COMPARISON OF ARTIFICIAL NEURAL NETWORKS AND FUZZY LOGIC APPROACHES FOR CRACK DETECTION IN A BEAM LIKE STRUCTURE

International Journal of Artificial Intelligence and Applications (IJAIA), Vol.9, No.1, January 2018
DOI : 10.5121/ijaia.2018.9103 35
COMPARISON OF ARTIFICIAL NEURAL
NETWORKS AND FUZZY LOGIC APPROACHES
FOR CRACK DETECTION IN A BEAM LIKE
STRUCTURE
B Prakruthi Gowd1
, K Jayasree2
and Manjunath N. Hegde3
1
Assistant Professor, Department of Civil Engineering,
Vidya Jyothi Institute of Technology, Hyderabad, India.
2
Assistant Professor, Department of Civil Engineering,
Vasavi College of Engineering, Hyderabad, India.
3
Professor and Dean (Academics), Department of Civil Engineering,
Dr. Ambedkar Institute of Technology, Bengaluru, India.
ABSTRACT
This paper proposes two algorithms of crack detection one using fuzzy logic (FL) and the other artificial
neural networks (ANN). Since modal parameters are very sensitive to damages, the first three relative
natural frequencies are used as three inputs and the corresponding relative crack location, relative crack
depth are used as the two outputs in the algorithms. The three natural frequencies for an undamaged beam
and different cases of damaged beam (Single crack at various locations with varying depths) were obtained
by modelling and simulating the beams using a finite element based (FEM) software. Results concluded
that both the approaches can be successfully employed in crack detection in a beam like structure but FL
approach performed better in determining relative crack depth whereas ANN approach performed better in
determining relative crack location. All the comparisons made in the study are based on the R2
values.
KEYWORDS
Damage detection, Modal Properties, Artificial Neural Networks, Fuzzy Logic, R2
values.
1. INTRODUCTION
Structures such as buildings, bridges, or any other, are prone to sudden damage, deterioration and
aging during their service life. The existence of damage in these structures presents a serious
threat to the longevity of these structures. Furthermore, the presence of damage can affect the
performance of these structures and influence structural integrity of other parts which are
associated with them. Detection of these damages at an early stage can allow engineers have
enough time to make careful decisions on whether to repair the part of the structure which is
damaged while the structure is still in operation or to temporarily keep the structure out of service
or to entirely discard the structure. This can possibly avoid many undesirable failures of damaged
structures, which if left unnoticed would lead to lack of public safety, reduction in the life span of
the structure and huge economic losses. Therefore, a health monitoring system or an effective
detection technique is required that will be able to detect and identify the damage in real time and

36
at the earliest stage possible, so as to maintain the structural stability, integrity and to maximize
the life span of the structure as much as possible.
Many techniques have been employed in the past for damage detection. Some of these are Visual
inspection, Non-destructive testing and Use of sensors to detect local damage. These methods
take a lot of time and do need the testing of the entire structure in minute detail to tell if the
structure is damaged. Furthermore, if a damage is deep inside the structure it may not be
detectable by these localized methods. Based on the changes in the modal parameters researchers
have developed many signal processing and artificial intelligence based techniques for damage
identification for any kind of damage detection. These methods have been designed with an aim
for faster and accurate estimation of the presence of damage in the structures even from a remote
place.
Motivated by the above reasons, this work aims at exploring the use of artificial intelligence
techniques such as artificial neural networks and fuzzy logic approaches in determining the
damages in engineering structures at an early stage by capturing the vibration parameters.
2. THEORETICAL BACKGROUND
The following section gives a brief theoretical insight of the two approaches which are used for
developing the damage detection algorithms individually and also their differences.
2.1. Fuzzy logic
Fuzzy Logic belongs to the family of many-valued logic. It focuses on fixed and approximate
reasoning opposed to fixed and exact reasoning. A variable in fuzzy logic can take a truth value
range between 0 and 1, as opposed to taking true or false in traditional binary sets. Since the truth
value is a range, it can handle partial truth. Beginning of fuzzy logic was marked in 1965, with
the introduction of fuzzy set theory by Lotfi Zadeh [1]. Fuzzy logic provides a method to make
definite decisions based on imprecise and ambiguous input data.
2.2. Artificial Neural Networks
ANN is a computational model that is developed based on the biological neural networks. An
ANN is made up of artificial neurons that are connecting with each other. Typically, an ANN
adapts its structure based on the information coming to it. A set of systematic steps called
learning rules needs to be followed when developing an ANN. Further, the learning process
requires learning data to discover the best operating point of the ANN. ANN’s can be used to
learn an approximation function for some observed data. But when applying ANN, there are
several factors one has to consider [2]. The model has to be carefully selected depending on the
data. Using unnecessarily complex models would make the learning process harder. Choosing the
correct learning algorithm is also important, since some learning algorithms perform better with
certain types of data.
2.3. Difference between Artificial Neural Networks and Fuzzy Logic
Fuzzy logic allows making definite decisions based on imprecise or ambiguous data, whereas
ANN tries to incorporate human thinking process to solve problems without mathematically
modelling them. Even though both of these methods can be used to solve nonlinear problems, and
problems that are not properly specified, they are not related. In contrast to Fuzzy logic, ANN
tries to apply the thinking process in the human brain to solve problems. Further, ANN includes a

