Microprocessors and Microsystems 36 (2012) 33–44

Contents lists available at SciVerse ScienceDirect

Microprocessors and Microsystems

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

Design automation of cellular neural networks for data fusion applications

Prodromos Chatziagorakis, Georgios Ch. Sirakoulis ⇑, John N. Lygouras
Democritus University of Thrace, Department of Electrical and Computer Engineering, Laboratory of Electronics, GR-671 00 Xanthi, Greece

a r t i c l e i n f o a b s t r a c t

Article history: In this study, a novel methodology for the design automation of cellular neural networks (CNNs) for dif-
Available online 16 August 2011 ferent applications is proposed. In particular, an evolvable algorithm has been developed providing the
ability to generate the netlist of the requested CNN in any desired dimension through a very simple pro-
Keywords: cedure, which greatly simplifies the network design process, without the requirement of any relative
Cellular neural networks design knowledge. Furthermore, the user is also granted with control over the selection of the overall
Design automation function of the network, in order to make it suitable for a variety of data fusion applications. Moreover,
Data fusion
the generated netlist can be imported in the SPICE Cad System, resulting in the automated generation of
Sensor network
the network schematic, which can be used for the circuit hardware implementation. More specifically, a
tutorial 10  10 CNN model is generated via the proposed methodology for use in a data fusion and con-
trol application. The produced model is tested by its application to a real distributed temperature sensor
network for an application involving the attainment and the conservation of the thermal stability of a
system. The data transmission is implied through the use of a set of wireless transmitters–receivers.
Finally, a series of experimental results on real world conditions are presented, proving the effectiveness
and the robustness of the generated CNN and respectively of the proposed methodology.
Ó 2011 Elsevier B.V. All rights reserved.

1. Introduction power consumption, and cost. This has enabled compact, autono-
mous nodes, each containing one or more sensors, computation
Nowadays, spatially distributed sensor network systems are and communication capabilities, and a power supply [11].
widely used in both domestic [1] and industrial domains [2]. In Specifically, in the case of data control and automation systems,
most cases, these networks consist of autonomous sensor units, distributed smart sensors comprise the most significant tool to-
which are cooperatively used in order to measure and monitor wards meeting the corresponding prospective goals in each appli-
physical or environmental properties, such as temperature, pres- cation. Smart sensors provide the feature of implying the sensing,
sure, motion, position, light intensity, etc. These properties act as the conditioning and the processing part of the system within
the stimulus to the sensor, the output of which is often conditioned the cell of a single device [12]. Some practical examples of such
and processed to provide the corresponding measurement of the smart devices can be observed in a variety of everyday applica-
physical property. Sensors are generally part of a larger system tions including smart houses [1,13,14], hotel automation systems
consisting of signal conditioners and various analog, or digital sig- [15], robotics [16], industrial process control systems [2,17,18],
nal processing circuits. That system could be a measurement sys- automation in aviation [4,19], environmental control systems
tem [3,4], a data acquisition system [5–8], or a process control [6,8,20,21], etc. The problems that these systems are usually
system [2,9,10], for example. meant to both confront and solve are often related to some global
Ubiquitous computing envisages everyday objects as being aug- system aspect. This is the exact reason why these sets of sensors
mented with computation and communication capabilities. While should not merely operate autonomously, but they ought to be
such artifacts retain their original use and appearance, their aug- part of a wide network, where each entity should interact with
mentation can seamlessly enhance and extend their usage, open- the rest, resulting in some sort of global controlling function
ing up novel interaction patterns and applications. Networks of [22,23].
wireless sensors are the result of rapid convergence of three key In order for that global function to be achieved, one of the avail-
technologies: digital circuitry, wireless communications, and Mi- able parallel computing logical models should be adopted. In that
cro-ElectroMechanical Systems (MEMS). Advances in hardware way, the entire spatially distributed sensor system would become
technology and engineering design have led to reductions in size, a parallel sensing network, which would provide us with the
expected global properties and behavior. Some of the most effec-
⇑ Corresponding author. Tel.: +30 25410 79547; fax: +30 2541 79540. tive and common computational models so far are neural networks,
E-mail address: [email protected] (G.Ch. Sirakoulis). cellular automata [24], genetic algorithms [25] and fuzzy logic

0141-9331/$ - see front matter Ó 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.micpro.2011.08.001
34 P. Chatziagorakis et al. / Microprocessors and Microsystems 36 (2012) 33–44

