Mechanical Systems and Signal Processing ()

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

The use of a non-probabilistic artificial neural network to

consider uncertainties in vibration-based-damage detection
Khairul H. Padil a, Norhisham Bakhary a,b,n, Hong Hao c
a
Faculty of Civil Engineering, Universiti Teknologi Malaysia, 81310 UTM Skudai, Johor, Malaysia
b
Institute of Noise and Vibration, Universiti Teknologi Malaysia, City Campus, Jalan Semarak, 54100, Kuala Lumpur, Malaysia
c
Centre of Infrastructure Monitoring and Protection, School of Civil and Mechanical Engineering, Curtin University, Kent St., Bentley WA
6120, Australia

a r t i c l e i n f o abstract

Article history: The effectiveness of artificial neural networks (ANNs) when applied to pattern recognition
Received 24 April 2015 in vibration-based damage detection has been demonstrated in many studies because
Received in revised form they are capable of providing accurate results and the reliable identification of structural
7 June 2016
damage based on modal data. However, the use of ANNs has been questioned in terms of
Accepted 10 June 2016
its reliability in the face of uncertainties in measurement and modeling data. Attempts to
incorporate a probabilistic method into an ANN by treating the uncertainties as normally
Keywords: distributed random variables has delivered promising solutions to this problem, but the
Vibration-based damage detection probabilistic method is less straightforward in practice because it is often not possible to
Artificial neural network
obtain unbiased probabilistic distributions of the uncertainties. Moreover, the probabil-
Uncertainties
istic ANN method is computationally complex, especially when generating output data. In
Non-probabilistic
Interval analysis this study, a non-probabilistic ANN is proposed to address the problem of uncertainty in
vibration damage detection using ANNs. The input data for the network consist of natural
frequencies and mode shapes, and the output is the Young's modulus (E values), which
acts as an elemental stiffness parameter (ESP). Through the interval analysis method, the
noise in measured frequencies and mode shapes are considered to be coupled rather than
statistically distributed. This method calculates the interval bound (lower and upper
bounds) of the ESP changes based on an interval analysis method. The ANN is used to
predict the output of this interval bound by considering the uncertainties in the input
parameters. To establish the relationship between the input parameters and output
parameters, a possibility of damage existence (PoDE) parameter is defined for the un-
damaged and damaged states. A stiffness reduction factor (SRF) is also used to represent
changes in the stiffness parameter. A numerical model and a laboratory-tested steel portal
frame demonstrate the efficacy of the method in improving the accuracy of the ANN in the
presence of uncertainties. The effect of different severity levels and the influence of dif-
ferent noise levels on the identification results are discussed.
& 2016 Elsevier Ltd. All rights reserved.

1. Introduction

The vibration-based damage detection method has been widely used to increase the safety of civil engineering

n
Corresponding author at: Faculty of Civil Engineering, Universiti Teknologi Malaysia, 81310 UTM Skudai, Johor, Malaysia.
E-mail address: norhisham@utm.my (N. Bakhary).

http://dx.doi.org/10.1016/j.ymssp.2016.06.007
0888-3270/& 2016 Elsevier Ltd. All rights reserved.

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i
2 K.H. Padil et al. / Mechanical Systems and Signal Processing ()

structures. Many approaches based on this method have been proven to be effective in addressing problems in both basic
and complex structures. This approach explains that damage can affect both the physical and dynamic characteristics of the
structural properties. Physical characteristics include the mass, stiffness and damping, while dynamic characteristics include
the frequency response functions (FRFs), natural frequencies, damping ratio and mode shapes. For example, changes in both
the mode shape and natural frequencies result when the stiffness of a beam structure is reduced. Vibration-based damage
detection methods range from conventional direct methods, such as measurements of changes in frequencies [1,2], mode
shapes [35] and FRFs [6,7], to the application of advanced computational methods such as fuzzy logic [8], genetic algo-
rithms [9] and artificial neural networks (ANNs) [10,11].
One of the best explored computational approaches to vibration-based damage detection is ANNs. The ANN technique
has proven effective in damage detection due to its capability to model the nonlinear relationship between the vibration
parameters and the damage location and severity [1217]. ANNs represent a type of computing that is inspired by the
structure and the information processing capability of the human brain. This model mimics the real life behavior of neurons
and the electrical messages they produce between the input processing of the brain and the final output from the brain.
Once trained, an ANN is capable of pattern recognition and classification.
The issues of uncertainty become more significant as civil engineering structures become more complex. There are two
types of unavoidable uncertainties in the application of ANNs in damage detection: modeling error and measurement noise.
Modeling error refers to the existence of uncertainties in the finite element model (FE model) due to the inaccuracy of
physical parameters, non-ideal boundary conditions, finite element discretization and nonlinear structural properties. This
may result in the FE model not representing the exact behavior of the modeled structures, leading to the trained ANN model
using data generated from the FE model being unable to represent the exact relationship between the input parameters
(vibration data) and the damage information. On the other hand, noise in measurement data that are normally used in the
testing phase may also lead to inaccurate damage identification. Bakhary et al. [10,18] demonstrated the effects of un-
certainties in training and testing data on the performance of ANNs in structural damage detection, and they suggested the
use of a probabilistic ANN method to consider the existence of uncertainties in the FE model and measurement data. The
study demonstrated that the probabilistic method is capable of facilitating accurate damage detection based on noisy data.
The probabilistic method was also proven efficient in handling noisy data in several other studies [10,17,19]. Despite its
efficiency in handling noisy data, the probabilistic method suffers from several disadvantages such as its assumption of
uncertain parameters as normally distributed random variables with a given variance. In practice, it is not possible to obtain
the probability density function due to the complexity of the sources of uncertainty [20,21]. Furthermore, insufficient data in
experimental studies also reduce the capability of obtaining an unbiased probability density function. Moreover, the
probabilistic method requires multiple sets of data for the probabilistic analysis. These data sets are normally generated
through an established FE model or ANN model based on a specific standard deviation. This process involves an iterative
process of simulation that is very time consuming.
There have been several attempts to apply non-probabilistic methods to overcome the drawbacks of statistical methods.
Most studies have used interval analysis to replace the uncertain parameters with intervals to produce an output that is also
an interval. For example, Gabriele et al. [22] demonstrated the applicability of interval analysis in considering uncertainties
in the context of model updating. The authors considered the uncertainties in frequencies to obtain intervals for the stiffness
values. Damage detection was performed by calculating the stiffness central values. Garcia et al. [23] further demonstrated
the efficacy of the interval analysis method by applying it to uncertainties in material properties. The existence of damage
was determined by observing the load at which the structure changes from exhibiting elastic behavior to nonlinear be-
havior. Recently Wang et al. [24] improved the method by providing a new damage measure index (DMI) that combines the
stiffness reduction factors (SRFs) and the possibility of damage existence (PoDE) using a non-probabilistic interval analysis
framework. In their study, both the physical and dynamic characteristics are measured using interval analysis. Using a Taylor
series expansion, the uncertainties of the interval bounds of the physical and dynamic characteristics of the damaged and
undamaged states are derived based on the results of the FE model. Compared to the conventional probabilistic analysis,
non-probabilistic interval analysis does not require any assumption about the uncertainties distribution in the calculation of
the PoDE. It requires only the upper and lower bounds of the uncertain parameters, which can generally be obtained in real
engineering problems. Thus, damage detection with noisy data becomes simpler and computationally less complex com-
pared to the approaches based on the conventional probabilistic method.
This study extends the applicability of the non-probabilistic method using ANNs to consider uncertainties in damage
detection. For this purpose, an ANN is trained using frequencies and mode shapes as inputs and elemental stiffness para-
meters (ESPs) as the output variables. The uncertainties are considered by calculating the lower and upper bounds of the
input parameters based on interval mathematics to produce the lower and upper bounds of the output parameters (ESP).
Therefore, two identical ANNs are used to determine the upper and lower bounds of the noisy input data. SRFs and PoDE are
later used to define the reduction in the ESP values and the existence of damage. To provide a better indicator of damage
existence, the DMI is adopted. The applicability of the proposed method is demonstrated through a numerical model and a
lab experiment on a steel portal frame. A sensitivity study on the effects of different levels of uncertainties on the input and
output parameters is also conducted. The results show that the proposed method is able to efficiently provide the location
and severity of damage.

