is fundamentally a
temperature controller in which large-scale changes in forced airflow produce significant changes in
heat load.

The initial labs use PID control techniques to solve the temperature control problem and to
demonstrate the problem that PID controllers have with large disturbances. The following labs
address the same problem using a neural network control strategy. An actual neural network
controller is built and used to perform the same temperature control as the classical PID system.
Capabilities and drawbacks of neural network control are demonstrated.

Introduction

The classical PID feedback control system is shown in Figure 1. The setpoint (SP) is the system’s
input and the process variable (PV) is the output. The PID controller uses the error signal (SP minus
the PV) to calculate a value for the manipulated variable (MV). In turn, the value of the MV then
determines the value of the process variable. Unexpected changes to the process output can be
caused by system disturbances, and it is the PID controller’s job to adjust the MV to account for
these changes.

Based on the characteristics of the process, the PID controller is tuned to achieve the desired system
response. Unfortunately, when the disturbance input varies over a large range, the PID controller’s
ability to achieve the desired system response is greatly reduced. For these situations, a more
advanced control scheme is necessary. One possible method to overcome large disturbance
variations is a neural network (NN)-based control strategy.
Page 6.65.1
 disturbance
manipulated process
setpoint variable variable
(SP) error (MV) (PV)
controller process

Figure 1. Classical Negative Feedback Control System

Neural Network Overview

A significant amount of literature exists regarding the operation of a NN. A detailed description of
NN systems is given in [1]. A brief overview is given here to define terms.

A NN is used to predict one or more outcomes based on the NN’s inputs. The basic element of a NN
is the neuron as shown in Figure 2. The neuron has 1 or more inputs but produces only 1 output.
The output is found by first calculating the weighted-sum of the inputs and then “squashing” that
sum according to some non-linear squashing function. One popular squashing function is the
sigmoid function:

1
F(x) = ----------------------
1+ exp (-sum)

The weights that are used in the sum are “learned” by the NN based on training data that it acquires
both during an initial training period and after the NN goes on-line. Page 6.65.2

“Proceedings of the 2001 American Society for Engineering Education Annual Conference & Exposition
Copyright  2001, American Society for Engineering Education”
 x1

F(x)= 1
-y
1+e
w1 where y = [w] • [x]

x2 w2 F(x)

[x] = input vector

w3

[w] = weight vector

x3

Figure 2. Typical Individual Neuron

A neural network (NN) is formed by connecting the neurons together. A typical configuration is
shown in Figure 3. A NN has an input layer, one or more hidden layers, and an output layer. The
number of neurons in the input layer corresponds to the number of NN inputs and the number of
neurons in the output corresponds to the number of outputs to be predicted. Squashing does not take
place in input neurons; the output is simply the input. The number of hidden layers and the number
of neurons in each hidden layer depend on the application; in general, the NN’s output prediction
improves as the number of hidden layers and hidden layer neurons are increased.

Once the architecture of the NN is determined, the initial neuron weights are randomly chosen. The
NN is then trained by presenting it with many sets of training data where a set consists of a value for
each input and the corresponding desired output(s). From this training data, the NN calculates
(learns) the input weights for each neuron so that the NN output closely matches the corresponding
input for each point in the training data. This learning process is an iterative back propagation
algorithm. The learning process is terminated when the NN’s predicted output(s) matches the desired
output(s) for each set in the training data to within some criteria. One such criterion is the sum-of-
the-errors-squared.

The more sets in the training data, the better the NN is able to predict the output(s). A rule of thumb
is that the number of sets should be at least 10 times the number of weights.
Page 6.65.3

“Proceedings of the 2001 American Society for Engineering Education Annual Conference & Exposition
Copyright  2001, American Society for Engineering Education”
 F(x)

NN
input # 1

F(x) F(x) NN
output # 1

NN
input # 2

F(x) F(x)
NN
output # 2

NN
input # 3

F(x)

input hidden output

layer layer layer

Figure 3. Typical Neural Network

Application of a Neural Network to Process Control

One possible application of a NN to process control is shown in Figure 4. This application is used
for this lab experiment. While this is not the best application in terms of response time, it is very
stable. The inputs to the NN are the MV and the process disturbance. Based on these inputs, the NN
predicts the value of the PV. This predicted PV is then used for feedback in place of the actual PV to
determine the error signal and thus the MV setting. A PI controller is needed to remove the steady
state error since the NN is essentially a type 0 system.

