Int. J. Mach. Learn. & Cyber.

DOI 10.1007/s13042-015-0453-3

ORIGINAL ARTICLE

Structural-damage detection with big data using parallel

computing based on MPSoC
Chi Tran1

Received: 8 April 2015 / Accepted: 22 October 2015

 Springer-Verlag Berlin Heidelberg 2015

Abstract Life of many people and vital social activities handles the volume and variety part of big data.
depend on good functioning of civil structures such as ‘‘MapReduce’’ is a programming model of Google and an
nuclear power plants, large bridges, pipelines, and others associated implementation for processing and generating
especially during and after high winds, earthquakes or large data sets using a group of machines. Despite con-
environmental changes. We can recognize the structural tinuous data (BigData) originated from GPS-RTK and
damage of existing systems through signals of the stiffness WSN is not studied in this paper; only one selected data is
reductions of their components, and changes in their used to determine technical state of structure in real time.
observed displacements under certain load or environ- However, bigdata approach is used in this paper to solve
mental conditions. It can be formulated as an inverse the following problems: Firstly, small number of available
problem and solved by the learning approach using neural data for neural network training may lead to strange, not
networks, NN. Here, network function is referred as an practical solutions, which are beyond the permissible area.
associative memory device capable of satisfactory diag- To overcome this problem, small available data is sup-
nostics even in the presence of noisy or incomplete mea- plemented by numerical data generated from integration of
surements. It is not new idea, however, to obtain, in real Monte Carlo simulation and finite element method—
time; an advanced warning of the onset of durability/ MCSFEM model. Also, the learning approach using neural
structural problems at a stage when preventative action is networks, NN, gives effect which is guaranteed only if this
possible has emerged creating an exciting new field within network has, among the various required conditions, a
civil engineering. It refers to the broad concept of assessing complete set of training data and so good (optimal) archi-
the ongoing in-service performance of structures using a tecture of NN. Integration of a Monte-Carlo simulation
variety of measurement techniques, e.g., global position with NN—MCSNN model in order to find the optimal
system—real time kinematic, GPS-RTK, wireless sensor (global) architecture of NN is selected and formulated to
networks, WSN, and others. It leads to, in this case, the so solve this problem. Both MCSFEM model and MCSNN
called big data. Big data is a term applied to data sets model require expensive computing, which lead to neces-
whose size is beyond the ability of commonly used tradi- sity of the use of parallel computing based on MultiPro-
tional computer software and hardware to undertake their cessorSystem on Chip, MPSoC, architecture. Secondly, the
acquisition, access, and analytics in a reasonable amount of use of MPSoC architecture led to the emergence of mul-
time. To solve this problem, we can use the Hadoop like tiple NN named net of neural network, NoNN, rather than a
Batch processing system using a distributed parallel pro- single NN. The simple approach ‘‘trial and error’’ for a
cessing paradigm called MapReduce, which like system single neural network is insufficient to determine an opti-
mal architecture of NN from NoNN. To deal with this
problem we develop, in this paper, a new distributed par-
& Chi Tran allel processing using the ‘‘Master-Slaver’’ structure, in the
[email protected]
‘‘MapReduce’’ framework rather than the Google’s
1
The Faculty of Geodesy, Geospatial and Civil Engineering, Hadoop like Batch processing system using a distributed
WM University in Olsztyn, Olsztyn, Poland parallel processing paradigm called MapReduce. Our

