Archiving pushed Inferences from Sensor Data Streams
rg Brunsmann
Faculty of Mathematics and Computer Science, Distance University of Hagen
D-58097 Hagen, Germany
Abstract. Although pervasively deployed, sensors are currently neither highly
interconnected nor very intelligent, since they do not know each other and pro-
duce only raw data streams. This lack of interoperability and high-level reason-
ing capabilities are major obstacles for exploiting the full potential of sensor da-
ta streams. Since interoperability and reasoning processes require a common
understanding, RDF based linked sensor data is used in the semantic sensor web
to articulate the meaning of sensor data. This paper shows how to derive higher
levels of streamed sensor data understanding by constructing reasoning know-
ledge with SPARQL. In addition, it is demonstrated how to push these infe-
rences to interested clients in different application domains like social media
streaming, weather observation and intelligent product lifecycle maintenance.
Finally, the paper describes how real-time pushing of inferences enables prove-
nance tracking and how archiving of inferred events could support further deci-
sion making processes.
1 Introduction
Since centuries everyone get used to sensors in the form of classical thermometers
that measure temperatures. Nowadays, more sophisticated sensors monitor all kinds
of physical world phenomena and they are pervasively deployed, e.g. in cars as acce-
leration detector, at the door as movement alarm, in the office as smoke detector and
in watches as health monitoring sensors. Some of these sensors are enhanced with
spatial, temporal and concept annotations to enable high-level reasoning and these
sensors are integrated into the semantic sensor web as they are accessible via the
internet. Since sensor networks provide a digital interface to real world phenomena,
they are part of the internet of things [7] paradigm. Unfortunately, these sensors are
not very intelligent by themselves and are not interconnected, because they produce
only raw time series data and the process of combining multiple sensors to identify
high-level inferences is not yet very common. In addition, the delivery of sensor infe-
rence streams is based on the pull model where it is necessary to actively monitor the
streams. This paper describes how to lift raw sensor data to more expressive layers
that can be understood both by machines and humans and it describes how to push
relevant inferences to interested parties for further processing.
Therefore, this paper combines the following different technologies to form a sen-
sor data processing network that creates, pushes and archives inferences from sensor
data streams:
Brunsmann J..
Archiving pushed Inferences from Sensor Data Streams.
DOI: 10.5220/0003116000380046
In Proceedings of the International Workshop on Semantic Sensor Web (SSW-2010), pages 38-46
ISBN: 978-989-8425-33-1
2010 SCITEPRESS (Science and Technology Publications, Lda.)
1. Executing continuous SPARQL queries [3] over streamed sensor data.
2. Deriving high-level inferences [20] from raw stream data with SPARQL
3. Pushing [13] these inferences to interested clients with sparqlPuSH.
4. Create an archive that accepts and manages stream inferences.
5. Preserving RDF based inferences during ontology evolution.
The remainder of the paper is structured as follows. The next section provides a
characterization of sensor networks, semantic sensor data streams, data stream man-
agement systems and describes the sensor data pyramid containing different layers of
data relationships. Section 3 presents a system architecture that enables generating,
pushing and archiving annotated sensor data along the different levels of the sensor
data pyramid and applies this architecture to different application domains. The last
section concludes with a description of future work.
2 Sensors
A sensor is a device with small memory, reduced computing capacities and limited
power supply that measures an observable physical quantity or environmental condi-
tion like temperature (thermal), sound (acoustic), vibration (seismic), visual light,
infrared light, pressure, magnetism, radar, pollutants or motion (e.g. traffic). A sensor
monitors a feature of interest (e.g. a product or lake) and reports observations as value
of some property representing a quality of a feature obtained using a specified proce-
dure. Sensors vary in terms of the sampling frequency and the amount of data they
produce. For example, a thermometer may output a single byte and may be sampled
every few seconds whereas a video camera that assists with reversing may output
several megabytes per second. A sensor data stream is a continuous and timely or-
dered sequence of observations. A sensor network consists of multiple spatially dis-
tributed autonomous sensor devices. A sensor network allows that many sensors coo-
perate and interact with each other. In the semantic sensor web, measured sensor data
is described by knowledge representations languages like the Resource Description
Framework (RDF) [15]. In order to express such semantics, concepts from widely-
used vocabularies and data from the Linked Open Data [5] cloud is used to allow
machine interpretation [2]. Special Data Stream Management Systems (DSMS) con-
nect to and monitor one or more stream sources. A DSMS will not archive permanent-
ly generated stream data. Such streamed data will rather be processed directly by
executing continuous queries which are used to produce alarms if some events occur.
