This paper describes an application of semantics to sensors to automatically create sensor compositions. Two main contributions of this paper are a sensor ontology and a composition system. The OWL ontology describes user goals, functional and non-functional properties of sensors and sensor composition, while the Region Connection Calculus and Allen's Interval Algebra are used for describing the spatial and temporal properties of sensors. The composition system composes sensors and sensor processes to satisfy a user specified goal. Through experiments, this paper demonstrates the feasibility of semantics for specifying sensor compositions with low reasoning times, and the composition approaches with low composition times in most cases. Some scalability issues with semantic sensor networks are resolved with reasoning, but new composition approaches are needed for large scale use requiring large compositions.
Sensors are becoming more prevalent and diverse [
The sensor ontology describes semantics for composition such as types,
processes, spatial and temporal properties, user goals, and sensor specific semantics.
The Web Ontology Language (OWL) (specifically OWL 1.1 [
The composition system, in the initialisation phase, takes syntactic sensor specifications, extracts the semantics, and asserts them into the sensor ontology. The derivation of RCC and AIA relations are performed by the composition system. In the composition phase, a matchmaking algorithm then analyses a set of sensors and sensor processes to create a set of possible matches based on the sensor and user goal constraints. A composition algorithm then creates composite sensors by building a binary tree of source to target sensors using the matches. The new composite sensors are then added to the current set of sensors. The process repeats until the goals are satisfied or there are no more compositions to be created. Every composition must end at data collecting sensors or base stations.
This paper discusses experiments that evaluate the capabilities of the composition system through three aspects of sensor networks: the number of neighbouring sensors, the number of sensors in the network, and the length of composition. The first aspect looks at wide composition trees, which translates to many short compositions at each sensor. The second aspect tests the filtering of relevant sensors and processes by the matchmaking algorithm. The third aspect looks at deep composition trees, or long sensor compositions. Favourable results are seen with the first two experiments with low execution and reasoning time. The third experiment shows some scalability issues with the matchmaking approach adopted by other sensor composition systems.
An evaluation of the sensor ontology and composition system is given to identify and discuss strengths and weaknesses of the system. The novel addition of RCC and AIA semantics in addition to OWL allows relations about spatial and temporal properties be asserted in the ontology. These additions are exploited in the composition system through the fast reasoning to find spatially and temporally collocated sensors. The composition system allows composition to be created automatically based on user specified goals. While short compositions can be found quickly, long and large sensor compositions are still problematic because of the matchmaking approach. However, the application of semantics to sensors shows feasibility for composition, which provides semantic interoperability of sensor resources, and automation of many tasks, allowing users to interact with sensor networks on a higher level by expressing queries and goals.
The rest of this paper is organized as follows. Section 2 summarises related
work. Section 3 describes the semantic constructs of the sensor ontology with
visualisations. Section 4 describes the composition system. Section 5 shows a
few experiments with the system. Section 6 gives an evaluation of the ontology
and the system. Section 7 provides concluding remarks. The sensor ontology and
other artefacts from this paper, including the original thesis [
This section looks at the related research fields of semantic Web service composition (SWSC), sensor ontologies and sensor composition.