systems [26]. Each one of these models implies a variety of proper- The paper is organized as follows: in Section 2, all the necessary
ties and features, that can be desirable or not depending on the background on cellular neural networks is provided. Section 3 pre-
specific requirements of the current application. For that reason, sents a CNN implementation approach based on the utilization of
an extensive study of these requirements is often of utter signifi- operational amplifiers. The proposed design automation methodol-
cance, so as to be able to make the right selection. Alternatively, ogy of CNNs lies in Section 4, whereas in Section 5, the experimen-
a combination of the properties of the foretold models could result tal test and its corresponding results are thoroughly described. In
in a hybrid computing system, which could include only the de- the end, the conclusions drawn and future work are presented in
sired features. Section 6.
In this study, the main interest is the fusion and the control of
data acquired by a network of multiple distributed sensors via
the utilization of an analog cellular neural network (CNN) [27–29]. 2. Cellular neural networks
CNN is a hybrid logical model, which implies a lot of great features
that makes it appropriate for use in control applications. CNN con- The cellular neural network (CNN) model which was proposed
sists of identical structural computational entities, the cells, which by Chua [27] from University of California at Berkeley as a practical
are connected locally to each other. However, despite the total lack circuit alternative to Hopfield [36,37] and other type of recurrent
of global connections, this network of cells presents a global logic networks, involves an attractive parallel computational structure,
function and is capable of dealing with very complex computa- particularly from the perspective of implementation in various mi-
tions. In specific, except for the fast parallel computing ability, cro and nanoelectronic technologies. The CNN paradigm includes
CNN implies additional features such as: continuous time global cellular automata (CAs) [24] as a particular case and in addition
dynamic behavior, great flexibility regarding to the shaping of its it borrows many ideas and techniques from the field of Neural
behavior pattern, simple and smart structure, great stability, low Computation. Computation in CNNs is brain-like rather than ‘‘clas-
cost implementation and great robustness. However, one of its sic’’ in the sense of the widespread computing architectures based
most significant features refers to its high accuracy, which comes on microprocessors. Unlike CAs which are mostly used to prove
as a result of its analog circuit nature. In [30], a CNN architecture various theories or to model physical processes [38,39], CNN was
has been proposed for realizing multi-sensor data fusion and intended from the beginning to be also a useful signal processing
distributed control systems, whereas in [31] CNN is proposed as paradigm.
a decentralized control paradigm for robotic locomotion. In Due to the great number and variety of available architectures,
addition, in [32] a CNN approach is discussed as far as distributed it is relatively difficult for a strict CNN definition to be synthesized.
sensor networks are concerned. The specific computational tool As far as circuit topology is concerned, CNNs are systems that com-
implies the local connectivity and the global data processing prise from a finite set of locally interconnected non linear and often
features, however, it is only limited on one-dimensional circular identical structural units. These entities, which include one single
topology. In the end, another remarkable feature of the CNN is output and multiple inputs, are named as ‘‘Neurons’’, or ‘‘Cells’’.
the flexibility that it provides in the setting of its total functional- Mathematically, the cell entity can be modeled by a non linear sys-
ity. Some examples of template programming of CNN can be found tem that utilizes the incoming data, its initial condition and other
in [33–35]. parameters in order to create its output [27,39].
In the length of this manuscript, a novel methodology is pro- There is a great difference between cellular and classic compu-
posed for the design automation of CNN circuits in any desired tation. Suppose an array of memory cells (i.e. a Random Access
dimension, in association to the current application needs. In other Memory). In a classic computer, the cells are sequentially updated
words, an algorithm has been developed, which takes advantage of and located by the central processing unit via some external buses
the already known circuit structure and topology of the proposed while in a cellular computer each memory cell exchanges informa-
CNN and subsequently generates a new network according to the tion locally only with the neighboring cells. A ‘‘gene’’ associated
declared desired dimension. That means that efficient computing with each cell controls the exchange of information with the cells
tool greatly simplifies the design and generation of the CNN in a neighborhood. The cells are updated in parallel and there is no
through a short process, which just requires the declaration of central processing unit to control the cells. Like in a classical com-
the desired network dimension, the range of the local interaction, puter, the array of cells starts from an initial state, which contains
as well as the selected local behavior which are some elements that the problem, and the solution will be found in the same array of
are going to be presented below. In addition to that, in order to cells after a period of time during which computation emerges.
prove the effectiveness and the robustness of both the methodol- While the designer of a classic computer focuses on the central
ogy and the CNN model in real world data, an experimental test processing unit, on data coding, address buses and instruction sets,
of a generated 10  10 network is made through the utilization the designer of a cellular computer has to focus mostly on the cell.
of a distributed high accuracy analog temperature sensors net- The ‘‘program’’ is now coded in what Leon Chua called ‘‘cells’ gene’’
work. The constraint that ought to be taken into consideration re- (i.e. the entire set of parameters defining the cell). Quite often all
fers to the nature of the sensors that could accompany the cells have identical structure and parameters and various tasks
proposed CNN, which must be analog (direct cell connection), or can be ‘‘programmed’’ on the same CNN chip by simply changing
digital via the utilization of a digital-to-analog converter (indirect the genes [40].
cell connection). An additional limitation that has to be taken into Next, a more specific and formal definition of the CNN model is
account as far as the proposed CNN network is concerned, involves going to be presented by referencing some of its basic descriptive
the maximum input voltage allowed. Due to the utilization of the mathematical equations. As it is already aforementioned, the basic
TL084CP Operational Amplifier model as a basic structural unit of structural unit of the CNN is the cell. The cell circuit is usually a set
the CNN, each one of the available cell input voltages (and there- of both linear and non linear elements, such as resistors, capacitors,
fore each one of the sensors output) should not exceed the barrier vccs (voltage controlled current sources) and independent sources.
of ±15 V. The selected sensor model that satisfies all the above Although CNN architecture can be extended into three dimensional
requirements is the low voltage analog temperature sensor spaces, for simplicity reasons the case of a two dimensional circuit
TMP36. In this test, the final goal is to maintain the temperature is currently studied. A paradigm of a 4  4 network is depicted in
parameter within the desired levels, so as to ensure the global sta- Fig. 1. A standard CNN architecture consists of an M  N rectangu-
bility and the proper function of the system. lar array of cells C(i, j), where i and j are the corresponding
 P. Chatziagorakis et al. / Microprocessors and Microsystems 36 (2012) 33–44 35

Fig. 3. The simplified circuit of the CNN cell.

Although the cell units have identical internal structure, the CNN
has some boundary conditions. Namely, there are Regular and
Boundary cells, where a cell C(i, j) is called a regular cell with re-
spect to Sr(i, j) if and only if all neighborhood cells C(k, l) e Sr (i, j)
exist. Otherwise, C(i, j) is called a boundary cell.
A typical example of the simplified cell circuit can be found
Fig. 1. A two dimensional 4  4 CNN. The square patterns represent the cell units,
whereas the straight lines the direct connections and interactions between them.
in Fig. 3, where a random C(i, j) cell is depicted. It consists of
two vccs, a linear capacitor and two linear resistors. Parameters
u, x and y stand for cell input, state and output, respectively. In
Cartesian coordinates of each cell (i = 1, 2, . . ., M, j = 1, 2, . . ., N). In particular, the voltage Vuij corresponds to the cell input, Vxij
particular, the specific network consists of M  N cells with cell refers to the state of the cell, whereas Vyij to its output. The vccs
C(i, j) placed in ith line and jth row. Ixy(i, j; k, l) and Ixu(i, j; k, l) receive some input signal from the
As it has already been foretold, every single cell is directly con- cell neighbors and condition it according to their conductance
nected to its neighbors and interacts locally with them. However, values.
there is an indirect connection and interaction between the total The state equation of the basic cell circuit is given by:
set of the network cells, due to the continuous time dynamics that X
dv xij ðtÞ 1
consists one of the most significant features of the CNN model. In C ¼  v xij ðtÞ þ Aði; j; k; lÞv ykl ðtÞ
dt Rx
this way, in spite of having total lack of global connections, CNN Cðk;lÞ2Sr ði;jÞ
X
present a common dynamic behavior throughout its magnitude. þ Bði; j; k; lÞv ukl ðtÞ ð2Þ
Moreover, the concept of the neighborhood itself includes the Cðk;lÞ2Sr ði;jÞ
sphere of the direct influence Sr(i, j) of each cell unit, and is ex-
pressed by the following statement: where 1 6 k 6 M and 1 6 l 6 N.
The above equation derives from the application of the Kirch-
Sr ði; jÞ ¼ fCðk; lÞ; where maxfjk  ij; jl  jjg 6 rg ð1Þ
hoff Current Law to the basic circuit. The voltages Vykl and Vukl rep-
where r is a positive integer that is equal to the size of the neighbor- resent the cell inputs, which are the output and the input of the
hood, 1 6 k 6 M and 1 6 l 6 N. It is also important to be noticed that neighbor cells in respect, whiles the coordinates M and N, are used
the coordinates (i, j) correspond to the current cell, whereas the to specify the network dimension. The interactive parameters A(i, j;
coordinates (k, l) correspond to its neighbors, which are directly k, l) and B(i, j; k, l) include the specific conductance value of each
connected to it. The above equation states that the sphere of influ- one of the vccs, which corresponds to a separate discrete input
ence Sr of each cell includes the total set of the neighboring cells, from a different neighbor. So, the number of the vccs is equal to
which coordinates satisfy the condition max{|k  i|, |l  j|} 6 r. The the one of the available cell inputs. The set of these values com-
specific statement prevents the sphere of influence Sr of each cell poses a matrix, the dimension of which is equal to the one of the
from expanding beyond the borders of its predefined neighborhood. cell neighborhood (sphere of influence Sr). Its central value refers
In other words, the size of the neighborhood depicts the set of to the cell self feedback, whereas the peripheral ones refer to the
the local interconnections of each cell and remains fixed through- corresponding outputs of the neighbors. The matrix A(i, j; k, l) is
out the entire network. Fig. 2 depicts three different neighborhood called Feedback Template, whereas the matrix B(i, j; k, l) is called
examples with radius values equal to r = 1–3, respectively. Control Template and their combination defines the dynamic
Additionally, it is of utter significance to observe the feature of behavior of the CNN. A typical example for a better understanding
symmetry that the neighborhood system includes. In specific, if C(i, is given below. It concerns the case of the cell C(2, 2) of a random
j)e Sr(k, l), then C(k, l) e Sr(i, j) for every C(i, j) and C(k, l) in a CNN. CNN that has neighborhood radius equal to r = 1.