37
learning process that involves learning algorithms and requires training data whereas fuzzy logic
includes development of membership functions and rules to relate them [3].
The brief overview of the work done so far related to the above work is as follows: The
relationship between the physical damage to a structure and changes in the dynamic characteristic
has been studied by Richardson [4]. In his survey he has focused on modal analysis, modal testing
method, and relation between the extent of damage and changes in the modes of vibration. During
a review process of structural health monitoring, Rytter [5] has classified the damage detection
technique as four levels. The level 1 determines the presence of damage in structures. The level 2
determines the geometric location of the damage. The level 3 quantifies the extent of damage and
level 4 predicates the life of the structure. The local flexibility that affects the vibration signatures
has been suggested by Dimarogonas [6] in his review paper. The crack opening and closing
happen in time that depends upon the vibration and rotation amplitude. He has suggested that the
local stiffness matrix at the cracked section of a shaft leads to a coupled system and for uncracked
shaft, the system is decoupled. A fuzzy rule-based system [7] is developed for the blade of a BO-
105 helicopter rotor modeled as a cantilever beam and demonstrated that the fuzzy system
performs accurately even in the existence of noisy data. De Miguel and Blázquez [8] developed a
decision making module based on fuzzy logic for model based fault diagnosis applications. A
fault detection and isolation system based on the input and output parameters was successfully
applied in laboratory equipment to reduce the uncertainties of the output parameter. Bakhary et al.
[9] have explained a statistical approach to take into account the effect of uncertainties in
developing an ANN model. Chandrasekhar and Ganguli [10] showed that geometric and
measurement uncertainty causes considerable problem in the damage assessment. They used
Monte Carlo simulation to study the changes in the damage indicator due to uncertainty in the
geometric properties of the beam. The results obtained from the simulation were used for
developing and testing the fuzzy logic system. They addressed the uncertainty associated with the
fuzzy logic system for structural damage detection. Das and Parhi [11] have extended the
application of neural network technique for studying the fault diagnosis of a cracked cantilever
beam.
3. FINITE ELEMENT MODELLING AND SIMULATION
3.1. Methodology
The methodology followed here is to use a FEM package software namely ANSYS 12.1 to model
beam like structures (flats) with fixed end condition which include both damaged and undamaged
once. Furthermore, modal analysis is carried out in order to obtain the natural frequencies. The
variation of these natural frequencies have been analysed for different cases of the damaged
structures with respect to the undamaged one so that they can be used further in developing the
two damage detection algorithms.
3.2. Model properties and Cases studied
The material and geometric properties of the model are as shown in Table 1 and Figure 1. An
element type of Solid-Brick-2 noded 188 and element mesh division of 0.0025 is used during
FEM modelling. A total of 84 different damaged cases were studied and are described below in
Table 2.

