Undefined 0 (2017) 1 1IOS Press

Capturing the contributions of the semanticweb to the IoT: a unifying visiona IRIT, UMR 5505, UT2J, Departement de Mathematiques-Informatique, 5 allees Antonio Machado,F-31058 Toulouse Cedex, FranceE-mail: firstname.lastname@irit.frb LAAS-CNRS,7 avenue du Colonel Roche,F-31400 Toulouse, FranceE-mail: firstname.lastname@laas.frc Univ de Toulouse, INSA, LAAS, F-31400, Toulouse, France

Nicolas Seydoux a,b,c, Khalil Drira b,c Nathalie Hernandez a Thierry Monteil b,c

Abstract. The Internet of Things (IoT) is a technological topic with a very important societal impact. IoT application domainsare various and include: smart cities, precision farming, smart factories, and smart buildings. The diversity of these applicationdomains is the source of the very high technological heterogeneity in the IoT, leading to interoperability issues. The semantic webprinciples and technologies are more and more adopted as a solution to these interoperability issues, leading to the emergence ofa new domain, the Semantic Web Of Things (SWoT). Scientific contributions to the SWoT are many, and the diversity of archi-tectures in which they are expressed complicates comparison. To unify the presented architectures, we propose an architecturalpattern, LMU-N. LMU-N provides a reading grid used to classify processes to which the SWoT community contributes, and todescribe how the semantic web impacts the IoT. Then, the evolutions of the semantic web to adapt to the IoT constraints aredescribed as well, in order to give a twofold view of the convergence between the IoT and the semantic web toward the SWoT.

Keywords: Semantic Web of Things, Architectural pattern, Internet of Things, Semantic web, survey

1. Introduction

The internet has evolved toward a globally con-nected environment. The interconnection of devices,smartphones, services, etc, form what is called the In-ternet of Things (IoT). A detailed definition of theIoT is provided in [1], embracing the diversity of na-ture and purpose of the so-called Things. In this paper,the term IoT refers to the technologies and principlesenabling the deployment of devices and services net-works. IoT networks are based on very diverse techno-logical stacks, for hardware, software and communi-cations. The integration of traditional technologies ofthe web in the IoT, such as URIs for resource nam-ing, HTTP connectivity, or REST interfaces, is calledthe Web Of Things (WoT), defined by [2] as ”A wayto realize the IoT where (physical and virtual) things

are connected and controlled through the World WideWeb”. This definition is used as a reference by theW3C Web of Things Working Group1.

Indeed, the IoT and WoT ”Things” are now an ev-ery day life reality for many people. To support thisclaim, one might consider the increasing number ofsmart cities including but not limited to Dublin (IR)2,Santander (ES)3, Milton Keynes (UK), San Francisco(US), New York (US), Yokohama (JA), that are beingequipped in order to monitor their environment and tooffer innovative services. It is estimated that by 2020,26 billion devices will be connected by Machine-to-Machine (M2M) technologies, compared to the 0.9 bil-

1https://www.w3.org/WoT/WG/2http://smartdublin.ie/3http://www.smartsantander.eu/

0000-0000/17/$00.00 c© 2017 – IOS Press and the authors. All rights reserved

arX

iv:1

709.

0357

6v1

[cs

.CY

] 3

0 A

ug 2

017

2 N. Seydoux et al. / A unifying vision for the Semantic Web of Things

lion connected in 2009[3]. The exponential multiplica-tion of connected devices is correlated with a dramaticincrease in the volume of exchanged data.

The growing integration of connected Things to hu-man activities is partly driven by the wide scope of ap-plication domains covered by the IoT: environmentalmetering, transportation, home automation, e-health,agriculture, manufacturing...

However, the diversity of IoT-based applications isrestricted by the approach of the industry so far, whichhas been oriented toward vertical silos, as [4] pointsout. Proprietary systems are designed with a specificpurpose, and on top of the devices the vendor also dis-tributes the application serving the purpose. This ap-proach cannot match the diversity of possible scenariosdriven by user requirements, and raises interoperabil-ity issues. Users should be allowed to combine theirconnected devices in a personalized fashion, and an ap-plication developers should be able to deploy genericapplications that adapt to the available devices, whichis totally opposed to vertical integration. Furthermore,such customization of one’s IoT network requires de-vices to be able to understand each other: they must besemantically interoperable. [5] and [6] make a dis-tinction between syntactic interoperability, or the abil-ity for systems to exchange content, and semantic in-teroperability, the ability to understand the exchangedmessages. The extreme technological diversity[7], aswell as the multiplicity of standards (including butnot limited to oneM2M4, OIC5, AllJoyn6, LWM2M7),makes syntactic interoperability at a large scale a non-trivial issue. However, solving syntactic interoperabil-ity issues only is not sufficient to achieve the completepotential of the IoT, and it is not in the scope of thispaper.

Semantic interoperability is an important enablerfor the future development of the IoT [8]. Indeed,the core of the IoT is M2M communication. For in-stance, the observations gathered by sensors are meantto be distributed to other devices, and not to be humanreadable without processing. Since no human can re-contextualize exchanges between devices, messagesshould be understandable from a machine to an-other.

As [9] or [8] points out, semantic web principlesand technologies can provide solutions to the issues the

4http://www.onem2m.org/5https://openconnectivity.org/6https://allseenalliance.org7http://www.openmobilealliance.org

IoT is facing. The use of dereferencable vocabulariessuch as ontologies enables the capture of metadata ina machine-understandable way, easing M2M commu-nication. Many recent research contributions from thesemantic web community inject semantic web capa-bilities (rich content description, reasoning...) into theWoT in order to develop the so-called Semantic WebOf Things (SWoT), an evolution of the WoT wherethe IoT is extended with semantic web principles andtechnologies. The publication of resources from an IoTnetwork to the SWoT can feed the Linked Open Data(LOD), in order to make the collected data available.

This paper depicts a landscape of the SWoT, andprovides a reading grid to classify the research contri-butions transforming the IoT into the SWoT. First, tostructure the analysis, section 2 defines Lower, Mid-dle and Upper Node (LMU-N), a node-centric IoT ar-chitectural pattern. LMU-N provides a framework tosupport the identification of recurrent patterns in ex-isting research. It is used to describe how the seman-tic web principles and technologies contribute to theIoT in section 3. On the other hand, section 4 focusesthe challenges the semantic web faces to be compliantwith the IoT constraints, and how recent semantic webcontributions propose to adapt to them. Finally, section5 concludes this paper and proposes some perspectivesfor the future of SWoT.

2. Unifying the heterogeneity of architectures withLMU-N

After having motivated the need for a unifying ar-chitectural pattern, this section presents LMU-N andits main components: nodes and messages flows. Othersurveys are listed as related work, before showing howLMU-N can be used as a reading grid for our survey.

2.1. Motivations

The papers associating the semantic web and the IoTare many, and their publication rate is increasing: from15 publications in 2003, to 494 publications in 20168.The total count of publications amounts to 1426. Wefocused on 71 scientific publications for the survey,chosen for their quality, their innovative aspect, and

8After a study on http://ieeexplore.ieee.org, http://www.sciencedirect.com and http://dl.acm.org,searching for the exact keywords ”semantic web” and ”internet ofthings”

N. Seydoux et al. / A unifying vision for the Semantic Web of Things 3

Fig. 1. The LMU-N pattern

for the balance in their content between semantic weband IoT. An extra attention was given to publicationsproposing semantic web contributions in explicitly andprecisely defined IoT architectures. These publicationsdescribe contributions that can be competing or com-plementary, mutually exclusive or not, etc. In order tobe able to capture a structured landscape of the contri-butions of the semantic web to the IoT, one needs tobe able to compare these contributions, and to organizethem with respect to one another. To do so, we proposeLMU-N, a unifying architectural pattern that aimsat identifying the core components of an IoT networkin a generic manner, and to characterize their relation-ship. LMU-N is then used as a framework to contextu-alize semantic web contributions to the IoT.

2.2. Constituting elements of LMU-N

LMU-N has two major components: the nodes, orthe ”Things” communicating on the IoT, and the flows,representing the communications between the nodes.LMU-N is the result of a bottom-up analysis, and itsdescription is linked to a set of pre-existing architec-tures from which it was deduced. The pattern is rep-resented as an UML model in figure 1. The pattern iscomposed of three main classes: the Node, specializedin three subclasses, the Message flow, and the Process.The nature of these classes are respectively describedin sections 2.2.1, 2.2.2 and 2.3.

2.2.1. Representing physical device and virtualservice as Nodes

Definition of an IoT node The IoT is a network ofThings connected together, and can be seen as a graphwhere the Things are vertexes and their connec-tions are the edges. So far, the notion of Thing hasbeen used to indifferently refer to devices (sensors and

actuators) and to services. This choice is justified bythe similarity of the roles devices and services have onIoT networks: a service can be seen as the endpointof a device for the other nodes of the IoT, and a de-vice can be seen as the physical implementation of aservice, depending on the chosen perspective.

That is why we introduce in this section the node, anabstraction covering both device and service concepts.Fundamentally, a node is an active entity that can beaddressed on the network. An active entity is able tosend and/or receive requests, which is the case for bothdevices and services. This notion of node is alreadypresent in architectures described in [10] or in [11].[10] describes a contribution based on oneM2M9, astandard architecture structured around different nodeswhich interfaces are formally defined to enforce inter-operability. Blurring the line between physical objectand virtual service is also at the core of propositionssuch as [12], [13] or [14], where ”virtual entity” is usedas an abstraction to enable the composition of servicesor devices.

Section 3 lists approaches that are either directed todevices or services, and their seamless integration andcomparison is enabled by the adoption of the node asthe central architectural element. Moreover, present-ing nodes over devices or services serves the purposeof interoperability by easing the homogeneous mod-elling of their capabilities. For instance, virtual entitiesin [15] are used to capture the characteristics of a nodemaking it relevant to its neighbours.

Abstracting devices and services into a genericnode also aims at focusing on their intrinsic char-acteristics, such as processing power or communica-tion abilities. Since the early studies on semantic sen-sor networks, the description of devices and serviceshas been a primary concern of the SWoT community.Contributions like [16], [17] or [18] propose ontolo-gies to define both the capabilities of a device, and toadd metadata to the content it handles. Section 2.3.2details how node descriptions are used to integrate anIoT deployment in the SWoT.

Classification of LMU-N nodes All nodes of anIoT are not equivalent: some devices are very con-strained, whereas the servers providing analytic capa-bilities are powerful machines. For instance, in its vi-sion of urban IoT, [19] uses the notion of node, andproposes to use two different protocol stacks accordingto the capabilities of each node (constrained or not).

9http://onem2m.org/

4 N. Seydoux et al. / A unifying vision for the Semantic Web of Things

In this paper, different characteristics are used to clus-ter nodes into meaningful classes constituting LMU-N, represented as attributes of the Node class in fig.1. The core characteristic of a node is its processingpower, i.e. its ability to apply treatments of varyingcomplexity to content. The processing power also de-termines the ability of the node to process content ofa varying expressivity, from the very simple agreed-upon byte array to the much more complex Knowledgebase (KB) instantiations. The higher a node’s process-ing power is, the more expressive content it can pro-cess, and the more complex operations it can achieve.Nodes are also characterized by their memory, i.e. thequantity of information it can hold at a given time, andstorage capability. Storage is the available space giv-ing access to persistent content. The notion of IoT isinseparable from the notion of connectivity, and a nodecan also be classified according to its communicationcapabilities. These capabilities include the protocolsit supports, its general availability on the network, andits bandwidth. Nodes also differ by the nature of theirenergy source: while some nodes are attached to tradi-tional power grids, other nodes, deployed in the field,are reliant on batteries, or on renewable energy sourceslike solar panel, or energy harvesting.

