Towards a Novel Engine to Underlie the Data Transmission of Social

Urban Sensing Applications

Carlos Oberdan Rolim

, Anubis Graciela de Moraes Rossetto

, Valderi R. Q. Leithardt

Guilherme A. Borges

, Tatiana F. M. dos Santos

, Adriano M. Souza

and Claudio Geyer

Institute of Informatics, Federal University of Rio Grande do Sul (UFRGS), Porto Alegre, RS, Brazil

Postgraduate Program in Production Engineering, Federal University of Santa Maria (UFSM), Santa Maria, Brazil

Keywords:

Urban Sensing, Smart Cities, Opportunistic Networks, Machine Learning, Prediction, Situation Awareness.

Abstract:

Social urban sensing is a new paradigm which exploits human-carried or vehicle-mounted sensors to ubiq-

uitously collect data for large-scale urban sensing. A challenge of such scenario is how to transmit sensed

data in situations where the networking infrastructure is intermittent or unavailable. In this context, this pa-

per outlines the early stages of our research which is concerned with a novel engine that uses Opportunistic

Networks paradigm to underlie the data transmission of social urban sensing applications. It applies Situation

awareness, Neural Networks and Fuzzy Logic for routing and decision-making process. As we know, this is

the ﬁrst paper to use such approaches in Smart Cities area with focus on social sensing application. As well as

being original, the preliminary results from our simulations signals the way that further research can be carried

out in this area.

1 INTRODUCTION

Smart Cities are urban systems that use Information

and Communication Technologies (ICT) to provide

an infrastructure and public services within a more in-

teractive, accessible and efﬁcient city (Pellicer et al.,

2013). As a result, researchers are seeking alterna-

tives to serve citizens with new services to improve

their quality of life and to fulﬁll the criteria of energy

efﬁciency and sustainability. In this way, urban sens-

ing applications emerges as a promising way to im-

prove the comprehension of such urban ecosystems in

order to assist the decision-makers in the organization

of the city and the welfare of its residents.

In this context, an important question is how to

foster citizen participation and community involve-

ment to achieve a better interaction with the urban

ecosystem. Among these initiatives, social urban

sensing applications are a promising way to pro-

mote the sensing of data about different aspects of

cities, bringing the computational world and commu-

nity closer together.

A challenge for social urban sensing applications

is how to transmit sensed data in situations where the

networking infrastructure is intermittent or unavail-

able. We argue that Opportunistic Networks is an

alternative to overcoming such limitations. Oppor-

tunistic Network is a recent and promising mobile net-

working paradigm that stem from research into con-

ventional Mobile Ad Hoc NET-works (MANET) and

uses contact between mobile nodes to transmit data

(Boldrini et al., 2010).

In this paper, we outline the early stages of our

research on a novel engine that uses Opportunistic

Networks paradigm to underlie the data transmission

of social urban sensing applications. It applies Situ-

ation awareness, Fuzzy Logic and Neural Networks

for routing and decision-making process. This en-

gine will be used as an internal Communication com-

ponent in our Ubiquitous Service-Oriented Architec-

ture for Urban Sensing called UrboSenti (Rolim et al.,

2014). We are based on the hypothesis that a non

“IP-Centric” paradigm like Opportunistic Networks

in conjunction with contextual and intelligent ap-

proaches could supply the requirements of data trans-

mission in wide-scale urban scenarios. In summary,

the main contributions made by this paper are the ar-

chitecture of the engine and the conceptual models

used as guidance for its development. As we know,

this is the ﬁrst paper to use such approaches in Smart

Cities area with focus on social sensing application.

The rest of this paper is structured as follows: The

662

Oberdan Rolim C., Graciela de Moraes Rossetto A., R. Q. Leithardt V., A. Borges G., F. M. dos Santos T., M. Souza A. and Geyer C..

Towards a Novel Engine to Underlie the Data Transmission of Social Urban Sensing Applications.