38
Table 1. Material and Geometric properties.
Young’s Modulus
(E), GPa
Poisson’s
Ratio (µ)
Density (d),
Kg/m3
Length
(l), mm
Width
(b), mm
Depth
(d), mm
End
Conditions
210 0.3 7850 3000 50 200 Fixed
Figure 1. Undamaged Beam like structure (Flat)
Table 2. Cases Studied.
Crack Depths (cd) Varies from 10mm to 120mm (increasing at an interval of 10 mm) (12
varying depths)
Crack Locations (cl) Varies from L/2 to L/16 from the fixed end (7 different locations)
Note: Thickness of the crack is taken as 1mm for all the cases.
3.3. Results of Modal Analysis
The natural frequencies and mode shape deflections of the undamaged flat model obtained from
ANSYS are as shown in Table 3 and Figure 2. In order to study the variation of the natural
frequencies for different damaged cases the relative natural frequencies are calculated with the
help of the results obtained from ANSYS.
Table 3. Natural Frequency of Undamaged Flat.
First Natural
Frequency, Hz
Second Natural
Frequency, Hz
Third Natural
Frequency, Hz
112.49 299.39 562.35

39
Figure 2. Mode shape deflections obtained from ANSYS for 1st
, 2nd
and 3rd
modes.
The variation of the first relative natural frequencies verses relative crack location from the
fixed end and relative crack depth are as shown in Figure 3 and 4.
Figure 3. Relative First Natural Frequency variation for same Depth but different Location.
Figure 4. Relative First Natural Frequency variation for same Location but different Depths.
It is evident from Figure 3 that the relative first natural frequency for a flat increase as the
distance of the crack is increasing from the fixed end and is maximum at L/2 location (i.e., a
relative crack location of 0.5). Also it can be observed that the relative first natural frequency
decreases as the depth of the crack increases and is minimum for 120mm crack depth (i.e., a
relative crack depth of 0.6). Further Figure 4 shows that the relative first natural frequency for a
flat decrease as the relative depth of the crack increases and is minimum at a crack depth of
120mm (i.e., a relative crack depth of 0.6). and also the relative first natural frequency for a flat
decrease as the location of the crack is getting nearer to the fixed support and is minimum at L/16
location (i.e., a relative crack location of 0.0625).

40
4. DEVELOPMENT OF FUZZY LOGIC AND ARTIFICIAL NEURAL
NETWORKS ALGORITHMS FOR SINGLE CRACK DETECTION
4.1. Methodology
A single crack identification algorithm using both Fuzzy Logic and Artificial Neural Networks
has been developed and the performance is evaluated. Both the approaches use vibration response
of the structure in its development. These Algorithms for crack detection has been designed with
three inputs (first three relative natural frequencies) and two outputs (relative crack locations,
relative crack depths). A number of different Fuzzy and ANN algorithms have been used to
develop the proposed crack detection methodology. The performance of the best Fuzzy and ANN
algorithm used for crack detection out of the developed models has been compared with the
results obtained from FEA. To check its suitability for the detection of damage accurately
Regression Analysis is performed. Sensitivity Analysis is carried out further to test the robustness
of the results including uncertainties obtained from both the approaches. Finally, a comparative
study is carried out to identify which approach is best suited for crack detection.
4.2. Input and Output Parameters for Fuzzy and ANN algorithms
The Fuzzy and ANN models in this work are developed with three input parameters and two
output parameters.
The input parameters used are as follows:
• Relative first natural frequency = “rfnf”;
• Relative second natural frequency = “rsnf”;
• Relative third natural frequency = “rtnf”;
The output parameters used are as follows:
• Relative crack location = “rcl”
• Relative crack depth = “rcd”
4.3. Fuzzy and ANN Model Properties
4.3.1. Details of Fuzzy Logic Models
The fuzzy models developed in this work are based on Triangular, Trapezoidal and Gaussian
membership functions with three input and two output parameters as described in section 4.2.
The pictorial view of the Triangular, Trapezoidal and Gaussian membership fuzzy models is
shown in Figure 5(a), (b) and (c) respectively.
Figure 5(a): Triangular fuzzy model Figure 5(b): Tripezoidal fuzzy model

41
Figure 5(c): Gaussian fuzzy model
The membership functions used in developing the Triangular fuzzy inference system are shown in
Figure 6(a), (b), (c), (d) and (e). For all the three fuzzy models, thirty membership functions for
each input parameters have been used. Whereas, twelve and ten membership functions have been
used for the two output parameters respectively.
Figure 6(a), (b) and (c): Triangular membership function for relative natural frequency for first, second and
third mode of vibrations respectively