123
 Int. J. Mach. Learn. & Cyber.

method is based on the ‘‘single instruction multiple data’’, the mechanical aspects of the finite element method in
SIMD, technology using computer with multiprocessor order to ensure uniform structural properties of numerical
systems on a chip (or computer with multiple cores, CPUs) data and relationship of them to characteristics of the
without necessarily Google’s computer program and a studied structures.
group of machines. It is used for both MCSFEM model and On the other hand, NN effect is guaranteed only if this
MCSNN model in order to generate virtual dataset and to network has, among the various required conditions, an
find out the optimal architecture of NN from NoNN. It optimal architecture (good enough). To solve this opti-
enables us to create quickly many numerical data and mization issue, we usually can use the gradient method;
eighty thousand architectures (using computer with eight however, it gives us only a local optimal solution, i.e. the
CPUs) can be created instead of eight thousand (using optimal NN architecture is not guaranteed. We can use, in
computer with one CPU) only in the same computation this case, two global approaches: integration of the multi-
times. Effect obtained is presented in the numerical ple neural networks model with either genetic algorithm or
example for the local damage detection in real time of the Monte-Carlo simulation (global optimization methods).
plane steel truss structures so that one can react quickly The use of genetic algorithm, according to studies of many
when problems appear or to predict new trends in the near authors, is not suitable for NN computing. Then, integra-
future. tion of a Monte-Carlo simulation with NN, MCSNN
model, used to find the optimal (global) architecture of NN
Keywords Damage detection  Big data  Finite element is selected and formulated.
method  Monte-Carlo simulation  Artificial neural Also, to obtain an advanced warning, in real time, of the
network  Parallel computing  Single instruction multiple onset of structural problems at a stage when preventive
data  Multiprocessor systems on chips action is possible has emerged creating an exciting new
field of civil engineering. We need new tools such as global
position system and real time kinematic technology, GPS-
1 Introduction RTK, (see [5] or wireless sensor networks, WSN. It is
continuous observation giving us information about dis-
As we know, changing in any one of the mechanical and placements of structural elements in real time. Information
geometrical properties of the building structures leads to originated from mentioned tools can be a collection of data
change their responses. Many applications of artificial sets in its ‘‘large size’’ framework, referred as the ‘‘big
Neural Network, NN, have been developed, as in [3, 5, 13, data’’ (see [7, 14]). Big Data, according to James [10], is a
16], in application of a PSO-based neural network; [4], in dataset that is large, fast, dispersed, unstructured and
the use of adaptive-network-based fuzzy inference systems beyond the ability of available hardware and software
models and others, and applied to a variety of structures. facilities to undertake their acquisition in a reasonable
Monitoring of the building structures referred as the way to amount of time and space.
compare actual structural responses and values derived Hence, both MCSFEM model and MCSNN model
from designed process in order to obtain an early warning require expensive computing due to mathematical models
of threats at a stage when preventive action is possible. with the following characteristics: Big data, real-time,
Usually, it consists the observation of the system over a multitasking and a lot of parallelism. The performance
period of time using periodically spaced measurements, demanded by this computing leads to necessity of the use
extraction of features from these measurements and anal- of Multi-Processor-System architectures, in a single chip,
ysis of these features to determine the current state of MPSoC, endowed with complex communication infras-
health (or damage-detection) of this system. A supervised tructures, such as networks on chip, NoC. We need, in this
learning mode using of NN, for the structural local damage case, to integrate hardware and software—hardware
detection can be used. It is now not new idea in the field of dependent software rather than the sum of traditional
structural health monitoring. hardware and software (hardware independent software). It
On the one hand, incomplete, insufficient data/small is worthless from a decision making perspective in a range
available data may lead, in the NN training process, to of areas, including business, science, engineering, defense,
solutions beyond the permissible area. Then, we must, in education and others. Computer architecture MPSoC, (see
addition to small available data, generate the so called Katalin [9] in which, embedded applications are changing
virtual/numerical data belonging in the desired range by, in from single processor-based systems to intensive data
this paper, integration of Monte Carlo simulation and the communication and parallel programming techniques with
finite element method, (MCSFEM model). Here, we per- efficient automation of mapping parallel software to par-
form a combination of much creative mathematical possi- allel hardware, enables us to solve our mathematical
bilities of Monte Carlo simulation with high accuracy in model. However, the use MPSoC, architecture led to the