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i
 K.H. Padil et al. / Mechanical Systems and Signal Processing () 3

Fig. 1. Multilayer perceptron ANN architecture.

2. Methodology

2.1. Artificial neural network model

In this study, the input parameters are the natural frequencies ( ) and mode shapes ( ), and the outputs are the ESPs ().
Because the responses of structural vibrations are measured only along a limited number of axes in practice, the mode
shapes are not complete; thus, the direct relationship between the vibration parameters and damage parameters is difficult
to establish. Because ANNs have been proven to be efficient for establishing the nonlinear relationship between vibration
parameters (frequencies and mode shapes) with structural damage parameters [2527], in this study, a multilayer per-
ceptron LevenbergMarquardt back-propagation learning algorithm is used as the ANN model. The reason for this is that
multilayer perceptron networks have been applied successfully to many different problems [28] and have been proven to be
a universal approximator. Such networks can approximate any continuous multivariate function to any degree of accuracy
[29]. In vibration damage detection, the Levenberg-Marquardt back-propagation learning algorithm has been proven to be
superior in relating vibration parameters with structural parameters compared to many other types of learning algorithms
[30]. The ANN architecture is shown in Fig. 1.
An ANN requires training to establish the relationship between input and output parameters. In the training phase, a
series of damage cases are randomly generated using a FE model. The damage cases are idealized by reducing the ESPs of
selected elements. To assess the reliability of the trained ANN model, several damage cases are generated to test the effi-
ciency of the trained ANN model. The stiffness reduction ratio (SRF), as shown in Eq. (7), represents the changes in the
stiffness parameter for each segment.
Tangent sigmoid transfer functions are employed for all layers because the input and output vectors are normalized
between
1 and 1. Early stopping method is used as the stopping criteria due to its superiority in avoiding overfitting
problems in ANN training [31]. Thus, the training data are divided into 60% for training and 40% for validation. A series of
trial and error processes are performed to select the best number of hidden neurons.
Once the ANN is trained, the testing data are fed into the ANN model to predict the location and severity of damage. For a
numerical example, the testing cases are generated based on the FE model, whereas for experiment, the testing data are the
measured data. However, in reality, both the FE model and the testing data inevitably contain noise due to uncertainties in
the measurements and modeling, which will lead to inaccurate outputs. Therefore, in this study, modeling errors and
measurement noise are considered using non-probabilistic interval analysis. Through this method, the uncertainties are
considered by calculating the lower and upper bounds of the input parameters as well as the ANN outputs (ESPs) with lower
and upper bounds. Once the lower and upper bounds of the ESPs are obtained, the possibility of damage existence (PoDE)
can be calculated, followed by the calculation of the damage measure index (DMI), which is used to indicate the damage
severity.
Based on this method, only two ANN models are needed to predict the upper and lower bounds of the ESP values.
Compared to the conventional statistical ANN method [10], the proposed non-probabilistic method requires fewer ANN
models. In the conventional statistical ANN method, which is based on a point estimate method [32], four ANN models are
required to calculate the statistical behavior of the E values, whereas in this method, only two ANN models are involved.
Because ANN predictions are normally not exact and are contaminated with and
prediction errors, the small number of
required ANN models leads to lower prediction error and greater computational efficiency. This is experienced in the sta-
tistical ANN model, where the prediction error normally increases the standard deviation of the predicted E values. This
higher standard deviation may result in lower probability of damage existence (PDE) when detecting damage in structures.