In addition, a DSMS is used to build aggregated and semantic annotated data from
raw stream data. To do so, a DSMS must lift raw sensor data to higher layers of un-
derstanding to support both human and machine interpretation. The sensor data py-
ramid is a model that illustrates these different layers.
2.1 Sensor Data Pyramid
The sensor data pyramid [17] that is depicted in figure 1 is a concept for modeling
relationships between sensor data, information, knowledge and wisdom.
Fig. 1. The sensor data pyramid includes four different layers including raw sensor data, feature
metadata and ontology metadata.
In the pyramid, data and associated concepts build information. If meaning is at-
tached to information (e.g. semantic annotation) then knowledge is created. Finally, a
human being can have insight into knowledge to gain wisdom. Data at the lower le-
vels will arrive with higher frequency than on higher levels and higher layers have
more expressiveness than lower layers.
2.2 Streams of Stream Inferences
Raw stream data that is lifted to higher expressive layers (e.g. via ontology annota-
tions) build new streams (e.g. RDF streams) that serves as input for other consumers.
Fig. 2. Raw stream data is lifted to more expressive layers which is then used as input for other
stream consumers.
For example see figure 2: rule 1 reads facts (triples) from two different sensors and
a database (triple store) to create new facts. These newly created facts (fact 4) are the
input for another rule 2 that itself leads to an action.
2.3 Stream Inference Pushing and Archiving Architecture
The sparqlPusH approach [13] broadcasts changes made to RDF data in real-time to
clients as Atom or RSS feeds. The system architecture of sparqlPuSH consists of
publishers, hubs and subscribers that can be combined to form a network of multiple
actors. The publisher is the source of updates and the hub pushes the newly published
data to multiple interested clients that have previously been registered. This architec-
ture can be extended with regard to the data (inferences rather than updates) that is
broadcasted. In such a stream inference pushing and archiving architecture the
workflow is executed in two different steps:
Registration. First, the client registers the inference query at the endpoint of the pub-
lisher. The publisher redirects the request to the hub and the client registers for a spe-
cific feed published by the hub.
Broadcasting. Second, the publisher executes continuous queries over streams and
the publisher executes the sensor data pyramid loop (see below). If registered infe-
rence queries generate new results, a notification is sent to the hub. The hub itself
broadcasts these inferences to registered feed clients.
Fig. 3. The broadcasting of inference streams to subscribers after the subscribers has been
registered at the hub.
The DSMS handles large amounts of streaming data and executes an endless sen-
sor data pyramid processing loop (figure 3). While executing this loop, the DSMS
pushes new inferences to well known hubs which broadcast these inferences to inter-
ested clients. Since the DSMS does not archive all data, one of the clients includes a
long-term archive that archives and preserves important facts.
Sensor Data Pyramid Loop. The DSMS execute a sensor data pyramid loop that
includes reading, analyzing, annotating, reasoning and pushing of stream data. This
loop contains the execution of continuous queries like C-SPARQL [3], the generation
of inferences and the pushing of newly annotated stream data. In other words, each
time new stream data arrives at the DSMS, rules (inference queries) are executed.
