Temperature prediction of a two stage pulse tube cryocooler by neural network

International INTERNATIONAL Journal of Advanced JOURNAL Research OF in Engineering ADVANCED and Technology RESEARCH (IJARET), IN ISSN ENGINEERING
0976 – 6480(Print),
ISSN 0976 – 6499(Online) Volume 5, Issue 9, September (2014), pp. 63-69 © IAEME
AND TECHNOLOGY (IJARET)
ISSN 0976 - 6480 (Print)
ISSN 0976 - 6499 (Online)
Volume 5, Issue 9, September (2014), pp. 63-69
© IAEME: www.iaeme.com/ IJARET.asp
Journal Impact Factor (2014): 7.8273 (Calculated by GISI)
www.jifactor.com
IJARET
© I A E M E
TEMPERATURE PREDICTION OF A TWO STAGE PULSE TUBE
CRYOCOOLER BY NEURAL NETWORK
Asif Sha.A1, Shaleena Manafuddin2
1(Assistant Professor, Mechanical Engineering, IHRD- College of Engineering, Kollam-691531,
Kerala, India)
2(Assistant Professor, Electrical and Electronics Engineering, TKM College of Engineering,
Kollam- 691005, Kerala, India)
63
ABSTRACT
Cryocoolers are refrigerating machines which are able to achieve and to maintain cryogenic
temperature, i.e. temperature below 120K. The presence of moving parts in the cold area of
cryocooler makes it difficult to meet the required efficiency. Thus a new concept of cryocooler, the
single stage pulse tube refrigerator (PTR) satisfied many of the requirements, but the temperature
attained at the cold end side is only 30K. Thus for attaining a temperature below 30K, multistaging
of PTR is an attractive option. A lot of numerical models for the two stage pulse tube have been
developed during the last few decades. Unfortunately, not a single simulated model has been
published that can give all the design data related to the prediction of the two stage pulse tube cold
end temperature. Thus the thought of a new concept for prediction of temperature of two stage pulse
tube lead to the approach of artificial neural network. The objective of this work is to train an
artificial neural network to learn and predict the lowest temperature attained by a two stage pulse
tube cryocooler. The training is done with input as the experimental data and its output temperature
as the target. The data presented as input were, the diameter and length of pulse tubes, frequency and
orifice diameters. The network output is the minimum temperature attained by the cryocooler. After
successful training, the validation of the network is done with validation input. When the percentage
error is within the tolerance limit, the network weights and biases are used for the prediction
purposes and the predicted temperature obtained by neural network approach is within the tolerance
limit
Keywords: Cryocooler, Neural Network, Simulation, Temperature, Validation.