2.2. Non-probabilistic artificial neural network

To consider uncertainties from the measurements and FE model, the proposed method adopted the basic concept of
interval mathematics by providing the upper and lower bounds of the ANN input parameters ( and ). Through this
method, the ANN models that are trained with the upper and lower input parameters will provide the upper and lower

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i
4 K.H. Padil et al. / Mechanical Systems and Signal Processing ()

bounds of the output parameters. Due to the nature of the ANN model used to establish the relationship between the input
and output via a black box procedure, the basic equation of interval analysis proposed by Polyak and Nazin [33] can be
directly applied to frequencies and mode shapes to produce the intervals of the output parameters (ESP values). The in-
tervals of the ESPs, natural frequencies and mode shapes for the undamaged and damaged states can be formulated as
follows:

[ ] ; = ESP value lower bound (1)

[ ] ; = ESP value upper bound (2)

The upper and lower bars denote the upper and lower bounds of , and , respectively.
Therefore, the interval bounds for each parameter can be derived as

cI = cI , cI = = ciI , ciI
T
{ I
c1, cI2, , ciI } , I
ci (3)

I I
cI = cI , cI =
T
{ ( I T
c1) ,
(cI2 )T , , (cjI )T } , I
cij = cij , cij (4)

cI = cI , cI = = ck I
T
{ I
c1, cI2, , ck
I
} , I
ck
I
, ck (5)

where
c number of damage cases.
i number of modes.
j number of structural nodes.
k number of segments of structures.
and the middle values of the input and output are denoted as
( x + x )
x c = m ( x) =
2 (6)

where x indicates the exact values of the input parameters (frequencies and mode shapes) and output parameters (ESPs).
The upper and lower bars denote the upper and lower bounds of x, respectively.
Based on (Eqs. (1)6), the uncertainties are coupled with and in terms of the interval bounds. These and are used
as input parameters, while the ESPs () are used as outputs. Thus, two ANN models, which include the lower bound and
upper bound analyses, are established. The upper and lower levels of uncertainties from the measurement and the FE model
are considered through plus and minus errors in the training input and testing input, respectively, as shown in Table 1.
where the superscripts Ir and Ie represent the interval variables of training and testing, respectively. The variable
indicates the uncertainty level for the input data by which the values of the uncertainties differ for the natural frequencies
and mode shapes.
Based on the above table, the boundaries (lower and upper bounds) of the input parameters are applied through the
and values of the uncertainties in two different ANN models: ANN1 and ANN2. and are the outputs of the ANN models
and represent the lower and upper bounds of the predicted ESPs of damage case c. Once the lower and upper bounds of the
stiffness parameters are obtained, the PoDE can be calculated.

2.3. Damage index

Three damage indexes, SRF, PoDE and DMI, are used to indicate damage in this study. For the deterministic model, the
damage indicator is defined as the ratio of the ESP change to the initial ESP value, as idealized below:
d
SRF = 1
u (7)

Table 1
Training and testing input and output variables.

Model Training Input Testing Input Output

ANN1 ciIr = ciIr ciIr ( ) ciIe = ciIe ciIe ( ) ck

Ir Ir Ir Ie Ie Ie
cij = cij cij ( ) cij = cij cij ( )

ANN2 ciIr = ciIr + ciIr ( ) ciIe = ciIe + ciIe ( ) ck

Ir Ir Ir Ie Ie Ie
cij = cij + cij ( ) cij = cij + cij ( )

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i
 K.H. Padil et al. / Mechanical Systems and Signal Processing () 5

Fig. 2. Space for damaged and undamaged ESP.

where d ESP value in the damaged state and u ESP value in the undamaged state
To incorporate the changes in the stiffness parameter into the proposed non-probabilistic method, the PoDE introduced
by Wang et al. [34] is applied. The PoDE is calculated by comparing the vectors of the ESPs in the damaged and undamaged
states. The vectors are the interval bounds (lower and upper bounds) of the ESPs, which are the outputs of ANN1 and ANN2
(refer to Table 1), respectively. The expressions are as follows:
T
uI = { I
u1, uI2, , uk
I
} (8)

T
dI = { I
d1, dI2, , dk
I
} (9)

where uI denotes the interval bound for the undamaged ESP ([auk , auk ]) and dI denotes the interval bound for the damaged
ESP ([adk , adk ]).
Fig. 2 shows the spaces of the damaged (dk )and undamaged ESPs (uk ) on two different axes. A solid rectangle denotes
the variation of both intervals. It is being crossed by the failure plane uk dk . The damaged region is hatched, whereas the
undamaged is unshaded. Because the damaged ESP is smaller than the undamaged ESP, the PoDE is defined as the ratio
between the area of the damaged region and the area of the entire region as follows:
A damage
PoDE = possibility (dk < uk ) = 100%
Atotal (10)

The PoDE ranges between 0 and 100%, with 100% indicating a high possibility of damage and 0% indicating that no
damage occurred at that specific element. The PoDE 100% occurs when the interference between uk anddk does not exist,
the upper bound of the damaged ESP (dk ) is equal to or less than the lower bound of the undamaged ESP ( uk ). The state
whereby dk uk is shown in Fig. 3. Fig. 4(a) and (b) show two states of PoDE 100% on one axis, where both are re-
presented by shaded regions. The first states correspond to the condition dk uk , where the distance between the two

Fig. 3. Space for damaged ESP.