The SPARQL CONSTRUCT pattern allows expressing such rules because the
WHERE clause is the condition that must be fulfilled by current available facts and
the CONSTRUCT pattern allows to express the reasoned facts. These reasoned facts
are then being used as stream input for other rules. The DSMS will push the infe-
rences to a hub which broadcasts the inferences to clients. One potential client is a
long-term archive that archives specific observation events.
The endless sensor data pyramid loop is executed in the DSMS. The loop reads mul-
tiple sensor streams, mine and correlate sensor data, annotate one or more streams
with semantics, pushes these streams, reason over newly generated facts and pushes
potential new generated inferences
while true do
Inferences for Decision Making Processes. The archive client collects all inferences
and their provenance. Once all this data has been collected, it can be used by a human
to make decisions which can be assisted by machine readable rules.
Preserving Stream Inferences. The inference stream consists of RDF statements that
use a specific version of an ontology. These vocabularies evolve over time and arc-
hived inferences are only consistent with a specific ontology version. The archive
must be able to handle this ontology evolution (threat of semantic obsolescence) by
migrating inference instances via ontology mapping. In addition, an inference archive
must also have data integration facilities to cope with the heterogeneity of ontologies
that describe sensors and sensor data streams.
3 Example Use Cases
In this section several example use cases from different application domains (social
media, weather observation and engineering) are described in more detail.
3.1 Social Media Data
Although semantic streams originating from social media do not monitor physical
phenomena, they can also be used to infer new knowledge. Microblogging sites like
Twitter generate streams of short messages. Unfortunately, these streams are not an-
notated with semantics. Therefore, services described in [18] and [10] and [21] or the
ShreddedTweet [19] service can be used to enhance twitter messages to include RDF
triples for “hashtags”. Those RDF triples can then be used to infer knowledge (inter-
ests, activity profile and social network analysis) about message posters.
3.2 Weather Observations
One example for the combination of multiple sensors in weather observation is a
temperature sensor and a wind speed sensor. Both sensors can be used to calculate the
wind chill factor. Another example which is described below makes use of a rain and
temperature sensor to derive higher-level knowledge about occurrences of potential
freezing rain. In this scenario, the following RDF data (in N3 representation) is
created after reading raw stream data. Here, a sensor delivered a concrete value for
rain of 0.2 cm in the last hour and a current temperature of minus 1.1 degrees.
After reading the data streams, the following data is attached with semantic information and
pushed as new stream (to save space, the prefixes have been left out).
ex:obsveration a obs:WheatherObservation ;
obs:degreesCelsius "-1.1"^^xsd:double ;
obs:rain "0.2"^^xsd:double ;
obs:time "2010-06-20 22:30:03.0" .
These facts can be used to infer new facts. That means, after producing the above
semantic stream, the following SPARQL CONSTRUCT query can be used to infer
new facts.
Rule to infer freezing rain which occurs if the temperature is below zero and if rain has fallen.
The CONSTRUCT pattern creates new facts if the condition in the WHERE clause is fulfilled
CONSTRUCT { obs:freezingrain prov:where geo:hagen .
obs:freezingrain prov:why ex:reason .
obs:freezingrain prov:when ?time .
ex:reason obs:degreesCelsius ?temp .
ex:reason obs:rain ?rain . }
WHERE { ?s obs:degreesCelsius ?temp .
?s obs:rain ?rain .
?s obs:time ?time .
FILTER ( ?rain > 0 && ?temp < 0) }
If the condition in the WHERE clause applies, new facts will be created that in-
clude the inference provenance. According to [9] the following metadata are of inter-
est while creating observations: Where is the geographical location of the captured
data? When did the observation begin and when did it end? What was measured?
How the observation was executed (instruments)? Who is the originator of the data?
To describe this provenance, the W7 provenance model [14] can be used to attach
provenance to specific events. The following code shows the inferred facts.
A freezing rain inference attached with provenance information produced from a
obs:degreesCelsius "-1.2"^^xsd:double ;
obs:rain "1.2"^^xsd:double .
prov:when "2010-06-20 22:30:03.0" ;
prov:where geo:hagen ;
prov:why ex:reason .