In SWSC, semantic interoperability of Web services provides additional
capabilities that can be automated: (semantic) searching, invocation, composition,
interoperation, execution monitoring and recovery [
The research on sensor ontologies has produced many ontologies for modelling
of sensors and their capabilities, but they are domain specific for each research
community [
There are two sensor composition systems that relates to this paper by
focusing on the data flow in sensor networks: a Semi-Automatic Sensor Composer and
Semantic Streams. The Semi-Automatic Sensor Composer is an extension of the
Semi-Automatic Web services Composer [
Semantic Streams offers a fully automated semantic sensor composition
system based on Constraint Logic Programming Reals (CLP-R) [
These two sensor composition systems have a similar composition approach of identifying the user goals and build sensor composition trees towards that goal. Composing sensors mainly differs from composing Web services by the additional sensor, spatial and temporal properties. These properties add additional complexities to the composition problem, which are not addressed by Web services composition. This paper follows the composition approach of the two sensor systems above. 3
This section describes the construction of the sensor ontology. Each of the semantics used for composition can exist independently, but each is required in the composition system. 3.1
The top level of the COMETMAN energy classification scheme of sensors is
used in the sensor ontology: Chemical, Optical, Mechanical, Electrical, Thermal,
Magnetic, Acoustic, and Nuclear [
The semantics for sensor processes are modelled after the CSIRO Sensor
Ontology [
The matching of outputs to inputs on the semantic level allows analysis beyond precise or identical definitions required on the syntactic level. The added semantics allows determining the degree of match of outputs to inputs, which is exploited with the DataT ype concept. The hasDataT ype property maps to a string representing the DataT ype concept. The composition system takes the strings and interprets them as concepts3. 3.3
Sensors have spatial information required in forming networks that can benefit
from a semantic interpretation, Figure 3. The sensing/communication regions of
sensors can be approximated with polygons [
Some advantages of using RCC relations are that the RCC relations only
change when necessary, whereas numeric descriptions require frequent
computation. The view of the spatial and temporal world as relations is more natural
to users. The RCC relations can simplify the creation of sensor compositions
through RCC reasoning, but this is not explored in this paper. The
complexity of RCC can be varied by using more or fewer relations to describe spatial
properties [
Sensors operate over one or more time intervals that can be described
qualitatively using the thirteen exhaustive relations of Allen’s Interval Algebra (AIA) [
Fig. 4. The temporal semantics in the sensor ontology. 3.5
Sensors are the focus, which is reflected in Figure 5 showing the sensor semantics
in the sensor ontology. The AtomicSensor concept is described by the semantics
above. The CompositeSensor concept connects the sensors and defines the order
of execution and data flow.
Goals specify the desired results and constraints that the composition system
must try to satisfy. The semantics shown in Figure 6 are modelled after the
Sensor concept in a similar way that Lin et al. [
The changeStream data property is used in the composition system to determine if the most general data type of Stream should be instantiated. Instantiating Stream to a more specific type allows sensors that can forward data to forward it directly to sensors that can take that specific type. This allows for sensors not required in the goal to be used and so extend the length of compositions. The sensingRegionOverlap data property indicates whether the sensing regions of the sensors must overlap. 4
A conceptual model of the composition system, with the flow of information, is given in Figure 7. This section covers the parts of the system surrounding the sensor ontology. Syntactic descriptions of sensors and goals can be easily created or generated. The composition system extracts semantics from these syntactic descriptions and populates the sensor ontology. The following tools used are not discussed: OWL 1.14, Pellet 2.0.0 RC75, and OWL API6. 4 http://www.w3.org/Submission/owl11-overview/ 5 http://clarkparsia.com/pellet/ 6 Version that came with Pellet 2.0 RC7 was used. http://owlapi.sourceforge.net/ Matching sensors is filtering the required sensors based on user goals and then filtering on sensor properties as compositions are created. Filtering sensor types matching the goal is a simple query. Filtering on spatial and temporal properties is made simpler by the asserted RCC and AIA relations. The canSendDataT o relation in the sensor ontology is derived from the RCC and AIA relations to suggest sensors in range and that operate in the same time frames of each other.
For sensor processes, the matching is based on their inputs and outputs. The
ontology allows determining the degree of match of the concept(s) of the
outputs to the concept(s) of the inputs. There are four degrees of matches: exact
(outputClass ≡ inputClass), subsumed by (outputClass v inputClass),
subsumes (outputClass w inputClass) and fail [
Creating new matches is performed by a matchmaking algorithm. The algorithm creates a set of possible matches from the current working set of sensors for the composition algorithm. The matches are recorded based on the Binding concept with the structure of the properties of the concept. The composition algorithm creates a new set of composite sensors on each iteration and the matchmaking process is repeated on these new composite sensors. The composition algorithm takes the bindings created by the matchmaking algorithm and builds a binary tree of processes and sensors to create new composite sensors. Base stations are defined as sensors having a process marked to be a data collector. Composition is directed towards these sensors and all compositions end with a base station.
New composite sensors are checked against the goal requirements. If the new composite sensors satisfy the goal, the system stops and adds the new sensors into the sensor ontology; otherwise, the composite sensors are added to the set of working sensors and the process repeats. The new composite sensors are compared to the old sensors to reduce the number of comparisons. Circular compositions resulting from composite processes on the same sensors or in a chain of compositions are checked against the current composition trees.