Fig. 2. The cell neighborhood with radius values equal to r = 1–3.

36 P. Chatziagorakis et al. / Microprocessors and Microsystems 36 (2012) 33–44

2 3 2 3
Að2; 2; 1; 1Þ Að2; 2; 1; 2Þ Að2; 2; 1; 3Þ X
6 7 v max ¼ 1 þ Rx jIj þ Rx max 4 jAði; j; k; lÞj þ jBði; j; k; lÞj5
Aði; j; k; lÞ ¼ 4 Að2; 2; 2; 1Þ Að2; 2; 2; 2Þ Að2; 2; 2; 3Þ 5 16i6M;16j6N
Cðk;lÞ2Sr ði;jÞ
Að2; 2; 3; 1Þ Að2; 2; 3; 2Þ Að2; 2; 3; 1Þ
ð7Þ
and As far as the network stability is concerned, it is proved via the
2 3 Lyapunov Method [27]. In order for the cell to converge to a fixed
Bð2; 2; 1; 1Þ Bð2; 2; 1; 2Þ Bð2; 2; 1; 3Þ and finite equilibrium point after the end of the transient, the fol-
6 7 lowing condition must be met:
Bði; j; k; lÞ ¼ 4 Bð2; 2; 2; 1Þ Bð2; 2; 2; 2Þ Bð2; 2; 2; 3Þ 5 ð3Þ
Bð2; 2; 3; 1Þ Bð2; 2; 3; 2Þ Bð2; 2; 3; 1Þ 1
Aði; j; i; jÞ > ð8Þ
Additionally, the relation between the above templates and the vccs Rx
currents Ixy(i, j; k, l) and Ixu(i, j; k, l) is described in the following Furthermore, the magnitude of the stable equilibrium points
equation. must be greater than 1. In other words, the following properties
must be realized:
Ixy ði; j; k; lÞ ¼ Aði; j; k; lÞv ykl
and ð4Þ lim jv xij ðtÞj > 1; 1 6 i 6 M; 1 6 j 6 N and
t!1
ð9Þ
Ixu ði; j; k; lÞ ¼ Aði; j; k; lÞv ukl lim v yij ðtÞ ¼ 1; 1 6 i 6 M; 16j6N
t!1
From all the above, it can be easily derived that the selection of
the partial circuit elements can alter the final cell behavior. In con-
clusion, although there are no global connections and all the cells 3. CNN implementation with operational amplifiers
are indirectly connected to each other, the entire CNN adopts a
common function that is described by the behavior template. In the previous section, some of the most fundamental theoret-
According to the available inputs number, more than one template ical aspects of CNN were presented. In this section, taking into con-
may exist. sideration the CNN theory, some novel circuit architecture
In addition to the above, the output equation of the cell circuit is specifically designed for use in data fusion and control applications
given by the following statement: is proposed [43]. The basic cell circuit is depicted in Fig. 4.
This basic cell circuit consists of two discrete parts - circuits.
1 The first part is indicated by the blue dotted line and includes
v yij ðtÞ ¼ ðjv xij ðtÞ þ 1j  jv xij ðtÞ  1jÞ ð5Þ
the resistors R1–R5 and the operational amplifier A1 is a vccs.
2
The vccs drives the cell input to the capacitor C (where its cur-
where 1 6 k 6 M and 1 6 l 6 N. In other words, the output voltage rent state is stored) and then to the second and final part of
of the cell circuit can vary between 1 and +1 V. This feature is the circuit, which is confined by the red dotted line. That part
often used in image processing applications [41,42]. Of course, is a combination of a Schmitt trigger (comparator) in sequence
the value range of the cell output can be easily customized with a voltage follower. It is responsible for trimming the voltage
through the final vccs, so as to agree with the current application off, so as for it to attain some value within the desired range (in
needs. our case 1 V to +1 V). Each cell is equipped with a vccs per
Finally, the network has some constraint condition as far as the input. Therefore, the total number of the vccs indicates the set
cell initial state is concerned. In particular, it is expressed by the of the inputs. The conductivity of these sources determines the
following equation: effect that they have on their input and respectively influences
the interaction between the cells, as well as the behavior of
jv xij ð0Þj 6 1 ð6Þ
the entire network. That conductivity depends on the values of
It is clearly understood that the initial state must be less than or the resistors R1–R5. In this way, the proper selection of these
equal to one. Furthermore, the time constant of the circuit dynamic resistors has great effect on the determination of the total func-
is equal to CRx and is often selected to have a value between 108 s tion of the network.
and 105 s. In the specific case, i.e. data control applications, all the nec-
The set of the cell states are bounded for t > 0 and the limit vmax essary adjustments were made to the cell circuit, in order for
can be calculated through the following statement: the CNN to be capable of implementing data fusion and control

R2

R1
. - R3 Ixy(i,j;k,l) Vxij
-
.
Vykl A1 - A3
+ A2 + Vyij
R4 + R8

Rx C R7 R9
R5 R6

Fig. 4. An implementation of the cell circuit via the utilization of operational amplifiers. The first part of the circuit is confined by the blue dotted small circle on the left,
whereas the second part is confined by the red large circle on the right. (For interpretation of the references to colour in this figure legend, the reader is referred to the web
version of this article.)
 P. Chatziagorakis et al. / Microprocessors and Microsystems 36 (2012) 33–44 37