Three homogeneous types of nodes are identified,constituting the three layers of LMU-N. These sub-classes of Node are represented on fig. 1, and are de-scribed in tables 2 and 1 regarding their typical char-acteristics and roles on the network.

– Upper Node (UN) are the ”weakly constrainednodes” of [19], or Infrastructure Node in [10]. Inthis category, we classify cloud or local serversas well as powerful mobile devices, includingstandard laptops and smartphones and domain-specific mobile robots or machines. They havehigh processing power, extended communicationcapabilities, and large storage capabilities. For in-stance, in [20], UN are in charge of applying datafusion operators to data streams. In [21] or [1],the UN are situated in the cloud.

– Middle Node (MN) are very often referred to inthe literature as gateway ([22], [10], [4]), becausethey are bridges between powerful UN and moreconstrained nodes. They are usually dedicated tocontent transformation and protocol bridging: in[23], MN are presented as intelligent nodes wherecontent can be converted from its raw representa-tion to a richer one. [4] proposes an architecturewhere the gateway is a contact point between

Fig. 2. An instantiation of LMU-N

the IoT and the SWoT, performing both anno-tation and protocol proxying. In [24] and [19],MN are proxies for wireless devices networks.MN typically have medium processing power, ex-tended communication capabilities, and restrictedmemory storage: they are nodes where contentcollected by sensors or sensing services is col-lected, transformed, and redistributed.

– Lower Node (LN) are typically connected de-vices, with very limited power source, processingand communication capabilities, and very limitedto no storage capabilities. These are by definitionpresent in every IoT architecture, and in directcontact with the physical world. In some paperssuch as [20] or [25], these nodes are not directlypresent, their representation is wrapped by a MN.The LN were in early studies mainly sensors, likein [22], and evolved toward diverse nodes includ-ing actuators, displays and composite devices inmore recent work such as [10].

Figure 2 represents an instantiation of the LMU-N pattern based on the work described in [18]. Twolamps, a temperature sensor, and a pressure sensorsinstantiate lower nodes. Their data is collected in arequest-response fashion by two gateways. The com-munication is initiated by the gateway, therefore thearrows are directed from the gateways to the devices.The two upper nodes represent the server hosting theenrichment application, and the robot which processesthe content generated by the sensors. The gatewaysthemselves have push capabilities, and can also bequeried directly by the server: the communication goesboth ways. SmartSantader10 gives another simple ex-

10http://www.smartsantander.eu/

N. Seydoux et al. / A unifying vision for the Semantic Web of Things 5

ample of real-world deployment conform to the LMU-N pattern.

All the architectures presented in this section canbe described by LMU-N, supporting its unifying sta-tus. A notable exception to the three-layers architec-ture model is presented in [26], where the authors pro-mote a direct connection between devices and applica-tion servers. They propose to this end an architecturalapproach where each node is equipped with a high-efficiency RDF serializer and tuple store. However, inour opinion, this approach requires the devices to beable to run such a stack, and reduces the possibility toinclude non semantic-aware, legacy devices in an IoTsystem. Moreover, we think that directly connectingLN to UN in a flat architecture is a source of scalabilityissues in large systems. That is why LMU-N is struc-tured with an intermediary level to mediate betweenpotentially very constrained nodes and the rest of thenetwork.

2.2.2. Representing exchanges between nodes asmessages flows

In LMU-N’s representation of an IoT network as agraph, the edges are communications between nodes,instantiated by messages flows. We will refer to thesemessages’ expressivity according to the classificationprovided by [27] as data, information or knowledge,organized in the Data, Information, Knowledge andWisdom (DIKW) hierarchy. Typically, sensors pro-duce data, and this data is later on contextualized tobecome information, and enriched using semantic webprinciples and logic in order to produce knowledge.

Furthermore, the IoT relies on already existing in-ternet technologies and paradigms. Nodes hosting ser-vices are servers, and they offer their services to ap-plications or other nodes that are clients in a classicclient-server architecture. This model implies asym-metry in the communications among nodes: in a client-server model, the interaction is initiated by the client.This entails that on the graph of nodes, the edges areoriented, from the client to the server node.

As shown on fig. 1, flows can be directed in threegeneral directions: horizontal for nodes of the samelevel, such as the work done in [28], upstream (as in[29] or [30]), when the source node is of a lower levelthan its destination node, and downstream (as in [31])otherwise. The notions of upstream and downstreamcommunication in IoT architectures are presented by[31]. The content exchanged in these flows can eitherbe application-dedicated, for instance in [20], [30] or[31], specific to the function of the devices that col-

lected it (temperature observations, user requests...),or system-dedicated, describing the nodes constitutingthe network, like in [24].

2.3. Using LMU-N to classify SWoT contributions

By describing IoT systems with the LMU-N pattern,different recurring processes are identified, agnos-tic to the underlying technology stack and to the ap-plication domain. The aim of this paper is to identifyand describe the processes that can benefit from an in-tegration into the SWoT, and how they are instantiatedin research contributions.

The processes are instantiated by exchanges of mes-sages between nodes, supported by flows as describedin section 2.2.2. Therefore, our analysis is based onthe assumption that (i) processes are directed (hori-zontal, upstream or downstream) and (ii) dependanton the characteristics of the nodes at stake. Contri-butions brought to the SWoT are dedicated to certainprocesses, and they can be located in LMU-N depend-ing on the nodes they involve. For instance in [32] or[33], the process of content enrichment is an upstreamprocess, as the content is produced in lower nodes andenriched later on in middle or upper nodes. Processescan also be associated with specific flows content: forinstance, in [34], the node exposition process is basedon node metadata, and not with applicative content.

All the processes introduced in this paper are iden-tified from published papers after their projectionon LMU-N. Two categories of processes are distin-guished, presented in the remainder of this section.

2.3.1. Content-related processesContent-related processes are the activities that are

specific to an application domain, and that focus on thetransformation, transport and processing of contentrelevant to an application. Tables 3 and 4 classify pa-pers contributing to these processes, and situate theminto LMU-N. Section 3.1 details the processes of thiscategory.

2.3.2. Node-related processesNode-related processes are activities not specific to

an application domain, but common to the IoT do-main. The development of such processes goes againstvertical fracturation. The messages exchanged in theseprocesses focus on the description of nodes character-istics, capabilities, preferences, etc, in order to allownodes to be aware of their neighbours on the graph,and to offer a homogeneous representation. Table 5references the papers contributing to these processes,and they are described in section 3.2.

6 N. Seydoux et al. / A unifying vision for the Semantic Web of Things

Table 1LMU nodes physical characteristics

Energy supply Processing power Communication capabilities Memory Storage

UpperTraditional powergrids

High to very high Standard web protocolsHigh to very high(Several Go)

Large to very large(internal HD todisk bay)

MiddleMixed, dependanton the deployment

MediumExtended, both ad-hoc andstandard web protocols [4]

Medium to low(Up to 1Go)

Medium to limited(Internal HD toSD card)

LowerOften limited: battery,renewable source [19]

Very limited,often microcontroller

Constrained, Ad-hoc, potentiallyshort range (BLE, Z-Wave)[4] [19]

Very low(Under 500Ko)

Limited (Flashmemory, SD card)

Table 2LMU nodes roles and examples

Role in the network Example

UpperContent processing, decisionmaking, and user interface(display or API)

Cloud server,laptop, tablet

MiddleLocal content aggregation,adaptation of ad-hoc com-munication to the web

Domestic box,gateway,micro-pc

LowerLink to the physical world :measurement and action

Sensor, actuator

2.3.3. Ontologies of the SWoTIn order to integrate IoT deployments in the SWoT,

their resources must be described using ontologies.Depending on the type of process, the ontologies re-quired for the SWoT integration is different.

In the case of content-related processes, there isa need for ontologies capturing domain knowledge.The LOV4IoT11 initiative mimics the principle of theLOV12, and references ontologies that have a connec-tion with the IoT domain, either describing IoT con-cepts, or domain knowledge in a field impacted bythe IoT, e.g. agriculture, home automation, or e-health.These ontologies are diverse, and disconnected fromeach other: the vertical fracturing between domains isnot bridged.

Node-related processes are more focused on inter-operability, and require the modelling of IoT systemscommon characteristics despite their diverse domainsof interest. IoT ontologies capturing these connecteddevice network characteristics exist. Several ontolo-gies have been proposed in order to describe nodes:

11http://www.sensormeasurement.appspot.com/?p=ontologies

12http://lov.okfn.org/dataset/lov/

Semantic Sensor Network (SSN)13, saref14[35], iot-ontology 15, IoT-lite 16[36], Spitfire 17, IoT-S18, SA19,the oneM2M base ontology20 and IoT-O 21. Papers like[22], prior to the apparition of SSN, or [1], much morerecent, survey these ontologies.

In both cases, the diversity of ontologies at stake canbe an issue. That is why good practices must also befollowed in the conception and the integration of theseontologies, described for instance in [5] or [18].

To sum up, LMU-N is an architectural pattern builtfrom a bottom-up study of IoT architectures. It is com-posed of nodes and messages flows, two abstractionsthat can be described using ontologies. LMU-N is notthe core contribution of this paper, but it provides areading grid that is used in the next section. It is a rel-evant classification framework: 72% of the contribu-tions listed in section 3 can be at least partially situatedwithin LMU-N.

13http://purl.oclc.org/NET/ssnx/ssn14https://w3id/saref15http://ai-group.ds.unipi.gr/kotis/

ontologies/IoT-ontology16http://iot.ee.surrey.ac.uk/fiware/

ontologies/iot-lite17http://sensormeasurement.appspot.com/ont/

sensor/spitfire.owl18http://personal.ee.surrey.ac.uk/Personal/

P.Barnaghi/ontology/OWL-IoT-S.owl19http://sensormeasurement.appspot.com/ont/

sensor/hachem_onto.owl20http://www.onem2m.org/ontology/Base_

Ontology/21http://www.irit.fr/recherches/MELODI/

ontologies/IoT-O

N. Seydoux et al. / A unifying vision for the Semantic Web of Things 7

2.4. Related work

Previous work has been done to survey the conver-gence between the IoT and the semantic web:

– Early work in the SWoT focused on semantic sen-sor networks. For instance, [22] surveys sensorontologies and observation representations. Thescope of this work is especially on models, andeven if it proposes an overview of technologiesenabling semantic sensor networks, it does notpresent specific applications. Similarly, [1] givesan overview of the semantic web stack applied tothe IoT, and surveys IoT ontologies. It goes be-yond semantic sensor networks, but is still lim-ited to models analysis. We propose in this paperto focus on how the ontologies and the technolo-gies of the semantic web are used to develop theSWoT, rather than on identifying exhaustively themodels used.

– [37] is a survey of the IoT domain, proposing adefinition for the notion of IoT and listing appli-cation domains and enabling technologies for theIoT. The IoT paradigm is described as the con-vergence of Internet technologies, electronic de-vices, and semantic web technologies. However,the paper itself does not cover how semantic webtechnologies are integrated into the WoT, whilewe intend to analyze in detail and compare differ-ent contributions to the SWoT.

– [38] studies the roles semantic web technologiescan play in the IoT, as well as the challenges theyrepresent. This paper identifies some processessimilar to what is presented in this paper in sec-tion 3. However, we situated the contributions tothese processes within LMU-N in order to en-able a deeper understanding of how an IoT net-work is integrated into the SWoT, and to give afiner-grained analysis grid: we want to identifythe impact of the nodes on the processes they areinvolved into. The explicit identification of pro-cesses, as well as the characterization of nodesbased on identified criteria, also allows futurestudies to be classified within the analysis gridwe propose.