International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN 0976 – 6480(Print),
64
1. INTRODUCTION
Cryogenics is referred to the technology and science of producing low temperature, below
120K. Cryocoolers are refrigerating machines which are capable of achieving cryogenic temperature.
An increased need in cryogenic temperature in many areas of science and technology during the last
decades caused a rapid development of cryocoolers. But the moving parts in the cold end side of the
cryocoolers makes it difficult to meet high efficiency, high reliability, low cost, low noise level etc
caused for the new concept of cryocooler, the pulse tube refrigerator (PTR). This concept was
introduced by Gifford [1] in the late 1964’s. The minimum temperature attained by this configuration
is only 60K. It is really impossible to reach very low temperature by a single stage system. So
cascading is an important option to reach very low temperature. Thus for temperature below 30K it is
better to split the system into two.
The concept of two stage was first introduced by Zhou [2] and his coworkers in 1988. After
that a number of experiments were taken place on this type of PTR. In 1990 E. Tward et al [3]
developed a two stage pulse tube test cooler which reach 26K while rejecting heat above 300K.
C.Wang and Y.L.Ju [4] developed a two stage pulse tube refrigerator with a rotating valve in 1996.
The minimum temperature attained by the tube is 11.5K and produce a cooling load of 1.3W at 20K.
Wang .C.[5] in 1997 developed a computer program for the numerical simulation of a 4K pulse tube
cooler, which is used for the liquefaction of 4He. In 2001 Y.L.Ju [6] uses a mixed Eulerian-
Langrangian numerical model for simulation and visualizing the internal processes of PTR operating
at 4K. J.L.Shi et al [7] in 2006 developed a design model for a two stage pulse tube cryocooler. Here
the differential equation associated with energy and mass conservation is solved in order to provide a
detailed but one dimensional model of the pulse tube. Thus a lot of numerical models for the PTR
have been developed during the last decades. Unfortunately not a single simulated theoretical model
has been published that can give all the design data related to the pulse tube. Thus the unavailability
of the design procedure for a two stage pulse tube causes the attempt for using neural network
toolbox as a design tool for the prediction of the lowest temperature attained by a particular
configuration of a two stage pulse tube cryocooler.
To date, neural network have been applied successfully to a number of engineering problems.
Several researchers have demonstrated that they can be more reliable at predicting energy
consumption in a building other than traditional statistical approach because of their ability to model
non-linear patterns. In 2000 C.Renottee and M.Remy [8] introduced neural network for modeling
and control of a heat exchanger. T.T.Chow and C.L.Song [9] introduced the use of neural network
and genetic algorithm for the chiller system optimization in 2001. It was A.K Soteris and N.S
Christs. [10] in 2005 introduced neural network method for computing the thermal comfort index for
the heating ventilating and air conditioning (HVAC) systems . Unfortunately the researchers of the
two stage pulse tube cryocoolers do not explore the utility of the neural network as a design tool for
the prediction of the temperature at the cold end side and its cooling load. The aim of this study is to
investigate the suitability of the neural network as a tool for the prediction of the lowest temperature
attained at the cold end side of the PTR using the minimum possible set of input data.
The main aim for using a two stage pulse tube cryocooler is to attain a very low temperature
and the process for attaining the very low temperature is a time consuming one and the cost for the
fabrication of the two stages PTR is high. Moreover there is not a numerical equation or design
method for determining the lowest temperature attained at the cold end side of the PTR. But the
temperature is an important factor for the estimation of the cooling load of the PTR. Therefore the
need of an alternative method for determining the temperature of the PTR is helpful for avoiding the
time wastage and the huge cost for the fabrication of the PTR. Thus the recently developed
technology neural network offers such an alternative method. It is widely accepted as a design
technology offering an alternative way to tackle complex and ill specified problems. They can learn

from the given sets of examples and are able to deal with nonlinear problems. The network which
has once trained can perform the prediction and generalization at very high speed.
65
2. ARTIFICIAL NEURAL NETWORKS
Artificial neural network represents a non-algorithmic, black box computational strategy. It is
composed of interconnected artificial neurons; each has an input/output (I/O) characteristic and
implements a local computation. Figure 1 shows an artificial neuron model with ‘p’ number of
inputs. A weight ‘w’ is assigned to each input ‘p’ to describe its influence (strength). The sum of the
weighted inputs and the bias ‘b’ forms the input to the activation function ‘f’, which can be either
linear or non-linear differentiable. The output ‘a’ from the neuron is then given by
a = f (Wp+b)
p w r a

b
f
Figure 1: A single input neuron model
Neural network can be described as a machine learning technique which modifies the
numerical values of its connection weights and biases through certain training algorithm that causes
the network to approach the solution of a system model. Numerous researchers under different
constraints have shown multilayer feed forward (FF) ANN which is capable of approximating any
finite function to any degree of accuracy [11]. Also the multilayer feed forward (FF) [11] network
structure as is so far one of the most popular and effective ANN structure The learning ability of a
neural network depends on the arbitrary choice of its architecture as well as the training algorithm.
The choice of activation ( f ) may significantly influence the applicability of the training algorithm.
Lack of success in application is likely attributable to faulty training, faulty architecture or lack of
functional relationship between inputs and outputs. One of the biggest shortcomings of FF network is
the limited availability of suitable training algorithms. So far back propagation (BP) [12] has been
found highly successful. The standard BP algorithm is a gradient descent algorithm, which adopts an
error correction based learning procedure. The Levenberg-Marquardt [11] algorithm is an alternative
method for achieving fast optimization. In this process the performance of the network can be
evaluated by the mean square error.
The main objective in ANN design and training is to produce network that are able to apply
correctly to new unforeseen inputs .By partitioning the available data sets some testing sets are
reserved for accessing the generalization performance. This is known as cross validation. Over-training
might lead to memorization, and therefore pure performance when applying the testing sets.