42
Figure 6(d) and (e): Triangular membership function for relative crack depth and relative crack location
respectively
Similar to the above Triangular membership function both Trapezoidal and Gaussian membership
functions were also developed.
The fuzzy linguistic terms and examples of some fuzzy rules (Ten) used to design the fuzzy
interface systems are represented in Table 4 and Table 5 respectively.
Table 4: Description of fuzzy linguistic terms
Membership Functions
Name
Linguistic
Terms
Description And Range Of Linguistic Terms
LF1, LF2 up to LF10 rfnf 1-10
Low ranges of relative natural frequency for first mode of
vibration in ascending order respectively.
MF1, MF2 up to MF10 rfnf 11-20
Medium ranges of relative natural frequency for first mode of
HF1, HF2 up to HF10 rfnf 21-30
High ranges of relative natural frequency for first mode of
LS1, LS2 up to LS10 rsnf 1-10
Low ranges of relative natural frequency for second mode of
MS1, MS2 up to MS10 rsnf 11-20
Medium ranges of relative natural frequency for second mode
of vibration in ascending order respectively.
HS1, HS2 up to HS10 rsnf 21-30
High ranges of relative natural frequency for second mode of
LT1, LT2 up to LT10 rtnf 1-10
Low ranges of relative natural frequency for third mode of
MT1, MT2 up to MT10 rtnf 11-20
Second ranges of relative natural frequency for third mode of
HT1, HT2 up to HT10 rtnf 21-30
High ranges of relative natural frequency for third mode of
SD1, SD2, SD3, SD4 rcd 1-4
Small ranges of relative crack depth in ascending order
respectively.
MD1, MD2 rcd 5,6
Medium ranges of relative crack depth in ascending order
respectively.
LD1, LD2, LD3, LD4 rcd 7-10
Large ranges of relative crack depth in ascending order
respectively.
NL1, NL2, NL3, NL4 rcl 1-4
Nearer ranges of relative crack depth in ascending order
respectively.
ML1, ML2, ML3, ML4 rcl 5,6
Intermediate ranges of relative crack depth in ascending order
respectively.
FL1, FL2, FL3, FL4 rcl 7-10
Farther ranges of relative crack depth in ascending order
respectively.

43
Table 5: Examples of ten fuzzy rules used in the fuzzy models
Sl. No. Examples of Some rules used in the fuzzy model
1 If fnf is HF9, snf is HS10, tnf is HT9 then rcd is SD1 and rcl is FL4
2 If fnf is HF8, snf is HS10, tnf is HT8 then rcd is SD2 and rcl is FL4
3 If fnf is HF9, snf is HS4, tnf is HT5 then rcd is SD4 and rcl is ML1
4 If fnf is HF3, snf is HS10, tnf is MT10 then rcd is MD1 and rcl is FL4
5 If fnf is MF10, snf is HS6, tnf is HT9 then rcd is MD2 and rcl is NL1
6 If fnf is LF8, snf is HS4, tnf is HT7 then rcd is LD4 and rcl is NL2
7 If fnf is MF9, snf is HS9, tnf is HT3 then rcd is LD2 and rcl is NL3
8 If fnf is HF6, snf is HS5, tnf is MT2 then rcd is LD1 and rcl is NL4
9 If fnf is HF5, snf is HS3, tnf is LT4 then rcd is LD3 and rcl is NL4
10 If fnf is HF9, snf is HS7, tnf is HT7 then rcd is SD3 and rcl is ML1
4.3.2. Details of ANN Models
The ANN models developed in this work are based on feed forward backpropagation network
type with three input and two output parameters as described in section 4.2. The back propagation
neural network has one input layer and one output layer and different amount of hidden layers.
The input layer contains three neurons and the output layer contains two neurons corresponding
to the input and output parameters. Whereas the hidden layers contain varying layers of neurons
as shown in figure 7. The neurons associated with the input layer of the network represents the
first three relative natural frequencies and the neurons associated with the output layer of network
represents the relative crack depths and relative crack location.
Figure 7. Neural Network Model.
Six different ANN Algorithm were developed using different number of layers and different
number of neurons in each layer and the best of them is chosen for further discussion. The
network properties and the training parameters of the ANN algorithm are as shown in Table 6.
Table 6. Details of ANN Model.
NETWORK PROPERTIES
Network Type Feed Forward Backpropagation algorithm
Training Function TRAINLM (Levenberg-Marquardt backpropagation)
Adaption Learning Function LEARNGDM
Performance Function MSE (Mean Squared Normalised)
Transfer Function TANSIG
Number of Hidden Layers 1 in first 3 Cases and 2 in the next 3 Cases
Number of Neurons in each layer 8, 10 and 12 respectively