2 3 2 3
procedures. Thus, each cell is coupled to a sensor, which is c11 c12 c13 c1 c1 c1
responsible for scanning – detecting the desired external condi- 6 7 6 7
C ¼ 4 c21 c22 c23 5 ¼ 4 c1 c0 c1 5 and
tions (e.g. temperature, air pressure, etc.) and providing the cell c31 c32 c33 c1 c1 c1
with all the necessary data. In other words, the output of the sen- 2 3 2 3 ð10Þ
sor forms an input for the cell itself and its neighbors. Further-
b11 b12 b13 b1 b1 b1
6 7 6 7
more, an additional input for the cell circuit is the state of its B ¼ 4 b21 b22 b23 5 ¼ 4 b1 b0 b1 5
neighbors, as well as its own state. Thus, every cell receives b31 b32 b33 b1 b1 b1
two incoming voltages per neighbor, which represent its state
where c1 and b1 are the corresponding neighbors coefficients, while
and its sensor output. Moreover, it receives some feedback from
c0 and b0 are the central values of the templates that refer to the cell
its own state and sensor output. This is the reason why, during
itself.
the designing process every cell is considered to be a neighbor
The sole condition that has yet to be met for the asymptotic sta-
of its own self. Next, every source drives its input to the capacitor
bility of the two dimensional (or even three dimensional) system
C and in the end that voltage reaches the final part of the circuit.
of symmetric (or asymmetric) topology to be guaranteed, derives
There, a Schmitt trigger trims that voltage off, in order to
from the CNN theory bibliography [27] and is here expressed by
condition its value between 1 V and +1 V. Then, the voltage fol-
the following statements.
lower drives it to some mechanism (actuator), which implies the
necessary counteraction. 1
c0 P  þ Sn jc1 j
In sort, the entire CNN is connected to a sensor network, Rx
which measures the external conditions. In case of some external
ð11Þ
1
disturbance being detected, the stability of the CNN becomes de- and  þ Sn jc1 j < 0
Rx
cayed. Then, the corresponding sensor sends the necessary feed-
back to the network, where it is propagated from cell to cell. where Rx is the value of the corresponding cell resistor unit, Sn is the
Finally, a great number of cells (depending on the network size total number of the cell neighbors and c0, c1 are the previously pre-
and the signal propagation, it could be the aggregate of the cells) sented coefficients of the state template C. The adopted matrices C
is aware of that abnormality. Following, via their output voltage, and B, which come to agreement with the current application needs
each one of them enables respectively their attached mechanism and also satisfy the above stability condition, are stated below.
(actuator), which provides the system with the necessary coun-
2 3
0:5 0:5 0:5
teraction and restores its stability. In other words, the proposed 6 7
C ¼ 4 0:5 0:1 0:5 5  103 X1
CNN has the attribute of being globally aware of any external
disturbance and thus efficiently enables its control throughout 0:5 0:5 0:5
its entity. Consequently, someone could observe that there are and 2 ð12Þ
3
two types of vccs; the one that receives the states of the neigh- 1 1 1
6 7
bors and the one that involves the sensors outputs. In correspon- B ¼ 4 1 1 1 5  103 X1
dence to all the above, there are two kind of templates that 1 1 1
characterize these sources; the State template C and the Control
template B. These templates are mathematical matrices, which As is has already been mentioned, each value of these templates
contains the conductivity value of each vccs of the neighbor- correspond to the conductivity of a different vccs. The peripheral
hood. The central value concerns the vccs of the cell itself, values c1 of C have all the same weight value that is equal to 0.5,
whereas the peripheral ones those of the neighbors. These values whereas the central value c0 is equal to 0.1. The negative sign
are formed from the cell circuit elements properties. Thus, it is of c0 depicts the tendency of the cell circuit to reduce the ampli-
relatively easy for someone to achieve a variety of different net- tude of the incoming signal. Next, the values of B are all equal to
work behaviors by simply altering some of those properties. The 1, resulting to the data being driven to the comparator, where it
aforementioned basic cell circuit in Fig. 4 depicts the implemen- is conditioned as previously described. However, due to the satis-
tation of these templates. In particular, the first part of the cir- faction of the condition for the asymptotic stability of the cell cir-
cuit (vccs) implies the State template C (blue dotted line), cuit, the CNN always converges to an equilibrium point, which here
whereas the second and last circuit part implies the Control tem- is equal to 0. Finally, the values of the resistor Rx and the capacitor
plate B (red dotted line). C are 100 X and 1 nF, in order to satisfy the time constant con-
In this study, the aforementioned adaptability feature of the straint condition, and the asymptotic stability criterion.
CNN has been utilized and the corresponding template matrices
have been properly selected, in order to adopt the desired 4. Methodology for design automation of CNNs
network dynamics, which implement the necessary data fusion
and control procedures. In specific, the cell should uniformly In the previous chapter, a theoretical aspect of the CNN was pre-
interact with its neighborhood. In order for that cell- sented. Furthermore, a cell architecture for use in data fusion and
neighborhood interaction to be symmetric, every cell in the net- control was proposed. The above cell circuit was designed with
work should interact instantly with all its neighbors in exactly the use of the program Spice, which is CAD system that enables
the same way. For this reason, the corresponding peripheral its user to design and simulate electric, electronic and logical cir-
coefficients of both the State and the Control template should cuits. Furthermore, Spice produces an output source code file
be equal in each case. Though, the central value of these known as netlist. This specific file contains every single detail of
matrices should not essentially be equal to the rest of the coef- the circuit, such as its components, their parameters, the connec-
ficients, as long as it is related to the cell itself and the process- tions between them, the various conditions, etc. A similar netlist
ing of the incoming data within its circuit. This cell property file was created containing all the essential information of the cell
would result to a CNN with equally weighted direct connections circuit.
between its partial units. This feature of the symmetric interac- Following, a novel methodology for the design automation of
tion between the cell and its neighborhood could be imple- the CNN is proposed. It is based on the fact that the network con-
mented, by applying the following transformation to the state sists of identical cell units. The exact parameters and information
and control templates: of the circuitry, the elements and the connections of the cell are in-
38 P. Chatziagorakis et al. / Microprocessors and Microsystems 36 (2012) 33–44

Table 1
Methodology for the design automation of the CNN.

Step Action Remarks

1 Enter a value for the dimension M The user is prompted to declare the desired network dimensions,
Enter a value for the dimension N as well as the neighborhood radius
Enter the neighborhood radius R value
2 N_angular = (2R + 1)  (R + 1) The total number of neighbors of each cell is calculated
in accordance to the declared radius value
2
N_external = (R + 1)
N_internal = (2R + 1)2
3 Enter a value for the desired conductivity for each of the vccs below. ‘‘C’’ stands for state vccs
Cself = . . . ‘‘B’’ stands for control vccs
Cneighbor = . . . ‘‘self’’ refers to the cell itself
Bself = . . . ‘‘neighbor’’ refers to the neighboring cells
Bneighbor = . . .
4 Create State Vccs according to the values of the State Tenplate C Here, the values of the circuit elements of each vccs are defined
Create Control Vccs according to the values of the Control Tenplate B according to the containing conductivities of Tenplates C and B
5 for (i = 1;i++;i = M) In the end, the proper cell type is chosen and placed
for (j = 1;j++; j = N) in reference to the current topology
according to the current i, j values
create AngularCell(i, j)
or
create InternalCell(i, j)
or
create ExternalCell(i, j)
end
end

cluded in its netlist file. The following methodology takes full neighbors in total; the external cells would have five neighbors,
advantage of the available data which can be imported in Matlab, whereas the internal would have eight neighbors. These sums are
in order to compose a CNN of any dimension. In that way, the user calculated via the use of the following equations.
could generate it in any desired dimension fitting his current
needs, bypassing in this way a great part of the calculating and N angular ¼ ð2R þ 1Þ  ðR þ 1Þ
designing process. N external ¼ ðR þ 1Þ2 ð13Þ
At first, as listed in Table 1, the total size of the CNN as well as N internal ¼ ð2R þ 1Þ 2