3.3 Sensors in Product Lifecycle Management
Product lifecycle management (PLM) systems allow the management of product
related information for the complete lifecycle of a product from its conception
through design and manufacturing to service and disposal. During product operation
and service valuable data can be obtained from products by using sensors [16] [1].
Sensors enable to continuously monitor the conditions of a product. If such data is
broadcasted, it can be collected and shared as PLM data among lifecycle actors. Data
obtained from sensors during product operation can be used to detect needs for main-
tenance, service or repair in advance. After broadcasting such conditions, spare parts
can be ordered and service personnel can be scheduled. In addition, sensor data help
to proactively identify and report conditions before a failure occurs. Other functional
aspects of collecting product data from sensors include:
Product Tracing. Products may be equipped with low cost radio frequency identifi-
cation (RFID) and wireless communication that enable automated tracking and trac-
ing the origin, location, movements, physical properties, environmental conditions
and usage history of a product. The tracing of products help to fight counterfeit prod-
Product Diagnostics. Modern vehicles contain multiples sensors that monitor various
aspects of the engine’s operation, such as thermometers or fuel flow rate which are
not only interesting for the driver but also for the service company. Such data can be
broadcasted in real-time or collected in a memory to be passed later to a remote node.
In addition, product diagnostics data improve the way that products are recycled when
they arrive to their end-of-life.
Product Manufacturing Processes. Sensor data can help to improve existing busi-
ness processes and help to adapt processes in real time. By giving real-time product
status information, manufacturing procedures can be improved. For example, the car
sensor data can be broadcasted to the vehicle manufacturer. If the vehicle manufactur-
er could collect such real-use information from a big amount of cars, maintenance
scheduling can be planned more efficiently.
Another notion of sensor enriched products is intelligent products [11] that collect
usage information and react on it proactively. Intelligent products can be divided into
three categories: an intelligent product should at least be able to manage its own in-
formation given by sensors, RFID-readers (information handling aspect). A more
intelligent product is a product which can report (problem notification aspect) when
there is a problem (e.g. temperature is too high). Finally, the most intelligent product
is the product which can completely manage its own life (decision making aspect).
4 Conclusions and Outlook
This paper presented a system architecture that combines different technologies in
order to enable sensor stream processing together with ontology representations and
standard SPARQL inference formalisms to derive high-level understandings of raw
streamed sensor data. It has been described how reasoning is achieved in the seman-
tic sensor web with standard technologies like RDF based linked data and continuous
inference queries based on the SPARQL CONSTRUCT pattern. In addition, this pa-
per described how inferences are pushed to interested clients via sparqlPuSH technol-
ogy. Then, three different example use-cases where described in more detail from
social media, weather observation and engineering.