– [39] gives an overview of the evolution from theIoT to the WoT and toward the SWoT. It is fo-cused on the role of standards in interoperability,and the integration of semantic web technologiesin standards. It also provides an overview of tech-nologies at stake in the IoT. This paper is oriented

toward projects and industrial consortiums, whichis out of the scope of our study. We focus on thecontributions of the SWoT to IoT issues, and onlyintegrate standardization concerns when they arerelated to this domain.

3. Contributions of the semantic web to IoTprocesses

In this section, we describe for each process the con-tributions of associated papers to point out the addedvalue of semantic web principles and technologies.Each category of process (content-related and node-related) are separated into subcategories of general re-curring patterns among processes in order to clusterthem into coherent sets: complementary processes, orprocesses implicated in similar scenarios.

The complete classification is summarized in tables3, 4 and 5. In these tables, each column represents aprocess, and each line represents the situation of thecontribution in LMU-N: if applicable, the contribu-tion is precisely situated between a source and a tar-get node type. However, due to the lack of informationin some paper, the positioning of some contributionsis only partial (in the lines ”unspecified”). Moreover,some contributions are described completely decorre-lated from any architecture, or in an architecture notcompatible with LMU-N, they are placed in the ”Non-LMU/Not specified” line.

3.1. Content-related processes

Content-related processes classification is summa-rized in tables 3 and 4. Three main categories ofcontent-related processes are distinguished, shown inthe headers of the tables: representation transforma-tion, transport, and processing, each described in theremainder of this section.

3.1.1. Content representation transformationThese processes are dedicated to the transformation

of content according to the DIKW hierarchy. The coremeaning of the content is not changed by these pro-cesses, but the expressivity of its representation varies.

Enrichment is the process of transforming contentrepresentation upward in the DIKW hierarchy. For in-stance, the description of data with meta-data to trans-form it into an information is an enrichment: content isdescribed with ontological entities to unambiguouslydefine its meaning, capture its context and increase its

8 N. Seydoux et al. / A unifying vision for the Semantic Web of Things

reusability and its overall value to an application. En-richment is one of the earliest processes that used se-mantic web technologies to contribute to the IoT withcontributions such as [16] or [29]. This matter of factis explained by the predominance in early IoT workof sensor networks, a subset of IoT networks whereall the LN are sensors. Enrichment is performed whencontent is exchanged upward in LMU-N: all the en-richment contributions in table 3 are directed upward,which makes enrichment an upstream process, increas-ing content complexity and processing cost.

In [29], the enrichment is performed by annotatingthe content with RDFa22. This approach does createsemantically enriched content, but it requires the LNto produce complex, structured content (here compli-ant with the OGC’s O&M model23). This might be im-possible for some LN, as it requires to exchange heavyXML documents over communication links that canbe very restricted (LPWAN networks or CoAP for in-stance). To overcome this issue, [23] proposes to pushthe enrichment task to the MN level: the LN pro-duces raw data, in any format, this data is transportedto the MN by a dedicated network, and only then theMN uses its own knowledge of the producing nodeto enrich the data. [40] instantiates enrichment at thesame level as [23] (as shown on table 3), from LN toMN. However, contrary to [23], [40] proposes an ap-proach where the LN offers a ”semantic interface”, im-plementing enrichment functionalities: content is en-riched on the LN side. Unfortunately, the paper doesnot give any additional precision to describe its enrich-ment approach. In a different approach, enrichment toa semantic format is performed in [6] and [18] on theUN side when data is collected from the MN. Con-tent collected from sensors is stored in a standardizedstructure on the gateway, but it is explicitly describedwith an ontology only on the server side.

The importance of the enrichment process is under-lined by the important number of publications con-tributing to it: 17 papers in this survey. It comes fromthe discrepancy between the LN, hardly able to pro-duce high-level content, and the UN having applicativeneed for homogeneous content represented in expres-sive formats. Moreover, transforming data into knowl-edge is a common practice in the semantic web do-main, even outside the IoT. This is an explanation tothe fact that 47% of the contributions to this process

22https://www.w3.org/TR/xhtml-rdfa-primer/23http://www.opengeospatial.org/standards/om

cannot be situated precisely into LMU-N: their ap-proaches are more centered on content than on the ar-chitecture around it.

Lowering is the process exactly opposed to enrich-ment: it is the transformation of content from a se-mantically rich format to a less expressive, more con-strained representation. The need for such transforma-tion arose from the integration not only of sensors, butalso of actuators as LN in the IoT. These devices arecontent consumers, but their constrained nature pre-vents them from being able to consume some type ofcontent. Actions represented in rich formats by UNneed to be adapted to the target device so that it caninterpret it and act as intended. This transformation de-prives the content from part of its expressivity and con-text, but the transformed, simpler content is interpretedin a known context, leading to a trade-off for consis-tency. Lowering requires a node to have a represen-tation of the capabilities and expectations of a remotenode: it is a process that needs the source node to bemore powerful than the destination node, and that isnecessary because of the restrictions of the destinationnode. It is therefore a downstream process.

Lowering is opposed in the literature to ”lifting”,a synonym for enrichment. [41] describes SAWDSL,a language aiming at making it possible to lift XMLto RDF and to lower RDF to XML thanks to XMLschema annotations. However, in the studies surveyedhere, most of the approaches to lowering are not ex-plicitly based on annotations, but rather on ad-hoc ap-proaches. For instance, in [18], an autonomic controlloop is instantiated, and high-level representation ofactions are transformed into service calls, but the trans-formation process ad-hoc: RDF individuals and theirOWL descriptions are consumed by a software that callthe procedures based on an a priori interpretation ofthe expressed content.

Overall, the transformation of content from a high-level representation to a form that can be processed bya constrained node is still a challenge for the SWoT. Inthe publications referenced in this paper, the proposedmechanisms are either manual or ad-hoc. A major-ity of the studies cannot be clearly situated in LMU-N, denoting shallow contributions. The lack of interestin this process comes from its contradiction with theusual practices in the semantic web community. Con-tent is generally moved upward in the DIKW pyramid,because it gains value this way. However, the presenceof constrained content consumers (actuator nodes)on the IoT, combined to the need for high-level con-

N. Seydoux et al. / A unifying vision for the Semantic Web of Things 9

tent in UN, makes lowering a process as necessary asenrichment for a real SWoT.

3.1.2. TransportIn processes listed under the transport category, con-

tent itself is not transformed. These contributions focuson how content is distributed across an IoT network(e.g. routing), or on how the content is used (e.g. con-trol). Content itself is not necessarily expressed usingthe semantic web technologies, but its management onthe network is based on these technologies.

Notification/Dissemination is the process of for-warding a piece of content to a distant node in a pushmanner (initiated by the content collector). It can bedriven by the interest of the receiver, or on applicativelogic of the emitter. Content is disseminated from thenode that generated it, either by collection (typicallyby a LN), or by a higher-lever process (e.g. enrichmentby a MN or processing performed by a UN) towardanother node. The diversity of nodes being sources ortargets for this process makes it, in theory, a mixed pro-cess, going both upstream from LN and downstreamfrom UN. This mixed aspect is present for instance in[31]: on the one hand, sensor data is sent upstream,toward a control center. The control center processesthe data, and produces alert messages that are dissem-inated downstream.

[42] proposes a dissemination approach based on aad-hoc interest expression for content consumers rep-resented by a query. Each consumer query is regis-tered on a content matching component in the form ofpredefined triple pattern. Triples matching a registeredpattern will be redirected to the relevant content con-sumer. However, this approach is oriented toward per-formance, and reduces the expressivity of queries, lim-ited to a unique triple where each element can be fixedto a constant value. Approaches based on stream pro-cessing, such as [33] or [43], use the full expressivityof SPARQL coupled to specific streaming operators,based on a time window or a number of events. Theresult of a stream query is an event stream, which isdirected toward the querying node.

Even if fundamentally, dissemination is a mixedprocess, in the surveyed papers notifications are mainlysent upstream. This can be related to the architecturaltendency of IoT networks to a hierarchical organiza-tion, with complex processings performed in the uppernodes. Therefore, it makes sense that UN, driven byapplicative needs, express their needs to nodes, lowerthan them, that will collect content in a general pur-pose approach. The absence of goal-awareness in con-

tent collection by LN is motivated by the will to avoidstovepiped deployments.

Control is the process where a node sends a com-mand to a remote node for it to execute. Commands areeither issued by the user, or by high-level applicationsexecuted on UN, and are eventually executed by a LN,making control a downstream process. In this process,the use of semantic web technologies and principlesallows the target node to have a deeper understandingof the command.

[44] propose an approach based on rule evaluation:if certain conditions (expressed in RDF) are matched,the rule entails an action. This approach allows thesame node to have different behaviours upon the recep-tion on the command, if the rule head relies on agentscontextual knowledge. In [18] and [6], a representationof the actions to be taken is inferred from the represen-tation of the user requirements and from the observa-tions of the environment. These actions are associatedto service description that are processed by a remotenode to actually make the service calls.

The necessity for the control process arises fromthe integration into IoT networks of actionable nodes.The reduced number of actionable nodes deployedcompared to sensors explains that only three papersdiscuss contributions to this process.

Routing is the process where a route among nodesis determined for a content instance. Routing policiesare at the core of the internet, and not only of the IoT.The integration of the semantic web technologies en-ables the definition of new routing policies. Routingpolicies are not specifically directed toward a directionin LMU-N, therefore routing is a mixed process. Rout-ing is different than dissemination in the sense that itis implemented by nodes on content they did not nec-essarily generate. Moreover, routing tends to be drivenby network topology and nodes description, whereasdissemination by content characteristics.

[45] proposes a distributed approach to routing,where each node computes the optimal route from it-self to a destination broker. Each nodes knows the dis-tance of its neighbours to the destination broker, andto break ties, the semantic distance between the nodeprofile and the profile of its neighbours is used. A nodeprofile instantiates a dedicated ontology, and it is seri-alized into a binary string. Each node profile is unique,and used as an identifier. The evaluation function com-putes the energetic cost of the produced routing tree,and shows a clear reduction of energy consumption.In this approach, the semantic web principles are used

10 N. Seydoux et al. / A unifying vision for the Semantic Web of Things

to reduce energy consumption at runtime. However,the algorithm used to serialize the semantic descrip-tion of the node output is not clearly described. It out-puts results that are suitable for constrained nodes, butit is unclear whether it supports the full expressivityof the ontology or not. In a different, hierarchical ap-proach, [46] describes the construction of routing ta-bles for service provisioning based on node clustering.Servicing nodes are described with an ontology, andthe are clustered by the gateway they are connected tobased on a semantic similarity measure. The obtainedclusters are provided to the upper-level gateway, whichuses the clusters from the nodes to build a routing ta-ble, and reproduces the process recursively. [47] alsouses a semantic description of sensor nodes to createclusters of nodes satisfying a certain query that repre-sents the interest of the user, for instance ”Nodes de-tecting a temperature over 50 degrees”. These clustersthen aggregate the content they collect, before send-ing it to a sink node. The authors used the number ofnodes alive against time as a metrics to show the ef-fectiveness of their approach: the economy of energyinduced by the reduction of communications allowsnodes to communicate longer. However, the authorsdo not detail how semantic technologies are used pre-cisely, and they make the assumption that the nodes aresufficiently powerful to match their own profile againsta user query.

In these three approaches, the routes to forward con-tent from a node to another are computed based on asemantic description of nodes. However, to the best ofour knowledge, no contribution was made toward theintegration of knowledge about the content itself in therouting process.

Querying is the process where an application ex-plicitly accesses content on a remote node, in a re-quest/response manner.

