You are on page 1of 6

Use of Artificial Neural Networks for Improving Fiber Optic

Microbend Sensor Performance


H. S. Efendioglu, T. Yildirim, and K.Fidanboylu

Abstract This paper presents experimental results


related with the behavior of fiber optic microbend
sensors based on different configurations. Different types
of deformer sets having different mechanical
periodicities, corrugation size and number of
deformations cycles have been used to test the validity of
the proposed technique. Normalized output intensity of
the microbend sensor as a function of applied force is
later used in the prediction of desired sensor response
using Artificial Neural Networks (ANNs). It is shown
that, ANNs can detect measurement errors and can be
used in the development of intelligent and robust sensors
that can monitor and detect the abnormalities in the
sensors state.
I. INTRODUCTION
Recent developments in fiber optic sensors have opened a
new era in sensor technology because of their advantages
such as immunity to electromagnetic interference,
lightweight, small size, remote sensing and multifunctional
sensing capabilities, etc.
Microbend sensors are one of the earliest fiber optic
sensors developed and have been employed by several
researchers for over thirty years. Fiber optic microbend
sensors have been used in several industrial applications
such as pressure measurement in gear interface systems [1],
temperature and pressure measurement in aircraft fuel tanks
[2], investigation of characteristics of various grades of
cement [3], pH sensing [4], acoustic and displacement
sensing [5].
Because of their fast real-time operation and learning
abilities, Artificial Neural Networks (ANNs) have been
preferred as a computational paradigm in which a
deterministic description of the computation is either
complex or difficult to identify [6, 7]. Due to these
characteristics, ANNs have been successfully used in optical
fiber technology in the development of intelligent fiber optic
sensors [8], calibration [9] and the prediction of sensing
parameters [10, 11].
During long-term sensing, its possible that problems may
occur in electronic or mechanic parts of the fiber optic
microbend sensor system such as deformation of the fiber,
breakdown of the detector or the light source. If the output
of the microbend sensor can be predicted and compared with
the measurements, fault tolerance of the sensor can be

978-1-4244-8126-2/10/$26.00 2010 IEEE

decreased and more robust sensors can be designed. The


ability of monitoring and predicting possible abnormalities
and imminent failures in the system enables immediate
condition awareness and ability to take action against
changing conditions. With the proposed approach the
microbend sensor becomes an intelligent system.
In order to predict the sensor response of fiber optic
microbend sensors, it is necessary to carry out lengthy and
complex mathematical computations. Since, many
parameters such as the properties of the optical fiber and the
microbend sensor system affect the computation of the
sensor response, changing sensor configuration or type of
the fiber bring extra complexities. To overcome these
problems ANNs can be used, because they can generate
appropriate outputs for given inputs without any necessity to
mathematical formulations between input and output data.
Hence, this can greatly simplify the sensor response design
and prediction problem.
In this paper, experimental results related with the
behavior of fiber optic microbend sensors with different
configurations are presented. The optical intensity is
measured as a function of applied force to the microbend
deformer sets. Different types of deformer sets having
different mechanical periodicities (wavelength), corrugation
size and number of deformations cycles (bends) have been
used to test the validity of the proposed technique.
Normalized output intensity of the microbend sensor as a
function of applied force has been used in the prediction of
desired sensor response using ANNs.
II. OVERVIEW OF MICROBEND SENSORS
When a fiber is subjected to small bends or perturbations,
a certain portion of light propagating in the core of the fiber
is coupled into radiation modes and is lost. Mode coupling
can be achieved by employing corrugated plates that deform
the fiber into a series of bends. Therefore, microbending
causes the light intensity to decrease [11]. By monitoring
and correlating the loss of light intensity, different types of
microbend sensors can be designed.
Sensing region of the microbend sensor is composed of
two corrugated plates, called deformer plates. The optical
fiber is squeezed under applied force due to the corrugations
as shown in Figure 1.

corrugation diameters of 6 mm and 10 mm with mechanical


periodicities of 14 mm and 16 mm, respectively were
constructed. All sensors had three deformation cycles.

Fig. 1. Schematic of the microbend sensor.

The deformer in response to a force change F to the bent


fiber causes the amplitude of the fiber deformation X to
change by an amount X. The transmission coefficient T, for
light propagating through the bent fiber is in turn changed by
an amount T so that :

T
X

T =

AsYs

F k f +

ls

(1)

where T/ X is the sensitivity, As is the area, ls is thickness


of deformer, Ys is the Youngs modulus, and the kf is the
effective spring constant of the optical fiber [13]. The
effective spring constant in turn can be expressed as follows:

k f 1 =

3
3 Yd 4

(2)

Fig. 2. Schematic of the experimental setup.