For experiments, the composition system is split into the “Initialisation Phase”, encapsulating the creation and extraction of semantics from syntactic descriptions (left half of Figure 7), and “Composition Phase” (right half of Figure 7), encapsulating the composition system. Data collection and ontology consistency checks of the ontology are performed at the start, middle, and end of the two phases of the system. The data collected are summarised in next section. 5
Experiments were conducted to evaluate the system for performance based on
automatically generated random syntactic sensor descriptions. The evaluation uses
randomly generated location, sensing and communication regions, and operation
intervals. The same properties are also randomised in the goal descriptions. Data
points were recorded after three data points had been generated to “warm-up”
the Just-In-Time compiler [
∗ Outputs wind speed • For Experiment 3 only: additional Message Forwarding Process ∗ Input is a generic stream of data and output is also a stream – n Temperature Sensors with • Temperature Process (outputs the ambient temperature) • Wind Chill Temperature Process ∗ Inputs of wind speed and temperature ∗ Output of wind chill temperature • For Experiment 3 only: additional Message Forwarding Process ∗ Input is a generic stream of data and output is also a stream – 2 × n/5 Base Stations with
• Capture Messages Process – Randomising • Location • Sensing Region
∗ Maximum size: 100 × 100 • Communication
∗ Maximum size: 200 × 200 • Operation Interval
∗ Operation times varies over the same day
The number of wind and temperature sensors remains the same, and the number of base stations is in proportion to the other sensors in a ratio of (number of windsensors + number of temperature sensors) : 1 basestation. The ratio ensures that satisfiable compositions are likely to be found. The results for the three experiments are summarised in Figures 8 and 9 for the average time over 50 iterations of the algorithm, and Figures 10 and 11 for the number of individuals created for each part of the algorithm. The key of the results figures means windSensors temperatureSensors baseStations (e.g. 10 10 4 means 10 Wind Sensors, 10 Temperature Sensors, and 4 Base Stations were used in the experiment).
Fig. 8. Average execution times of experiments 1, 2 and 3 for compositions satisfying the goal.
This experiment looks at the performance of the composition system as more sensors are added to a fixed region. As the density increases, it is expected that the likelihood of compositions satisfying the goal is higher, but more sensors may need to be checked in a targeted region. This experiment tests the composition system as the breadth of the composition trees increases.
Results The results show an increase in the execution time of all parts of the algorithm as more sensors are added. The Initialisation Phase dominates Fig. 10. Number of Individuals generated in the ontology for experiments 1, 2 and 3 for compositions satisfying the goal. The rightmost set of results is not the number of individuals.
Fig. 11. Number of Individuals generated in the ontology for experiments 1, 2 and 3 for compositions not satisfying the goal. The rightmost set of results is not the number of individuals. No individuals are created for an unsatisfied goal. the execution time and the reasoner takes the least time. Comparing the two cases for the goal, the composer takes less time in the unsatisfiable case, the Initialisation Phase remains about the same, and the reasoner takes less time. For the unsatisfiable case, the composer does not add more individuals to the ontology as there are no suitable sensors, hence the reasoner takes less time.
The increasing number of individuals with the increasing number of
sensors is observed. The size of the ontology and the number of individuals
naturally affects the reasoner’s capabilities to check for consistency and answering
queries [
The small number of individuals for composite sensors is ideal as more compositions are likely to be added, but initialisation only needs to be performed once (or partially when new sensors are added). This experiment shows composite sensors do not make the ontology unreasonable complex and that most of the time taken is in the initialisation and composition phase. The number of composite sensors increases, suggesting more sensors evaluated in the composition phase and the likelihood of finding a satisfying composite sensor decreases. 5.3
This experiment looks at the performance of the system as more sensors are added to a region of space that increases in ratio with the number of sensors. Since the density remains constant, it is expected that the likelihood of compositions satisfying the goals in any region remains the same and similarly with the running time. This experiment tests the filtering capabilities of the composition system to remove irrelevant sensors.
