The IoT is a domain in exponential growth: both the number of connected devices and the quantity of data they produce are increasing. The heterogeneity of technologies involved, and the diversity of domains impacted raise interoperability concerns. The semantic web principles and technologies are semantic interoperability providers, and ontologies like SSN have been used in several IoT projects. However, many existing IoT ontologies fail to comply with the good practices of the semantic web. After detailing such good practices, this paper proposes IoT-O, a modular core-domain IoT ontology. IoT-O is then showcased in a home automation use case: it is used to semantically describe the devices of the system, and to guide the decisions of an autonomic agent.
compliant with these guidelines. This is why we propose IoT-O4, an IoT coredomain modular ontology engineered for reusability and extensibility. IoT-O is also available on the LOV5.
In the remainder of this paper, section 2 introduces a motivating use case that will serve to instantiate portions of IoT-O. Section 3 presents the design process of IoT-O, and gives an overview of the ontology. Finally, section 4 details how IoT-O is instantiated in the use case. 2
The miniaturization of devices has made it possible to disseminate multiple lowpower devices in an everyday environment, such as the home. The automation of the home, or domotics, is a domain of the Internet of Things (IoT) with direct impact on citizens. At LAAS-CNRS, the ADREAM project6 aims at conducting research thanks to an instrumented, energy-positive building. This building is equipped with more than 4500 sensing devices, producing up to 500,000 measures a day. Inside the building, there is a mock-up apartment equipped with commercial devices from diverse vendors. Deployed devices include sensors (temperature, luminosity, humidity, pressure), actuators (fan, space heater, diverse lamps), which communicate using di erent technologies (phidget, ethernet, zigbee) with gateways connected to a central server (see g. 1).
Our use case is de ned as follows: the user should be able to de ne simple high-level policies to manage its environment ("the temperature in the living room should stay between 19oC and 25oC"), without having to select speci c 4 http://www.irit.fr/recherches/MELODI/ontologies/IoT-O 5 http://lov.okfn.org/dataset/lov/vocabs/ioto 6 http://www.laas.fr/public/en/adream sensors or actuators to perform the task. The user should also be able to extend the capabilities of the apartment by adding devices without restarting the system.
To ful ll these requirements, both syntactic and semantic interoperability among devices are required. Syntactic interoperability is ensured using OM2M7, an open-source horizontal integration platform implementing the oneM2M standard. On top of OM2M, another platform, SemIoTics, is in charge of ensuring semantic interoperability and of implementing the policies de ned by the user. SemIoTics is guided by a knowledge base containing information about the devices, described with our ontology, IoT-O. This use case is applied to home automation and is described in a dedicated knowledge base extending IoT-O, ADREAM-Model8, but the genericity of IoT-O makes it relevant to any domain impacted by the IoT. 3
The design of IoT-O is compliant with the NeOn methodology, presented in [
The rst step of the NeOn process is to de ne requirements. We split them in two types: conceptual, regarding the concepts that should be present in the ontology (detailed in section 3.1), and functional, regarding the ontology structure and design principles (detailed in section 3.2).
These requirements are used to analyze existing IoT ontologies: Semantic
Sensor Network (SSN)9, Smart Appliance REFerence (SAREF)10, iot-ontology
11, IoT-lite 12, Spit re 13, IoT-S14, SA15 and the oneM2M base ontology16.