44
DATA DIVISION (RANDOM)
Training Data 70% (58 Samples)
Validation Data 15% (13 Samples)
Test Data 15% (13 Samples)
TRAINING PROPERTIES
Show Window TRUE
Show Command Line FALSE
Show 25
Epochs 1000
Time Inf
Min Grad 1e – 05
Max Fail 1000
Mu 0.001
Mu_Max 10000000000
4.4. Results and Discussions
4.4.1. Fuzzy Results
The results obtained from ANSYS, Fuzzy Triangular, Trapezoidal and Gaussian model after
defuzzification are compared in Table 7 (Only few cases).
Table 7. Comparison of results between Fuzzy Triangular, Trapezoidal and Gaussian models.
It can be seen for the Table 7 that the results obtained from the Fuzzy Triangular, Trapezoidal and
Gaussian models are in close proximity with the outputs obtained from the FEA results. Thus the
developed fuzzy crack detection model/systems can be effectively used in damage detection and a
continuous structural health monitoring system can be developed for any system.
To determine the best suited model out of the three models developed, an analysis of R2
value
(Regression plots) has been done and results obtained are as tabulated in Table 8. Also the
comparison of the values is shown in Figure 8(a) and (b) for relative crack depths and relative
crack locations respectively.
ANSYS Fuzzy Models
Relative
Crack
Depth
Relative
Crack
Location
Triangular Model
Trapezoidal
Model
Gaussian Model
rcd rcl rcd rcl rcd rcl rcd rcl
0.05 0.5 0.0751 0.208 0.0751 0.139 0.0762 0.22
0.1 0.25 0.0883 0.25 0.092 0.25 0.0926 0.248
0.15 0.1667 0.113 0.278 0.125 0.28 0.114 0.28
0.2 0.125 0.126 0.119 0.13 0.13 0.13 0.161
0.25 0.1 0.249 0.102 0.325 0.275 0.227 0.0985
0.3 0.0833 0.299 0.0504 0.325 0.275 0.3 0.0527
0.35 0.0625 0.35 0.0501 0.35 0.0501 0.35 0.051
0.4 0.5 0.4 0.5 0.4 0.5 0.4 0.499
0.45 0.25 0.45 0.25 0.45 0.25 0.45 0.25
0.5 0.1667 0.5 0.2 0.5 0.2 0.5 0.2
0.55 0.125 0.55 0.15 0.55 0.15 0.55 0.15
0.6 0.1 0.598 0.102 0.375 0.275 0.598 0.0999
0.6 0.0833 0.325 0.275 0.375 0.275 0.555 0.0756
0.6 0.0625 0.325 0.275 0.375 0.275 0.597 0.0528

45
Table 8. Comparison of R2
values between original output and the outputs obtained from Fuzzy Triangular,
Trapezoidal and Gaussian models.
SL.No Type of Fuzzy Model
R2
rcd rcl
1 Triangular Fuzzy Model 80.30% 73%
2 Trapezoidal Fuzzy Model 75.90% 53.40%
3 Gaussian Fuzzy Model 98.70% 88.40%
Note: A 0.05% level of significance i.e., a 95% level of confidence is assumed in determining the R2
values
(a) rcd (b) rcl
Figure 8: R2
values between original output and output obtained from fuzzy models for Relative Crack
Depth and Relative Crack Location.
It can be seen from Table 8 and Figure 8(a) and (b) that the results obtained from the Fuzzy
Gaussian model has the highest R2
value when compared to the Fuzzy Triangular and Trapezoidal
models for the determination of both relative crack depths and relative crack locations. Thus
Fuzzy Gaussian model can be effectively used in damage detection since it produces the best
results for both finding both rcd and rcl.
4.4.2. ANN Results
Six different ANN models were developed and the results obtained are as shown in Table 9. The
best ANN Algorithm out of the six networks developed is the one that has the highest percentage
of R2
value.
Table 9. Structure and Performance of different ANN models developed.
Sl.no.
Number of
Hidden Layers
Numbers of
Neurons
ANN
structure
Iteration
number
Training
Time (sec)
R2 value
in %
1 1 8 3--8--2 24 1 97.82
2 1 10 3--10--2 79 1 98.23
3 1 12 3--12--2 83 1 97.5
4 2 8 3--8-8--2 164 1 95.97
5 2 10 3--10-10--2 92 4 97.3
6 2 12 3--12-12--2 781 7 97.6
The 3-10-2 ANN Structure was found to be the best suitable ANN model for crack detection with
a 98.23% R2
value.
The results obtained from the ANN Model developed are in Table 10 (only few cases). Also the
R2
value are as shown in Table 11.