DOI: 10.5220/0005457406620667

In Proceedings of the 17th International Conference on Enterprise Information Systems (ICEIS-2015), pages 662-667

ISBN: 978-989-758-097-0

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

next section describes the motivational scenario and

raises some of the current computational challenges;

Section 3 describes the proposed architecture; Section

4 presents preliminary results from simulations, and,

ﬁnally, in Section 5 some conclusions are reached, to-

gether with recommendations for future research.

2 PROBLEM SCENARIO

Our research has been driven by the problem-scenario

that is shown in Fig. 1. This scenario includes a city

with several data sources that are being used for sens-

ing. Human-carried, ﬁxed or vehicle-mounted sen-

sors are applied for obtaining sensing maps of tran-

sits, air quality, noise levels, temperature, CO2 con-

centration, etc. Moreover, data from social networks

in conjunction with sensors data are crucial to under-

stand the behavior of the city and to provide a holistic

view about it. To collect, analyze and give feedback

of sensed data acquired from several sources scattered

along the city, we are using our Ubiquitous Service-

Oriented Architecture for Urban Sensing called Ur-

boSenti. The main function of UrboSenti is to pro-

vide support for overall process of urban sensing. It

splits into two key modules: the Backend module and

Sensing module.

Figure 1: Problem scenario.

The Backend module runs in a data center infras-

tructure and, in short, is responsible for receiving

sensed data, processing it and giving feedback to the

citizens and other systems.

The Sensing module is responsible for social and

traditional sensing and encompasses activities of in-

tentional and non-intentional sensing. It runs in mo-

bile devices (e.g. mobile phones, embedded in ve-

hicles, etc) and in ﬁxed sensors scattered around

the city. It has a several components that could be

plugged “on demand” and a micro-kernel with set of

components that are responsible for essential features.

Our focus is the internal micro-kernel Communica-

tion component. It provides methods to send and re-

ceive data by means of the available network infras-

tructure, such as IEEE 802.11b/g/n (structured and

ad-hoc), GPRS/EDGE/3G and Ethernet as the under-

lying system for TCP/UDP communications. When

the network infra-structure is intermittent or unavail-

able, it supports data communication using alternative

ways, not based in end-to-end paths (like used in IP

communications). In summary, the Communication

component is the “power-horse” of all communica-

tions tasks in the Sensing module.

Hence, we are seeking for an engine to be used

as underlying for communication of such component.

It should handle with fuzziness of wide-scale urban

scenarios and provide suitable support for data trans-

mission of urban sensing applications running atop of

UrboSenti. For such task, the requirements are: (i)

to use non “IP-Centric” paradigm for communication;

(ii) adapts itself the transmission parameters accord-

ing to device used for sensing and the current con-

text; (iii) made adaptation decisions proactively; (iv)

perform decisions using uncertain data; (v) concerns

with processing and power restrictions of devices.

In this way, we are based on the hypothesis that

non “IP-Centric” paradigm like Opportunistic Net-

works could be used for data carrying, satisfying req-

uisite (i); Situation awareness could be applied to

deal with context adaptations, satisfying requisite (ii);

Neural Networks could make predictions to support

adaptations, satisfying requisite (iii) and; Fuzzy Logic

could be used for decision-making about routing and

internal adjusts, satisfying requisite (iv). We highlight

that all approaches are suitable to run in low powered

devices, satisfying requisite (v).