the neighborhood radius is selected by the user, who is initially

prompted to declare his preferences according to his current needs where N_angular, N_external, N_internal are the corresponding sets
and the application type. Namely, the values of the dimensions M of the three cell types and R is the predefined network radius.
and N respectively must be specified, since the network size de- Following, in step 3 the user must specify the desired conduc-
pends on these two parameters. At this point, it would be wise tivities of the vccs in order for the CNN behavior templates to be
to note that although the proposed methodology makes use of a created. This is probably the most significant step of the methodol-
two dimensional topology, it could be quite easily evolved into ogy, because the user defines the local interaction rule, which af-
three dimensions in future work. After having denoted the selected fects the total functionality of the CNN. Due to the fact that the
magnitude of CNN, in the end of step 1 the user is also prompted to network consists of a set of identical structural units, the user is
enter the value of another crucial parameter. That parameter is the only required to enter the values for one cell unit. The parameters
desired CNN neighborhood size, which is defined by the value of Cself and Bself that the user must firstly enter are the central values
the variable R. Previously, in Section 2, the concept of the cell for each template, which refer to the feedback that the cell receives
neighborhood has been explicitly described, whereas it has been from its own circuit. These parameters have been previously pre-
made clear that it plays a very significant role in both the function- sented as c0 and b0. In addition, CNeighbor and BNeighbor are the
ality and the complexity of the network. peripheral values of both templates that refer to the inputs that
Next, in step 2 follows a general separation of the network cells the cell receives from its neighbors. Those inputs have been for-
into different groups through a calculation that takes into account merly named as c1 and b1. Following, since all the necessary data
their relative position into the network. In sort, the cell units are has been declared, the State template C and Control template B can
categorized into three different types: angular (placed in the four be created.
angles of the network), external (placed in the network edges) Moreover, during step 4, following to the composition of the
and internal (placed in the network interior space) cells. That dis- two templates, the corresponding vccs of the cell circuit are cre-
tinction has actually nothing to do with the cell circuit structure it- ated. This procedure implies the calculation of the exact values of
self, as long as it remains unchanged from unit to unit. This the circuit components of the cell. This calculation is realized by
categorization comes indeed as a result of the boundary conditions equating the transfer function of the vccs circuits with the defined
of the CNN. In other words, these three types of cells have been conductivity values. Since, these circuits consist of a variety of ele-
separated in correspondence to the total number of their neigh- ments, the majority of them has an already known value and the
bors, which eventually affects the exact total number of their direct computation procedure calculates the value of the rest of the ele-
connections with the rest of the network. In specific, the relative ments according to the given state and control templates values.
position of each cell in the network is reflected in its neighbors In the end, during step 5 the entire CNN is created through a
number, which depends on the value of the neighborhood size R. repetitive circular procedure, which firstly checks the current cell
For example, in the specific CNN topology, in case that the neigh- topology and then places it in the network according to the above
borhood size R is equal to 1, the angular cells would have three data. In specific, a loop function checks the coordinates of the
 P. Chatziagorakis et al. / Microprocessors and Microsystems 36 (2012) 33–44 39

6000mm

Fig. 5. An example of the topology of the CNN in an office floor plan.

current cell unit and after taking into consideration its relative net- In conclusion, the proposed methodology clearly offers great
work position, it creates the corresponding cell type (angular, inter- flexibility and some significant features most of them have already
nal, or external). This procedure is presented by the following been mentioned. Firstly, despite the fact that it was initially de-
descriptive code, where the variables M and N correspond to the signed for use in data fusion and control; it could also be used
CNN dimensions, the variables i, j are the current cell coordinates, for a variety of other applications. In order for the CNN to adopt
whereas AngularCell, InternalCell and ExternalCell refer to the cell some different function, the user should manually alter the con-
type. ductivity value of each vccs, so as to create the desired C and B tem-
plates. These new templates will eventually alter the network
for(i = 1; i + +; i = M)
dynamic behavior, making it suitable for a variety of different
for(j = 1; j + +;j = N)
applications. In addition to that, one of the most significant fea-
according to the current i, j values
tures that the current methodology provides is the great flexibility
create AngularCell(i,j)
as far as network size, neighborhood radius and cell topology are
or
concerned. The user is provided with the ability to select the values
create InternalCell(i,j)
of these characteristics and create a CNN that fits to current needs.
or
Moreover, the presented methodology could be easily evolved in
create ExternalCell(i,j)
many aspects. For example, the adopted topology could be quite
end
easily evolved into three dimensional space, suiting the needs of
end
more complex applications. Finally, as long as CNN hardware pro-
duction is concerned, the above methodology could be imple-
In the end, the methodology for the automatic generation of the mented via some robotic mechanism, which could include a
CNN is depicted as a whole, accompanied by some of the already simple user interface. That could result in a fully automated CNN
mentioned remarks, in Table 1. production in the future.
40 P. Chatziagorakis et al. / Microprocessors and Microsystems 36 (2012) 33–44

5. Application be properly conditioned and driven to the CNN circuit, bearing in

mind that in this case the benefits of the analog nature of the
After developing the previously described methodology, a CNN could not be fully exploited. The overall efficiency and func-
10  10 CNN was generated for use in data fusion and control tionality of the CNN is boosted through the adoption of analog con-
applications, adopting the State and Control Templates that have nections, which does not always consist an option due to increased
been presented in Section 2, as well as neighborhood radius equal network complexity and spatial restrictions. An additional point of
to 1. In particular, the 10  10 network has been utilized in a tem- interest refers to the maximum permitted cell input voltage. Due
perature monitoring and conditioning application. The generated to the fact that the proposed cell architecture includes the utiliza-
CNN has been used in conjunction with a distributed temperature tion of operational amplifiers (model TL074CP), cell input voltage
sensor network in inducing the thermal stability throughout the should not exceed the bounds of ±15 V in order to remain within
network area. Prior to selection of the appropriate sensor model, their operating range. Finally, another parameter that has been ta-
some limitations as far as the network circuit nature is concerned ken into account was the total number of the sensor outputs. The
have been taken into consideration. specific CNN circuit supports only single output sensors, although
The first network parameter to be noticed is its continuous time according to the current application demands additional vccs could
dynamic behavior, which is a result of its analog circuitry. This is be included in the basic cell circuit, thus enabling the adoption of
the exact reason why the use of analog sensors network reflects multiple output sensor models.
the prerequisite for direct connection to the CNN. Alternatively, In agreement with the aforementioned limitations, a network of
in the case of an indirect connection, digital sensors could be used distributed high accuracy low voltage analog temperature sensors
combined with some D/A converter, so as for their output signal to TMP36 was selected for testing with the generated 10  10 CNN.

(a)

j
1 2 3 4 5 6 7 8 9 10
i
1 0.73 0.75 0.74 0.73 0.74 0.75 0.75 0.75 0.74 0.73
2 0.73 0.75 0.75 0.78 0.77 0.75 0.75 0.76 0.75 0.74
3 0.74 0.75 0.75 0.78 0.97 0.96 0.96 0.93 0.76 0.75
4 0.75 0.75 0.75 0.77 0.97 1.20 1.15 0.94 0.76 0.75
5 0.75 0.75 0.74 0.78 0.97 1.20 1.10 0.93 0.75 0.74
6 0.74 0.75 0.75 0.79 0.96 0.96 0.95 0.93 0.76 0.75
7 0.73 0.75 0.74 0.78 0.76 0.77 0.75 0.75 0.77 0.75
8 0.73 0.74 0.73 0.75 0.74 0.75 0.73 0.74 0.75 0.75
9 0.73 0.74 0.74 0.75 0.75 0.75 0.74 0.75 0.75 0.75
10 0.73 0.75 0.74 0.75 0.73 0.73 0.73 0.74 0.74 0.74

(b)
Fig. 6. The initial network state diagram in the form that it was firstly monitored and driven by the distributed TMP36 network to the 10  10 CNN input (a) three
dimensional diagram and (b) arithmetic table.
 P. Chatziagorakis et al. / Microprocessors and Microsystems 36 (2012) 33–44 41