46
Table 10. Comparison of results obtained from ANN Model and FEA.
Case No. Original (FEA) ANN
rcd rcl rcd rcl
1 0.05 0.5 0.07519 0.24083
2 0.1 0.25 0.07943 0.27894
3 0.15 0.16667 0.11177 0.21126
4 0.2 0.125 0.17028 0.07495
5 0.25 0.1 0.23316 0.08305
6 0.3 0.08333 0.29177 0.08197
7 0.35 0.0625 0.34208 0.06834
8 0.4 0.5 0.40433 0.5
9 0.45 0.25 0.45149 0.25281
10 0.5 0.16667 0.54498 0.17558
It can be seen from the Table 10 that the results obtained from the Artificial Neural network
model are in close proximity with the outputs obtained from the FEA results. Thus, the developed
ANN crack detection model can be effectively used in damage detection and a continuous
structural health monitoring system can be developed for any system.
Table 11. Structure and Performance of different ANN models developed.
Type
R2
rcd rcl
ANN Model 98.30% 92%
Note: A 0.05% level of significance i.e., a 95% level of confidence is assumed in determining the R2
values.
It can also be seen from the Table 11 that the results obtained from the ANN model has very good
R2
value for the determination of both relative crack depths and relative crack locations. Thus
ANN model can be effectively used in damage detection since it produces the best results for
finding both rcd and rcl.
4.5. Sensitivity Analysis on the Fuzzy and ANN Models Developed
In order to know the robustness of the Fuzzy and ANN models developed a sensitivity analysis
has been made. This is carried out by changing the values of relative first natural frequency,
relative second natural frequency and relative third natural frequency by 0.1% one after the other
respectively. The details of the results obtained in terms of the R2
values are tabulated in Table 12
and 13. Also the comparison of these values are shown in Figure 9(a) and (b), 10(a) and (b), 11(a)
and (b), 12(a) and (b) for rcd and rcl respectively.
Table 12. R2
value after sensitivity analysis for the ANN model developed.
SL
No
Type
R2
rcd rcl
1 ANN Model 98.30% 92%
2 0.1% increase in rfnf 98.20% 85.70%
3 0.1% increase in rsnf 98.20% 90.70%
4 0.1% increase in rtnf 98.20% 89.20%

47
Table 13. R2
value after sensitivity analysis for the ANN model developed.
(a) rcd (b) rcl
Figure 9: Comparison of R2
Values for Relative crack depth (a) and Relative crack location (b) by
increasing rfnf, rsnf and rtnf by 0.1%.
(a) rcd (b) rcl
(a) rcd (b) rcl
SL.
No.
Type of
Fuzzy Model
R2
value – for 0.1%
increase in RFNF
R2
value – for 0.1%
increase in RSNF
R2
value – for 0.1%
increase in RTNF
rcd rcl rcd rcl rcd rcl
1 Triangular 54.80% 53% 80.30% 73% 80.30% 73%
2 Trapezoidal 46.90% 39.80% 75.90% 53.40% 75.90% 53.40%
3 Gaussian 98.00% 84.40% 98.70% 88.40% 98.70% 88.40%