On the IoT, content might not necessarily be en-riched and stored in a knowledge base, preventing itfrom being queried using semantic web technologies.[43] proposes a streaming Ontology-Based Data Ac-cess (OBDA) layer to access data in a mixed approach,where stream semantic queries are processed throughR2RML mappings to be transformed into queries overfederated sensor networks, and to transform the re-sponse in order to answer the original SPARQL streamquery. [48] proposes a similar approach, with trans-formation from SPARQL to SQL via RML, with andwithout the streaming extension to SPARQL. In otherpapers, where querying is not the core contribution,

direct querying in SPARQL is also used after storingRDF content in a knowledge base, as it is proposedin [12] for instance. [31] proposes a mixed approach,where part of the data is stored in a traditionnal SQLdatabase, but its associated metadata is stored as RDFfor more expressive SPARQL queries.

The request/response interrogation of RDF graphsusing SPARQL queries is not specific to the IoT, andthat is why few contributions in the survey cover it.This is also a reason why none of the contributions forthis process can be described precisely with LMU-N:the focus is not on the IoT architecture. Transforma-tion from SPARQL to SQL is not the main practice toquery IoT content, contrary to what this survey maysuggest at a first glance: it is a research topic receivingcontributions, as it is presented in section 4.2.2, dedi-cated to the evolutions of the semantic web.

3.1.3. Processing contentProcessing is an important part of the IoT value cre-

ation, as is described in [49]. It covers a broad rangeof processes (we identified 5) where content is lever-aged in an application, and semantic web principlesand technologies can be used in order to provide inno-vative services.

Abstraction has similarities with Complex EventProcessing (CEP), as [50] points out: low level symp-toms are extracted from content and correlated to-gether in order to be transformed into a more abstractdiagnosis. It can be based on reasoning, rules, pattern-recognition, etc. Abstraction requires both the collec-tion of content, and resource-intensive processing, it istherefore an upstream process. Abstraction is a processalready present in early studies such as [29] or [30],where an abstraction is defined as the ”representationof an environment derived from sensor observation”.It is a process that has been widely implemented: 11papers in this survey are dedicated to abstraction. Eachcontribution will not be presented in details in this pa-per, but they all have common characteristics: they usecontent described with an ontology (ssn for [51], orSenML for [52]), and infer new content described witha domain-specific vocabulary. The production of ab-stract content is necessarily driven by an applicativeneed.

[29] proposes an SWRL-based approach to infernew content from sensor observations, illustrated bythe example of the deduction of road conditions fromtemperature and precipitations observations. In [30],the authors propose a logic-based approach to abstrac-tion, and uses OWL-DL reasoning to derive the most

N. Seydoux et al. / A unifying vision for the Semantic Web of Things 11

probable abstraction from a set of observations. [11]takes another approach to abstraction, proposing a sur-vey on context-awareness in the IoT, and how seman-tic web technologies can support the modelling of con-text, as well as contextual reasoning. The authors alsopropose several step to abstraction, with a granularitythat could vary according to the layers involved in theprocess: content transferred from LN to MN is pro-cessed a first time to get a primary context, and it isprocessed a second time for more complex treatmentson larger content instances when transferred to UN.

Inferring new knowledge from a KB, optionallycombined to a set of rules, is a logical process not lim-ited to the SWoT. Traditionally, it is perform horizon-tally on a powerful machine: the KB, the reasoner andthe means to insert new content in the KB are all hostedon the same server. However, in the case of the IoT, theprocess is not horizontal on UN, but rather upstream,and performed on the UN side. This is related to thenature of inference in IoT applications: it is fed bycontent produced by LN.

Aggregation is a process where multiple instances ofcontent are used in order to produce a new content in-stance of the same level in the DIKW hierarchy, as op-posed to abstraction where the inferred content is ofa nature different from the pieces of content used forinference. For instance, computing the average (or themaximum) of several values is aggregation, while us-ing the same values to infer the occurrence of a mete-orological event is abstraction.

In [53], the aggregated elements are similarity ma-trices used for ontology alignment. The authors tacklethe heterogeneity of application fields for the IoTdomain, and the diversity of ontologies it entails,by proposing an ontology alignment technique. Theproposition of these authors is not designed to be runonline, and therefore it does not fit the LMU-N readinggrid. The proposed aggregation approach, driven by aconsistency measure, is applicable to any content an-notated with an ontology. In [20] and [33], aggregationof content is performed using operator from SPARQL,like COUNT, MIN, etc, from a customized extensionof SPARQL adapted to streaming queries to performaggregation over time or space. In these papers, ag-gregation is a driving mechanism for sensor mash-up.However, content aggregation is not separated fromcontent enrichment for instance, these two processesbeing both covered by the term ”data fusion” in [20].In these papers, the aggregated content is both pro-

duced and used in UN, making aggregation a horizon-tal process.

In the papers of the survey, data fusion is not rep-resented as an upstream process, which is counter-intuitive. It is explained by the fact that all the pa-pers relevant to the survey perform fusion operationson content after it has been enriched. Therefore, thecontent is enriched in an upstream process, but no con-tribution we surveyed performed the enrichment fromlower to middle nodes, and then enrichment from mid-dle to upper nodes. In this hypothetical case, fusionwould be performed upstream.

Visualization is the display of content in a human-readable manner. When produced, IoT content is typi-cally a numerical value with some metadata, and visu-alization processes propose a visual interpretation of itinstead of its raw numerical form.

For instance, studies such as [29] or [54] use geo-graphical metadata to display the location of contentsources on a map, and to give access to these sourcesvia a map interface. In [16], an additional processing isexecuted to represent not only the geographical loca-tion of entities, but also to extract further meaning fromthe content and to attribute a color code to the sensorsdepending on their nature (i.e. the class they instantiateclass in an ontology). In [33] (an extension of [54]),content is aggregated prior to the visualization phase,in order to display refined content, like heat maps orgraphs, complementary to the map information.

Consistency enforcement is the process of ensuringthe consistency of the content, that is to say the ab-sence of contradictions among the facts expressed inthe content, and the absence of wrong assertions (e.g.sensor measures different from the reality of the phys-ical world).

[11] discusses the role of context in the identifica-tion of inconsistency issues. [55] proposes an ontologydesign pattern to capture the semantics of an indus-trial, non-standard, tagging vocabulary defined for theproject Haystack24 with the semantic web formalisms.The obtained ontology is then automatically populatedusing the instances of Haystack tags to generate thecorresponding individuals. However, the transforma-tion from a vocabulary with limited expressivity to arich ontology can lead to inconsistencies, for instanceif individuals bear two tags from the vocabulary thatare disjoint in the ontology. The authors use DL rea-

24http://project-haystack.org/

12 N. Seydoux et al. / A unifying vision for the Semantic Web of Things

soning to detect inconsistencies in the generated KB,and identify modelling issues in the ontology or incon-sistencies in the individual tags.

Consistency is not instantiated in details in the con-tributions we surveyed, even if [9] identifies consis-tency as a challenge for the SWoT. Contradictions be-tween several nodes can be dealt with in an aggrega-tion process where data fusion techniques are applied,but consistency enforcement also includes the detec-tion of logical issues in models and their instantiations.More generally, consistency enforcement is not a topiclimited to the SWoT, and generic consistency mech-anisms could be applied to IoT datasets.

Decision support is a process where content is usedas the input of a decision-making process. Some smartcity scenarios use sensor networks and IoT deploy-ments for decision support, e.g. Dublin’s Smart EnergyDemand Analysis25.

[31] proposes an Early Warning System (EWS) ar-chitecture enabled with semantic functionalities. Con-tent collected by sensors is enriched and disseminatedupstream, before being used by UN in order to be ab-stracted into high-level events. Once the events iden-tified, the system takes operational decisions (whereto send rescue, what places to evacuate...), and theseoperational decisions are implemented in downstreamprocesses. [56] proposes the modelling of policies byontologies in order to manage the services offered bydiverse nodes, dynamically created according to con-text. The policies capture directives for decision mak-ing.

Decision support is a process with two aspects: onthe one hand, the supported decision-making processcan be performed by human beings, and on the otherhand it applies to autonomic systems. The former caseis not specific to the IoT, but can be instantiated with asensor network, and the system only provides guidanceto a human decision maker.

3.2. Node-related processes

In node-related processes, the messages exchangedbetween nodes do not focus on the content these nodescollect or process, but on the nodes themselves. Node-related processes classification is summarized in tables5. They are separated in two sub-categories: processesdedicated to awareness among nodes, and processesdedicated to node heterogeneity management.

25http://smartdublin.ie/smartstories/spatial-energy-demand-analysis/

3.2.1. Process related to neighbourhood awarenessIn order to communicate, IoT nodes need to be

aware of the existence of each other, and to haverespective addresses to exchange messages. Further-more, nodes may be only intermittently available, tosave battery life for example, or due to failure, main-tenance operations, etc. In this context, a dynamicawareness of a node’s surroundings is required, and thefollowing processes contribute to it.

Discovery is the process of gathering descriptions ofthe nodes that can be reached from the node host ofthe process. In a typical hierarchical IoT architecture,nodes of a given level are connected to multiple nodesof an inferior level, and to a few nodes of superior level(often only one). Therefore, LN can be deployed withthe address of the middle node they communicate withhard-coded, while MN or UN need to discover under-lying nodes as they come and go. That is why discov-ery is a downstream process.

[34] proposes a discovery method based on google’sPhysical Web (PW)26 and extending it to a so-calledPhysical Semantic Web (PSW). Discovery of nearbynodes is enabled by the protocols and technologiesof the PW, but the PSW extracts from the identifica-tion messages semantic annotations in order to decidewhether the node is relevant or not to the client, run-ning on a mobile device. [40] proposes a discovery forMN based on a semantic description issued to a cen-tral repository which address is known a priori by theissuer of the discovery request. Similarily, SPARQLquerie can be issued to the MN describing LN charac-teristics in order to collect LN identifiers. The ontol-ogy proposed in [57] is developed in order to ease nodediscovery by modelling contextual knowledge, such aslocation, domain knowledge, and policy knowledge.[58], [38], [32], [59] and [60] all propose a discoverybased on a node specification sent to a repository con-taining nodes descriptions.

Discovery is a widely implemented process (10 con-tributions in this survey), because it is related to thedynamic nature of an IoT network: it allows nodes’representation of the graph topography to remainconsistent with the reality. Discovery is also a pro-cess that is necessary to perform other processes. Dis-covering the capabilities of a distant nodes is requiredfor processes such as lowering for instance.

26http://google.github.io/physical-web/

N. Seydoux et al. / A unifying vision for the Semantic Web of Things 13

Exposition is the process complementary to discov-ery, in which a node makes its own description avail-able in order to be discoverable. However, expositionis initiated by the node target of the discovery, thatis why it is an upstream process.

In [34], exposition is supported by the PSW, al-ready described for discovery. In [10], the authors pro-pose to enhance the registration mechanisms offeredby the oneM2M standard to support a semantic de-scription exposition. A LN can register itself onto aMN if it knows its adress, and the registration querycontains a standard description of the LN. The authorspropose to include a semantic description, or a deref-erencable URI, in the registration request. Since then,the oneM2M standard included a ”semantic descrip-tor” resource, carrying an RDF/XML description of itsparent node. It allows a node to expose its capabilitiesto its target when registering onto it. Another aspectof exposition is proxying: after having discovered aset of LN, a MN can act as a proxy and perform ex-position of said LN in their stead, as [24] proposes.The authors propose an architecture where the expo-sition/discovery process can be specific to a type ofnetwork or technology. This way, the process is bothadapted to the constraints of the LN and understood bythe MN, which performs an enrichment phase on thenode metadata in order to redistribute them in a moregeneric format.

All the contributions implementing these processescan be precisely situated in LMU-N, because they allspecifically describe technical content with a detailedarchitecture. There is significantly fewer contributionsto the exposition process than to the discovery process,which seems to be counter-intuitive regarding the de-pendency of discovery upon exposition. However, ex-position is a process requiring the lowest node to beactive, while in discovery, the burden is on the highestnode, making it easier to implement.