3 PROPOSED ENGINE

A challenge for social urban sensing applications is

how to transmit sensed data in situations where the

networking infrastructure is intermittent or unavail-

able. We argue that Opportunistic Networks paradigm

could fulﬁll such gap. There are some initiatives in

Opportunistic Networks are

Our proposed engine is lying in Opportunistic

Networks area. There are some initiatives in such

area, like Epidemic, Spray&Wait (and the Spray

TowardsaNovelEnginetoUnderlietheDataTransmissionofSocialUrbanSensingApplications

663

variants like Spray&Focus, Fuzzy-Spray and others),

Prophet, BubbleRap, MobySpace, AFRON, Cartoon,

CAR, HiBOp, CiPRO, Propicman and the most recent

Prophet improvement called DRAFT. Due to space

limitation and paper scope we will not present more

information about these initiatives – for further infor-

mation see (Jedari et al., 2013). However, we re-

mark that none could handle all requirements pre-

sented above and cannot be used “as is” in our sce-

nario. Thus, in this section we will present our engine

architecture to overcome such limitations.

3.1 Situation Awareness Model

Situation-awareness (SAW) is a computing paradigm

by which applications use context data to sense and

comprehend the current situation and to project the

future demands. The ﬁrst formalization of SAW was a

3-levels model proposed by Endsley(Endsley, 1995).

To deal with context adaptation in a proactive fashion,

our engine implements a Situation awareness model.

It is based in 3-levels model proposed by Endsley and

is outlined in Fig. 2.

Figure 2: Situation model.

This model uses internal and external context

about node to derive the low-level context that char-

acterizes current situation. A set of rules and past sit-

uations are used to project future situations that will

be used for routing decisions.

3.2 Prediction Model

The Prediction model (Fig. 3) is used to project fu-

ture situations. A recurrent Neural Network (NN) is

used for this purpose. We have chosen NN due to its

capacity to solve non-linear problems, it are universal

functions aproximators being suitable for prediction.

Each node is responsible to train and runs its own

instance of NN. With this approach we ensure that

Figure 3: Prediction model.

each node have a suitable NN to its needs. The cur-

rent and past low-level context data are used as input

for NN. It starts the Training phase, testing several

conﬁgurations from a conﬁguration repository trying

to ﬁnd the optimal Network (with lower Root Mean

Squared Error – RMSE). When a optimal Network is

found, it is used for prediction in Exploitation phase.

The outputs of this process are new predicted low-

level context data that probably characterize a future

situation of node. These data are used in decision-

making process. An essential requirement for NN in

this case, is a low computational cost due to power

and processing constraints of mobile nodes.

3.3 Decision Making Model

The decision-making process encompasses precise

and non precise context data (from current and pro-

jected future situation). So, conventional logic may

produce completely wrong decisions due to uncertain

of such data. An alternative to deal with imprecise

data is Fuzzy Logic. According to (Zadeh, 2008),

Fuzzy Logic is a viable alternative to reason and make

rational decisions in an environment of imprecision,

uncertainty, incompleteness of information, conﬂict-

ing information, partiality of truth and partiality of

possibility. Thus, to perform decision-making pro-

cess, a Fuzzy Inference System (FIS) is used (Fig.

4). Context values (current and future) are pumped as

input to FIS and after internal calculation (by usage

of membership functions and fuzzy rules) the output

is the potential of a node to be a good “data mule”

(a term used in Opportunistic Network to designate

a node that will carry a message). This information

will be used to decide when to forward messages to

encountered node. The decision-making process also

runs to decide if some internal parameters needs to be

adjusted.

ICEIS2015-17thInternationalConferenceonEnterpriseInformationSystems

664

Figure 4: Decision making model.

3.4 Engine Architecture

The above presented models are used by our engine.

Its internal architecture features are outlined in Fig. 5

and its behavior is explained below. Again, we high-

light that due to space limitation we will not fully ex-

plain each component.

Figure 5: Engine architecture.

It starts with the Contextual Information that rep-

resenting information about the context of node. At a

constant time interval, Context Collector collects the

data and stores them in a layered structure called Con-

textual Graph, thus creating a new Layer 2 vertex.