Training data is obtained from the process every time the process reaches a steady state operating
point. For training purposes, the NN compares its predicted value for the PV with the actual value of
the PV and then adjusts the weights using the back-propagation algorithm.

Note that the system trains itself. For a given disturbance input and MV input, a prediction for the
PV is made. Even if a bad prediction is made, a new steady state operating point is reached and a
new training set is loaded into the training data.
Page 6.65.4

“Proceedings of the 2001 American Society for Engineering Education Annual Conference & Exposition
Copyright  2001, American Society for Engineering Education”
 disturbance

SP MV process PV actual Neural

PI controller
Network

predicted

Figure 4. Block Diagram of a Possible NN Control System

Lab Equipment Descriptions

The simulation hardware is shown in Figure 5. It is based on temperature and flow simulators
designed by Jim Rehg [2]. The temperature simulator consists of an enclosed box with a 23-Watt
soldering iron. Varying the duty cycle of the iron’s 120-Volt ac power input controls the soldering
iron’s heat output. This is accomplished by connecting the 4-20 mA control signal to a pulse control
module (PCM) that is connected to a solid-state relay. The relay connects the 120 volts to the iron
and therefore, the larger the control signal, the longer the 120 volts is applied to the iron. A type-J
thermocouple connected directly to the tip of the iron measures the temperature. The box has a
squirrel cage fan connected to one end and is open on the other end. A damper mounted in the inlet
tube controls the airflow through the box, and a muffin fan connected at the end of the tube works as
a generator to measure airflow.

The system’s SP is the desired box temperature, the MV is the 4-20 mA control signal to the
soldering iron, the PV is the thermocouple temperature, and the disturbance is the airflow.

The NN algorithms are implemented on a Control Micro Systems’ SCADAPack programmable logic
controller (PLC). It is programmable in relay ladder logic and multitasking C. It also provides up to
32 PID controllers. In its basic configuration, it provides 20 digital inputs (3 of which also function
as counter/accumulator inputs), 12 relay outputs, and 8 analog inputs. Analog outputs are not
included in the basic configuration and so those are added.
Page 6.65.5

“Proceedings of the 2001 American Society for Engineering Education Annual Conference & Exposition
Copyright  2001, American Society for Engineering Education”
 Figure 5. Simulation Hardware

Neural Network Implementation

The architecture of the NN has 2 input neurons (heater setting and airflow), one hidden layer with 4
neurons, and one output neuron (predicted temperature). Thus, there are 12 neuron weights to be
determined.

The NN output and the PID control are implemented using ladder logic as a foreground task. The
NN training takes place using a program coded in C running as a background task.

To keep the experiment manageable, the amount of training data is limited. The training data is
based on the NN inputs and is stored in a 10 by 10 training matrix. The rows of the training data
matrix correspond to heater settings that are broken down into 10% ranges of 0% to 9%, 10% to
19%, etc. All inputs are clamped to 99%. The columns of the matrix correspond to airflow and are
broken down in the same manner as the heater setting. Entries in the matrix are the desired
temperature for the particular heater setting and airflow inputs. Note that the 100 sets of training data
Page 6.65.6

are less than the desired 120.

“Proceedings of the 2001 American Society for Engineering Education Annual Conference & Exposition
Copyright  2001, American Society for Engineering Education”
The NN is initially “trained” by placing the system in manual. Then, each of the 100 possible
combinations of heater setting and airflow is individually set and the process allowed to come to
steady state. When steady state is reached, the program retrieves the temperature and stores it in the
training matrix as the desired output for that set of inputs. The definition of steady state for this
process is when the heater setting and airflow don’t change more than 3% and the temperature
doesn’t change more than 3 degrees over a 30 second period.

After the initial training, the NN is put on-line to control the process. When a new steady state is
reached during on-line operation, the new entry is averaged with the old entry. This demonstrates the
ability of a NN to adjust for a process that might be changing with age.

This training period points out a drawback of NN control. The initial training period took several
hours to complete just to input the training data. After the training data was input, it took several
days for the back-propagation algorithm to converge on a set of usable neuron weights. In practical
applications, if the disturbance cannot be varied easily, the training period could take considerable
time.