123
Int. J. Mach. Learn. & Cyber.

emergence of multiple NN named Net of Neural Network, 2 Structural damage detection of truss structures
NoNN, rather than a single NN. The simple approach ‘‘trial
and error’’ for a single neural network is insufficient to Generally, from the static displacements U and prescribed
determine an optimal architecture of NN from NoNN. One load F, the static equilibrium of structures in the FEM
can use the Hadoop like Batch processing system using a framework, is represented in the form:
distributed parallel processing paradigm called MapRe- KU ¼ F ð1Þ
duce, which is a programming model of Google and an
associated implementation for processing and generating K denotes a stiffness matrix; K = f(E, S, L, h), where each
large data sets (BigData or discretized stream data [15]) structural element has modulus of elasticity E, cross-sec-
using a group of machines. A new approach presented in tional area S and length L, which is inclined with an angle
the Hadoop like Batch processing system (see [17]) has h measured counterclockwise from the positive global X
evolved and matured over past few years for excellent axis. To find out a value, U, from available information of
offline data processing platform for Big Data, which, K and F we can use a direct problem presented as a Feed-
according to [2], can crunch a huge volume of data using a forward approach for structural damage detection. It is
distributed parallel processing paradigm called Google’s demonstrated by the direct mapping W as:
MapReduce. Here, application developers specify the W : S 2 Rm ! U 2 Rn ð2Þ
computation in terms of a map and a reduce function, and
the underlying MapReduce job scheduling system auto- where, n expresses a number of nodes; m denotes a number
matically parallelizes the computation across a cluster of of structural members. With constant geometry of struc-
machines. It provides a standardized framework for ture, an approach to introduce damage is through variations
implementing large-scale distributed computation. How- in the cross-sectional areas, S, of structural members. The
ever, it does not have the mechanism to identify parallelism structure is described as a set of finite elements categorized
of NN computations integrated with Monte Carlo simula- into undamaged and damaged states in different degrada-
tion, MCS, and finite element method, FEM. tion levels. The structural damage is defined as the maxi-
To solve this problem, we develop a new distributed mum stiffness reduction at one or more local elements or
parallel processing model based on the single instruction the loss of stiffness of structure through a reduction in the
multiple data, SIMD, technology using the ‘‘Master-Slaver’’ cross-sectional areas of any element used in computing kij
structure. The SIMD technology enables us to split big data of the stiffness matrix K. A disturbance process using
on n datasets, processed at the same time on q processors, Monte-Carlo simulation is carried out through variations in
CPUs, rather than processed on single processor CPU by the the cross-sectional areas, Si, i = 1, 2,…, m, which is pre-
single instruction single data, SISD technology. Results will sented in the form:
be obtained from single laptop with multicores, CPUs,
Si ¼ S0  0:5  S0  f ð3Þ
through a combination of multiprocessing systems and par-
allel programming techniques in order to efficient automa- where S0 denotes the initial cross-sectional area of all of
tion of mapping parallel software into parallel hardware, elements of the studied structure; f denotes the uniform
without Google’s computer program and a cluster of random numbers distributed in the interval [0, 1]. The
machines. The computing time using the SIMD technology relations between the static displacement values Uj, j = 2,
will be about n times shorter in compare with the use of SISD 3,…, n and the different values of cross-sectional area Si,
technology for same problem. Multiprocessing systems, i = 1, 2, …, m of each structural element (or stress in each
MPS, according to [1], helps us to solve very large problems element of the structure, ri, i = 1, 2, …, m), is presented in
such as weather modeling, designing systems that satisfy the form:
limited budgets, and improving programmer productivity. It Uj , Si or Uj , ri ; i ¼ 1 : m; j ¼ 1 : n ð4Þ
is used, in this paper, for both MCSFEM model and MCSNN
model in order to generate virtual dataset and to find out the Assuming that a number of experimentally observable
optimal architecture of NN from NoNN. It enables us to displacements, Uobs., are available for different conditions
create quickly many numerical data and eighty thousand of structural damage. The damage-detection problem is
architectures (using computer with eight CPUs) can be cre- formulated as follows: for a given vector Uobs., find the
ated instead of eight thousand only (using computer with one vector S*, such that KUobs. = F, where S* denotes a threat
CPU) in the same computation times. element of structure, which is determined by:

123
 Int. J. Mach. Learn. & Cyber.

 

or may not be trustworthy or uncertain. These characteristic
S ¼ Si : max S0i  Si ð5Þ properties of big data are summarized in Fig. 1b.
i¼1:m
Artificial Neural Network (NN) According to the
It is mathematically equivalent to obtaining the mapping, demands, feed forward and feedback, of our Structural
U, that shows a feed back problem between the known Damage Detection problem, we use the layered feedfor-
response of a structure and the physical state of them. ward network with nonlinear differentiable transfer func-
U : U 2 Rn ! S 2 Rm ð6Þ tions, Tan-Sigmoid and Log-Sigmoid, which enables us to
use back propagation training manner on a representative
Finally, a process of early damage detection is presented in set of input/target pairs rather than training the network on
the following figure. all possible input/target pairs. Generally, NN consists of
interconnected processing elements called nodes or neu-
2.1 Big data (BD) rons that work together to produce an output function. The
two-layer network is shown in Figs. 2 and 3.
Data originated from global position system and real time Where, q denotes a number of input; m—neuron number
kinematic, GPS-RTK, and wireless sensor networks, WSN, of the first layer; h—neuron number of the second layer;
leads to the so called big data. Big Data, according to the k—output number; ai, i = 1, 2, …, k denotes the outputs of
Wikipedia, ‘‘is a collection of data sets so large and com- network. Each input is weighted with an appropriate w.
plex that it becomes difficult to process using on-hand Neurons use Tan-Sigmoid/Log-Sigmoid transfer function
database management tools’’. Dimensions which have to generate their output.
become common for characterizing big data, and which A multiple layers of neurons can have several layers, in
together with volume, according to Gottfried Vossen [12], all cases, it is not obvious how many neural layers and
are called the ‘‘4 Vs of big data’’, are the velocity or the neurons of each to include in the network. As, two neural
speed with which data is produced and needs to be con- layers provide the greater flexibility necessary to model
sumed, the variety data can have, and the veracity the data complex-shaped solution surfaces, and are recommended
comes with. Velocity refers to the fact that data often as a general function approximation. It can approximate
comes in the form of streams which do not give the any function with a finite number of discontinuities, arbi-
respective consumer a chance to store them for whatever trarily well, given sufficient neurons in the hidden (first)
purpose, but to act on the data instantly. Variety means that layer. Two-layer network used in this work is shown in the
data can come in different forms such as unstructured (e.g., abbreviated form as:
text), semi-structured (e.g., XML documents), or structured The network mapping I/O, p $ a, is fit to a given data
(e.g., as a table), and veracity refers to the fact the data may set of q exemplar input vectors {p(q)} and their associated
training output vector {a(k)} by adjusting the weight
vector, W, and L, of coefficients so that a(k) = t(k) (ap-
proximation, t).
(a)
Neural number of each layer Generally, there is no
direct and precise way of determining the most appropriate
number of neurons to include in each neural layer, and this
problem becomes more complicated as the number of
hidden layers in the network increases. A range of different
configurations of neurons is usually considered, and that
with the best performance is accepted. Failures in the use