3. TEMPERATURE PREDICTIONS USING NEURAL NETWORK
The two stage pulse tube cryocoolers is mainly used for attaining temperature less than 30K.
Here the output temperature of the PTR mainly depends on the eleven input parameters of the PTR
given in the table1. Thus for the training of the network data from the past experimental works were
used. The training was executed systematically over a family of architecture, including single and
multi layer network. It was found that all ANN architecture with single layer could not achieve in
satisfactory results. The architecture from those tested, that gave the best result and finally adopted
consists of a three layer network- an input layer, an output layer and a hidden layer. The input layer
is mainly for receiving the input data of the training set which consists of eleven neurons, based on
the eleven input parameters. The output layer of the network consists of one neuron which is mainly
for giving the output of the network; here the temperature of the PTR is the output of the network.
The hidden layer of the network is mainly for remapping of the input/output relation of the network,
which also connects the input and output layers of the networks. The brief outline of the temperature
prediction is given in Flow chart 1. The number of neurons of the hidden layer of the networks is
fixed as five, after a number of training processes. The training of the network is mainly done with
randomly selected learning rate coefficient and epochs. Once a satisfactory training is done with
varying the
Table.1: Input variable
Sl.No Input variables
1 Frequency
2 First orifice diameter
3 Second orifice diameter
4 First stage pulse tube diameter
5 First stage pulse tube length
6 First stage regenerator diameter
7 First stage regenerator length
8 Second stage pulse tube diameter
9 Second stage pulse tube length
10 Second stage regenerator diameter
11 Second stage regenerator length
66

Flow Chart. 1
Brief Outline of Temperature Prediction
Start
Give input and target (actual output) to the neural network
Create initial weights and biases
Perform ANN training by learning law
If stopping criteria met
met
Network output and compare with actual output
% error within tolerance
Save weights and biases
Provide validation input
Compare validation output with actual output
% error in tolerance limit
Save weights and biases for prediction
67
No
No
No
YES
Stop
learning rate coefficient and epochs, the weights and the biases are fixed. Then the network having
fixed weights and biases are used for the prediction purposes whose results and training results are
given in the next section.

68
IV. RESULTS/VALIDATION
The training of the network is mainly done by the three layer network by varying the learning
rate and epoch. Initially the training is done with learning rate less than 0.1 and various epoch. Thus
for a learning rate of 0.0988 and epoch 18 the network output is obtained. The percentage error
between the target and network output of the above training are less than 10%. When validation
input is given to the trained network, the network predicts the output but the maximum error between
the target temperature and the network output temperature is 11.2% and the aim was to get an error
less than 10% in the validation process. Thus it causes for taking the decision for training the above
network further. Then the network training is done with learning rate 3. For the above learning rate
the network gives a better output for an epoch 14 where the maximum percentage error between the
target and network output is around 5%. Thus the network weights and biases are used for the
validation purposes but the result obtained from the validation makes the percentage error greater
than 10. Therefore the network is not yet suitable for the prediction purposes because the learning
rate greater than one makes the weights and biases vector to overshoot from the ideal position there
by the result of the validation is not in the accurate form. Now the training of the network for the
input and its target is done with learning rate which is randomly selected from 0.1 to 1. Thus for a
learning rate of 0.920 and epoch 24, the network gives the output temperature which is close to the
target temperature, where two of the percentage error is zero and most of the percentage error was
very small. These are because of the overtraining of the network which causes for the memorization
of input data’s by the network. So the network does not give the exact output for the simulation. The
training is continued for various learning rate and epoch and finally for a learning rate of 0.96985
and epoch 13, a good result is obtained which is shown in the table 2. Here the maximum percentage
error is 3.5 and that of minimum is 0.3. This percentage error of the result is within the tolerance
limit of +5% to-5%. Now the input weights, layer weights, input biases and layer biases are used for
the validation purposes whose result is given in table.3. The percentage errors obtained by the
validation are within the tolerance range +10% to -10%; also it is one of the best results of the ever
training process for this work and also for the validation. Thus this results shows that the neural
network approach is more suitable for the prediction of the lowest temperature attained by a two
stage pulse tube cryocoolers.
Table.2: Training Result
Target
Tem:
(K)
7.7
4.6
6.2
7.2
14
5.1
4.5
6.6
4.3
4.4
4.8
8.5
3.3
12.3
N/O
Tem:
(K)
7.68
4.76
6.14
7.15
13.88
5.07
4.39
6.53
4.35
4.38
4.82
8.55
3.25
12.27
% error 0.30 -3.5 1.0 0.7 0.86 0.6 1.11 1.1 -1.2 0.5 -0.42 -0.6 1.5 0.3
Table.3: Validation Result
Target Tem:(K) 7.9 5.5 8.8 3.7
N/O Tem: (K) 8 . 6 4 5.10 7.94 3.91
% error -9.3 7.2 9.8 -5.6