TMP36 provides a voltage output that is linearly proportional to is spatially scattered throughout an office floor, occupying some of
the Celsius (centigrade) temperature, with a scale factor equal to the most critical positions. For clarity reasons, the wireless sensors
10 mV/°C. The specific model comes to a complete agreement with have been colored and enlarged. It becomes clear that the CNN
all the above conditions. Furthermore, some of their noticeable fea- topology can be easily adapted in order to satisfy the current appli-
tures are high accuracy, high linearity (±0.5 °C), stability with large cations needs. For example, consider a building of multiple similar
capacitive loads, wide thermal operative range (40 °C to +150 °C) floors each one having its own wireless network. Those networks
and low self heating. The connection and data transfer between the could simply be integrated to one CNN by adding a third dimension
distributed sensor network and the CNN has been carried out to the neighborhoods of floor networks. In other words, each cell-
through the use of Wi–Fi receivers–transmitters for testing sensor-actuator set would acquire its corresponding vertical neigh-
purposes. bors from the adjacent floors resulting in a three dimensional wire-
More specifically, the implementation of these transceivers has less sensor network with the property of a global data processing
been realized through the combination of a microprocessor PIC16- function.
F877A and a module 802.11b Digi Connect Wi-ME. A bidirectional Following, a series of experiments has been carried out in order
communication channel has been established between the two to provide the necessary information of the network efficiency and
units via the direct connection of their serial ports. The IEEE robustness. Below, a representative test-case is described. In Fig. 6,
802.11b wireless local area network (WLAN) standard has been the initial network state is depicted in the form of a three dimen-
adopted, which provides a maximum raw data rate of 11 Mbit/s, sional diagram (a) and an arithmetic table (b). The following re-
exploits the 2.4 GHz frequency band and utilizes the Carrier Sense sults represent the CNN final state that was firstly initialized by
Multiple Access Collision Avoidance (CSMA/CA) data access mode. the sensors outputs and the network initial condition and was
Moreover, the dynamic range of the 802.11b standard is far from eventually formed through the function of the CNN according to
adequate as far as the distance requirements of the specific appli- the theory described in previous section.
cation are concerned. The operational radius of the transceivers is The state of every cell unit in this network is correlated with the
approximately equal to 38 meters for indoor and 140 m for out- actual temperature value that its sensor has measured. It is quite
door environments. These values are acceptable due to the local easily observed that cells C(4, 6), C(4, 7), C(5, 6) and C(5, 7) present
interaction feature of the proposed CNN, which does not require an unusually high temperature level, affecting all the adjacent
any global connections. The microprocessor PIC16F877A consti- neighbors. The measured temperatures in the center of the distur-
tutes a control unit for the transferring data and enables the re- bance result to output voltages that vary from 1.1 (60 °C) to 1.2
mote management of the wireless network data through a (70 °C) V, disrupting in that way the network thermal stability.
distant computer that has installed the appropriate software (Real The goal in the specific experiment is to achieve the thermal stabil-
Port), which creates a virtual serial port in order to provide this ity of the network, which is translated into a relatively flat measur-
useful feature. Eventually, this remote management feature also ing diagram. Following, in Fig. 7, the final network state is
offers the capability of saving the sensory data and thus provides presented after the interference of the CNN and the end of the nec-
the user with the option to realize any further analysis. essary interval of time needed for the convergence of the system to
Before continuing further, an aspect that should be commented an equilibrium point. Analyzing and comparing the two diagrams,
refers to the real topology of the CNN. Despite the fact that the cell one can observe that the CNN has achieved to spread the initial
units are considered to be placed in a rectangular grid, in reality external local disturbance. In that way, the state voltage values
they could be freely positioned in space. Apparently, this assump- of the neighboring cells has been augmented, thus enabling the
tion has been made for convenience. As long as the direct connec- corresponding actuators. Therefore, a more global and decisive
tions between the neighboring cells, as well as the total number of control is applied by the set of the affected cells. Furthermore,
the neighbors of each cell are maintained intact, the CNN-sensor through a comparison between the initial and the final condition,
system is not going to present any difference in its behavior or it is easily understood that the voltage values of the set of the cell
functionality. An example of a possible CNN topology is depicted units have been greatly reduced from the range of V to some mV,
in Fig. 5, where a wireless sensor network consisting of 20 sensors which is a fact that reflects the CNN total control efficiency. From

Fig. 7. The final network state diagram after the interference of the developed 10  10 CNN.
42 P. Chatziagorakis et al. / Microprocessors and Microsystems 36 (2012) 33–44

Fig. 8. The final network state diagram after the interference of the developed 10  10 CNN (critical scenario).

a macroscopic view, the network state is approximately equal to The final practical conclusion is that the generated 10  10 cir-
zero, as long as the maximum cell state value does not overcome cuit has successfully managed to confront the local thermal distur-
0.08 V. These high state values are located in the external and bance, by applying a global control function that was registered in
angular cells. This phenomenon was something quite expected. its local behavior templates. In addition to that, the 10  10 net-
These boundary cell units do not have any further neighbors to re- work has been able to properly function during the case of the crit-
lieve their internal state. As the data of the initial disturbance is ical scenario, despite the presence of some faulty cells. All the
transmitted throughout the network like a water ripple, its magni- above are some serious indications that the generated 10  10
tude is gradually diminished. So, it is quite logical that the external CNN is an effective, reliable and robust computation tool that could
cells will eventually have the higher state values. Would the net- find use in a variety of other control applications which might also
work be wider, the disturbance would finally be completely de- require global data processing integrated into a compact and low
cayed. However, it should be noticed that the CNN presented a cost computing system.
global homogenous function and even the boundary cells have de-
tected the distant disturbance. It is clear that in the end the devel- 6. Conclusions and discussion
oped 10  10 CNN has both successfully managed to cause the
reduction of the local thermal peak and ensure the total network In this paper, a novel methodology for the design automation of
stability. a CNN model in any desired dimension has been proposed. More
In addition to all the above, another case that was thoroughly specifically, a Matlab algorithm has been developed, providing
examined was a critical scenario that involves the CNN effective- the automatic generation of the CNN netlist through a simple pro-
ness during the malfunction of a finite number of cell units cedure that greatly simplifies the designing process. The corre-
throughout the network. More specifically, an experiment has been sponding file, once obtained, can be imported to Spice, which
carried out having 5 out of 100 cell units in non operative mode will automatically produce the schematic of the corresponding cir-
due to some unknown reason. More specifically, the cells C(5, 4), cuit. Next, a 10  10 network CNN model has been generated via
C(5, 5), C(7, 5), C(8, 4) and C(9, 8) had been disabled. The adopted the proposed methodology and has been used in conjunction with
initial network state was the same as in the previous case. A point a real distributed analog sensors network in a data fusion and con-
of utter significance is the fact that two of the faulty cells (C(5, 4) trol application that involved the thermal stability achievement of
and C(5, 5)) had been located within the center of the thermal peak. a system. Following, a series of experimental testing has been car-
In Fig. 8, the final network state for the critical scenario is pre- ried out and the efficiency, the reliability and the robustness of the
sented. It can be observed that the malfunction of the above cells generated CNN model has been proved. It is quite remarkable that
has not resulted in preventing the CNN from effectively handling even in the case of a partially malfunctioned CNN circuit, the cor-
the situation. The main difference between the two diagrams lies responding results have been far from disappointing. On the con-
in the zero state values of the faulty cells and the different color trary, the CNN has finally achieved in inducing the desired
code of the legend, due to the wider voltage variation. The explana- thermal stability of the system, without greatly deviating from
tion of this phenomenon is quite simple. Despite the fact that the the ideal scenario. Due to the synchronized parallel character of
above cells had been disabled, their direct connections with their the applied control the generated model succeeded in decisively
neighborhood had kept on working. In this way, they had kept confronting the thermal abnormality.
on transmitting their zero state value to their area of influence. In addition, it should be noticed that the effectiveness of the pro-
That zero value could have some great influence on the states of duced CNN is similar to the one of other implementations men-
the adjacent cells, but eventually it did not. The reason why this tioned in literature [30]. However, the generated CNN offers
has happened is that every cell has a maximum of eight neighbors. higher flexibility and evermore greater parameterization. In partic-
So, in case that one of them is continuously sending a zero feed- ular, it adopts a two-dimensional structure instead of the circular
back, the rest of its neighbors manage to balance that error at some one-dimensional found before. Moreover, the usage of the opera-
degree. Consequently, although the CNN is not 100% functioning, tional amplifiers for the construction of the cell circuit makes it more
the final result does not deviate far from the ideal scenario. versatile than the previously suggested implementations, while
 P. Chatziagorakis et al. / Microprocessors and Microsystems 36 (2012) 33–44 43