(b)
Input
p1 W Layer 1 Layer 2 Output
V a1
P2
. . a2
. . .
. . .
.
ak
pq k
q m h
Fig. 1 a Early structural damage detection. b The defining of Big
Data Fig. 2 Two-layer network

123
Int. J. Mach. Learn. & Cyber.

Input 4. Check if there is another task to be calculated, if yes,

Layer1 Layer2
tell slave to continue until all tasks are done (NN
Output computing) and the desired [MSE] value is reached.
W ∑ f ∑ f
L In the mentioned model, assumptions: Two-layer net-
k work and a fixed number of first layer, m* = 50, are
b b determined by trial and error method through many cal-
q culations. Their sensitivities are shown in the numerical
m h
resultset presented in the tables presented in the numerical
Fig. 3 The abbreviated form of two-layer feedforward network example. There we can see clearly changing of MSE
depending on value of m. There are no solutions, from
of NNs may be attributable to inappropriate number of
many calculations, better than the results derived from
neurons. We can usually start by training a relatively large
these assumptions. The best solution is shown in the
network and later reduces in size by removing the neuron
numerical example presented below.
numbers or formulate a mathematical model used to opti-
We want to keep the program as flexible as possible with
mize the NN architecture, q-m-h–k.
respect to the configuration of the MPSoC architecture,
Optimization of the NN architecture In this paper for
including 8 processors, for example, referred as 8 slaves.
layered feedforward network with equality: input num-
Let’s say we want to check 800000 NN architectures, we
ber = output number = number of neurons of second
will create and assign 100000 NN architectures to each
layer = q. Then, q–m–q–q is referred as architecture of
slave.
NN. Optimization of architecture of NN is formulated as:
Resilient back propagation training As, the sigmoid
determining the number of neurons of first layer, m [ M,
transfer functions used in the hidden layers in order to
such that during training, the weights and biases of the
compress an infinite input range into a finite output range
network are iteratively adjusted to minimize the network
are characterized by their slope, which must approach zero
performance function, the average squared error between
as the input gets large. When using steepest descent to train
the network output, a, and the target output, t (mean square
a multilayer network with sigmoid functions, the gradient
error MSE); M denotes the acceptable neurons number (a
can have a very small magnitude caused small changes in
fixed number of first layer, M = 50 for example); N
the weights and biases. The purpose of the resilient back
denotes a number of slaves (NN number). For the feed-
propagation (RBP) training algorithm is eliminating men-
forward NN learning rule, the error is calculated as the
tioned harmful effects of the magnitudes of the partial
difference between the target output and the network out-
derivatives. Only the sign of the derivative is used to
put. Hence, the default performance function for feedfor-
determine the direction of the weight update. Size of the
ward networks is mean square error, MSE; we want to
weight change is determined by a separate update value.
minimize the average of the sum of differences between the
The update value for each weight and bias is increased by a
target output and the network output, which is presented by
factor delt_inc (see MATLAB, v. [8], Neural Network
MSE. The MSE algorithm created, in the NN computing,
Toolbox) whenever the derivative of the performance
adjusts the weights and biases of network so as to minimize
function with respect to that weight has the same sign for
this error.
two successive iterations. The update value is decreased by
This demand is presented by:
  a factor delt_dec whenever the derivative with respect that
9qmqq min ðMSEÞ  ½MSE ð7Þ weight changes sign from the previous iteration. It the
m2M;n2N derivative is zero, then the update value remains the same.
If the weight continues to change in the same direction for
To solve the above problem we use the integration of
several iterations, then the magnitude of the weight change
Monte-Carlo simulation and NN computing based on
will be increased.
MPSoC, i.e., looking for the optimal architecture, q–m*–q–
q, from a net of NNs, NoNN, through parallel computing
using SIMD technology. We use the following,
3 Training data problem of NN
1. Determining Configuration of the MPSoC architecture
in order to perform openMP program. The distribution of the training data within the problem
2. Send a range of m values and the desired MSE value, domain can have a significant effect on the learning and
([MSE] = 5e-6), to slaves. generalization performance of a network. Failures in the
3. Start as many different slave processes as there are use of NNs may be attributable to inadequate of the
physical machines. training data range. For any problem, e.g.,