Selection is a process where a node decides whichother node should perform a task. It is especially rele-vant when multiples nodes are offering similar servicesyet having different capabilities, characteristics, costs,etc. Node selection is a downstream process, drivenby high-level policies (e.g. energy saving or time effi-ciency), based on criteria to compare nodes, enablingthe computation of a score and ranking. The use of se-mantic web technologies and principle are used to de-fine these criteria.

[25] refers to the node selection process as the as-sessment of relevance metrics. The authors use ab-

stracted nodes, or Virtual Object (VO), and the seman-tic description of their characteristics (not preciselydescribed in the paper), in order to perform a selec-tion. The proximity criteria are defined dynamicallydepending on applicative requirements. This work fo-cuses on the ability to perform a selection despitethe heterogeneity of underlying objects thanks to theabstract representation of nodes. In [61], the authorsidentify two types of criteria for node selection: non-negotiable criteria, representing the ability of the nodeto provide a service, and negotiable requirements, overwhich selected nodes will be ranked. The filteringphase is performed using SPARQL queries represent-ing user requirements, and the ranking phase is basedon a multi-dimensional criteria aggregation. [62] de-scribes a service selection process to create a compos-ite service out of existing services. The actual node se-lection is performed at binding time, where servicesdescriptions stored in a service cache by the middlenodes is used to associate the composite service withactual nodes. In [46], node selection is a process drivenby a service search initiated by the user. This searchstarts from a top gateway which is the entry point ofthe user to a building network. Node selection is per-formed recursively in a gateway hierarchy, based onthe semantic description of services clusters. At thelowest level of the hierarchy, the final gateway, actu-ally connected to the lower nodes, returns the actualcharacteristics of the nodes matching the service re-quest. In [44], the node selection is based on an a prioriprediction of the entities that would potentially matcha user query: only sensors that are associated to theseentities are queried.

Similarly to discovery, selection is a downstreamprocess where the highest node perform most (if notall) the computation. Node selection can be performedto initiate a notification process: in the case of up-stream notification, the lowest node has to decidewhether to send a content to an upper node or not. Toreduce the load on the lowest node, the process can beswitched into a node selection, where the upper nodeselects appropriate lower nodes to provide it with con-tent. However ,in this approach, one can see that theinterest is switched from the content description in no-tification to the node description in node selection.

3.2.2. Processes providing homogeneity among nodesAs stated in the introduction, heterogeneity, and the

interoperability issues it brings, is one of the leadingreasons for the introduction of semantic web technolo-gies into the IoT. The three processes dedicated to ho-

14 N. Seydoux et al. / A unifying vision for the Semantic Web of Things

mogeneity are focus on the management of this diver-sity.

Abstraction is the process of representing a node by avirtual entity. In a context of heterogeneous nodes, ab-straction aims at focusing on the modelling of genericnodes representations to describe their character-istic in a unified way. It allows applications to dealwith a set of homogeneous nodes, breaking the ver-tical silos between application domains. Nodes ma-nipulate abstracted representations for nodes of infe-rior level, making abstraction an upstream process. Onthe SWoT, ontologies are used to describe abstractnodes.

In [12], the behavioural patterns of nodes are usedin order to automatically attach a description to un-classified nodes. Similarity is computed between theoutput of the nodes already described in the KB andthose of the newly introduced nodes, and the unde-scribed nodes is annotated using the description of thesimilar nodes. This approach suffers a cold start is-sue: when only a few sensors are annotated, the clus-tering is less efficient, and the system needs to growin order to propose a wider variety of sensors andmore representative clusters. [25] proposes the notionof VO to describe how nodes can be abstracted, anddescribes how this approach tackles heterogeneity is-sues. The authors then describe how abstracted nodescan be used in other processes, such as composition orselection. These contributions are described in the cor-responding paragraphs. In [15], physical nodes are as-sociated to avatars, a virtual representations describedin OWL. The authors identify requirements for a WoTplatform, and show how their proposed avatar architec-ture meets these requirements, such as interoperabil-ity, reactivity, safety... Avatars are also given introspec-tion capabilities, in order to support collaboration andservice composition, described more specifically in theparagraphs associated to these processes.

Abstraction is an important process, contributed toby 14 papers in this survey. Among the processessituated in LMU-N, more contributions are situatedfrom lower to middle nodes than from middle to uppernodes. A possible explanation is that the LN class isthe most heterogeneous: end devices have the widestvariety of functionalities. Therefore, being able to ab-stract their representation as soon as possible in the IoThierarchical network makes their management easierfor MN.

Composition is the process of associating nodes be-tween them in order to create new nodes, offering ser-vices that were unavailable on the network before. It isa process often connected to node abstraction, becausean abstracted node representation enables the creationof virtual composite nodes.

The notion of VO presented in [25] and in [13] isassociated to the notion of Composite Virtual Object(CVO). The authors of these contribution describe aprocess to build CVOs on top of homogeneous VOs,themselves being abstractions for real-world objects.Based on a specification of applicative needs, CVOsare dynamically created. In [62], node composition isperformed using the semantic description of differentservices. A plan is then computed, and services se-lected are called sequentially according to the plan.The process is dynamic, and the composite nodes de-scribed by the execution plan is not stored in the KB.[57] uses the same terminology of service offered by aCVO composed of several VO. They propose a mod-ular ontology architecture, and associate the ontologymodules with the entities they define: only the VO aredescribed using the Resource ontology module, but itis the CVO and the services that are annotated with theService ontology module for instance. However, theauthors do not give insight regarding the creation ofthe CVO.

Composing nodes is only possible provided thatthese nodes are described by abstractions, and thatthese abstractions can be manipulated. This is why amajority of papers are present in both abstraction andcomposition process: they first define how they ab-stract physical nodes, and then demonstrate how theseabstractions can be manipulated separately from thenodes they initially represent. The semantic web prin-ciple and technologies provide flexible models thatsupport such approach.

Specification/Configuration is the process oppositeto abstraction: generic representations are adapted tospecific deployments, and physical nodes are config-ured in order to match their virtual representation.[44]proposes a different definition of configuration, fo-cused on node introspection and pattern learning: newnodes cluster with existing nodes in order to automati-cally annotate the data they produce. It does not qualifyas configuration as it is defined in the present paper.

[40] proposes an architecture where smart agentsconsume semantic messages from a broker, and con-trol a actuators by changing their virtual representa-tion. This process is different from control as defined

N. Seydoux et al. / A unifying vision for the Semantic Web of Things 15

in the content-centric process because the focus is noton the message that is sent in order to change the stateof the device, but rather on the representation by thesystem of the state of the device as it should be. How-ever, the paper does not provide further details on theimplementation.

Globally, specification process as the lowering of ageneric, virtual representation into potentially severalimplementation-specific commands has not been im-plemented, to the best of our knowledge.

3.3. Identified trends

Four architectural characteristics emerge from exist-ing studies:

– (i) Communication between nodes of differentlevels are limited to adjacent layers. Processes be-ing supported by message flows, processes instan-tiations are therefore specifically represented inthe tables 3, 4 and 5 between nodes of adjacentclasses, e.g. from upper to middle node.

– (ii) A node of a given level has contacts with alimited number of higher level nodes, and multi-ple contacts with lower level nodes.

– (iii) Horizontal contacts between nodes of thesame level multiply as the level of said nodesgets higher: there is little to no direct contact be-tween lower nodes, and upper nodes communi-cate with each other frequently. The five contri-bution to horizontal processes presented in the ta-bles are all instantiated between UN.

– (iv) The IoT being a system tightly coupled withthe physical world, nodes differ by their ”prox-imity” with the environment. For a given node,this characteristic can be measured by the num-ber of edges that separate it from a node having adirect contact (sensing or acting) in the environ-ment. As one would expect, the lower a node is inthe LMU-N hierarchy, the closer it is to the envi-ronment: LN directly observe or act on the envi-ronment, while UN only manage a digital repre-sentation of the physical world.

Situating SWoT contributions in LMU-N showshow nodes constraint the processes they interactwith: contributions dedicated to the same process butat different levels in LMU-N do not expect similar out-comes. However, it is worth noting that some contribu-tions are only partially situated in LMU-N because thepapers describing these contributions do not provideenough implementation details. This either reveals that

Tabl

e3:

:Pap

ers

cont

ribu

ting

toL

MU

cont

ent-

cent

ric

proc

esse

s(1

/2)

Rep

rese

ntat

ion

tran

sfor

mat

ion

Tran

spor

t/Pr

ovis

ionn

ing

From

ToE

nric

hmen

tL

ower

ing

Not

ifica

tion/

Dis

sem

inat

ion

Con

trol

Rou

ting

Que

ryin

g

LM

[23]

,[4]

,[40

],[6

3]M

U[2

0],[

6],[

18]

Ups

trea

mU

nspe

cifie

d[2

9],[

32],

[64]

,[54

],[3

3],

[12]

,[22

],[6

5][4

2],[

54],

[33]

UM

[18]

[6],

[18]

ML

[40]

[46]

Dow

nstr

eam

Uns

peci

fied

[44]

,[66

],[6

5][4

4][4

9],[

54]

U MH

oriz

onta

lL

Mix

ed[3

1][4

5]N

on-L

MU

/Non

spec

ified

[43]

,[55

],[6

7][5

5][4

7][4

3],[

47],

[48]

16 N. Seydoux et al. / A unifying vision for the Semantic Web of Things

Table4:Papers

contributingto

LM

Ucontent-centric

processes(2/2)

ProcessingFrom

ToA

bstractionC

onsistencyenforcem

entA

ggregationV

isualizationD

ecisionsupport

LM

[11]M

U[11],[6],[52],[63],[51]

Upstream

Unspecified

[50],[38],[49][38],[54],[33]

UM

ML

Dow

nstreamU

nspecifiedU

[31],[20],[33][31]

MH

orizontalL

Mixed

Non-L

MU

/Non

specified[29],[30],[67]

[11],[55][53],[68]

[29],[16][69],[56]

Table5:Papers

contributingto

LM

Unode-centric

processes

Awareness

Hom

ogeneityFrom

ToD

iscoveryE

xpositionSelection

Abstraction

Com

positionSpecification

/Configuration

LM

[10],[34][14],[24],[40],[46]

[14]

MU

[70],[24][13],[6]

[20],[13]U

pstreamU

nspecified[58],[61],[25],[15],[57],[66],[12]

[25],[15],[57],[62]

UM

[14],[40],[34][62]

ML

[40][46]

[40]D

ownstream

Unspecified

[58],[38],[38],[32],[57]

[20],[61],[25],[38],[12],[51],[44]

U[59]

MH

orizontalL

Mixed

Non-L

MU

/Unspecified

[60][47]

[71]

N. Seydoux et al. / A unifying vision for the Semantic Web of Things 17

the deployment into a concrete architecture is not in thescope of the paper, or that the process presented is notspecific to the IoT but rather uses the IoT as an appli-cation domain, for instance in the case of the contribu-tions to the querying process. The constraints broughtby the nodes justifies the presence of [40] twice forthe discovery process: discovery from upper to mid-dle nodes being differs from discovery between middleand lower nodes.

The superior capabilities of nodes of higher levelallows system designers to push the constraintsbrought by the limitations of lower nodes up. Thisis why, regardless of the direction of the process, thehighest nodes between the target and the source per-forms the more computation for content transforma-tion and processing processes, as well as for node-centered processes. In both cases of discovery and en-richment, which are opposed processes, this observa-tion is verified: UN are responsible for the discoveryof MN for instance, and content produced by LN isenriched by MN. Moreover, architectural characteris-tic (ii) is a consequence of this shift of constraints: amiddle node is able to manage the processing requiredfor content enrichment that multiple lower nodes couldnot perform.