69
5. CONCLUSION
Once trained, the network predicts the lowest attainable temperature of a particular
configuration of a two stage pulse tube cryocooler. At this stage the work was confined at primarily
investigating the suitability of artificial neural network for the prediction of temperature. In order for
the network to be of significant use of cryogenic engineers it needs to be enriched with more training
cases and diverse constructional and environmental parameters. Furthermore, it is estimated that the
performance will improve with use, since the network has the capability of learning from moderate
data. As these become available and more data, they may be used to retain the network and hence to
improve the accuracy. This method may also be applied for the cooling load estimation, a job which
is much harder.
REFERENCES
[1] Gifford W.E and Longsworth. R.C, Pulse tube refrigeration, Trans ASME, 1964, pp264-267.
[2] Zhou et al,” Two stage pulse tube refrigerator with axial curvature”, Advances in Cryogenic
Engineering Vol. 6, 1988, pp137-144.
[3] Tward E, Chan C K and Burt W W, “Pulse tube refrigeration”, Advances in Cryogenic
Engineering Vol. 35, 1990.
[4] Wang c, Ju Y L and Zhou, “Experimental investigation of a two stage pulse tube
refrigerator”, Advances in Cryogenic Engineering ,Vol. 36, 1996, pp 605-609.
[5] Wang C “Numerical analysis of 4K pulse tube cryocoolers, Part 1, Numerical simulation”,
Advances in Cryogenic Engineering Vol. 37, 1997, pp 207-213.
[6] Ju Y L, ‘Computational study of a two stage pulse tube cooler with mixed Eulerian-
Lagrangian method”, Advances in Cryogenic Engineering Vol.41, 2001, pp 49-57.
[7] Shi J L, Pfotenhauer and Nellis G F, ‘Design model for a two stage pulse tube cryocooler”,
Transactions of the Cryogenic Engineering Conference, Vol.51, 2006.
[8] Renotte.C and Remy.M “Neural modeling and control of a heat exchanger based on. SPSA
techniques”, The American Control Conference, Chicago, June, 2000.
[9] Chow .T .T and Song C. L “Applying neural network and genetic algorithm in chiller system
optimization”, Seventh International IBPSA Conference, Brazil, August 13-15, 2001.
[10] Soteris.A.K and Christs.N.S. “The application of neural network in HVAC system”
Volume-8, IEEE, August 2005, pp 4800-4804.
[11] Hagun Demuth.B, Neural network design (Thomson Learning: Singapore, 2002).
[12] Laurence Fauselt, Fundamentals of neural network (Pearson Education India, 2005.
[13] James. A. Anderson. An introduction to neural networks (Prentic-Hall-India. 2002).
[14] Lakhwinder Pal Singh and Jagtar Singh, “Effects of Cryogenic Treatment on the Cutting Tool
Durability”, International Journal of Mechanical Engineering Technology (IJMET),
Volume 3, Issue 1, 2012, pp. 11 - 23, ISSN Print: 0976 – 6340, ISSN Online: 0976 – 6359.
[15] Sweety H Meshram, Vinaykeswani and Nileshbodne, “Data Filtration and Simulation by
Artificial Neural Network”, International Journal of Electronics and Communication
Engineering Technology (IJECET), Volume 5, Issue 7, 2014, pp. 46 - 55, ISSN Print:
0976- 6464, ISSN Online: 0976 –6472.

Temperature prediction of a two stage pulse tube cryocooler by neural network

More Related Content

What's hot (20)

Similar to Temperature prediction of a two stage pulse tube cryocooler by neural network (20)

More from IAEME Publication (20)

Recently uploaded (20)

Temperature prediction of a two stage pulse tube cryocooler by neural network