The Contextual Graph component underlies all

the data storage. Its main function is to store instan-

taneous and predicted context information. In Con-

textual Graph, the vertexes of the graph are struc-

tured in layers: Layer 1 stores basic information

about the node (i.e. node name, address, network

interfaces). Layer 2 stores instantaneous context in-

formation about the node (i.e. node power, current

position, buffer usage, number of messages, current

time, speed, distance traveled from last point, num-

ber of reachable neighbors) that will be used as his-

torical values to prime the Forecaster. Layer 3 stores

predicted context values from Forecaster. Moreover,

we used edges to represent the contacts between the

nodes.

The Situation Manager component, implements

our Situation awareness model. It draws on data from

Contextual Graph to build, analyze, project and cre-

ate a repository of situations. The information gener-

ated by this module will be used later by the Decision

Maker. It also runs maintenance routines like pruning

old data and invoking Forecaster for prediction. At

time intervals Situation Manager retrieve context data

from Contextual Graph and using a set of rules stored

in its internal situation repository it tries to identify

(build) the current node’s situation. The identiﬁed sit-

uation (e.g. “node is sensing with low battery power

and high buffer usage”) is analyzed, and if it indicates

that some action needs to be done it is reported to De-

cision Maker. If a situation could not be identiﬁed, an

unknown situation is found. Thus, a new set of rules

that characterize this situation is created “on the ﬂy”

and stored in repository for future use. When the Situ-

ation Manager component detects a sufﬁcient amount

of context information, it could project a future situ-

ation. For this purpose, the Forecaster component is

invoked.

Forecaster component implements our prediction

model, in order to predict the probable values that

will characterize a future situation. For such task, we

applied a lightweight network called Echo State Net-

work (ESN) proposed by Jaeger(Jaeger, 2001) is ap-

plied. ESN are a kind of three-layered recurrent net-

work with sparse, random, and (crucially), untrained

connections within the recurrent hidden layer (for fur-

ther and mathematical foundation please read original

Jaegers paper). The main difference between the ESN

and other neural networks is that only the weights of

the reservoir output signals are trained. The weights

of the connections within the reservoir are not trained

but are randomly generated. This approach signiﬁ-

cantly reduces the learning process when compared

with other algorithms (e.g. back propagation through

time) resulting in low computational cost to imple-

ment it(Fink et al., 2013). The Forecaster compo-

nent, uses context values stored as Layer 2 in Contex-

tual Graph as historical data to train ESN (i.e. node

power, current position, etc.). Due to its low com-

putation cost to train the neural network, we are able

to make each node of the network to builds its own

ESN with the most appropriate conﬁguration for its

context. This is carried out by testing different inter-

nal parameters of ESN (i.e. size of reservoir, spar-

sity of the reservoir, spectral radius and leaking rate)

with different values until the best one (i.e. the one

TowardsaNovelEnginetoUnderlietheDataTransmissionofSocialUrbanSensingApplications

665

with minimal MSE) is found. When optimal neu-

ral network is found, its conﬁguration is stored. At

this point, optimal network is ready to predict future

values in the exploitation phase. During exploitation

phase, the structure with historical data is used to

“pump” the best network (found and saved in previous

phase) with some data steps and thus to activate the

internal reservoir. Some stages later, the input from

the historical data is switched off to allow the network

to predict values by itself. The predicted values are

stored at Contextual Graph as Layer 3 vertices. At

this point, the component sets an internal variable to

indicate that the node is now running in smart mode.

In smart mode, all decisions done by Decision Maker

are made using past, current and predicted data in or-

der to improve data transmission. In “dummy” mode,

just current situation is used in decisions.

Decision Maker runs at constant time intervals to

decide if some internal parameters needs to be ad-

justed (such as buffer scheduling policy, maximum

size of messages, time to live of new messages, etc.)

or if a “trap” should be triggered to require atten-

tion of an external component of micro-kernel (e.g to

change conﬁguration of network interface, to perform

some adaptation action, etc.). All decisions are made

using decision making model presented above in sec-

tion 3.3. Decision Maker is also invoked when current

node contact another node to decide if some buffered

message should be forwarded, delivered or remain at

the local buffer. In simple terms, Decision Maker de-

cides if the encountered node is a good “data mule”.