123
 Int. J. Mach. Learn. & Cyber.

• we have, for example, available data (training data): 4 Multi-processor system-on-chip (MPSoC)
x = [1, 2, 3, 4, 5, 6, 7, 8, 9]
Parallel computing Parallel computing is a form of com-
y = [1.22, 3.16, 4.42, 2.28, 0.18, 2.01, 7.9, 8.3, 10.1].
putation based on either the computer-internet integration
• Any measurable function on x(9) to y(9) can be or multiprocessor/multicore computers, in which many
approximated as closely as desired, in the range x [ calculations are carried out simultaneously. It operates on
[1, 9], through regression by p = polyfit(x,y,7); the principle that large problems can be divided into smaller
f = polyval(p,x) derived from MATLAB program. ones, which are then solved in parallel. MPSoC is an inte-
We obtain the approximated function: grated circuit (IC) or a system on a chip (SoC) that inte-
y ¼ f ðxÞ ¼ 0:0030x7  0:0947x6 þ 1:1666x5  7:2207x4 grates all components of a computer or other electronic
system into a single chip. Software plays a critical role in
þ 24:0056x3  43:3274x2 þ 41:6380x  14:9533
MPSoC design because the chip won’t do anything without
ð8Þ software. Beyond its hardware architecture, an MPSoC
Graphical results described ‘x – y’ relation according to system is generally running a set of software application
available data (‘o’) and regression results (‘*’) are shown divided into tasks and an operating system devoted to
in the following Fig. 4a. manage both hardware and software through a middleware
Using this approximated function (Eq. 8) for the layer, e.g. drivers. A heterogeneous multiprocessor provides
required data: the computational concurrency required to handle concur-
x1 = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15] rent real-world events in real time via reducing conflicts
y1 = [1.22, 3.16, 4.42, 2.28, 0.18, 2.01, 7.9, 8.3, 10.1, among processing elements and tasks. In an MPSoC, either
13.0, 15.0, 9.0, 8.0, 11.0, 35.0]. hardware or software can be used to solve a problem. An
We get graphical results shown in the Fig. 4b: abstract view of the Software-Hardware integration of
The fit is good for this data on the interval x [ [1, 9]. MPSoC can be presented in the following figure:
Beyond this interval the graph shows that the polynomial An ideal software design flow allows the software
behavior takes over and the approximation quickly deteri- developer to implement the application in a high-level
orates (on the interval out of the available data, x [ (9, 15]). language, without considering the low-level architecture
We have, in this case, insufficient training data. details. The software design for MPSoC is more complex
Bigdata To generate in addition to available data in than a simple software compilation. As, an ideal software
order to reach a sufficient training data, we perform a design flow allows the software developer to implement the
combination of much creative mathematical possibilities of application in a high-level language, without considering
Monte Carlo simulation with high accuracy in the the low-level architecture details. In an ideal design flow,
mechanical aspects of finite element method, MCSFEM. the software generation targeting a specific architecture
Consequently, we can obtain a lot of data named bigdata— consists of a set of automatic final application software
massive datasets. Many different hardware architectures code generation, and hardware-dependent software, HdS,
apart from traditional Central Processing Units, CPUs, can code generation (Fig. 5c); the classical approaches is
be used to process data. The common point of these shown in Fig. 5b.
architectures is their massive inherent parallelism as well The HdS is made of lower software layers that may
as a new programming model rather than the classical von incorporate an operating system, OS, communication
Neumann CPUs. management, and a hardware abstraction layer to allow the

Regresja n=7 Zakres regresji [1:9]

12 100
data1
(a) 90 (b) data2
10
80
8 70
6 60
Y

50
y

4 40
2 30
20
0
10
-2 0
1 2 3 4 5 6 7 8 9 0 5 10 15
x x

Fig. 4 Fitting the required data by the function, f(x), for x, y (a) and x1, y1 (b)