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i
6 K.H. Padil et al. / Mechanical Systems and Signal Processing ()

Fig. 4. Situation of uncertain ESPs (a) PoDE 100% with small SRF (b) PoDE 100% with large SRF (c) PoDE is between 100% and 0%.

middle values is small, indicating a small SRF; however, the two intervals separate completely due to their small variations
in uncertainties. The latter figure shows that the variations in uk and dk are larger but still completely separate due to the
large SRF. On the other hand, Fig. 4(c) shows the states where the PoDE is between 100% and 0%, where an overlapping
occurs between the two intervals with significant SRF.
It is seen that using the SRF and the PoDE separately may not be able to provide a direct indication of the damage severity
of the structure. Therefore, this study employed the DMI parameter suggested by Wang et al. [24] to provide a quantitative
measure of damage severity. Moreover, in most applications of the statistical method for damage detection [10,35,36], the
outcomes are given in terms of probabilities, which are unable to indicate the exact damage severity. As an improvement,
the DMI is derived by multiplying the SRF by the PoDE as follows:
DMI =SRF PoDE (11)

where the SRF is calculated based on Eq. (7), with u and d , the middle values, calculated using Eq. (6).

3. Numerical example

For demonstration purposes, a numerical model of the single-span steel portal frame used in Bakhary et al. [10] is
analyzed to verify the proposed approach. Fig. 5 shows the model of the frame. The numerical model is constructed using 30
beam elements, with 10 elements in each member. Rigid connections are applied between the beam and the columns, and
the supports are assumed to be fixed. The material properties are E 2.1  1011 N/m3, 7.67  103 kg/m3, and 0.2. The
cross section of the beam is 40.5  6.0 mm2, and that of the column is 50.5  6.0 mm2. For damage detection purposes, the
frame is divided into 6 segments, as shown in the figure.
The numerical model of the frame is used as a baseline model to generate the training data. In this example, 2000
damage cases with damage levels between 0.1E and 0.9E are applied randomly to each segment. The corresponding vi-
bration frequencies and mode shapes are generated as the training data. The first three modes of the frequencies and mode
shapes are used as the input parameters, and the output parameters are the ESP values of each segment. The mode shapes
are measured at each node of the structure.
Two damage scenarios are simulated to assess the ANN prediction performance. Scenario 1 consists of damages to two
segments of the frame (1 and 4), and scenario 2 has damages in four segments (1, 3, 5 and 6). Both damage scenarios are
simulated by reducing the ESPs of the corresponding segments. Table 2 shows the ESPs for both damage scenarios.
Selecting the optimal number of hidden neuron is important to obtaining the best ANN model. For this purpose, the
number of hidden neurons is varied from 4 to 36 at increments of two, and the best ANN model is determined based on the
best training performance. Fig. 6(a) illustrates the training and validation performance for different numbers of neurons. It is
seen that the training mean squared error (MSE) values decrease when the number of hidden neurons is increased to 20
hidden neurons (MSE 0.0026) and then remain around the same magnitude up to 22 hidden neurons before increasing to

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i
 K.H. Padil et al. / Mechanical Systems and Signal Processing () 7

Fig. 5. Finite element model of the frame.

Table 2
ESPs of damage for scenario 1 and scenario 2.

Segment 1 2 3 4 5 6

Scenario 1 0.4  E 1.0  E 1.0  E 0.2  E 1.0  E 1.0  E

Scenario 2 0.4  E 1.0  E 0.3  E 1.0  E 0.4  E 0.3  E

0.0126 at 36 hidden neurons. The same trend is also observed in terms of validation performance, where the minimum error
occurred at 20 hidden neurons. Based on this result, the ANN with the smallest error (20 hidden neurons) is selected. Fig. 6
(b) shows the training performance of the ANN over different numbers of epoch with 20 hidden neurons using error-free
training data. The figure shows that the training stops at the 34st epoch, with the minimum MSE value indicating that the
ANN is converged and that a satisfactory relationship between input and output parameters is established.
To assess the applicability of the proposed method, damage detection is first performed using a deterministic ANN model
that is trained using error-free data generated from the FE model. This ANN model is tested using two sets of testing data.
The first set is assumed to be measurement noise free, whereas data in the second set are contaminated with measurement
noise of 2% and 15% in the frequencies and mode shapes, respectively. The results in Fig. 7 demonstrate that the ANN is able
to provide satisfactory results in identifying damage in both scenarios when noise-free testing data are used. However,
when noisy data are used, the prediction outputs are no longer accurate as shown in Fig. 8. For example, in case 1, damage is
falsely predicted at segment 3 and the damage severity at segment 4 is underestimated. In case 2, there are significant
prediction errors at segments 2 and 4, and the damage severity at segment 3 is also underestimated. The results indicate
that the ANN is unable to provide a reliable result if used with noisy data. It should be noted that, in these examples, the
assumed damage levels are very high. For example, 0.4E corresponds to a 60% stiffness loss in the respective segment, which
implies a frequency decrease of approximately 37% if all the elements experience the same level of stiffness loss. The current
examples demonstrate that even if the noise level is smaller than the structural damage level, the trained ANN model still
could not provide accurate damage predictions.
Because the established ANN model is trained based on data generated from the FE model and tested with the measured
data, the existence of errors in the FE model and noise in the measured data provide more significant contributions to the
failure in detecting damage rather than significant changes in modal data. Therefore, it is essential to consider the existence
of uncertainties in both the FE model and the measured data when performing damage detection using an ANN model.
Based on the proposed approach, the uncertainties in the FE model and the measurement data are considered. Two
models of ANNs are constructed to predict the lower and upper bounds of the ESPs. Errors of 2% and 15% are respectively
applied to the frequencies and mode shapes of training and testing data based on Table 1. For consistency, the non-prob-
abilistic ANN models are trained using the same damage cases as the deterministic model, and the same ANN architecture
and parameters are applied. Once both models are trained, the frequencies and mode shapes of the same damage scenarios
are fed to the ANN models to predict the lower and upper bounds of the ESPs.
Fig. 9(a) and (b) show the interval situations of every segment in scenarios 1 and scenario 2, respectively. For scenario 1,
the intervals of the uncertainties of the damaged ESP and the undamaged ESP for segment 1 and 4 are smaller than the SRFs,
which make them completely separated, providing 100% PoDE. The intervals of the uncertainties for the remaining

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i
8 K.H. Padil et al. / Mechanical Systems and Signal Processing ()

Fig. 6. (a) ANN performance with different number of neurons (b) ANN performance with increasing number of epochs.