We used the term “potential” to represent the capacity

of the node to be a good data mule. The strategy used

is quite simple: if the potential of the contacted node

is greater than the potential of the current node, then

the message is forwarded; otherwise, the message re-

mains at the local buffer (obviously the message is

delivered if the encountered node is its destination).

The question arising from this approach is: how to

calculate the potential of each nodes? For this task,

all context values (current, historical and predicted)

of the current and contacted node from Contextual

Graph and use it as input for the Fuzzy Inference Sys-

tem (FIS). FIS uses its internal components and rules

to calculate the potential of each node. The potential

values of nodes are used by Decisor in decision mak-

ing process.

4 PRELIMINARY RESULTS

To verify the functionality and performance of pro-

posed engine we implemented some main modules

and carried out some simulations using ONE (Oppor-

tunistic Network Environment) Simulator. For simu-

lation setup we adopted 6 hours for all scenarios with

a different number of nodes (10 for the ﬁrst, 25 for

the second, and 50, 75 and 100 for each consecu-

tive group). We used two groups: pedestrians and

cars with ShortestPathMapBasedMovement as mo-

bility model. Pedestrian nodes moved between 0.5

and 1.5 Km/h, and had a Bluetooth device with a ra-

dio range of 20 meters and transmission speed of 2

Mbit/s. The Car nodes moved between 10 and 50

Km/h and had a Wi-Fi interface with a range of 50

meters and transmission speed of 10 Mbit/s. On aver-

age, the nodes generated about one message every 25

to 35 seconds (total of 711) and the message lifetime

was set at 24 minutes (1440 seconds). We used mes-

sage sizes that were uniformly distributed between

100 KB and 2 MB.

In implementation of engine, we used ESN-

Java software

to build Forecaster and JFuzzyLogic

library

in the Fuzzy Inference System used by

Decision-Maker. The following context data was

used: current power, current speed, total distance

traveled from last point, overall distance traveled, cur-

rent coordinates, last coordinates, current buffer us-

age, current number of carried messages, total num-

ber of forwarded messages, current number of neigh-

boring nodes, and total number of connections. To

calculate the variable “potential” which is used as the

output of FIS, three Triangular membership function

are used. The COG (Center Of Gravity) was used as

a defuzziﬁcation method. The fuzzy inference rules

were deﬁned in compliance with Fuzzy control lan-

guage (FCL).

We conducted a set of experiments using simula-

tion setup presented above with different scenarios.

The results are displayed in Fig. 6. This shows that

in general, there is an increment in the number of de-

livered messages that corresponds to the increment of

the number of nodes.

Figure 6: Performance in scenarios with different number

of nodes.

http://www.wsi.uni-tuebingen.de/lehrstuehle/cognitive-

modeling/code/overview.html

http://jfuzzylogic.sourceforge.net/

ICEIS2015-17thInternationalConferenceonEnterpriseInformationSystems

666

The scenario with 100 nodes had a increment of

245% in number of delivered messages in relation to

scenario with 10 nodes, but with just 134% of more

overhead (an assessment of bandwidth efﬁciency in

relation of the number of relayed and delivered mes-

sages). This overhead percentage is less than in sce-

nario with 50 and 75 nodes. We believe that this ra-

tio can be attributed to our strategy of just relaying

messages to data mules with a good potential to de-

livery the message. Furthermore, we can note that

with increment of number of nodes, each one could

store more historical context data to be used by Fore-

caster for prediction. With this, the predictions be-

comes more accurate over the time. In other words,

the engine becomes smarter when run more time and

with more neighbors nodes. We need to investigate

why with 50 nodes the number of delivered messages

was less than with 25 nodes. I As ﬁnal remark, one

factor not reported in the chart is the computational

cost of ESN. Even when each node used in the sim-

ulation testing, had, on average, 810 different conﬁg-

urations to ﬁnd the best network, the impact of the

processor load was minimal. This lightweight fea-

ture was the main differential of ESN when compared

with all the other machine learning approaches that

we have tested in our previous work.