123
Int. J. Mach. Learn. & Cyber.

(a) machines exploit data level parallelism, which is a form of

parallelization of computing across multiple processors in
parallel computing environments according to the ‘‘Master-
Slaver’’ structure. The SIMD technology for multi-core
computing is used for MPSoC and the ‘‘Master-Slaver’’
(b) (c) structure are illustrated in Fig. 6a, b.
MapReduce The assignment of tasks to a target processor
that will execute them is called mapping. The mapping
represents the association between the tasks and the pro-
cessing elements on which they are executing and the
association between the buffers used for the communication
between tasks and hardware communication resources. In
history, Google’s developers create a style of programming
that they called MapReduce that is surprisingly effective at
processing very large amounts of data and yet are able to
express a wide range of algorithms. Google’s MapReduce is
a programming model and a software framework for large-
scale distributed computing on large amounts of data. Given
Fig. 5 The Software-Hardware integration with Hardware-SubSys-
a set of data, a series of operations (kernel functions) is
tem (HW-SS), SoftWare-SubSystem (SW-SS) and Net on Chip, NoC.
Software design flows: b classic software design flow and c ideal applied to each element in the dataset and local on-chip
software design flow memory is reused to minimize external memory bandwidth
and to automate and optimize on-chip management ‘‘map-
reduce’’ process. Generally, it parses the input, splits to
OS functions to access the hardware resources of the plat- each slave and produces records. Intermediate results gen-
form—MPSoC architectures. The classical approaches for erated in the map phase are shuffled and sorted by the
the software design use programming models to abstract the MapReduce system and are then fed into the slaves in the
hardware architecture (Fig. 5b). These generally induce reduce phase. Despite that MapReduce gains popularity for
discontinuities in the software design, i.e., the software its excellent performance when implementing a large
compiler ignores the processor architecture (e.g., interrupts spectrum of applications. However, it does not have the
or specific Input/Output system, I/Os). To produce efficient mechanism to identify parallelism of NN computations
code, the software needs to be adapted to the target archi- integrated with Monte Carlo Simulation, MCS, and Finite
tecture by using specific libraries, such as system library for Element Method, FEM. Besides, current embedded appli-
the different hardware components or specific memory cations are migrating from single processor-based systems
mapping for the different CPU and memory architectures. to intensive data communication requiring multi-processing
SIMD Technology (Single Instruction, Multiple Data): systems. In this paper, we perform a new mapreduce model,
various efforts have been spent on finding alternative ways based on MPSoC using SIMD technology for a hardware
of Single Instruction, Single Data, SISD, to perform mas- architecture made of several identical processors, CPUs,
sive amounts of computations. Different programming called homogeneous symmetrical multiprocessing, SMP,
models have been introduced for high-level programming architecture, in which the communication between the dif-
of heterogeneous multiprocessor architectures. The pro- ferent CPUs is made through global shared memory. This
gramming model specifies how parts of the running parallel new distributed parallel processing enables us to perform a
communicate information to one another and what syn- new approach, like Google’s ‘‘mapreduce’’ idea, based on
chronization operations are available to coordinate their the ‘‘Single Instruction Multiple Data’’, SIMD, technology
activities. A number of MPSoC-specific programming using multiprocessing systems on a chip, MPSoC, archi-
models have been defined recently, e.g., message-passing tecture without necessarily other computer program and
interface, MPI (see [6]; compute unified device architec- results are obtained from one laptop with multiple cores
ture, CUDA (see [11]; open multi-processing, OpenMP without the use of a group of machines. Quantitative effects
(see [9], itd. A result of those efforts was SIMD, a pro- of this approach in compare with the ‘‘Single Instruction
gramming paradigm which allowed applying one instruc- Single Data’’, SISD, technology are presented in the
tion to multiple instances of data. The SIMD idea refers numerical example.
as a class of parallel computers and describes computer General model is shown in Fig. 7a, b.
with multiple processing elements that perform the same Determining the optimal architecture of NN from NoNN
operation on multiple data points simultaneously. Such by MCSFEM and MPSoC in the MapReduce framework in

123
 Int. J. Mach. Learn. & Cyber.

Fig. 6 a An idea of the (a)

systems-on-chips (SoCs). b The
‘‘Master-Slaver’’ structure

(b)

(a) • A two-layer feedforward network is used to solve

this problem. There are m neurons in the first layer
and (h = q) neurons in the second layer. The transfer
function in the first layer is tan-sigmoid, and the
second layer transfer function is log-sigmoid. The
training function is the batch gradient descent (see
MATLAB program, Neural Network Toolbox),
(b) which enables us to update the weights in the
direction of the negative gradient of the performance
function.
• The desired behavior is summarized by a set of input,
output pairs: p1, t1, p2, t2, …, pq, tq, where p is an input
to the network (the package cross-section of the
elements belonging to the structure) and t is the
Fig. 7 a A high-level illustration of the ‘‘MapReduce’’ programming
model. b The idea of ‘‘reduce’’ activities
corresponding correct (target) output (the package of
displacements of nodes of the structure). Thus, two
training datasets in the matrix form, P and T with the
order to solve the structural damage detection problem is data size, (ixj), where i denotes a number of the
shown in Fig. 8. elements belonging to the structure and j denote a
Eight processors are used in this paper. It is resulted number of processor of available computer used to
from the practical conditions of our personal computers. generate the virtual data, are referred as the input to the
Effect of them is shown in the numerical example. network.
• One assumed data, t*, referred as selected from GPS-
RTK data (displacements of nodes) is used to monitor
5 Numerical example the actual response of the structure in real time.
• The cross-sectional areas are taken as design variables
In this paper, the plane truss used for damage identification with side constraints: (2.0e-3)m2 B Si B (0.005)m2,
is shown in Figs. 9 and 10. i = 1, 2, …, 9, which are referred as the input p to
Each structural element has a modulus of elasticity the network. From the originate state, (S0i , = 0.005 m2,
E = 210 GPa. The designed cross-sectional area i = 1, 2,…, 9). It is sole data set and closely related to
S0 = 0.005 m2 for all elements. Key parameters of the the movements of structural elements observed in order
model are listed as follow: to monitor their necessary technical conditions.