Fig. 7. ANN predictions compared to the actual values using noise-free input data.

segments are within the boundaries of the undamaged ESP, indicating that the segments are undamaged. The same trend is
also observed for scenario 2, where the uncertainty intervals for the damaged segments (1, 3, 5 and 6) are completely
separated from the undamaged segment (PoDE100%), while the uncertainties intervals for the undamaged ESP and da-
maged ESP for the undamaged segments (2 and 4) are approximately equal.
The calculated PoDEs and DMIs are shown in Table 3. In scenario 1, higher PoDE values are obtained at segments 1 and
4 compared to the undamaged segments. The DMI value of segment 4 is also higher than segment 1, thereby indicating that
more severe damage occurred at segment 4 compared to segment 1; both of these segments are true damage locations with
different severity conditions. The same situation is observed for scenario 2, where higher PoDE values observed at segments
1, 3, 5 and 6 compared to other segments indicate that the damage location is correctly identified. The DMI values are also
higher at the segments with more severe damages. Moreover, the DMI values are close to the actual ESP values and are able
to indicate the severity of the damaged segments. Therefore, undamaged and damaged segments can be distinguished, and
the severity levels of the damaged segments can be assessed. These results demonstrate that the proposed method provides

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i
 K.H. Padil et al. / Mechanical Systems and Signal Processing () 9

Fig. 8. ANN prediction for scenarios 1 and 2 compared to the actual values using noisy input data.

Fig. 9. Interval situation for (a) Scenario 1 (b) Scenario 2.

Table 3
PoDEs and DMIs for the six segments of the frame.

Segment no. Scenario 1 Scenario 2

PoDE (%) DMI (%) PoDE (%) DMI (%)

1 100.00 60.32 100.00 60.86

2 0.00 0.00 0.00 0.00
3 0.00 0.00 100.00 68.68
4 100.00 80.91 0.00 0.00
5 0.00 0.00 100.00 57.22
6 0.00 0.00 100.00 67.42

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i
10 K.H. Padil et al. / Mechanical Systems and Signal Processing ()

Table 4
PoDEs and PDEs for the six segments of the frame.

Segment no. Scenario 1 Scenario 2

PoDE (%) PDE [10] (%) PoDE (%) PDE [10] (%)

1 100.00 96.00 100.00 99.70

2 0.00 38.50 0.00 1.00
3 0.00 3.00 100.00 100.00
4 100.00 99.80 0.00 13.50
5 0.00 0.00 100.00 99.70
6 0.00 0.00 100.00 99.90

an improvement over the conventional statistical method by providing a more meaningful damage severity indication
compared to the statistical ANN method.
To further prove the capability of the proposed method, Table 4 compares the PoDE values obtained by the proposed
method with the PDEs values for the same damage cases calculated using a statistical method developed by Bakhary et. al
[10]. From the table, the PoDE is shown to be a more accurate damage indicator, generating smaller errors in both scenarios.
For example, for scenario 1, in segment 2, which is undamaged, the proposed method shows a 0% PoDE value compared to a
38.5% PDE value. The same situation occurred in scenario 2, where undamaged segment 4 shows 0% damage, and the
statistical method indicates a 13.5% probability of damage. It is also observed that for both scenarios, the proposed method
provides higher PoDE values at the damaged segments compared to the PDE value. The main reason is because, as men-
tioned earlier, the proposed non-probabilistic method used fewer ANN models than the statistical method, which generated
smaller prediction errors.

3.1. Parametric study

A more detailed study is conducted to investigate the sensitivity of the proposed method to different levels of damage
and different levels of uncertainties. To assess the sensitivity of the proposed method to different levels of damage, the
damage severities of scenario 2 are varied while the damage locations remain the same. The damages are located at seg-
ments 1, 3, 5 and 6, with SRFs equal to
25%,
20%,
10%,
7%,
5% and
2%. The uncertainty is fixed at 2% and 15% for
the frequencies and mode shapes, respectively. Table 5 shows the values of the PoDEs and DMIs with different damage
severities. The PoDE and the calculated DMI values are higher at the damaged segments, with values close to the damage
severity, thus indicating that the method is able to provide a reliable indication of damage severity. However, some false
identification is also likely. For example, at the
10% damage level, the PoDE value of the undamaged segment 2 is 100%,
although the DMI value is small, implying false identifications. Moreover, at the
2% damage level, the PoDE and DMI value
at the damaged segment 1 are approximately 71% and 0.38%, respectively, implying that the method failed to identify this
true damage with 100% confidence and also underestimated the damage level. This is because when the damage is small,
there are no significant changes in the modal parameters, and the influences of the uncertainties become prominent.
Nonetheless, the above examples demonstrate the superiority of the proposed method, which correctly identified most
damages, even when the damage level is lower than the noise level.
Next, to study the influence of different noise levels on the identification results, five sets ANN model are developed. The
models are trained and tested with different levels of uncertainties. The uncertainty levels for the frequencies and mode
shapes, respectively, are (1) 0% and 0%, (2) 0.5% and 5%, (3) 1% and 10%, (4) 2% and 15% and (5) 5% and 40%.
Table 6 provides the PoDE and DMI values for different combinations of uncertainties in the training and testing data. The
table shows that higher PoDE and DMI values occurred at the damaged locations, thereby indicating that the proposed
method is capable of detecting damage using noisy data. The table also shows that the accuracy of the ANN model decreases

Table 5
PoDE (%) and DMI (%) for different damage severities in scenario 2.