The influence of nodes on processes is not limited tophysical constraints: the highest nodes also holds thelargest part of the application logic. Having LN col-lecting content without a specific applicative goal lim-its vertical fracturation: only the nodes processing con-tent should be application-oriented. According to thearchitectural characteristic (iv), reducing vertical frac-turation requires (among other measures) to reduce theapplication logic executed on nodes with a high prior-ity with the environment.

A side effect of this is the predominance of up-stream processes compared to downstream processes.When comparing enrichment to lowering, or abstrac-tion to specialization, it is clear that upstream pro-cesses are studied more actively. An exception to thispredominance is the case of discovery and exposition,clearly visible in table 5: the downstream process ismore represented. This is coherent with the location inthe highest nodes of the application logic and the heav-iest processing. Our analysis of this matter of fact is thesimilarity between upstream processes in the spe-cific case of the IoT, and traditional research in thesemantic web domain: enriching content with meta-data, or inferring abstraction from knowledge base arewell-established topics. On the other hand, the needfor downstream processes such as lowering of special-

ization arises from the constrained nature of some IoTnodes, and therefore the application domain is muchmore specific than the one of upstream processes. Thepredominance of upstream processes is coherent withthe predominance of sensor nodes among IoT nodes:downstream processes are required for actionablenodes, which are a minority. The prevalence of up-stream processes also relates to the need for homo-geneity: when performing upstream operations such asnode or content abstraction, the result is expressed in arich model, providing an interoperable representation.

Moreover, in some surveyed architectures, central-ization is an issue. Even if the IoT is by nature a dis-tributed system, given its hierarchical topology, up-stream processes concentrate high-level content in afew UN. When services (such as visualization) are of-fered directly by the server to its clients, this modelis suitable, but when the network needs the capacityto make autonomous local decisions, this approachshows shortcomings. With the multiplication of nodes,and the increase data flow it involves, scalability isan issue for resource-intensive processes, and the nextsection surveys approaches to tackle this issue. More-over, the lack of downstream processes limits the abil-ity of MN to benefit from the content inferred by theUN that would enable them to make local decisions.

This section described contributions of the semanticweb to the IoT, with the identification of generic pro-cesses and the situation of the research contributionstoward these processes within LMU-N. Processes suchas enrichment or abstraction exist in the semantic weboutside of the IoT, and they are used to solve issuesspecific to the IoT. However, processes such as low-ering or control are more specific to the convergencebetween the IoT and the semantic web, and are partof evolutions of the semantic web driven by IoT con-straints. These evolutions are described in the next sec-tion.

4. How the semantic web evolves to face IoTconstraints

The issues tackled by emergence of the SWoT areissues affecting the development of the IoT and the de-ployment of processes within IoT architectures. Theseissues (e.g. heterogeneity, lack of interoperability, con-tent transformation) are recurrent concerns for the se-mantic web community, not necessarily related to theIoT or the WoT. However, the IoT also has intrinsic

18 N. Seydoux et al. / A unifying vision for the Semantic Web of Things

characteristics due to the constraints on its constitutingnodes, the distributed nature of its deployments, andthe dynamism of its topology. These constraints applyto any solution deployed in an IoT architecture. Thatis why contributions between the semantic web andthe IoT domain are not unidirectional: the semanticweb does contribute to the emergence of the SWoT, butsemantic web technologies and principles must also beadapted to meet the constraints of the IoT in order todevelop the SWoT.

4.1. Adapting to the constraints on the nodes

One of the most important constraint in the IoT isthe presence of constrained physical nodes, be it LNor, to some extent, MN. The constraints on these nodeswere presented in section 2.2.1. Among the resourcesthat can be limited for a nodes, we distinguished en-ergy, processing power, communication channels, andmemory. These resources are not independent, e.g. thelimitation of processing power or of time exchangingmessages over the communication channel saves en-ergy. In order to be suitable for an IoT network, a so-lution should be adaptative regarding the node runningit, and the integration of semantic web principles andtechnologies must be thought differently on each levelof LMU-N: a LN does not have the capability to runthe full semantic web stack.

Energy is a primary concern in the IoT domain, fortwo reasons. On the one hand, the dissemination ofnodes in the environment makes it hard to connectthem to a power grid (especially in non-urban area,such as fields or forests). Therefore, some nodes run onbatteries, and their lifetime is directly related to theirconsumption of energy. On the other hand, the mul-tiplication of nodes implies a multiplication of powerconsumers, and managing energy consumption at thescale of the node leads to energy saving at a moreglobal scale. The use of semantic web technologies toreduce the cost of content transport was already dis-cussed in the description of the associated processespresented by [47] and [45]. However, in these papers,the use of semantic web technologies is not modifiedto match the constraints of an IoT network.

To show how semantic web technologies can beadapted to IoT networks requirements, [72] comparesthe different serialization formats available for seman-tically rich data with respect to the size of the messagesencoded in each format, as well as the number of CPUcycle required to produce these messages. The energyconsumption associated to both the creation, the trans-

mission, the reception and the decoding of these mes-sages is then compared for each format. In this paper,two aspects of energetic consumption are pointedout: the processing required to handle the content, andthe communication channels required to exchange it.

Reducing the cost of the communication can beachieved by using protocols adapted to the needs andconstraints of the IoT, and by adapting existing plat-forms to these protocols. For instance, [73] proposesan extension of the Linked Data Platform (LDP)27

specification, a recommendation of the W3C. TheW3C links primitives of LDP to HTTP28, a proto-col rooted in TCP29, therefore requiring connectionbetween the communicating entities. HTTP is there-fore not adapted to IoT architectures, where morelightweight protocols are preferred. For instance, theauthors of this paper propose a mapping of LDP toCoAP30, based on UDP31. CoAP is a protocol espe-cially designed for constrained applications, with re-duced headers and limiter packet body. Such initiativeallows IoT nodes to be connected to the LOD, andtherefore to extend the WoT, while respecting the con-straints of IoT nodes.

Distributing content via adapted protocol also re-quires said content to be stored on the device distribut-ing it. Contributions such as [74] aim at allowing de-vices with limited memory to store semantically richdata. In this paper, the authors propose a method tostore compressed RDF data in memory while ensuringa certain level of efficiency regarding encoding and de-coding of data. This paper is not specifically targetedto the IoT domain, but its contribution matches the re-quirements of this domain. The ability to be able toaccess and modify efficiently the stored data is impor-tant in the case of streaming data, the authors pointout. In the case of IoT deployments, and especially ofsensor networks, the ability of storing and distributingup-to-date data is an important feature, and there mustbe a trade-off between memory optimization, and costof encoding/decoding. [26] also proposes a tuple storesuitable for embedded systems, as well as a protocol-independant RDF broker, that can be mapped to CoAPfor instance. The authors propose the adoption of aprotocol stack adapted to constrained nodes in order to

27https://www.w3.org/TR/ldp/28https://tools.ietf.org/html/rfc261629https://tools.ietf.org/html/rfc79330http://coap.technology/31https://www.ietf.org/rfc/rfc768

N. Seydoux et al. / A unifying vision for the Semantic Web of Things 19

Fig. 3. Example of sequential processes

include them into the SWoT without the need for smartgateways acting as proxies.

Another adaptation to the constraints of the nodes isthe splitting of an application into processes, and thespreading of these processes across the IoT depend-ing of the nodes able to support it. This approach isat the core of the LMU-N architecture, and explainswhy some papers appear in the tables at different levelsor for different contributions, such as [31] where theauthors propose contributions to both aggregation anddecision support. Figure 3 gives a very simple illustra-tion of the sequential spreading of processes, similarto [4] for instance. The content generated by the LN isenriched by the MN, and abstracted according to ap-plicative needs by the UN after it has been notified.

4.2. Managing characteristics of IoT content

Besides the constraints on nodes, the IoT also pro-duces content that has specific intrinsic characteris-tics. The usage of semantic web principles have to beadapted to this type of content as well.

4.2.1. Stream processingThe content dealt with on the SWoT is diverse.

The two families of processes identified in section 2.3(node-related and content-related) are separated by thenature of the focus of the messages they convey, andthe dynamism of the content these processes rely onvaries. Node-related content has both a static and a dy-namic part: endpoint or API description for instance donot vary often for a given node, but its battery level orits availability can be updated frequently. Content col-lected by IoT nodes, on the other hand, is associated toa time and a place of collection, and is by nature dy-

namic. Applications relying on such up-to-date con-tent do not need large historic storage.

This behaviour is associated to stream processing,an approach used in papers such as [43], [75] or [33].In these papers, part of the content, especially con-tent describing the system itself, is static, while con-tent generated by the system (e.g. sensor observations)is managed as an endless stream. This dual approachto content processing ensures scalability, due to thepotentially infinite nature of the content stream gen-erated by the nodes. SPARQL extensions, such as C-SPARQL32 or CQELS33, are design to execute of con-tinuous queries over RDF streams. Continuous query-ing is associated to windowing operators bounding thequeried portion of the stream. [75] distinguishes twotypes of windows: time-based, containing all eventsover a period of time, and event-based, containing afixed number of events. The repetitive, structured na-ture of IoT content makes it suitable for both windowstypes, the choice depending on the application.

4.2.2. Ontology-based Data AccessOBDA is an approach to content access where the

semantic web technologies are used in order to buildan abstraction layer on top of a content source [68].Queries are issued to the abstraction layer, and bene-fit from the expressivity of the semantic web technolo-gies. The abstraction layer then transforms the querybased on mappings to the destination content sourcelanguage, retrieves content, and enriches content ac-cording to the mappings before returning them to theissuer of the query. This approach trades storage ef-ficiency in large content instances for query languageexpressivity. Mapping languages such as R2RML34

or more recently RML35 propose a syntax to definemappings between a structured data format (relationaldatabases for both, XML, JSON and CSV for RML).

[75] propose to access timestamped data streams us-ing the SSN ontology to access streams of observationsproduced by sensors and stored in relational databasesfor the Swiss Experiment project36. This approach ismotivated by the heterogeneity of the underlying datamodels for the different data sources. [68] proposes anOBDA approach where static and streaming data areseparated, in order to identify clearly background

32http://streamreasoning.org/resources/c-sparql

33https://github.com/cqels34https://www.w3.org/TR/r2rml/35http://rml.io/36http://www.swiss-experiment.ch

20 N. Seydoux et al. / A unifying vision for the Semantic Web of Things

knowledge and varying observations. Their motiva-tion is also the homogeneity provided by the abstrac-tion layer over the very heterogeneous underlying sen-sor network. The authors extend the role of the OBDAlayer in their system, by adding content aggregationfunctions to the content enrichment. [48] details whyIoT content is adapted to the OBDA approach: the ob-servations produced by the sensors are strongly struc-tured, and storing only one field header for all the ob-servations is more efficient than storing the metadatafor each observation. The mapping from the metadatato the schema can be easily stored in memory, andqueries on traditional relational databases are more ef-ficient than queries on RDF stores. The limitation ofthis approach, however, is a reduction of the flexibilityof the schema compared to an ontology, and the possi-bility to reason over the dataset. It is worth noting thatOBDA and stream processing can be complementaryapproaches, for instance in [43] and [75].

Overall, this section underlines how the conver-gence between the IoT and the semantic web towardthe SWoT challenges the semantic web principles andtechnologies: the constraints on the nodes and speci-ficities of IoT content are sources of innovation.