Results The execution times have no significant difference to Experiment 1 for both cases (paired 2-tailed student’s t-test value is between 0.45 and 1.00 for each time component). The composition part has some differences in the execution time, performing slightly better with more sensors (i.e. less time) and slightly worst with fewer sensors (i.e. more time). The number of individuals created by composition reflects the changes in execution time.
The area where the sensors are placed seems to have an effect on the number of sensors created. Smaller areas means sensors are likely to be closer, so there more sensors for the composition system to check. The time taken by the composition system seems to be dependent on the number of sensors instead of the sensors required to measure the goal sensing region. Another possibility is the goal region covers an area common to many sensors. The density of sensors does not seem to have a significant difference in the performance of the composition system. 5.4
This experiment investigates the complexity of the compositions by adding a simple message forwarding process to each sensor type. The process has one input and one output, where both has the most general type Stream and unit of measurement measureAny. This process simply takes any input and forwards the message to another sensor (not at the same sensor). This experiment tests the composition system as the sensor network becomes more complex, requiring longer chains of compositions.
Some problems were encountered in this experiment, which revealed limitations of the system for future work. The problems come from the vast number of composite sensors created in the composition algorithm. An upper limit on the time taken is placed at 400 composite sensors. The results presented in this section are only for compositions not satisfying the goal and when the algorithm terminates because of a large set of composite sensors.
Results The number of compositions is low compared to the number of composite sensors created as seen in Table 1. When the algorithm creates more than 400 composite sensors the average size of the set of sensors is shown in the last column of Table 1, which can be very large. The large set of sensors in the last column for a small starting set of atomic sensors results from the composition system check for many more possible composite sensors, resulting in complex compositions. The experiments show the feasibility of applying semantics to sensors and of automatically creating compositions. From the execution times for experiments 1 and 2, the reasoner performed efficiently and showing only small increases with more sensors and many more individuals in the ontology. Comparing to the initialisation phase, in the composition phase few number of individuals for composite sensors are created when the goal can be satisfied. Experiment 3 shows a wide variation of execution times over 50 iterations, which is only partially seen in the figures. However, the number of composite sensors and individuals are also low, but higher than experiments 1 and 2 in some cases.
Table 1 shows the crippling number of sensors created in the last column for experiment 3. The creation of long compositions required many sensors to be created, which vary greatly on each iteration, similarly with the time taken in the composition phase. However, checking consistency and reasoning on the ontology are performed quickly, similar to experiments 1 and 2. So, the semantic overhead in searching for composition do not contribute greatly to the overall execution times.
Experiments 1 and 2 have very similar execution times and resulting number of individuals. This suggests the number of surrounding sensors and number of sensors does not affect the composition system significantly. The composition system stops early when the goals cannot be satisfied, suggesting the system can quickly determine unsatisfiable goals. The execution time of the initialisation phase and checking of consistency by the reasoner are similar across all three experiments.
The number of individuals created in the initialisation phase is numerous because the quantitative values for the regions and intervals. The values do not affect the reasoner, but can be removed to reduce the size of the ontology. Despite the large number of composite sensors created in experiment 3, the number of individuals representing compositions remains low (in the case of satisfied goal). This suggests the composite sensors have many common constituent sensors.
These experiments show that the composition system works well on wide composition trees, but struggles on deep composition trees. The number of individuals in the ontology do not affect the reasoning time. Comparatively few individuals are created in the composition phase than the initialisation phase. Semantic sensor composition is achieved with a sensor ontology and a composition system, but some weaknesses are apparent that limits the application of composition system to large and complex sensor networks. 7
This paper describes two contributions: a sensor ontology and a composition system. The ontology covers semantics of the sensor types, process semantics, spatial semantics, temporal semantics, sensor semantics, and goal semantics. The spatial and temporal semantics are a novel addition, providing semantic diversity and capabilities beyond OWL. The composition system provides automated and goal directed composition of sensors using the ontology. Experiments show the feasibility of the sensor ontology in describing sensors and sensor compositions, and the limitations of the composition approach. From the experiments, future work includes looking at pruning methods and AI search approaches that exploits semantic information. The sensor ontology and composition system provide semantic interoperability, semantic diversity, and automation, which allows users to query, interact, and extract knowledge directly from any sensor network. Acknowledgements This work was conducted as part of an Honours research project with a scholarship from the CSIRO ICT Centre.