These ontologies are IoT ontologies for which we have found information on
the web. Further details are available on the Linked Open Vocabularies for the
IoT (LOV4IoT)17, a recent initiative that lists IoT ontologies, even if they are
not referenced on the LOV because they fail to comply with its requirements
recalled in [
As recommended by NeOn, reusable ontologies satisfying parts of the requirements are analyzed and presented in section 3.3. The core-domain ontology we propose is nally described in section 3.4. 7 om2m.org 8 http://www.irit.fr/recherches/MELODI/ontologies/Adream-Model 9 http://purl.oclc.org/NET/ssnx/ssn 10 http://sites.google.com/site/smartappliancesproject/ontologies 11 http://ai-group.ds.unipi.gr/kotis/ontologies/IoT-ontology 12 http://iot.ee.surrey.ac.uk/ ware/ontologies/iot-lite 13 http://sensormeasurement.appspot.com/ont/sensor/spit re.owl 14 http://personal.ee.surrey.ac.uk/Personal/P.Barnaghi/ontology/OWL-IoT-S.owl 15 http://sensormeasurement.appspot.com/ont/sensor/hachem onto.owl 16 http://www.onem2m.org/ontology/Base Ontology/ 17 http://www.sensormeasurement.appspot.com/?p=ontologies 3.1
Conceptual requirements These requirements come from an analysis of the IoT domain, driven by the home automation use case introduced in section 2, but not limited to it: the use case is not seen as an end per se, but as an instantiation of the general domain of the IoT. To be reusable in a wide scope of domains, an IoT ontology should contain a set of key concepts. These are representative of IoT systems with no regard to the application domain. This approach facilitates the merging of data collected in di erent domains for horizontal applications, and allows the ontology to be an extendable core-domain ontology. We distinguish namely: { "Device" and "software agent" constitute the two basic components of an IoT system, composed of both physical and virtual elements. The devices can be of two principle types, not mutually exclusive, that are listed below. { "Sensor" are devices acquiring data, and "observation" describe the acquisition context and the data collected by the system. { "Actuator" are the devices that enable the system to act on the physical world, and "action" represents what they can perform. { "Service": In many cases, the IoT and the programmable web are very close. Connected devices can be seen as service providers and consumers, and by specifying a notion of service, every aspect of an IoT system can be represented. { "Energy": In the paradigm of pervasive computing, many distributed Things perform computations. Most of these Things being physical devices, a complete modelling of the system will include a description of their energy consumption. Energy management is a crucial topic in IoT systems. { "Lifecycle": Be it data, devices or services, IoT components are all included in di erent scales of lifecycles. Devices are switched on and o , services are deployed or updated, pieces of data become outdated... The evolution through a set of discrete states representing a lifecycle is an important concept for IoT systems.
Concept coverage by existing ontologies Table 2 sums up the assessment of existing IoT ontologies regarding the presence of key concepts. One star means that the concept is super cially represented (few specializations, data/object properties), two stars that the requirement is covered, and stars between parentheses indicate that the requirement is met by an included ontology. IoT-O, the ontology we propose, is also included for comparison. Note that we focus on connected device ontologies, and exclude, on purpose, the ontologies SSN is based upon, since they are only focused on sensors and observation, which is only a subset of the identi ed key concepts. We can observe that some of the IoT ontologies cover most of the key concepts but none of them covers them all. Moreover, the di erent concepts are not represented with the same level of expressivity. In iot-ontology and SAREF, key concepts such as Actuator or Action are present but their representation is limited. For example, an actuator is de ned as a device that modi es a property. This is less expressive than what can be expressed for a sensor with SSN which proposes a deep modeling of the sensors and the property they observe, but also of the relations between the sensors and their observations, and of the observations themselves. In eDIANA18, an ontology referenced by SAREF, some specializations of actuator are given, but the mappings from these specializations to the saref:Actuator concept are not available directly. This analysis highlights the fact that an ontology for Actuators and Actions is needed (c.f. section 3.3). This analysis also highlights the failure of existing IoT ontologies in representing correctly all IoT key concepts. As these concepts are not limited to the IoT domain, reusing ontologies dedicated to them (such as SSN for sensor) could help gain in expressivity, as is shown in section 3.2.
Reusability: One of the most important aspects of an ontology in such a broad
domain as IoT is reusability: if an ontology is ad-hoc to a project, the work done
in its de nition will not bene t further projects. It is a critical issue that can be
solved by di erent, non-mutually exclusive approaches:
{ Modularization: as stated in [
Level of formalism: To use the full advantages of the semantic description of
devices and data, the description should enable reasoning and inference. This
choice is motivated by the possibilities it opens:
{ Applied to data, it is a way to bring context-awareness, as presented in [
Identi cation of existing ontologies is part of the NeOn process. Some concepts,
which are part of the conceptual requirements are de ned by existing ontologies
that are imported in IoT-O to avoid rede nition. SSN is a widely used W3C
recommended ontology for sensors and observations. However, no ontology
describes the concept of actuator the way SSN describes the concept of sensor.