48
(a) rcd (b) rcl
It can be seen for the Table 12 and 13 that when a sensitivity analysis was done by changing the
input values in order to check whether the proposed model can incorporate uncertainties in it, the
results obtained from the ANN and Fuzzy Gaussian models with and without change in the input
values has the R2
value close to each other when compared to the Original ANN, Fuzzy
Triangular and Trapezoidal models for the determination of both relative crack depths and
relative crack locations respectively. Thus ANN and Fuzzy Gaussian models can be effectively
used in damage detection even with some uncertainties in the input data since it produces the best
results for finding both rcd and rcl.
5. COMPARISON OF FUZZY AND ANN ALGORITHMS
A comparison was made between the best Fuzzy Model and best Artificial Neural Networks
Model developed and the results obtained are as tabulated below in Table 14 and Figure 13(a) and
(b). The results tabulated in Table 15 and shown in Figure 14(a) and (b) were obtained after
sensitivity analysis.
Table 14: R2
values for ANN Model and Gaussian Fuzzy Model developed
SL.No Type
R2
rcd rcl
1 ANN Model 98.30% 92%
2 Gaussian Fuzzy Model 98.70% 88.40%
(a) rcd (b) rcl
Figure 13: R2
values for ANN Model and Gaussian Fuzzy Model developed for Relative crack depth and
relative crack location Respectively
It can be seen from the Table 14 that both ANN model and Gaussian Fuzzy model gave a close R2
value to each other. Out of which Gaussian Fuzzy Model performed better in determining
Relative crack depth and ANN model performed better in determining Relative crack location.

49
Table 15: R2
values for ANN Model and Gaussian Fuzzy Model developed before and after performing
sensitivity analysis
SL.No Type
R2
rcd rcl
1 ANN Model 98.30% 92%
Gaussian Fuzzy Model 98.70% 88.40%
2 0.1% increase in rfnf (ANN) 98.20% 85.70%
0.1% increase in rfnf (FGM) 98.30% 84.40%
3 0.1% increase in rsnf (ANN) 98.20% 90.70%
0.1% increase in rfnf (FGM) 98.70% 86.80%
4 0.1% increase in rtnf (ANN) 98.20% 89.20%
0.1% increase in rfnf (FGM) 98.30% 82%
Figure 14(a): R2
values for ANN Model and Gaussian Fuzzy Model developed for Relative crack depth
before and after performing sensitivity analysis
Figure 14(b): R2
values for ANN Model and Gaussian Fuzzy Model developed for Relative crack location
before and after performing sensitivity analysis.
It can be seen for the Table 15 that both ANN model and Gaussian Fuzzy model gave slightly
different R2
value on the results obtained after the sensitivity analysis. Out of which Gaussian
Fuzzy Model performed better in determining Relative crack depth and ANN model performed
better in determining Relative crack location.
6. CONCLUSIONS
Whenever there is a damage introduced in a structure there is a change in the dynamic properties
of the structure such as Natural Frequency and Mode Shapes these changes are less predominant
unless the damage is very severe. Natural frequency of a structure decreases as the depth of the
crack increases since there is a reduction in the stiffness of the structure due the development of
crack. Based on the above details this paper presents two crack detections algorithms one using
Fuzzy Logic and other using Artificial Neural Networks on a fixed beam like structure (flat).
Further a comparison has been made to identify which algorithm is best suited for a single crack
detection. Both ANN and Fuzzy logic can be used in developing an effective damage detection
algorithm based on the results obtained. Out of the three types of fuzzy models developed, the
Fuzzy Gaussian model gave the best results when compared to the Fuzzy Triangular and
Trapezoidal Models. Even while handling the input data with some uncertainties, both ANN and