Severity level (%) Segment

1 2 3 4 5 6

PoDE DMI PoDE DMI PoDE DMI PoDE DMI PoDE DMI PoDE DMI

25 100 24.67 70.97 0.41 100 24.73 0 0.00 100 24.65 100 24.70

20 100 19.58 63.13 0.41 100 19.96 79.97 0.51 100 19.66 100 19.62

10 100 9.81 100 0.35 100 9.99 0 0.00 100 9.97 100 9.85

7 100 6.54 77.41 0.25 100 7.06 49.05 0.32 100 6.16 100 6.60

5 100 4.12 33.25 0.09 100 4.91 0 0.00 100 3.87 100 4.32

2 70.91 0.38 2.71 0.01 100 2.18 0 0 100 1.39 100 1.40

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i
 K.H. Padil et al. / Mechanical Systems and Signal Processing () 11

Table 6
PoDE(%) and DMI (%) for different combinations of uncertainties in the training and testing data.

Train Uncertainties (, ) Segment Test Uncertainties (, )

0%, 0% 0.5%, 5% 1%, 10% 2%, 15% 5%, 40%

PoDE DMI PoDE DMI PoDE DMI PoDE DMI PoDE DMI

0%, 0% 1 100 60.60 100 68.03 100 87.50 100 63.26 58.07 33.32
2 0 0.00 0 0.00 22.25 1.80 0 0.00 94.62 2.43
3 100 66.24 100 65.36 100 52.83 64.75 32.73 100 91.29
4 0 0.00 55.33 2.69 2.94 0.13 34.52 3.49 62.36 13.56
5 100 54.77 100 60.97 100 63.43 100 75.06 100 18.75
6 100 67.22 100 60.50 100 65.56 100 83.48 100 72.13

0.5%, 5% 1 100 54.90 100 57.99 100 60.69 100 68.74 100 58.12
2 0 0.00 5.85 0.00 0 0.00 20.05 2.21 73.52 2.45
3 100 84.28 100 67.40 100 71.16 100 73.55 100 49.52
4 11.50 0.14 0 0.00 0 0.00 51.30 10.53 0 0.00
5 100 60.86 100 57.70 100 65.25 100 67.53 95.47 23.98
6 100 56.00 100 66.66 100 68.50 100 30.18 97.74 23.47

1%, 10% 1 100 40.96 100 61.61 100 61.55 100 56.55 100 15.03
2 0 0.00 34.91 2.77 0 0.00 81.55 11.68 0 0.00
3 100 69.54 100 68.30 100 65.10 100 54.68 100 80.41
4 35.65 1.80 0 0.00 18.20 0.09 34.67 1.74 0 0.00
5 100 40.24 100 57.08 100 58.55 100 60.21 100 88.80
6 100 91.50 100 75.28 100 70.21 100 56.97 33.05 30.95

2%, 15% 1 100 81.40 100 57.73 100 62.66 100 60.86 15.53 10.01
2 0 0.00 13.23 1.46 0 0.00 0 0.00 71.80 1.25
3 84.33 34.73 100 62.97 100 53.26 100 68.68 100 88.55
4 0 0.00 68.35 4.47 15.17 0.28 0 0.00 0 0.00
5 100 52.61 100 52.48 100 61.53 100 57.22 100 47.90
6 78.82 45.16 100 70.97 100 52.01 100 67.42 67.29 31.10

5%, 40% 1 37.14 4.06 100 51.89 100 23.84 100 71.32 100 59.19
2 0 0.00 0 0.00 13.57 2.30 20.81 0.43 0 0.00
3 36.42 14.32 72.47 32.34 88.16 34.59 100 81.94 100 65.95
4 39.80 6.22 45.37 14.08 0 0.00 25.27 2.89 0 0.00
5 50.99 13.88 100 48.59 100 8.57 100 69.55 100 59.80
6 16.36 6.39 29.88 17.44 100 98.79 100 41.21 100 67.15

as the uncertainty level increases. For example, for an ANN model trained with a 0% uncertainty level in both the frequencies
and mode shapes (deterministic model), the PoDE values decreased when tested using data with high noise levels (e.g.,
5 and 40%). The same trend was also observed for the DMI values, which are unable to provide a reliable indication of the
damage severity when the noise level is high. The same situation occurs when higher uncertainties are applied to the
training data. For example, the predicted PoDE and DMI values are less accurate when the ANN model is trained with 5 and
40% uncertainty levels in the frequencies and mode shapes, respectively, and tested with 0% noise in the testing data.
The results also reveal that the highest PoDE and the most reliable DMI values are obtained when the uncertainty levels
in the training data and testing data are at the same level. The reliability of the method decreases as the difference between
the uncertainty level in the training data and testing data increases. This is evident at every level of uncertainty of the
training data, where high values of PoDE and reliable DMIs are obtained if the testing data have the same uncertainty levels.
This is because when the uncertainty level in the training and testing data are different from each other, the true in-
formation is submerged in noise, and thus, the damage cannot be identified accurately. These observations indicate the
importance of defining the uncertainty levels for training the ANN model, which requires profound engineering judgments
and vibration testing experiences, and could be a limitation of the proposed method. However, as demonstrated above, the
proposed method still provides better damage identification results, even when the uncertainty levels used to train the ANN
model are different from the actual uncertainty levels in the testing data, compared to the probabilistic ANN model ap-
proach proposed by the authors in a previous study [10].

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i
12 K.H. Padil et al. / Mechanical Systems and Signal Processing ()

Fig. 10. Segmentation of the frame.

4. Experimental example

A steel frame, as in the numerical model, was fabricated and tested for further verification of the proposed method. The
cross sections of the beam and columns are 40.5  6.0 mm2 and 50.5  6.0 mm2, respectively. The beam and columns were
welded together to represent the rigid connection. The properties of the frame are the same as in the numerical example.
For damage detection purposes, the frame was divided into 15 segments, as shown in Fig. 10. The bottom ends of both
columns were clamped to represent fixed supports. Three damage scenarios were simulated by saw cuts. The cut is 10 mm
deep and 5 mm wide and is applied at both sides of the structural member. Fig. 11 shows the saw cut and the experimental
setup. The number of cuts was increased to impose more severe damage at specified segments. Table 7 lists the damage
cases.
A modal test was conducted using an impact hammer and a sensor placed at node 9, and a roving technique was
adopted. Modal tests were performed at every damage level. Table 8 shows the first three modal frequencies corresponding
to the undamaged and three damage scenarios. The frequencies clearly decreased as the damage severity increased. Fig. 12

Fig. 11. Frame Specimen.