Figure 3 shows the results obtained from the microbend


sensors having corrugation diameters of 10 mm with
periodicities of 14 mm and 16 mm, respectively. From this
figure it can be observed that the normalized output intensity
decreases as the applied force increases. Furthermore, the
sensitivity of the microbend sensor increases as the
mechanical periodicity decreases. This result is in agreement
with the theory and physical phenomena (Eq. 1 and 2), since
as the periodicity decreases the fiber becomes tighter
between the corrugations.
1

where Y is the effective Youngs modulus, d is the diameter


of fiber and is the number of deformation cycles [14].

0.98
0.97

III. MICROBEND SENSOR EXPERIMENTAL RESULTS

A. Effect of Corrugation Size and Mechanical Periodicity


on the Output Light Intensity
In order to observe the effect of corrugation size on the
output light intensity, different microbend sensors having

0.96

Normalized Intensity

Six different microbend sensor configurations with


different corrugation size, mechanical periodicity and
number of deformation cycles have been observed to see the
microbending effect and take the training data from
microbend sensor measurements. Normalized output
intensity of the microbend sensors have been measured as a
function of applied force. The experimental setup is
comprised of a 16 mW solid-state laser with a wavelength of
650 nm, a step index multimode silica fiber with a numerical
aperture (NA) of 0.37, and core/cladding/jacket dimensions
of 200/230/500 m, microbend deformer and a power-meter.
Schematic of the experimental setup is illustrated in Figure
2.
The solid state laser source is coupled into the bare fiber
with a 20X objective lens. Experiments were conducted by
increasing the force from 0 N to 58.66 N for different sensor
configurations. At the output of the fiber, the light intensities
were measured by using a power-meter. The behavior of
microbend sensor was determined by conducting
experiments with different mechanical periodicities, number
of deformation cycles and corrugation size.

14 mm
16 mm

0.99

0.95
0.94
0.93
0.92
0.91
0.9
0.89
0.88
0

4.905

9.81

14.715

19.62

24.525

29.43

34.335

39.24

44.145

49.05

53.955

58.86
60

Force (N)

Fig. 3. Normalized output intensity versus applied force (three deformation


cycles having corrugation diameters of 10 mm and mechanical periodicities
of 14 mm and 16 mm, respectively).

Figure 4 shows the results obtained from the microbend


sensors having corrugation diameters of 6 mm with
periodicities of 14 mm and 16 mm, respectively. From this
figure it can be concluded that the normalized output
intensity decreases linearly as the applied force increases.
Moreover, it can also be seen that the microbend sensor with
14 mm periodicity is more sensitive to the applied force than
the one with 16 mm periodicity. All of these results are in
agreement with the microbend sensor having corrugation
diameters of 10 mm.

IV. OVERVIEW OF ARTIFICIAL NEURAL NETWORKS

14 mm
16 mm

0.99
0.98
0.97

Normalized Intensity

0.96
0.95
0.94
0.93
0.92
0.91
0.9
0.89
0.88
0

4.905

9.81

14.715

19.62

24.525

29.43

34.335

39.24

44.145

49.05

53.955

58.86
60

Force (N)

Fig. 4. Normalized output intensity versus applied force (three deformation


cycles, corrugation diameters of 6 mm and mechanical periodicities of 14
mm and 16 mm, respectively).

Comparing the results shown in Figures 3 and 4, we can


see that the normalized output intensity of the microbend
sensors having corrugation diameters of 6 mm is more
sensitive than the ones with corrugation diameters of 10 mm.
This shows that the sensitivity of the microbend sensor
increases as the corrugation diameter size decreases.
B. Effect of Number of Deformation Cycles on the Output
Light Intensity
In order to observe the effect of number of deformation
cycles on the output light intensity, microbend sensors with
two deformation cycles instead of three deformation cycles
having corrugation diameters of 6 mm were constructed.
Figures 5 shows the results obtained from the microbend
sensors with two deformation cycles. Comparing the results
shown in figures 4 and 5, it can be observed that the
sensitivity of the microbend sensor increases with the
increasing number of deformation cycles. These
experimental results are also in agreement with the theory
(Eq. 1 and 2).
1

14 mm
16 mm

0.99
0.98
0.97

Normalized Intensity

0.96
0.95
0.94
0.93
0.92
0.91
0.9
0.89
0.88
0

4.905

9.81

14.715

19.62

24.525

29.43

34.335

39.24

44.145

49.05

53.955

58.86
60

Force (N)

Fig. 5. Normalized output intensity versus applied force (two deformation


cycles having corrugation diameters of 6 mm and mechanical periodicities
of 14 mm and 16 mm, respectively).