5. Conclusion, perspectives and future work

The IoT is a central scientific and technologicaltopic, with a great expected societal impact. In an ef-fort to solve interoperability issues, several IoT net-works are being connected to the WoT, and semanticweb principles and technologies are used to turn theWoT into the SWoT. In order to be able to comparethe existing IoT architectures presented by the SWoTcommunity, we proposed in this paper LMU-N, a uni-fying IoT architectural pattern issued from a bottom-up analysis. LMU-N describes a graph of nodes, wherethree types of nodes are separated based on their physi-cal capabilities and their roles on the network. Genericprocesses have been identified through the projectionof SWoT contributions on LMU-N, and the contribu-tions of the semantic web to the IoT domain have beensituated with respect to these processes and to LMU-N. Our unified approach allowed us to identify cer-tain trends: upstream processes are predominantly rep-resented over downstream processes, and the capaci-ties of the nodes influence the distribution of processesacross the network. To complement this study, evo-lutions of the semantic web, developed in particularto match constraints of the IoT, have been presented.

This twofold approach to the landscape of the SWoTshowed the reciprocity of the convergence of the se-mantic web and the IoT: not only does the semanticweb provide solutions to the interoperability and com-plexity issues of the IoT, but the IoT also challengesthe semantic web principles and technologies to evolveto be compliant with its constraints.

Even if tackled by some studies surveyed in thispaper, some issues remain open for scientific contri-butions. For instance, consistency of content acrossnodes networks has to be ensured, in spite of contenttransformations when it is enriched or lowered. Whilecontributions to content enrichment are many, table 3reveals a reduced number of papers in the domain oflowering, a process essential to consistency.

The IoT is a constrained domain, and it requiresa continuation of the research presented in section4. Formalisms adapted to constrained nodes and net-works are emerging, proposed for instance by theW3C’s WoT working group, and techniques to inte-grate legacy devices in semantic-aware networks needto be developed.

The increasing number of nodes is associated to anincrease in the generated content volume, requiringstudies such as those presented in 4.2.1 to be extendedin order to tackle scalability issues. The use of decen-tralized approaches to LMU-N processes is also cru-cial to scalability: if sub-parts of an IoT network areable to collect, enrich, and process data in order toprovide a service, the load of data exchanged on theglobal network is reduced, and decisions can be takenlocally. Enabling local autonomic behaviours also in-creases resilience. However, local decisions can leadto the emergence of inconsistent global behaviours, asit is presented in the system of system theory [76]. Se-curity, the non-functional requirement listed numberone concern by open-source IoT developers in a recentsurvey by the Eclipse foundation[7], also needs to beintegrated deeper in the SWoT in order to enable safeautonomic computing.

LMU-N will evolve to be adaptable to different ar-chitectural granularities, in an ”unfolding” approach.In the case of very large scale deployments, such assmart cities, the node abstraction is not only suitablefor devices and services, but also for self-sufficient net-works. A smart building is a node compared to the city,and the capacities used for nodes classification needsto be adapted to be adapted to this recursive approach.

N. Seydoux et al. / A unifying vision for the Semantic Web of Things 21

Furthermore, multiple perspectives are opened bythe development of the SWoT. Among them, a promis-ing opportunity is the the evolution of IoT traditionalmachine-centric content to a richer, more expressivecontent, described with vocabularies connected to nat-ural language resources.

These resources enable the next step in semantic in-teroperability on the SWoT, model alignment. Man-ual model mapping is possible in all cases, but auto-matic alignment requires ontologies’ expressivity andmachine-understandability. For instance, the FIESTA-IoT project37 gives access to federated datasets de-scribed with a unique ontology. One the one hand, thisapproach ensures complete semantic interoperability,but on the other hand, it restricts the expressivity avail-able for data providers: datasets already described withanother vocabulary need to be transformed to be com-pliant with the fiesta ontology38. The transformationcan be manual, but it can also be systematized usingautomatic ontology alignments approaches. However,automatic ontologies matching techniques still needevaluation on IoT ontologies. If simple alignments arenot sufficient, complex alignments should also be con-sidered.

Moreover, being associated to M2M communica-tion only, IoT content is mostly numeric, and has to beinterpreted by applications when displayed to a user.The enrichment of content with ontologies enablesnew approaches to user interaction, based on naturallanguage for both user requirements expression, andquery answering. This perspective would make IoT,WoT and SWoT applications easier to access for non-expert users, bringing services to humans at the core ofnodes networks.

References

[1] I. Szilagyi and P. Wira, “Ontologies and Semantic Web for theInternet of Things - a survey,” in IECON, IEEE, 2016.

[2] ITU, “Framework of the web of things,” tech. rep., Interna-tional Telecommunication Union, 2012.

[3] J. Rivera and R. van der Meulen, “Gartner says the internet ofthings installed base will grow to 26 billion units by 2020,”tech. rep., Gartner cabinet, 2013.

[4] P. Desai, A. Sheth, and P. Anantharam, “Semantic Gatewayas a Service architecture for IoT Interoperability,” in Kno.e.sisPublications, 2015.

37http://fiesta-iot.eu/38http://ontology.fiesta-iot.eu

[5] A. Gyrard, M. Serrano, and G. A. Atemezing, “Semantic webmethodologies, best practices and ontology engineering ap-plied to Internet of Things,” in 2015 IEEE 2nd World Forum onInternet of Things (WF-IoT), pp. 412–417, IEEE, 2015.

[6] F. Aıssaoui, G. Garzone, and N. Seydoux, “Providing interop-erability for autonomic control of connected devices,” in Inte-rIoT, pp. 1–7, 2016.

[7] I. Skerrett, “IoT Developer Trends 2017 Edition,” 2017.[8] P. Murdock and al., “Semantic Interoperability for the Web of

Things,” 2016.[9] O. Corcho and R. Garcıa-Castro, “Five challenges for the Se-

mantic Sensor Web,” Semantic Web, vol. 1, no. 1, pp. 121–125,2010.

[10] M. Ben-Alaya, S. Medjiah, T. Monteil, and K. Drira, “To-ward semantic interoperability in oneM2M architecture,” IEEECommunications Magazine, vol. 53, no. 12, pp. 35–41, 2015.

[11] C. Perera, A. Zaslavsky, P. Christen, and D. Georgakopoulos,“Context aware computing for the internet of things: A survey,”IEEE Communications Surveys and Tutorials, vol. 16, pp. 414–454, jan 2014.

[12] D. Pfisterer, K. Romer, D. Bimschas, O. Kleine, R. Mietz,C. Truong, H. Hasemann, A. Kroller, M. Pagel, M. Hauswirth,M. Karnstedt, M. Leggieri, A. Passant, and R. Richardson,“SPITFIRE: toward a semantic web of things,” IEEE Commu-nications Magazine, vol. 49, pp. 40–48, nov 2011.

[13] V. Foteinos, D. Kelaidonis, G. Poulios, P. Vlacheas,V. Stavroulaki, and P. Demestichas, “Cognitive managementfor the internet of things: A framework for enabling au-tonomous applications,” IEEE Vehicular Technology Maga-zine, vol. 8, no. 4, pp. 90–99, 2013.

[14] M. G. Kibria, “Knowledge based open IoT service provision-ing through cooperation between physical web and WoO,” in2015 Seventh International Conference on Ubiquitous and Fu-ture Networks, pp. 395–400, IEEE, 2015.

[15] M. Mrissa, L. Medini, J.-P. Jamont, N. Le Sommer, andJ. Laplace, “An Avatar Architecture for the Web of Things,”Internet Computing, IEEE, vol. 19, no. 2, pp. 30–38, 2015.

[16] M. Eid, R. Liscano, and A. El Saddik, “A universal ontologyfor sensor networks data,” Proceedings of the 2007 IEEE Inter-national Conference on Computational Intelligence for Mea-surement Systems and Applications, CIMSA, no. JULY, pp. 59–62, 2007.

[17] L. Lefort and C. Henson, “Semantic sensor network xg finalreport,” Tech. Rep. June 2011, 2011.

[18] N. Seydoux, K. Drira, N. Hernandez, and T. Monteil, “IoT-O, a core-domain IoT ontology to represent connected devicesnetworks,” in EKAW, 2016.

[19] A. Zanella, N. Bui, A. Castellani, L. Vangelista, and M. Zorzi,“Internet of Things for Smart Cities,” IEEE Internet of ThingsJournal, vol. 1, no. 1, pp. 22–32, 2014.

[20] D. Le-phuoc and M. Hauswirth, “Linked open data in sensordata mashups,” in Proceedings of the 2nd International Work-shop on Semantic Sensor Networks (SSN09), vol. 522, pp. 1–16, 2009.

[21] J. Liu, Y. Li, M. Chen, W. Dong, and D. Jin, “Software-definedinternet of things for smart urban sensing,” IEEE Communica-tions Magazine, vol. 53, no. 9, pp. 55–63, 2015.

[22] M. Compton, C. Henson, L. Lefort, H. Neuhaus, and A. Sheth,“A survey of the semantic specification of sensors,” 2009.

[23] P. Barnaghi, S. Meissner, M. Presser, and K. Moessner, “Senseand sens’ ability: Semantic data modelling for sensor net-

22 N. Seydoux et al. / A unifying vision for the Semantic Web of Things

works,” 2009.[24] S. Nikoli, V. Penca, and Z. Konjovi, “Semantic Web Based Ar-

chitecture for Managing Hardware Heterogeneity in WirelessSensor Network,” in International Journal of Computer Sci-ence and Applications, vol. 8, pp. 38–58, 2011.

[25] P. Vlacheas, R. Giaffreda, V. Stavroulaki, D. Kelaidonis,V. Foteinos, G. Poulios, P. Demestichas, A. Somov, A. Biswas,and K. Moessner, “Enabling smart cities through a cognitivemanagement framework for the internet of things,” IEEE Com-munications Magazine, vol. 51, no. 6, pp. 102–111, 2013.

[26] H. Hasemann, A. Kroller, and M. Pagel, “RDF provisioningfor the internet of things,” in Proceedings of 2012 InternationalConference on the Internet of Things, IOT 2012, pp. 143–150,IEEE, oct 2012.

[27] J. Rowley, “The wisdom hierarchy: representations of theDIKW hierarchy,” Journal of Information Science, vol. 33,no. 2, pp. 163–180, 2007.

[28] M. Ma, P. Wang, and C.-H. H. Chu, “Ontology-Based Se-mantic Modeling and Evaluation for Internet of Things Appli-cations,” in 2014 IEEE International Conference on Internetof Things(iThings), and IEEE Green Computing and Commu-nications (GreenCom) and IEEE Cyber, Physical and SocialComputing (CPSCom), no. iThings, pp. 24–30, IEEE, 2014.

[29] A. Sheth, C. Henson, and S. S. Sahoo, “Semantic Sensor Web,”in IEEE Internet Computing, vol. 12, pp. 78–83, 2008.

[30] C. Henson, A. Sheth, and K. Thirunarayan, “Semantic percep-tion: Converting sensory observations to abstractions,” IEEEInternet Computing, vol. 16, no. 2, pp. 26–34, 2012.

[31] S. Poslad, S. E. Middleton, F. Chaves, R. Tao, O. Necmioglu,and U. Bugel, “A Semantic IoT Early Warning System for Nat-ural Environment Crisis Management,” IEEE Transactions onEmerging Topics in Computing, vol. 3, no. 2, pp. 246–257,2015.

[32] Z. Ming, F. Hong, and M. Yan, “Semantic annotation methodof IOT middleware,” in 2013 Fourth International Conferenceon Intelligent Control and Information Processing (ICICIP),pp. 495–498, IEEE, 2013.

[33] D. Le-Phuoc, H. Nguyen Mau Quoc, H. Ngo Quoc, T. TranNhat, and M. Hauswirth, “The Graph of Things: A step to-wards the Live Knowledge Graph of connected things,” WebSemantics: Science, Services and Agents on the World WideWeb, 2016.