This is why we propose Semantic Actuator Network (SAN)22, the Semantic
Actuator Network ontology. Actuators are devices that transform an input signal
into a physical output, making them the exact opposite of sensors. SAN is built
around Action-Actuator-E ect (AAE)23, a design pattern we propose, inspired
from the Stimulus Sensor Observation (SSO) design pattern described in [
To de ne the notion of service, IoT-O imports Minimal Service Model (MSM), a lightweight service ontology which is generic enough to represent both REST and WSDL services (contrary to OWL-S24). The notion of energy consumption dedicated to the IoT is speci ed in PowerOnt, an ontology referenced by SAREF. The concepts of lifecycle are described using Lifecycle25, a lightweight vocabulary de ning state machines. We extended Lifecycle in the IoT-lifecycle26 ontology with classes and properties speci c to the IoT. Finally, to maximize extensibility and reusability, IoT-O imports DUL27, a top-level ontology, and aligns all its concepts and imported modules with it. 3.4
IoT-O, the core-ontology we propose is composed of several modules. IoT-O's architecture is summarized in gure 4. The names of the newly created resources are in red and highlighted, the names of the reengineered resources are underlined, and the arrows show dependencies. Solid arrows represent imports, and dashed arrows the reuse of concepts without import.
{ The Sensing module describes the input data. Its main classes come from SSN: ssn:Sensor and ssn:Observation. ssn:Device and its characteristics (ssn:OperatingRange, ssn:Deployment...) provide a generic device description. 22 https://www.irit.fr/recherches/MELODI/ontologies/SAN 23 http://ontologydesignpatterns.org/wiki/Submissions:Actuation-Actuator-E ect 24 https://www.w3.org/Submission/OWL-S/, which is more dedicated to WSDL-based services 25 http://vocab.org/lifecycle/schema 26 https://www.irit.fr/recherches/MELODI/ontologies/IoT-Lifecycle 27 http://www.ontologydesignpatterns.org/ont/dul/DUL.owl { The Acting module describes how the system can interact with the physical world. Its main classes come from SAN: san:Actuator and san:Actuation. It also reuses SSN classes that are not speci c to sensing, such as ssn:Device. { The Lifecycle module models state machines to specify system life cycles and device usage. Its main classes are lifecycle:State and lifecycle:Transition. { The Service module represents web service interfaces. Its main classes come from MSM: msm:Service and msm:Operation. Services produce and consume msm:Messages, and RESTful services can be described with hRest. { Energy module: IoT-O's energy module is de ned by PowerOnt. It provides the poweront:PowerConsumption class, and a set of properties to express power consumption pro les for appliances.
The core of IoT-O: IoT-O28 is both the name of the ontology and of the top module. It gives a conceptualization of the IoT domain, independent of the application, providing classes and relationships to link the underlying modules. Since many concepts are already de ned in the modules, IoT-O's core is limited: it denes 14 classes (out of 1126 including all modules), 18 object properties (out of 249) and 4 data properties (out of 78). IoT-O key class is iot-o:IoT Thing, which can be either an ssn:Device or an iot-o:SoftwareAgent. The power consumption 28 http://www.irit.fr/recherches/MELODI/ontologies/IoT-O.owl of ssn:Devices is associated to lifecycle:State and poweront:PowerConsumption. iot-o:IoT Thing is a provider of msm:Service, and an msm:Operation can have an iot-o:ImpactOnProperty on an ssn:Property, linking abstract services to the physical world through devices.
As a core domain ontology, IoT-O is meant to be extended regarding speci c applicative needs and real-life devices and services. This design, inspired by SSN, makes IoT-O independent of the application. 4 4.1
Autonomic computing is a programming paradigm proposed in [
The rst step of the MAPE-K loop is the monitoring of the controlled system. In the apartment, sensors produce data re ecting their observations. This data is enriched to become a reusable piece of knowledge. Enrichment of sensor data is performed using the SSN ontology, which is in the Sensing module of IoT-O. Each ssn:Sensor has an ssn:Observation stream composed of ssn:SensorOutput whose value is described by ssn:ObservationValue. For provenance purposes, a ssn:SensorOutput can be linked to its original representation (before enrichment) with the iot-o:hasRawRepresentation data property. The sensor's characteristics (ssn:MeasurementProperty, the ssn:Property of the ssn:FeatureOfInterest it observes) can be used to enrich the observation as well. IoT-O and SSN are generic ontologies, so they might need to be extended with application-speci c modules to be fully functional. Such extension is proposed in the Adream-Model module29.