Artificial Neural Networks (ANNs) are processing


elements which generate proper outputs for given inputs
without any formulation between input and output. An
artificial neuron model consists of input(s) with weight(s),
activation function(s) and output(s). The weights are
adjusted until the desired output is generated for a given
input. ANNs are one of the most effective tools that have
been widely employed due to their computational speed,
ability to handle complex functions, and great efficiency
even in cases where full information for the studied
problems is absent. ANNs are mainly used for classification,
function approximation, clustering and regression [10].
ANNs have different types of connections. Feed forward
neural networks, which are also called Multi Layer
Perceptron (MLP), are the most popular model used in many
applications. An input layer, one or more hidden layers and
an output layer are the parts of MLP [15]. Computation
nodes are arranged in layers and information feeds forward
from layer to layer via weighted connections. MLP is chosen
in this work since it has many useful properties for
prediction problems. It has a relatively simple structure and
backpropagation algorithms are implemented in many
problems. MLP can efficiently learn large data sets and it
can establish relationship between variables.
There are several training algorithms in literature. It is
complicated to know which training algorithm will be the
fastest or more precise for a given problem. The algorithms
used to train MLP in this work are Fletcher-Reeves Update
(FRU), One Step Secant (OSS), Scaled Conjugate Gradient
(SCG), and Variable Learning Rate Backpropagation
(VLRB).
According to the conjugate gradient back propagation
with Fletcher-Reeves updates, the FRU is a network training
function that updates weights and bias values. Among the
conjugate gradient algorithms it has the smallest storage
requirements. The SCG is the only algorithm using the
conjugate gradient methods that does not require any line
search [14]. The one step secant method is an algorithm to
connect the gap between the conjugate gradient algorithms
and the quasi-Newton algorithms. Complete Hessian matrix
is not stored by one step secant algorithm; it assumes that at
each iteration, the previous Hessian was the identity matrix.
This has additional advantage that the new search direction
can be calculated without computing a matrix inverse.
Variable Learning Rate Backpropagation combines adaptive
learning rate with momentum training. With standard
steepest descent, the learning rate is held constant
throughout training. The performance of the steepest descent
algorithm can be improved if we allow the learning rate to
change during the training process. An adaptive learning rate
will attempt to keep the learning step size as large as
possible while keeping learning stable. It has the momentum
coefficient as an additional training parameter [16].

V. APPLICATION OF ANNS TO FIBER OPTIC MICROBEND


SENSORS

Among the six different sets of experimental data


provided in Part III, two sets of experimental data were
randomly selected and used for training and testing the
ANN. The first selected set of experimental data was for the
microbend sensor having two deformation cycles with a
corrugation diameter of 6 mm and mechanical periodicity of
14 mm (see Fig. 5). The second selected set of experimental
data was for the microbend sensor having three deformation
cycles with a corrugation diameter of 10 mm and mechanical
periodicity of 14 mm (see Fig. 3). The microbend sensor
response was predicted using the Multi Layer Perceptron
(MLP) with four different algorithms mentioned in Part IV.
Force (F) and normalized intensity are the input and
output variables of ANN models, respectively. Sensor
response can be predicted using F values. The network was
trained by a training dataset consisting of randomly selected
F vs. normalized intensity. New responses are predicted for
unseen F values, after the training process.
Each network is trained with eight experimental sets of
data. The performance of the algorithms used in the network
is compared in terms of their mean square errors (MSEs).
The training algorithms and the number of hidden neurons
used in MLP network for the randomly selected sets of
experimental data are given in Table 1.

Multi
Layer
Perceptron

Number of
Hidden Neurons

Fletcher-Reeves
Update

One Step Secant

Scaled Conjugate
Gradient

Variable Learning Rate


Backpropagation

Mean Square
Error (MSE)

Algorithm
Fletcher-Reeves
Update

1.94E-07

One Step Secant

7.29E-09

Scaled Conjugate
Gradient

1.36E-07

Variable Learning Rate


Backpropagation

9.98E-08

Multi
Layer
Perceptron

Table 2. The MSEs of different training algorithms of MLP for for the
randomly selected sets of experimental data.
ANN Model Outputs

F(N)

Sensor
Response

FRU

OSS

SCG

4.905

0.986

0.9901

0.9876

0.9902

0.9864

14.715

0.980

0.9786

0.9805

0.9788

0.9656

24.525

0.974

0.9735

0.9742

0.9742

0.9522

44.145

0.963

0.9651

0.9663

0.9664

0.9302

53.955

0.955

0.9578

0.9542

0.9559

0.9225

MSE

VLRB

6.23E-06 2.83E-06 6.31E-06 1.65E-05

Table 3. Comparison of fiber optic microbend sensor responses and the


ANN outputs.
1

Experiment
OSS
VLRB

0.99
0.98
0.97
0.96

Normalized intensity

Algorithm

shown in Figure 6. From these results, it can be seen that all


the algorithms are able to predict sensor responses with
small errors.