50
Fuzzy Gaussian model gave almost accurate results when compared to the other two models.
Between the Fuzzy Gaussian Model and ANN Model developed the Fuzzy Gaussian Model was
more accurate in determining the relative crack depth whereas the ANN Model was more
accurate in determining the relative crack location. Therefore, a research can be further carried
out in developing a hybrid model combining the ANN and Fuzzy Logic which can be an effective
method to detect damage in structures. Care should be taken to select the best type of membership
function to be used in case of fuzzy logic. Whereas, in case of ANN determination of optimum
number of hidden layers and the number of neurons that should be present in each layer is crucial.
REFERENCES
[1] L. A. Zadeh, (1965) “L. A. Zadeh: Fuzzy Sets”, Information and Control, Vol. 8, pp338-358.
[2] Kamalanathan S and Selvaraju S, (2012), “Collaborative Approaches for Personalized Web Search
Using Fuzzy Neural Networks”. In: Krishna P.V., Babu M.R., Ariwa E. (eds) Global Trends in
Information Systems and Software Applications. Communications in Computer and Information
Science, Vol. 270, pp367-376.
[3] G.Vellidis, M.T. C.Perry, C.Kvien and C.Bednarz, (2008), “A real-time wireless smart sensor array
for scheduling irrigation”, Computers and electronics in agriculture, Vol. 61, pp44–50.
[4] M. H. Richardson, (1980) “Detection of damage in structures from changes in their dynamic (Modal)
properties-a survey,” Tech. Rep. NUREG/CR-1431, UCLR-15103, Structural Measurement Systems,
Washington, DC, USA, 1980.
[5] A. Rytter, (1993), “Vibration based inspection of civil engineering structures” [Ph.D. thesis], Aalborg
University, Copenhagen, Denmark.
[6] A. D. Dimarogonas, (1996) “Vibration of cracked structures: a state of the art review,” Engineering
Fracture Mechanics, vol. 55, no.5, pp831-857.
[7] R. Ganguli, (2001) “A fuzzy logic system for ground based structural health monitoring of a
helicopter rotor using modal data”, Journal of Intelligent Material Systems and Structures, vol. 12,
no.6, pp397–407.
[8] L. J. DeMiguel and L. F. Bl´azquez, (2005) “Fuzzy logic-based decision making for fault diagnosis in
a DC motor”, Engineering Applications of Artificial Intelligence, vol. 18, no. 4, pp423-450.
[9] Bakhary, N., Hao, H. & Deeks A.J, (2007), “Damage detection using artificial neural network with
consideration of uncertainties”, Engineering Structures, Vol. 29, pp2806-2815.
[10] M. Chandrashekhar and R. Ganguli, (2009) “Damage assessment of structures with uncertainty by
using mode-shape curvatures and fuzzy logic”, Journal of Sound and Vibration, vol. 326, no. 3-5,
pp939–957.
[11] Parhi, D.R. & Das, H.C, (2009), “Application of Neural network for fault diagnosis of cracked
cantilever beam”, World Congress on Nature and Biologically Inspired Computing, NABIC, pp1303-
1308.
AUTHORS
B. Prakruthi Gowd, Assistant Professor, Civil Engineering Department and Presently
pursuing Ph.D in the field of Structural Health Monitoring. Gold medalist in M.Tech,
Structural Engineering. The research area is Damage Detection in Structures using
Vibration based Damage Detection Techniques (Wavelet Transforms, Fuzzy Logic and
Artificial Neural Network).

51
Dr. K. Jayasree, Assistant Professor, Civil Engineering Department has obtained PhD
in Transportation Engineering from JNT University and has 12 years of teaching and/
research experience. She has also worked as a scientist in NATPAC, Kerala. She has
developed an operational tool for urban planners to address mobility problems in urban
areas. Her research area includes application of fuzzy logic and multi criteria
evaluation techniques to civil engineering problems, application of GIS in
transportation planning and infrastructure design.
Dr. Manjunatha N. Hegde, Dean-Academic, Professor of Civil Engineering, has
obtained PhD degree in Civil Engineering from Indian Institute of Science (IISc) and
has 31 years of teaching and/ research experience. His research area includes FEM in
Stochastic Structural Dynamics & Earthquake Engineering, Fibre Reinforced Concrete,
System Identification, Damage Assessment and health Monitoring of Civil
Infrastructures, and Life Cycle Energy Cost of Buildings. He is the member of
Professional bodies like IEI, ISTE, ISET, ICI, ACCE (I), and INSTRUCT. He is the
Member of Board of Studies and/ Examinations of Autonomous Engineering Colleges
under VTU.

COMPARISON OF ARTIFICIAL NEURAL NETWORKS AND FUZZY LOGIC APPROACHES FOR CRACK DETECTION IN A BEAM LIKE STRUCTURE

More Related Content

What's hot (18)

Similar to COMPARISON OF ARTIFICIAL NEURAL NETWORKS AND FUZZY LOGIC APPROACHES FOR CRACK DETECTION IN A BEAM LIKE STRUCTURE (20)

Recently uploaded (20)

COMPARISON OF ARTIFICIAL NEURAL NETWORKS AND FUZZY LOGIC APPROACHES FOR CRACK DETECTION IN A BEAM LIKE STRUCTURE