maintaining at the same time its cost at very low level. Its most sig- [15] D. Snoonian, Smart buildings, IEEE Spectrum 40 (2003) 18–23.
[16] J. Schmalzel, F. Figueroa, J. Morris, S. Mandayam, R. Polikar, An architecture for
nificant feature however, lies on the fact that it can be automatically
intelligent systems based on smart sensors, IEEE Transactions on
generated via the proposed methodology, which offers to its users Instrumentation and Measurement 54 (2005) 1612–1616.
the ability of specifying both the CNN size and its dynamic behavior. [17] V.K. Varadan, V.V. Varadan, Microsensors, microelectromechanical systems
Despite its great features, the proposed methodology has also (MEMS), and electronics for smart structures and systems, Smart Materials
and Structures 9 (2000) 953–972.
some drawbacks, which nonetheless are not capable of decaying [18] K.S. Low, W.N.N. Win, M.J. Er, Wireless sensor networks for industrial
its total usefulness and offer food for thought for the future studies. environments, in: Proceedings of the International Conference on
The proposed methodology does not yet include any mechanisms Computational Intelligence for Modeling, Control and Automation and
International Conference on Intelligent Agents, Web Technologies and
for the validation of the network proper functionality. In the spe- Internet Commerce (CIMCA–IAWTIC), vol. 2, 2005, pp. 271–276.
cific case, the unproblematic functionality of the generated net- [19] J. Demo, A. Steiner, F. Friedersdorf, M. Putic, Development of a wireless
work has been taken for granted as long as its integrity had miniaturized smart sensor network for aircraft corrosion monitoring, in: IEEE
Aerospace Conference, 2010, pp. 1–9.
already been thoroughly tested [43]. In other words, creating dif- [20] A.S.K. Pathan, S.H. Choong, W.L. Hyung, Smartening the environment using
ferent function CNNs first require some testing to be done, which wireless sensor networks in a developing country, in: 8th International
seems to be a quite time demanding task. For this reason, in the fu- Conference of Advanced Communication Technology (ICACT), vol. 1, 2006, p.
709.
ture, a validating function will be incorporated into the methodol- [21] G. Yuan, J. McNair, Fault tolerant three dimensional environment monitoring
ogy, in order to check the circuit functionality in parallel with its using wireless sensor networks, in: IEEE Military Communications Conference
generation. Another issue is that the current methodology is lim- (MILCOM), 2006, pp. 1–7.
[22] H. Qi, S.S. Iyengar, K. Chakrabarty, Distributed sensor networks – a review of
ited in the generation of two-dimensional CNNs. However, it could
recent research, Journal of the Franklin Institute 338 (2001) 655–668.
be easily evolved and eventually gain the ability to produce three- [23] I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, E. Cayirci, Wireless sensor
dimensional networks. networks: a survey, Computer Networks 38 (2002) 393–422.
Our future work is mainly focused on enriching the current algo- [24] J. von Neumann, Theory of Self-reproducing Automata, University of Illinois
Press, Urbana, IL, 1966.
rithm with some additional features and eventually upgrading both [25] M. Mitchell, An Introduction to Genetic Algorithms, MIT Press, Cambridge,
the usability and the effectiveness of the methodology. The first one 1996.
is the capability of adding multiple cell inputs and different behavior [26] L.A. Zadeh, G.J. Klir, B. Yuan, Fuzzy Sets, Fuzzy Logic, Fuzzy Systems, World
Scientific Press, 1996.
templates from cell to cell if needed. That could result in the gener- [27] L.O. Chua, L. Yang, Cellular neural networks: theory, IEEE Transactions on
ation of CNNs that adopt more complex neighborhood interactions Circuits and Systems 35 (1998) 1257–1272.
and thus more inhomogeneous types of behavior. Moreover, an [28] P.G. Tzionas, A cellular neural network modeling the behavior of
reconfigurable cellular automata, Microprocessors and Microsystems 25
additional feature could be the ability to select the neighborhood (2001) 379–387.
type. In particular, the user could have the ability to define the de- [29] R. Porter, J. Frigo, A. Conti, N. Harvey, G. Kenyon, M. Gokhale, A reconfigurable
sired number of the neighbors of each cell, resulting to denser or computing framework for multi-scale cellular image processing,
Microprocessors and Microsystems Special Issue on FPGA-based
sparser neighborhoods depending on her/his application needs. Reconfigurable Computing 31 (3) (2007) 546–563.
[30] P. Arena, S. Baglio, L. Fortuna, S. Graziani, Analog cellular networks for
References multisensor fusion and control, IEEE Transactions on Circuits and Systems 47
(2000) 1378–1382.
[31] M. Haenggi, Distributed sensor networks: a cellular nonlinear network
[1] A. Mainwaring, D. Culler, J. Polastre, R. Szewczyk, J. Anderson, Wireless sensor
perspective, International Journal of Neural Systems 13 (2003) 405–414.
networks for habitat monitoring, in: Proceedings of the 1st ACM International
[32] P. Arena, L. Fortuna, M. Frasca, L. Patane, A CNN-based chip for robot
Workshop on Wireless Sensor Networks and Applications, 2002, pp. 88–97.
locomotion control, IEEE Transactions on Circuits and Systems I: Regular
[2] W. Ning, Z. Naiqian, W. Maohua, Wireless sensors in agriculture and food
Papers 52 (9) (2005) 1862–1871.
industry – recent development and future perspective, Computers and
[33] K.R. Crounse, L.O. Chua, The CNN Universal Machine is as universal as a Turing
Electronics in Agriculture 50 (2006) 1–14.
Machine, IEEE Transactions on Circuits and Systems I: Fundamental Theory
[3] A. Milenkovic, C. Otto, E. Jovanov, Wireless sensor networks for personal health
and Applications 43 (4) (1996) 353–355.
monitoring: issues and an implementation, Computer Communications:
[34] V.M. Preciado, D. Guinea, J. Vicente, M.C. Garcia-Alegre, A. Ribeiro, Genetic
Wireless Sensor Networks and Wired/Wireless Internet Communications 29
programming of a CNN multi-template tree for automatic generation of
(2006) 2521–2533.
analogic algorithms, in: Proceedings of the 6th IEEE International Workshop
[4] H. Bai, M. Atiquzzaman, D. Lilja, Wireless sensor network for aircraft health
on Cellular Neural Networks and Their Applications (CNNA), 2000, pp. 381–
monitoring, in: First International Conference on Broadband Networks, 2004,
385.
pp. 748–750.
[35] J.A.M. Hernandez, F.G. Castaneda, J.A.M. Cadenas, A method for designing CNN
[5] C. Xianghui, C. Jiming, Z. Yan, S. Youxian, Development of an integrated
templates, in: 4th International Conference on Electrical and Electronics
wireless sensor network micro-environmental monitoring system, ISA
Engineering (ICEEE), 2007, pp. 153–156.
Transactions 47 (2008) 247–255.
[36] G. Joya, M.A. Atencia, F. Sandova, Hopfield neural networks for optimization:
[6] G. Werner-Allen, K. Lorincz, M. Welsh, O. Marcillo, J. Johnson, M. Ruiz, J. Lees,
study of the different dynamics, Neurocomputing 43 (2002) 219–237.
Deploying a wireless sensor network on an active volcano, IEEE Internet
[37] Ming-Jung Seow, Hau Ngo, Vijayan K. Asari, Systolic implementation of 2D
Computing 10 (2006) 18–25.
block-based Hopfield neural network for efficient pattern association,
[7] A. Deshpande, C. Guestrin, S.R. Madden, J.M. Hellerstein, W. Hong, Model-
Microprocessors and Microsystems 27 (2003) 359–366.
driven data acquisition in sensor networks, in: Proceedings of the Thirtieth
[38] G.Ch. Sirakoulis, I. Karafyllidis, A. Thanailakis, A CAD system for the
international Conference on Very Large Data Bases, vol. 30, 2004, pp. 588–599.
construction and VLSI implementation of Cellular Automata algorithms
[8] J.K. Hart, K. Martinez, Environmental sensor networks: a revolution in the
using VHDL, Microprocessors and Microsystems 27 (2003) 381–396.
earth system science?, Earth-Science Reviews 78 (2006) 177–191
[39] I.G. Georgoudas, P. Kyriakos, G.Ch. Sirakoulis, I. Andreadis, An FPGA
[9] P. Albertos, G.C. Goodwin, Virtual sensors for control applications, Annual
implemented cellular automaton crowd evacuation model inspired by the
Reviews in Control 26 (2002) 101–112.
electrostatic-induced potential fields, Microprocessors and Microsystems 34
[10] N. Xu, S. Rangwala, K.K. Chintalapudi, D. Ganesan, A. Broad, R. Govindan, D.
(2010) 285–300.
Estrin, A wireless sensor network for structural monitoring, in: Proceedings of
[40] L.O. Chua, Universality and Emergent Computation in Cellular Neural
the 2nd International Conference on Embedded Networked Sensor Systems,
Networks, World Scientific Publishing, Singapore, 2003.
2004, pp. 13–24.
[41] K.R. Crounse, L.O. Chua, Methods for image processing and pattern
[11] Anna Hac, Wireless Sensor Network Designs, John Wiley & Sons, Chichester,
formation in Cellular Neural Networks: a tutorial, IEEE Transactions on
2003.
Circuits and Systems I: Fundamental Theory and Applications 42 (1995)
[12] M. Roberta, J.M. Rivierea, J.L. Noizettea, F. Hermanna, Smart sensors in flexible
583–601.
manufacturing systems, in: Proceedings of Eurosensors VI, vols. 37–38, 1993,
[42] P. Arena, A. Basile, M. Bucolo, L. Fortuna, Fortuna, image processing for medical
pp. 239–246.
diagnosis using CNN, Nuclear Instruments and Methods in Physics Research
[13] D.H. Stefanov, B. Zeungnam, B. Won-Chul, The smart house for older persons
Section A: Accelerators, Spectrometers, Detectors and Associated Equipment
and persons with physical disabilities: structure, technology arrangements,
497 (2003) 174–178.
and perspectives, IEEE Transactions on Neural Systems and Rehabilitation
[43] P. Chatziagorakis, G.Ch. Sirakoulis, J. Lygouras, Automatic generation of
Engineering 12 (2004) 228–250.
cellular neural networks for distributed sensor data processing, in: 13th
[14] M. Chan, D. Esteve, C. Escriba, E. Campo, A review of smart homes – present
Panhellenic Conference on Informatics (PCI), 2009, pp. 35–39.
state and future challenges, Computer Methods and Programs in Biomedicine
91 (2008) 55–81.
44 P. Chatziagorakis et al. / Microprocessors and Microsystems 36 (2012) 33–44