0.95
0.94
0.93
0.92
0.91

Table 1. The training algorithms and number of hidden neurons in MLP for
the randomly selected sets of experimental data.

The training MSEs results of used MLP algorithms for


two deformation cycles having corrugation diameter of 6
mm and mechanical periodicity of 14 mm configuration are
given in Table 2. These results show that, although it has one
of the minimum numbers of neurons, OSS algorithm has the
smallest MSE value.
The prediction performances of different ANN models
proposed in this work for two deformation cycles having
corrugation diameter of 6 mm and periodicity of 14 mm
have been tested with five sets of experimental data obtained
from the fiber optic microbend sensor.
The comparisons of the sensor responses and the ANN
model outputs are given in Table 3. From this comparison it
can be perceived that OSS has the smallest MSE value. The
worst MSE value belongs to VLRB. The comparison of the
best and the worst ANN outputs with respect to MSEs are

0.9
0.89
0.88
0

4.905

9.81

14.715

19.62

24.525

29.43

34.335

39.24

44.145

49.05

53.955

58.86
60

Force (N)

Fig. 6. Comparisons of the best and the worst ANN outputs with the sensor
responses for two deformation cycles having corrugation diameters of 6 mm
and mechanical periodicity of 14 mm configuration.

The other microbend sensor configuration used for ANN


analysis having two deformation cycles with a corrugation
diameter of 6 mm and mechanical periodicity of 14 mm was
also analyzed. The MSEs of different training algorithms of
MLP are given in Table 4. From the results it can be seen
hat, OSS algorithm has the smallest MSE value.

Algorithm

Mean Square
Error (MSE)

Fletcher-Reeves
Update

1.42E-07

One Step Secant

6.35E-09

Scaled Conjugate
Gradient

2.26E-07

Variable Learning Rate


Backpropagation

9.39E-08

Multi
Layer
Perceptron

Table 4. The MSEs of different training algorithms of MLP for the


microbend sensor having three deformation cycles with a corrugation
diameter of 10 mm and mechanical periodicity of 14 mm.

The prediction performances of the ANN algorithms were


again tested with five sets of experimental data selected
randomly. The comparisons of the sensor responses and the
ANN model outputs are given in Table 5. From this
comparison it can be perceived that FRU has the smallest
MSE value. The worst MSE value belongs to VLRB. The
comparison of the best and the worst ANN outputs with
respect to MSEs are shown in Figure 7. Results show that all
the algorithms predict sensor responses with small errors.
F(N)

ANN Model Outputs

Sensor
Response

FRU

OSS

SCG

VLRB

4.905

0.992

0.9914

0.9931

0.9932

0.9864

14.715

0.971

0.9679

0.9720

0.9678

0.9656

24.525

0.949

0.9476

0.9505

0.9469

0.9522

44.145

0.930

0.9304

0.9246

0.9324

0.9302

53.955

0.919

0.9186

0.9388

0.9166

0.9225

MSE

2.46E-06 5.66E-06 8.54E-05 1.65E-05

Table 5. Comparison of the sensor responses and the ANN outputs.

Experiment
FRU
VLRB

0.99

based on different configurations. The optical intensity was


measured as a function of applied force to the microbend
deformer sets. Different types of deformer sets having
different mechanical periodicities, corrugation size and
number of deformations cycles were used to test the validity
of the proposed technique. Normalized output intensity of
the microbend sensor as a function of applied force was used
in the prediction of desired sensor response using ANNs.
Experiments performed on microbend sensors with
different corrugation sizes have shown that the sensitivity of
the microbend sensor increases as the corrugation diameter
size decreases. Also, the sensitivity of the microbend sensor
increases with the increasing number of deformation cycles.
Beside these, the sensitivity of the microbend sensor
increases as the periodicity decreases. All the experiments
show that the normalized output intensity decreases as the
applied force increases. Experimental results are in full
agreement with the theory.
Different ANN algorithms using MLP have been
proposed to predict the response of a fiber optic microbend
sensor. Performance comparisons demonstrate that all of the
ANN models used in this paper can predict the sensor
responses with considerable errors. ANN can be used in two
different ways. First of all, one can use multiple sensors and
feed their output to an ANN. In this way, more accurate
force, strain, etc. measurements can be done. The second
advantage of using ANNs is, when multiple sensors are used
one can use the method of voting to detect sensor failures.
For example, if five sensors are present and one gives a
force reading significantly different than the other four, it is
very likely that there is a sensor failure. In summary, use of
ANN with multiple sensors will both result more accurate
and failsafe sensor architectures. Accordingly, ANNs can
detect measurement errors and can be used to develop
intelligent and robust sensors that can monitor and detect the
abnormalities in the sensors state.