[34] M. Ruta, S. Ieva, G. Loseto, and E. Di Sciascio, “From thePhysical Web to the Physical Semantic Web: knowledge dis-covery in the Internet of Things,” in enth International Confer-ence on Mobile Ubiquitous Computing, Systems, Services andTechnologies, 2016.

[35] L. Daniele, M. Solanki, F. D. Hartog, and J. Roes, “Interoper-ability for Smart Appliances in the IoT World,” in ISWC, 2016.

[36] M. Bermudez-Edo, T. Elsaleh, P. Barnaghi, and K. Taylor,“IoT-Lite: A Lightweight Semantic Model for the Internet ofThings,” in IEEE International conference on Ubiquitous In-telligence and computing, no. November, 2016.

[37] L. Atzori, A. Iera, and G. Morabito, “The Internet of Things:A survey,” Computer Networks, vol. 54, pp. 2787–2805, oct2010.

[38] P. Barnaghi, W. Wang, C. Henson, and K. Taylor, “Semanticsfor the Internet of Thingss: early progress and back to the fu-ture,” International Journal on Semantic Web and InformationSystems, vol. 8, no. 1, pp. 1–21, 2012.

[39] A. J. Jara, A. C. Olivieri, Y. Bocchi, M. Jung, W. Kastner, and

A. F. Skarmeta, “Semantic Web of Things: an analysis of theapplication semantics for the IoT moving towards the IoT con-vergence,” Int. J. of Web and Grid Services, vol. 10, no. 2/3,2014.

[40] J. Kiljander, A. D’elia, F. Morandi, P. Hyttinen, J. Takalo-Mattila, A. Ylisaukko-Oja, J.-P. Soininen, and T. S. Cinotti,“Semantic Interoperability Architecture for Pervasive Comput-ing and Internet of Things,” IEEE Access, vol. 2, pp. 856–873,2014.

[41] J. Kopecky, T. Vitvar, C. Bournez, and J. Farrell, “SAWSDL:Semantic Annotations for WSDL and XML Schema,” IEEEInternet Computing, vol. 11, pp. 60–67, nov 2007.

[42] Y. Qin, Q. Z. Sheng, and E. Curry, “Matching Over LinkedData Streams in the Internet of Things,” IEEE Internet Com-puting, vol. 19, no. 3, pp. 21–27, 2015.

[43] J.-p. J. Calbimonte, H. Jeung, O. Corcho, and K. Aberer, “Se-mantic Sensor Data Search in a Large-Scale Federated SensorNetwork,” Semantic Sensor Networks, pp. 14–29, 2011.

[44] I. Chatzigiannakis, H. Hasemann, M. Karnstedt, O. Kleine,A. Kroller, M. Leggieri, D. Pfisterer, K. Romer, and C. Truong,“True Self-Configuration for the loT,” in 3rd InternationalConference on the Internet of Things (IOT), 2012.

[45] M. I. Ashraf, L. Goratti, J. Haapola, and C. Pomalaza-Raez,“Distributed semantic algorithm for power constrained pub-lish/subscribe routing,” in 2010 Sixth International Conferenceon Intelligent Sensors, Sensor Networks and Information Pro-cessing, pp. 91–96, IEEE, dec 2010.

[46] S. B. Fredj, M. Boussard, D. Kofman, and L. Noirie, “A Scal-able IoT Service Search Based on Clustering and Aggrega-tion,” in 2013 IEEE International Conference on Green Com-puting and Communications and IEEE Internet of Things andIEEE Cyber, Physical and Social Computing, pp. 403–410,IEEE, aug 2013.

[47] F. Bouhafs, M. Merabti, and H. Mokhtar, “A semantic cluster-ing routing protocol for wireless sensor networks,” in CCNC2006. 2006 3rd IEEE Consumer Communications and Net-working Conference, 2006., vol. 1, pp. 351–355, IEEE, 2006.

[48] E. Siow, T. Tiropanis, and W. Hall, “SPARQL-to-SQL on In-ternet of Things Databases and Streams,” in ISWC, 2016.

[49] Y. Qin, Q. Z. Sheng, N. J. Falkner, S. Dustdar, H. Wang, andA. V. Vasilakos, “When Things Matter: A Survey on Data-Centric Internet of Things,” Journal of Network and ComputerApplications, no. April, 2016.

[50] Z. Li, C. H. Chu, W. Yao, and R. a. Behr, “Ontology-drivenevent detection and indexing in smart spaces,” in Proceedings- 2010 IEEE 4th International Conference on Semantic Com-puting, ICSC 2010, pp. 285–292, 2010.

[51] D. Hussein, S. N. Han, G. M. Lee, N. Crespi, and E. Bertin,“Towards a dynamic discovery of smart services in the socialinternet of things,” Computers & Electrical Engineering, 2016.

[52] S. K. Datta and C. Bonnet, “Connect and Control Things : In-tegrating Lightweight IoT Framework into a Mobile Applica-tion,” in 9th International Conference on Next Generation Mo-bile Applications, Services and Technologies, no. SEPTEM-BER 2015, p. 6, 2015.

[53] Z. Wang, R. Bie, and M. Zhou, “Hybrid ontology matching forsolving the heterogeneous problem of the IoT,” Proc. of the11th IEEE Int. Conference on Trust, Security and Privacy inComputing and Communications, TrustCom-2012 - 11th IEEEInt. Conference on Ubiquitous Computing and Communica-tions, IUCC-2012, pp. 1799–1804, 2012.

N. Seydoux et al. / A unifying vision for the Semantic Web of Things 23

[54] D. Le-Phuoc, H. Q. Nguyen-Mau, J. X. Parreira, andM. Hauswirth, “A middleware framework for scalable manage-ment of linked streams,” Web Semantics: Science, Services andAgents on the World Wide Web, vol. 16, pp. 42–51, 2012.

[55] V. Charpenay, S. Kabisch, D. Anicic, and H. Kosch, “An on-tology design pattern for IoT device tagging systems,” in Pro-ceedings - 2015 5th International Conference on the Internetof Things, IoT 2015, 2015.

[56] C. Sarkar, A. U. Akshay, R. V. Prasad, A. Rahim, R. Neisse,and G. Baldini, “DIAT: A scalable distributed architecture forIoT,” in IEEE Internet of Things Journal, vol. 2, 2015.

[57] S. Nambi, C. Sarkar, R. V. Prasad, and a. Rahim, “A unified se-mantic knowledge base for IoT,” Internet of Things (WF-IoT),2014 IEEE World Forum on, pp. 575–580, 2014.

[58] S. De, P. Barnaghi, M. Bauer, and S. Meissner, “Service mod-elling for the Internet of Things,” in 2011 Federated Confer-ence on Computer Science and Information Systems (FedC-SIS), pp. 949–955, 2011.

[59] J. Li, N. Zaman, and H. Li, “A Decentralized Locality-Preserving Context-Aware Service Discovery Framework forInternet of Things,” in 2015 IEEE International Conference onServices Computing, pp. 317–323, IEEE, jun 2015.

[60] D. J. Russomanno, C. Kothari, and O. Thomas, “Sensor ontolo-gies: From shallow to deep models,” in Annual SoutheasternSymposium on System Theory, vol. 37, pp. 107–112, 2005.

[61] C. Perera, A. Zaslavsky, C. H. Liu, M. Compton, P. Christen,and D. Georgakopoulos, “Sensor search techniques for sens-ing as a service architecture for the internet of things,” IEEESensors Journal, vol. 14, no. 2, pp. 406–420, 2014.

[62] S. N. Han, G. M. Lee, and N. Crespi, “Semantic Context-AwareService Composition for Building Automation System,” IEEETransactions on Industrial Informatics, vol. 10, pp. 752–761,feb 2014.

[63] A. I. Maarala, X. Su, and J. Riekki, “Semantic data provision-ing and reasoning for the Internet of Things,” in 2014 Interna-tional Conference on the Internet of Things, IOT 2014, 2014.

[64] F. Ganz, D. Puschmann, P. Barnaghi, and F. Carrez, “A Prac-tical Evaluation of Information Processing and AbstractionTechniques for the Internet of Things,” IEEE Internet of ThingsJournal, vol. 2, no. 4, pp. 340–354, 2015.

[65] K. Kotis, A. Katasonov, and J. Leino, “Aligning Smart andControl Entities in the IoT,” in Internet of Things, SmartSpaces, and Next Generation Networking: 12th InternationalConference, NEW2AN 2012, and 5th Conference, ruSMART2012, St. Petersburg, Russia, August 27-29, 2012. Proceedings(S. Andreev, S. Balandin, and Y. Koucheryavy, eds.), vol. 7469

LNCS, pp. 39–50, Berlin, Heidelberg: Springer Berlin Heidel-berg, 2012.

[66] W. Wang, S. De, R. Toenjes, E. Reetz, and K. Moessner, “Acomprehensive ontology for knowledge representation in theinternet of things,” Proc. of the 11th IEEE Int. Conference onTrust, Security and Privacy in Computing and Communica-tions, TrustCom-2012 - 11th IEEE Int. Conference on Ubiqui-tous Computing and Communications, IUCC-2012, pp. 1793–1798, 2012.

[67] F. Ganz, P. Barnaghi, and F. Carrez, “Automated SemanticKnowledge Acquisition From Sensor Data,” Systems Journal,IEEE, vol. PP, no. 99, pp. 1–12, 2014.

[68] E. Kharlamov, Y. Kotidis, T. Mailis, C. Neuenstadt, C. Niko-laou, O. Ozcep, C. Svingos, D. Zheleznyakov, S. Lamparter,I. Horrocks, Y. Ioannidis, and R. Moller, “Towards AnalyticsAware Ontology Based Access to Static and Streaming Data(Extended Version),” in ISWC, pp. 1–22, 2016.

[69] Q. Wu, G. Ding, Y. Xu, S. Feng, Z. Du, J. Wang, and K. Long,“Cognitive Internet of Things: A New Paradigm Beyond Con-nection,” IEEE Internet of Things Journal, vol. 1, pp. 129–143,apr 2014.

[70] L. Nachabe, M. Girod-Genet, and B. El Hassan, “Seman-tic Techniques for IOT Data and Service Management on toSmart System,” International Journal of Wireless & MobileNetworks, vol. 8, no. 4, pp. 43–63, 2016.

[71] M. Compton, H. Neuhaus, K. Taylor, and K. Tran, “Reasoningabout Sensors and Compositions.,” 2009.

[72] X. Su, J. Riekki, J. K. Nurminen, J. Nieminen, andM. Koskimies, “Adding semantics to internet of things,” Con-currency Computation, vol. 27, no. 8, 2015.

[73] G. Loseto, S. Ieva, F. Gramegna, M. Ruta, F. Scioscia, E. D.Sciascio, and B. I, “Linked Data ( in low-resource ) Platforms: a mapping for Constrained Application Protocol,” in ISWC,(Kobe), 2016.

[74] H. R. Bazoobandi, S. De Rooij, J. Urbani, A. Ten Teije, F. VanHarmelen, and H. Bal, “A Compact In-Memory Dictionary forRDF Data,” in ESWC, vol. 9088, pp. 205–220, 2015.

[75] J.-P. Calbimonte, H. Jeung, O. Corcho, and K. Aberer, “En-abling Query Technologies for the Semantic Sensor Web,” In-ternational Journal on Semantic Web and Information Systems,vol. 8, no. 1, pp. 43–63, 2012.

[76] C. A. Lana, N. M. Souza, M. E. Delamaro, E. Y. Nakagawa,F. Oquendo, and J. C. Maldonado, “Systems-of-systems devel-opment: Initiatives, trends, and challenges,” in 2016 XLII LatinAmerican Computing Conference (CLEI), pp. 1–12, IEEE, oct2016.