PID Control Results

PID parameters and results were obtained using a West model 6100 temperature controller. Process
reaction curves for three different airflows were taken to demonstrate the effect of a process
disturbance on the PID parameters. These curves were obtained by first letting the temperature
reach steady state with the soldering iron control signal set to 40%. The control signal was then
changed to 60%, and the temperature was recorded every 10 seconds until it reached steady state.
Two of these curves are shown in Figure 6. The PID parameters were then calculated for each of
these curves using the open-loop Ziegler and Nichols method [3]. These parameters are summarized
in Table 1.

Air Flow
0% 50% 100%
Kp 7.0 15 20

Ti 52 sec 40 sec 40 sec

Td 13 sec 10 sec 10 sec

Table 1. PID Parameters for Various Airflow Disturbances

Page 6.65.7

“Proceedings of the 2001 American Society for Engineering Education Annual Conference & Exposition
Copyright  2001, American Society for Engineering Education”
 Process Reaction Curve For 0% Flow

800

775
Temperature (degrees F)

750

725

700
0 50 100 150 200
Time (seconds)

Process Reaction Curve for 100% Flow

310

305

300
Temperature (degrees F)

295

290

285

280

275

270
0 50 100 150
Time (seconds)

Figure 6. Process Reaction Curves for Different Disturbance Flows

Page 6.65.8

“Proceedings of the 2001 American Society for Engineering Education Annual Conference & Exposition
Copyright  2001, American Society for Engineering Education”
To show the effects of the disturbance on the process, two step-input simulations were run using the
100% airflow PID parameters.

1. The temperature was stepped from 300 degrees to 400 degrees with the airflow at 100%.
2. The temperature was stepped from 300 degrees to 400 degrees with the airflow at 2%.

The simulation results are shown in Figure 7. It is noted that the PID controller performs as
expected when operating at the designed airflow disturbance of 100%. However, when the
disturbance is much lower, the response is unstable.

This result also points out that prudent PID control should be tuned for the worst case disturbance.

Neural Network Results

For this lab experiment, the PI controller varies the heater setting until the NN predicts a temperature
that is equal to the setpoint.

The same two step-input simulations were run after the NN was trained. The NN results are shown
in Figures 8.

It is noted from these results that the NN takes more time to reach the final value but stability
problems are no longer present. It is also noted that none of the setpoints are exact because the NN is
unable to predict the MV necessary to achieve the desired temperature; this is because the NN does
not have enough training data.

Conclusion

This paper has described a simple lab experiment to demonstrate the application of neural networks
to process control. This lab demonstrates the difficulty that PID controllers have in environments
with large disturbance fluctuations. It then demonstrates the ability of a neural network to overcome
these problems. It also points out the problem a neural network can have when there is insufficient
training data and the drawback it has regarding lengthy training time.

Page 6.65.9

“Proceedings of the 2001 American Society for Engineering Education Annual Conference & Exposition
Copyright  2001, American Society for Engineering Education”
 PID Response

420

400

380
Temperature (degrees F)

360

2% Airflow
340 100% Airflow

320

300

280

260
0 50 100 150 200 250
Time (seconds)

Figure 7. PID Responses Using Parameters from the 100% Airflow Process Reaction Curve

Neural Network Response

420

400

380
Temperature (degrees F)

360

2% Flow
340
100% Flow

320

300

280

260
0 50 100 150 200 250 300 350
Time (seconds)
Page 6.65.10

Figure 8. NN Responses to Temperature Step

“Proceedings of the 2001 American Society for Engineering Education Annual Conference & Exposition
Copyright  2001, American Society for Engineering Education”
Bibliography

[1] Kosko, Bart, Neural Networks and Fuzzy Systems, Prentice Hall, 1992.
[2] Rehg, James A., Low cost temperature and flow process control systems, Proceedings of the ASEE 1999 North
Central Section Conference, April, 1999.
[3] Murrill, Paul W., Fundamentals of Process Control Theory, 3rd Edition, 2000.

ROBERT LYNN MUELLER

Robert Lynn Mueller is an assistant professor of Electro-Mechanical Engineering at the New Kensington Campus of
the Pennsylvania State University. He received his B.S. in Electrical Engineering from Wichita State University and
his Ph.D. in Electrical Engineering from the University of Pittsburgh. Dr. Mueller spent 30 years working with
industrial control systems before joining Penn State and still continues to work as a consultant providing advice
regarding industrial automation.

Page 6.65.11

“Proceedings of the 2001 American Society for Engineering Education Annual Conference & Exposition
Copyright  2001, American Society for Engineering Education”