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i
 K.H. Padil et al. / Mechanical Systems and Signal Processing () 13

Table 7
Number of cuts and positions for damage cases 1, 2 and 3.

Segment no. Scenario 1 Scenario 2 Scenario 3

(num. of cuts) (num. of cuts) (num. of cuts)

1 4 5 5
2 None None None
3 None None None
4 None 3 5
5 None None None
6 None None None
7 None None None
8 None None None
9 None None 5
10 None None None
11 None None None
12 None None 5
13 None None None
14 None None None
15 None None None

Table 8
First three frequencies of scenario 1, scenario 2 and scenario 3.

Undamaged Scenario 1 Scenario 2 Scenario 3

Mode 1 4.53 4.46 4.42 4.41

Mode 2 17.60 17.5 17.30 17.00
Mode 3 28.30 28.00 27.80 27.40

Fig. 12. First three mode shapes of the undamaged and damaged cases 1, 2 and 3.

shows the first three mode shapes.

4.1. Damage identification

First, damage identification is conducted using a deterministic ANN model. The ANN model is trained using 2400 random
damage cases generated based on FE model simulations using the same material properties as in the numerical example.
The input parameters are the first three frequencies and mode shapes, and the output parameters are the ESP values of each
segment. Once trained, the ANN model is fed with the experimental data of the three damage scenarios. Fig. 13 shows the
predictions of the three damage scenarios using a deterministic ANN model. The obtained results show that the determi-
nistic model is unable to provide an accurate identification of the three damage states. The existence of modeling error in
the FE model and measurement noise results in inaccurate damage identifications.
Next, based on the procedure described earlier, damage identification is performed using the proposed non-probabilistic
ANN model. The PoDE and DMI values are obtained by assuming that the noise levels are 2% for the frequencies and 15% for
the mode shapes. Table 9 and Fig. 14 shows the PoDE and DMI values respectively.
The results show that the high PoDE and DMI values for the three damage scenarios are at the correct locations. It is also
noted that when the severity level increases, the DMI values also increase. The results indicate that the proposed method is
able to detect damage existence and provide a reliable estimation of damage severity despite the existence of uncertainties
in the FE model and experimental data.

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i
14 K.H. Padil et al. / Mechanical Systems and Signal Processing ()

Fig. 13. ANN prediction using a deterministic ANN model.

Table 9
PoDE (%) for scenarios 1, 2 and 3.

Segment no. Scenario 1 Scenario 2 Scenario 3

PoDE (%) PoDE (%) PoDE (%)

1 100.000 100.000 100.000

2 0.000 0.000 0.000
3 0.000 0.000 0.000
4 0.000 98.724 98.724
5 0.000 0.000 0.001
6 0.000 0.000 0.000
7 0.000 0.000 0.000
8 38.887 0.000 38.887
9 38.196 43.845 87.626
10 0.000 0.000 0.000
11 0.000 0.000 0.000
12 39.461 0.000 80.309
13 0.000 55.436 82.226
14 0.000 0.000 0.000
15 0.000 0.000 0.000

5. Conclusions

In this paper, a non-probabilistic ANN model is proposed to consider the existence of measurement noise and modeling
error in damage detection. An interval analysis is adopted with the ANN to consider the uncertainties using the interval
bounds of the uncertainties in the input parameters of the ANN. The efficacy of this method is demonstrated throughout the
study using numerical and experimental examples. To prove the accuracy of the proposed method, SRF and PoDE are used to
identify the location and reduction in the ESP values. The results of both numerical and experimental examples show that
the proposed method enables reliable damage detection for structures with noisy data. In summary, the proposed non-
probabilistic method is able to provide more accurate damage detection results compared to the probabilistic method due to
the smaller prediction errors. The proposed method is also less time consuming due to the smaller number of ANN models
to be trained.

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i
 K.H. Padil et al. / Mechanical Systems and Signal Processing () 15

Fig. 14. DMI values using a non-probabilistic ANN.

Acknowledgment

The authors would like to thank the Ministry of Higher Education, Malaysia, and Universiti Teknologi Malaysia (UTM) for
their financial support through the Fundamental Research Grant Scheme (4F308 and 4F800) and Research University Grant
Scheme (05H21).