123
Int. J. Mach. Learn. & Cyber.

Fig. 8 SIMD idea of Monte- Training & Testing Data

Carlo simulation using vector of
neural networks j = j + 1;
Rand. generation if j < J;
Interface m = 1:R, k = 1:C, NoNN
(Host)

(Slaves)

CPU1 CPU2 CPUi CPUn

a1= net a2= net ai= net am= net

e1= |a1 - t1| … e2= |a1 - t1| … ei= |ai - ti| … em= |am – tm|

20 kN II 4 IV network. Many experiments have been done using

Monte-Carlo simulation. The impact of the parameters,
(clearly changing of MSE), are shown in the selected
yj 5 numerical resultset presented in the following tables.
1 7 9 7m
3 Numerical resultset that represent the optimal architec-
j xj I
ture of NN are shown in the following tables:
2 III 6 V8 VI
• NN = m = neuron number of the first layer.
5m 5m 5m

Fig. 9 Topology of the plane truss

x 10 -4
NN-MSE Relation x 10
-4

3.9 3.8
3.8 3.7
3.7
3.6
3.6
MSE

3.5 3.5

3.4 3.4
3.3
3.3
3.2
3.2
3.1 60
0 10 20 30 40 50 0 20 40 net{5} = Optimal architecture of Neural Network at
NN NN
CPU5: 9-46-9-9 (m* = 46)
Fig. 10 Relationship between NN numbers and MSEs In the computer program, EE denotes the mean square
error, MSE, i.e., EE = MSE.

• The number of neurons of first and second layer, m [ • EE obtained from different slaves after training process:
M, and h [ H (h = q for this example). They are Relationship between NN numbers and MSEs is repre-
required to seek the optimal structure of the neural sented graphically as follows.