Prodromos Chatziagorakis received the Diploma Chamber of Greece (TEE) for his Diploma Thesis in 1996. He is Associate Editor of
Degree from the Department of Electrical and Computer the Recent Patents on Electrical Engineering and a member of the Institution of
Engineering, Democritus University of Thrace (DUTH), Engineering and Technology (IET), of the European Geosciences Union (EGU), of the
Greece, in 2008. Currently, he is a M.Sc. student in the International Society of Computational Biology (ISCB) and a member of the TEE. He
Democritus University of Thrace. His current research was also founding member and Vice President of the IEEE Student Branch of Thrace
interests include theory and applications of cellular for the period 2000–2001.
neural networks and methodologies for their automated
generation, cellular automata applications in modeling
of physical phenomena, artificial intelligence and neural John N. Lygouras was born in Kozani, Greece. He
networks applications in photovoltaic systems. He is a received the Diploma degree and the Ph.D. in Electrical
member of the Technical Chamber of Greece (TEE). Engineering from the Democritus University of Thrace,
Greece in 1982 and 1990, respectively, both with hon-
ors. From 1982 he was a Research Assistant and since
2000 he is an Associate Professor at the Democritus
University of Thrace, Department of Electrical and
Georgios Ch. Sirakoulis received the Dipl. Eng. and Computer Engineering. In 1997 he spent 6 months at
Ph.D. degree in Electrical and Computer Engineering the University of Liverpool, Department of Electrical
from the Democritus University of Thrace, Greece, in Engineering and Electronics as a Honorary Senior
1996 and 2001, respectively. He joined the Department Research Fellow. He has published more than 100
of Electrical and Computer Engineering, Democritus research articles in international journals and confer-
University of Thrace, Greece, as a non tenured faculty ences. Also, he is the author of three books and is co-author of three book chapters.
member, in 2002, where he serves as Assistant Profes- His research interests are in the field of robotic manipulators trajectory planning
sor. He has published more than 75 technical papers and execution. His interests also include the research on analog and digital elec-
and he is co-editor of one book and co-author of two tronic systems implementation and advanced methods for position control of
book chapters. His current research emphasis is on underwater remotely operated vehicles.
automated electronic systems design, cellular automata
theory and applications, CAD systems, physical design
automation of VLSI systems, applied electronics, bioelectronics and molecular
electronics. Dr. Sirakoulis received a prize of distinction from the Technical