References

[1] F.B. Sayyad, B. Kumar, Identification of crack location and crack size in a simply supported beam by measurement of natural frequencies, J. Vib. Control.
18 (2) (2012) 183190.
[2] N.T. Khiem, L.K. Toan, A novel method for crack detection in beam-like structures by measurements of natural frequencies, J. Sound Vib. 333 (18)
(2014) 40844103.
[3] B.H. Kim, T. Park, G.Z. Voyiadjis, Damage estimation on beam-like structures using the multi-resolution analysis, Int. J. Solids Struct. 43 (14-15) (2006)
42384257.
[4] M. Cao, P. Qiao, Novel Laplacian scheme and multiresolution modal curvatures for structural damage identification, Mech. Syst. Signal Process. 23 (4)
(2009) 12231242.
[5] M. Cao, M. Radzieski, W. Xu, W. Ostachowicz, Identification of multiple damage in beams based on robust curvature mode shapes, Mech. Syst. Signal
Process. 46 (2) (2014) 468480.
[6] A. Furukawa, H. Otsuka, J. Kiyono, Structural damage detection method using uncertain frequency response functions, Comput. Aided Civ. Infrastruct.
Eng. 21 (4) (2006) 292305.
[7] R.P. Bandara, T.H.T. Chan, D.P. Thambiratnam, Frequency Response Function based damage identification using principal component analysis and
pattern recognition technique, Eng. Struct. 66 (2014) 116128.
[8] M. Chandrashekhar, R. Ganguli, Damage assessment of structures with uncertainty by using mode-shape curvatures and fuzzy logic, J. Sound Vib. 326
(3-5) (2009) 939957.
[9] H. Hao, Y. Xia, Vibration-based damage detection of structures by genetic algorithm, J. Comput. Civil Eng. 16 (3) (2002) 222229.
[10] N. Bakhary, H. Hao, A.J. Deeks, Damage detection using artificial neural network with consideration of uncertainties, Eng. Struct. 29 (11) (2007)
28062815.
[11] R.P. Bandara, T.H.T. Chan, D.P. Thambiratnam, Structural damage detection method using frequency response functions, Struct. Health Monit. 13 (4)
(2014) 418429.
[12] X. Wu, J. Ghaboussi, J.H. Garrett, Use of neural networks in detection of structural damage, Comput. Struct. 42 (4) (1992) 649659.
[13] S.V. Barai, P.C. Pandey, Vibration Signature Analysis using artificial neural network, J. Comput. Civil Eng. 9 (4) (1995) 259265.
[14] S.F. Masri, M. Nakamura, A.G. Chassiakos, T.K. Caughey, Neural network approach to detection of changes in structural parameters, J. Eng. Mech. 122
(4) (1996) 350360.
[15] C. Zang, M. Imregun, Structural damage detection using artificial neural networks and measured FRF data reduced via principal component projection,
J. Sound Vib. 242 (5) (2001) 813827.
[16] P.M. Pawar, K. Venkatesulu Reddy, R. Ganguli, Damage detection in beams using spatial fourier analysis and neural networks, J. Intell. Mater. Syst.
Struct. 18 (4) (2006) 347359.

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i
16 K.H. Padil et al. / Mechanical Systems and Signal Processing ()

[17] J.J. Lee, J.W. Lee, J.H. Yi, C.B. Yun, H.Y. Jung, Neural Networks-based damage detection for bridges considering errors in baseline finite element models, J.
Sound Vib. 280 (35) (2005) 555578.
[18] N. Bakhary, H. Hao, A.J. Deeks, Substructuring technique for damage detection using statistical multi-Stage artificial neural network, Adv. Struct. Eng.
13 (4) (2010) 619640.
[19] X.Y. Li, S.S. Law, Damage identification of structures including system uncertainties and measurement noise, AIAA J. 46 (1) (2008) 263276.
[20] Z. Tadesse, K.A. Patel, S. Chaudhary, A.K. Nagpal, Neural Networks for prediction of deflection in composite bridges, J. Constr. Steel Res. 68 (1) (2012)
138149.
[21] J. Shu, Z. Zhang, I. Gonzalez, R. Karoumi, The application of a damage detection method using artificial neural network and train-induced vibrations on
a simplified railway bridge model, Eng. Struct. 52 (2013) 408421.
[22] S. Gabriele, C. Valente, F. Brancaleoni, An Interval uncertainty based method for damage identification, Key Eng. Mater. 347 (2007) 551556.
[23] O. Garca, J. Veh, J. Campos e Matos, A. Abel Henriques, J. Ramon Casas, Structural assessment under uncertain parameters via interval analysis, J.
Comput. Appl. Math. 218 (1) (2008) 4352.
[24] X. Wang, Y. Xia, X. Zhou, C. Yang, Structural damage measure index based on non-probabilistic reliability model, J. Sound Vib. 333 (5) (2014)
13441355.
[25] L.D. Goh, N. Bakhary, A.A. Rahman, B.H. Ahmad, Application of neural network for prediction of unmeasured mode shape in damage detection, Adv.
Struct. Eng. 16 (1) (2013) 99103.
[26] C.Y. Kao, S.L. Hung, Detection of structural damage via free vibration responses generated by approximating artificial neural networks, Comput. Struct.
81 (28-29) (2003) 26312644.
[27] N. Bakhary, H. Hao, A.J. Deeks, Substructing technique for damage detection using statistical multi-stage artificial neural network, Adv. Struct. Eng. 13
(4) (2009) 619639.
[28] D.E. Rumelhart, G.E. Hinton, R.J. William, Learning internal representation by error propagation, in: D.E. Rumelhart, G.E. Hinton, R.J. William (Eds.),
Parallel Distributed Processing, MIT Press, Cambridge, 1986, pp. 318362.
[29] K. Funahashi, On the approximate realization of continuous mapping by neural networks, Neural Netw. 2 (3) (1989) 183192.
[30] L. Goh, N. Bakhary, A.A. Rahman, B.H. Ahmad, A comparison of artificial neural network learning algorithms for vibration-based damage detection,
Adv. Mater. Res. 163 (2011) 27562760.
[31] L. Prechelt, Automotic early stopping using cross validation: quantifying the criteria, Neural Netw. 11 (4) (1995) 761767.
[32] E. Rosenblueth, Point estimates for probability moments, PNAS 72 (10) (1975) 38123814.
[33] S.A. Nazin, B.T. Polyak, Interval parameter estimation under model uncertainty, Math. Comput. Modell. Dyn. Syst. 11 (2) (2005) 225237.
[34] X. Wang, Z. Qiu, I. Elishakoff, Non-probabilistic set-theoretic model for structural safety measure, Acta Mech. 198 (1-2) (2008) 5164.
[35] I. Yeo, S. Shin, H.S. Lee, S.P. Chang, Statistical damage assessment of framed structures from static response, J. Eng. Mech. 126 (4) (2000) 414421.
[36] Y. Xia, H. Hao, Statistical damage identification of structures with frequency changes, J. Sound Vib. 263 (4) (2003) 853870.

Please cite this article as: K.H. Padil, et al., The use of a non-probabilistic artificial neural network to consider
uncertainties in vibration-based-damage detection, Mech. Syst. Signal Process. (2016), http://dx.doi.org/10.1016/j.
ymssp.2016.06.007i