Future research will be targeted to the analysis of systems with similar capabilities
and the evaluation of possibilities to integrate the described solutions into existing
systems. Further investigations include evaluation of other rule languages like RIF
(Rule Interchange Format) or XChange. [6], N3Logic [4] or ECA-Rules [12]. In addi-
tion, research will be executed how decision rules can be formulated and executed
that help decision making [22]. Further investigations will also include the integration
of ontology mappings into archive systems the preserve ontology based stream infe-
rences. Finally, while several ontologies exists that describe sensors and sensor obser-
vation [8], domain specific sensor knowledge ontologies (e.g. product monitoring) are
missing. For example, a common vocabulary for product lifecycle metadata is re-
quired which is necessary to support exchanging of PLM data during whole product
This paper is supported by the European Union in the 7th Framework within the IP
1. Anke, J., Främling, K.: Distributed Decision Support in a PLM scenario. Proceedings of
Product Data Technology Europe 14th Symposium (2005)
2. Barbieri D. F., Valle E.D.: A Proposal for Publishing Data Streams as Linked Data. Linked
Data on the Web Workshop (2010)
3. Barbieri D. F., Braga D., Ceri S., Della Valle E., Grossniklaus M.: Continuous queries and
real-time analysis of social semantic data with c-sparql. Proceedings of Social Data on the
Web Workshop at the 8th International Semantic Web Conference (2009)
4. Berners-Lee, T., Connolly D., Kagal L., Hendler J., Scharf Y.: N3Logic: A Logical Frame-
work for the World Wide Web. Journal of Theory and Practice of Logic Programming
(TPLP), Special Issue on Logic Programming and the Web (2008)
5. Bizer C., Heath T. Berners-Lee T.: Linked Data - The Story So Far. International Journal
on Semantic Web and Information Systems (2009)
6. Bry, F., Eckert, M., Patranjan, P.-L., Romanenko, I.: Realizing Business Processes with
ECA Rules: Benefits, Challenges, Limits. Proc. Int. Workshop on Principles and Practice of
Semantic Web. LNCS, Springer, Heidelberg (2006)
7. Christin D., Reinhardt A., Mogre P. S., Steinmetz R.: Wireless Sensor Networks and the
Internet of Things: Selected Challenges. Proceedings of the 8th GI/ITG KuVS Fach-
gespräch Drahtlose Sensornetze (2009)
8. Compton, M., Henson, C., Lefort, L., Neuhaus, H.: A survey of the semantic specification
of sensors. Technical report (2009)
9. Hayes, J., O’Conor, E., Cleary, J., Kolar, H., McCarthy, R., Tynan, R., O’Hare, R., Smea-
ton, A., O’Connor, N., Diamond, D.: Views From the Coalface: Chemo-Sensors, Sensor
Networks and the Semantic Sensor Web. International Workshop on the Semantic Sensor
Web (2009)
10. Hepp, M.: HyperTwitter: Collaborative Knowledge Engineering via Twitter Messages.
Technical Report (2010)
11. Meyer, G.G., Främling, K. & Holmström, J.: Intelligent Products: A survey. Computers in
Industry, 60 (2009)
12. Papamarkos G., Poulovassilis A., Wood P.T.: Event-Condition-Action Rules on RDF Me-
tadata in P2P Environments. In Proc. 2nd Workshop on Metadata Management in Grid and
P2P Systems (MMGPS): Models, Services and Architectures, London (2004)
13. Passant, A., Mendes, P. N.: sparqlPuSH: Proactive notification of data updates in RDF
stores using PubSubHubbub. 6th Workshop on Scripting for the Semantic Web (2010)
14. Ram S. Liu J.: Understanding the Semantics of Data Provenance to Support Active Concep-
tual Modeling. Proceedings of the Active Conceptual Modeling of Learning Workshop
15. Rodriguez, A., McGrath, R., Liu, Y., Myers, J.: Semantic Management of Streaming Data.
Proc. Intl. Workshop on Semantic Sensor Networks (2009)
16. Seitz C., Legat C., Neidig J.: Embedding Semantic Product Memories in the Web of
Things. First International Workshop on the Web of Things (2010)
17. Sheth A., Henson C., Sahoo S. S. Semantic Sensor Web. ARC Research Network on Intel-
ligent Sensors, Sensor Networks and Information Processing – ISSNIP, Melbourne (2008)
18. Shinavier J.: Real-time #SemanticWeb in <= 140 chars. Proceedings of the Third Work-
shop on Linked Data on the Web (2010)
19. Shredded Tweet:
20. Stuckenschmidt H., Ceri S., Valle, E.D., van Harmelen F.: Towards Expressive Stream
Reasoning. Proceedings of the Dagstuhl Seminar on Semantic Aspects of Sensor Networks
21. Twarql – Software implementation realizing the Linked Open Social Signals vision:
22. Zangiacomi A., Fornasiero R.: Modelling decision support systems for Middle-of-life in
product lifecycle management. 14th International Conference on Concurrent Enterprising,
Lisbon, Portugal (2008)