0.98

REFERENCES

0.97

[1]

Normalized intensity

0.96
0.95

[2]

0.94
0.93
0.92

[3]

0.91
0.9
0.89
0.88
0

4.905

9.81

14.715

19.62

24.525

29.43

34.335

39.24

44.145

49.05

53.955

58.86
60

[4]

Force (N)

Fig. 7. Comparisons of the best and the worst ANN outputs with the sensor
responses having three deformation cycles with a corrugation diameter of
10 mm and mechanical periodicity of 14 mm.

[5]

VI. CONCLUSION

[7]

In this paper, we have presented experimental results


related with the behavior of fiber optic microbend sensors

[8]

[6]

T. McCollum, and G. B Spector, Fiber Optic Sensor for Detection of


Dynamic Fluid Pressure at Gear Interfaces, Review of Scientific
Instruments, Vol. 65 No. 3, pp. 724-729, 1994.
S. F. Knowles, B. E. Jones, S. Purdy, and C. M. France, Multiple
Microbending Optical-Fibre Sensors for Measurement of Fuel
Quantity in Aircraft Fuel Tanks, Sensors and Actuators, Vol. A 68,
pp. 320-323, 1998.
M. Rajesh, K. Geetha, M. Sheeba, P. Radhakrishnan, C. P. G.
Vallabhan, and V. P. N. Nampori, A Fiber Optic Smart Sensor for
Studying The Setting Characteristics of Various Grades of Cement,
Optics and lasers in Engineering, Vol. 44, pp. 486-493, 2006.
S. T. Lee, B. Aneeshkumar, P. Radhakrishnan, C.P.G. Vallabhan,
V.P.N. Nampoori, A microbend fiber optic pH sensor, Opt.
Commun., 205, 253256, 2002.
N. Lagakos, W. J. Trott, T. R. Hickman, J. H. Cole, J. A Bucaro,
Microbend fiber-optic sensor as extended hydrophone, IEEE
Journal of Quantum Electronics, 18, 1633-1638, 1982.
J.M. Zurada, Introduction to Artificial Neural Systems. St. Paul,
MN:West Publishing Company, 1992.
S Haykin, Neural Networks: A Comprehensive Foundation, Newyork:
Mcamillan and IEEE Computer Society, 1999.
I. Dias, R Oliveira, and O. Frazo, Intelligent optical sensors using
artificial neural network approach, Springerlink: Innovation in
Manufacturing Networks, 266, 289294, 2008.

[9]

[10]
[11]
[12]
[13]
[14]
[15]
[16]

J. Wojtek, E. Porada, M. Beaulieu, and T. A. Eftimov, Automatic


calibration of a fiber-optic strain sensor using a self-learning system,
IEEE Transactions on Instrumentation and Measurement, 43, 341 346,
1994.
H.S. Efendioglu, T. Yildirim, and K. Fidanboylu, Prediction of Force
Measurements of a Microbend Sensor Based on an Artificial Neural
Network, Sensors, Vol. 9, No. 9, 7167-7176, 2009.
O.G. Saracoglu, An artificial neural network approach for the
prediction of absorption measurements of an evanescent field fiber
sensor, Sensors , Vol. 8, 1585-1594, 2008.
J.W. Berthold, Historical review of microbend fiber-optic sensors, J.
Lightwave Technol., Vol.13, 11931199, 1995.
S. T. Lee, A.N. George, P. Sureshkumar, P. Radhakrishnan, C.P.G.
Vallabhan, and V.P.N. Nampoori, Chemical sensing with microbent
optical fiber, Optic Letters, Vol. 26, No. 20 , 1541-1543, 2001.
N. Lagakos, J.H. Cole, and J.A Bucaro, Microbend fiber optic
sensor, Applied Optics, Vol. 26, 2171-2180, 1987.
N. Cokun, and T. Yildirim The effects of training algorithms in MLP
network on image classification. International Joint Conference on
Neural Networks (IJCNN), 2003, 2, 1223-1226.
H. Demuth, M. Beale. A, and M. Hagan, Neural Network Toolbox 6
Users Guide. Mathworks, 2008.

You might also like