5 CONCLUSIONS

In this paper, we have described the early stages of our

attempt to build a novel engine that applies Oppor-

tunistic Networks paradigm to transmit sensed data in

situations where the networking infrastructure is in-

termittent or unavailable. It runs as an internal com-

ponent of a wide architecture called UrboSenti and

provides support for communication of urban sensing

applications running atop of it.

We have also outlined our initial design models

for the software modules and their internal compo-

nents. The development of engine has been started.

Currently we are mainly working to implement Situ-

ation awareness model and to plug it with other com-

ponents. The preliminary results, without situation

awareness, are acceptable and indicates that our initial

hypothesis could be exploited better. The low com-

putational cost to run it with satisfactory number of

delivered messages has shown that it works and have

a good potential to be used in UrboSenti. We believe

that its performance will be improved when imple-

mentation of Situation Manager is ﬁnished.

Thus, we claim that the proposed engine at-

tend our requirements and is able to ﬁll the gap of

data transmission presented in our initial problem-

scenario. Moreover, this should encourage us to con-

duct further research into the multidisciplinary area of

Smart Cities with the aim of improving services and

applications for urban sensing.

For future work, we are seeking alternative means

of constructing fuzzy sets and rules “on the ﬂy”, de-

pending on the situation in which the node is im-

mersed and to explore the application of a Deep Belief

Network (DBN) or Restricted Boltzmann machines

(RBMs) as underlying for prediction.

ACKNOWLEDGEMENTS

We would like to give special thanks to Johannes

Lohmann of Universit

at W

urzburg for his help with

source code of ESNJava.

This research was supported under UbiArch project

– Ubiquitous Architecture for Context Management

and Application Development at UFRGS.

REFERENCES

Boldrini, C., Conti, M., and Passarella, A. (2010). Design

and performance evaluation of ContentPlace, a social-

aware data dissemination system for opportunistic

networks. Computer Networks, 54(4):589–604.

Endsley, M. R. (1995). Toward a Theory of Situation

Awareness in Dynamic Systems.

Fink, O., Zio, E., and Weidmann, U. (2013). Anticipating

Railway Operation Disruption Events Based on the

Analysis of Discrete-Event Diagnostic Data. Ehem-

ical Engineering Transactions, 33:715–720.

Jaeger, H. (2001). The echo state approach to analysing

and training recurrent neural networks. Technical re-

port GMD report 148. Technical report, German Na-

tional Research Center for Information Technology.

Jedari, B., Xia, F., and Member, S. (2013). A Survey on

Routing and Data Dissemination in Opportunistic Mo-

bile Social Networks. IEEE Communications Surveys

and Tutorials.

Pellicer, S., Santa, G., Bleda, A. L., Maestre, R., Jara, A. J.,

and Skarmeta, A. G. (2013). A Global Perspective

of Smart Cities: A Survey. 2013 Seventh Interna-

tional Conference on Innovative Mobile and Internet

Services in Ubiquitous Computing, pages 439–444.

Rolim, C., Rossetto, A., Leithardt, V. R. Q., Borges,

G., dos Santos, T. F. M., Souza, A., and Geyer,

C. F. R. (2014). Towards a Ubiquitous Service-

Oriented Architecture for Urban Sensing. In ASE

BIGDATA/SOCIALCOM/CYBERSECURITY Confer-

ence. Academy of Science and Engineering (ASE),

USA, ASE 2014.

Zadeh, L. a. (2008). Is there a need for fuzzy logic? Infor-

mation Sciences, 178(13):2751–2779.

TowardsaNovelEnginetoUnderlietheDataTransmissionofSocialUrbanSensingApplications

667