In our use case, the temperature sensor produces raw observations in the form
of XML documents standardized according to the oneM2M Content Instance
resource type. The enrichment process requires an approach speci c to the data,
either by writing a dedicated enrichment script, or by using semantic annotations
in the data as in [
In the Analysis step, the enriched observations are compared to the needs expressed by the user (represented by rules). User preferences are represented using the concepts de ned in yet another module: Autonomic30. The user creates autonomic:PropertyConstraints (seamlessly through a graphical interface), transforming a ssn:Property into a autonomic:ConstrainedProperty. In our use case, the ss::Property temperature of the ssn:FeatureOfInterest living room air has two constraints, instances of autonomic:MaximumValue (25oC) and autonomic:MinimumValue (19oC). The last ssn:ObservationValue of the autonomic:ConstrainedProperty is out of the bounds de ned by the autonomic:PropertyConstraint (26oC instead of 25), so the temperature is classi ed by the reasoner as an autonomic:OutOfBoundsProperty thanks to custom rules. 4.4
In the planing phase, the autonomic agent uses the inferred symptoms and policies de ned by the user or by the administrator beforehand to de ne a series of actions that have to be implemented on the system. The description of the actions is performed using SAN, the actuator ontology that also describes the actuators in the system. The agent, with successive queries to the knowledge base, will look for san:Actuator instances that san:actsOn the autonomic:OutOfBoundsProperty, and which san:receivesActuation an actuation that iot-o:hasImpact an autonomic:ImpactOnProperty that is coherent with the symptom. In the example, since the temperature is too high, the adream-model:fan can be used, but also the adream-model:spaceHeater, since its adream-model:turnO operation has a adream-model:NegativeImpact on the temperature. 29 http://www.irit.fr/recherches/MELODI/ontologies/Adream-Model 30 http://www.irit.fr/recherches/MELODI/ontologies/Autonomic
The orchestration of these actions (if need be) are determined using the Lifecycle module of IoT-O, which represents the devices as state machines by integrating the Objects with States (ows)31 ontology design pattern. ssn:Device (superclass of both ssn:SensingDevice and san:ActuatingDevice) are objects that ows:hasState exactly 1 ows:State, because objects should only be in one state at a time. The ows:State is equivalent to the lifecycle:State (from the Lifecycle32 vocabulary, extended by the IoT-Lifecycle33 ontology), and lifecycle:State are connected by lifecycle:Transition instances. Thanks to this vision of state machines, stateful transitions (that are only available in certain states of the device) can be represented. Only msm:Operation instances that iot-o:isGroundedBy a san:Actuation that iot-lifecycle:triggersTransition a lifecycle:Transition that is a lifecycle:possibleTransition of the device current lifecycle:State can be called at a given time. For instance, the space heater adream-model:turnO operation will only be available if the space heater is on. In our example it is o , so the agent plans to turn on the fan and creates the corresponding san:ActuationValue.
The selection of devices and their operations is driven by necessity (only the devices impacting the right property are selected), but it can also be driven by policies based on knowledge about the system, to minimize energy consumption, to optimize reaction time... 4.5
With the monitoring step, the execution is the moment when the agent is in
direct contact with the controlled system: it implements the di erent actions
composing the plan. The agent can execute a san:Action if it iot-o:isGroundedBy an
msm:Operation. During the execution, the agent transforms the
san:ActuationValue in a representation which is suitable for the corresponding operation. The
translation of knowledge into a simpler data format (the opposite process of
enrichment) can be driven by the semantic description of Operations, or dedicated
annotations as in [
This paper introduces IoT-O, a modular core-domain IoT ontology designed to be compliant with identi ed requirements. After a detailed presentation of its modules, an instantiation of IoT-O is presented in a home automation use case. 31 http://delicias.dia. .upm.es/ontologies/ObjectWithStates.owl 32 http://vocab.org/lifecycle/schema 33 http://www.irit.fr/recherches/MELODI/ontologies/IoT-Lifecycle The system implements the MAPE-K loop, an element of autonomic computing, and uses IoT-O at each step of the loop to describe knowledge about the connected devices and about the data they produce and consume.
In this paper, enrichment and translation techniques (allowing the transformation back and forth from data to knowledge) have been overviewed. Such techniques are essential to include constrained devices into the IoT: enriched data is more reusable than raw data, but it is heavier to exchange and process, so transformation is required between the end devices (sensors, actuators) and the more powerful nodes of the IoT, e.g. gateways, servers, laptops... We are currently working on such an approach. Other perspectives of our work will be to de ne an intuitive way to help end users/administrators express constraints and policies to drive the system.