123
 Int. J. Mach. Learn. & Cyber.

Finally, integrating of NN computing and FEM based on The heart of any prediction system is prediction model.
MPSoC enables us to recognize the threat of this structure, There are various machine learning algorithms available
which is located in the seventh element having the cross- for different types of prediction systems. Any prediction
sectional area, S7 = S* = 0.0028 m2 that does not belong system using NN will have higher probability of correct-
to the training dataset (compatible with FEM computing ness if model is built using good training samples and good
with error, e = 0.0032 - 0.0028 = 0.0004 m2). Note that (optimal) NN architecture. We would like to emphasize, in
this error will be reduced depending on the increase of this work, on an advantage of the use of the mapreduce
training dataset and engineering experiences. model based on MPSoC to solve bigdata and multitasking
problems (NN, NoNN, FEM and MCSFEM) for building
these models. It turns out also that combining multicore
6 Conclusions and coprocessor technology provides extreme computing
power for highly CPU-time-computing applications.
Today, we can talk about terabytes and petabytes of data As, the improvement in performance gained by the use
rather than a gigabyte of data that was considered to be of MPSoC depends on the software algorithms used and
imposing. In the engineering problems, we must adapt to hardware. A new model of the mapreduce system used in
data originated from a host of new sources such as: GPS, this paper for a hardware architecture made of eight iden-
wireless devices, sensors, Streaming communication gen- tical processors, CPUs, called homogeneous symmetrical
erated by machine-to-machine interactions. Data is con- multiprocessing, SMP, architecture.
tinuous, unstructured and no constrained to rigid structures Some issues should be considered in the future:
of rows and columns. It is proving to be very hard to
• real time data processing challenges are very complex.
process by traditional computation methods. This perfor-
Handling the velocity of continuous data is not an easy
mance demanded by such applications requires the
task. It would like to handle, in the future, the parallel
migration from single processor-based systems to intensive
processing of this data in order to extract the meaningful
data communication and the use of multiprocessor archi-
information from this moving stream, i.e., we need an
tectures in a single chip (MPSoC). To solve the bigdata
another model used to allow high-level programming of
problems, one can use the Hadoop like Batch processing
heterogeneous multiprocessor architectures;
system using a distributed parallel processing paradigm
• more processor may lead to better results, but how
called MapReduce, which is a programming model of
many for this case should be considered in the future.
Google and an associated implementation for processing
• It is an increasing interest in the development of smart
and generating large data sets (BigData) using a group of
structures with built-in fault detection systems that
machines. It provides a standardized framework for
would provide damage and/or threat warnings in the
implementing large-scale distributed computation. Despite
fields of Civil Engineering.
continuous data (BigData) originated from GPS-RTK and
WSN is not studied in this paper; only one selected data is
used to determine technical state of structure in real time; References
bigdata approach is used in this paper to solve the problems
concerned with both MCSFEM and MCSNN model. 1. Almasi GS, Gottlieb A (1994) Highly parallel computing, 2nd
Hence, Google’s MapReduce model does not have the edn. Benjamin Cummings, Redwood City
2. Bhattacharya D, Mitra M (2013) Analytics on big fast data using
mechanism to identify parallelism of NN computations real time stream data processing architecture. EMC Proven Pro-
integrated with Monte Carlo Simulation, MCS, and finite fessional Knowledge Sharing
element method, FEM. Then, we develop a new distributed 3. Chau KW (2007) Application of a PSO-based neural network in
parallel processing model based on the single instruction analysis of outcomes of construction claims. Autom Constr
16(5):642–646
multiple data, SIMD, technology using the ‘‘Master-Sla- 4. Cheng CT et al (2005) Long-term prediction of discharges in
ver’’ structure. It enables us to split big data on n datasets, Manwan Hydropower using adaptive-network-based fuzzy infer-
processed at the same time on n processors, CPUs, rather ence systems models. Lect Notes Comput Sci 3612:1152–1161
than processed on single processor CPU by the single 5. Hwang J, Yun H, Park S-K, Lee D, Hong S (2012) Optimal
method of RTK-GPS/Accelerometer integration to monitor the
instruction single data, SISD, technology. This paper gives displacement of structures. Sensor 12:1014–1034. doi:10.3390/
a brief description that indicates the enormous potential, s120101014 ISSN 1424-8220
which can be seen in parallel processing, through an 6. Laemmer L, Sziveri J, Topping BHV (1997) MPI for the trans-
example and use case, in the framework of developments tech parastation. Advances in computational Mechanics with
Parallel and Distributed Processing Ltd., Edinburgh
in computer science. It indicates that, multicore computing 7. Mayer-Schonberger V, Cukier K (2013) Big data: a revolution
is the issue of how to fit multiple processing cores onto a that will transform how we live, work and think. Houghton
single chip to produce a functional and efficient whole. Mifflin Harcourt Publishing Company, New York

123
Int. J. Mach. Learn. & Cyber.

8. Neural Network ToolboxTM User’s Guide, v. 2010b. https://www. 14. Yang LT, Chen J (2014) Guest Editorial, Special Issue on Scal-
mathworks.com/help/pdf_doc/nnet/nnet_ug.pdf. Accessed 2 Nov able Computing for Big Data; Big Data Research 1, 2014, 3–5.
2015 http://www.elsevier.com/locate/bdr
9. Popovici K, Rousseau F, Jerraya AA, Wolf M (2010) Embedded 15. Zaharia M, Das T, Li H, Hunter T, Shenker S, Stoica I (2013)
software design and programming of Multiprocessor System-on- Discretized streams: fault-tolerant streaming computation at
Chip. Springer, Heidelberg scale. http://dx.doi.org/10.1145/2517349.2522737
10. Tien JM (2013) Big data: unleashing information. J Syst Sci Syst 16. Zang C, Imregun M (2001) Combined neural network and
Eng 22(2):127–151. doi:10.1007/s11518-013-5219-4 reduced FRF techniques for slight damage detection using mea-
11. Tran C, Srokosz P (2010) The idea of PGA stream computations for sured response data. Arch Appl Mech 71:525–536
soil slope stability evaluation. Comptes Rendus Méc 338(9):499–509 17. Zhao Y, Wu J, Liu C (2014) Dache: a data aware caching for big-
12. Vossen G (2014) Big data as the new enabler in business and data applications using the MapReduce framework. Tsinghua Sci
other intelligence. Vietnam J Comput Sci 1:3–14. doi:10.1007/ Technol 19(1):39–50. ISSN 1007–0214 05/10
s40595-013-0001-6 Pub. Springer
13. Weber F (2004), Research Activity in Smart Structures at EMPA.
In: Guran A, Valasek M (eds) 3rd International Congress on
Mechatronics MECH2K4 CTU in Prague, Czech Republic

123