This document discusses deriving semantic metadata from raw sensor measurements. It presents a method to classify time series sensor data using piecewise linear approximation of the signals. Segment slopes are used to generate distributions for classification via k-nearest neighbors. Generating observed property and other metadata is also discussed. Evaluation on real sensor datasets shows promising results. Addressing metadata gaps and combining with other techniques for online usage are topics for future work.

1. 5th International Workshop on Semantic Sensor Networks at
ISWC2012
Deriving Semantic Sensor Metadata
from Raw Measurements
Jean-Paul Calbimonte1, Hoyoung Jeung2,
Zhixian Yan3, Oscar Corcho1, Karl Aberer3
1Ontology Engineering Group, Universidad Politécnica de Madrid
2SAP Research, Birsbane
3LSIR, EPFL Ecole Polytechnique Fédérale de Lausanne
jp.calbimonte@upm.es
Date: 14/11/2012

2. Sensor Observations
Semantic metadata
Annotations
Sensor Raw Measurements
Representation
Characterize
Classify
Publish
2

3. Streaming Sensor Data
Data streams
Continuous evaluation
Timestamped tuples
Time windows
Do we know what we are sensing?
3

4. Sensor Web
4

5. The Sensor Web
Universal, web-based access to sensor data
5

6. Semantic Sensor Web
“too much (streaming) data but not enough (tools to
gain and derive) knowledge”*
LinkedSensorData
Sensor data publishing
Linked Data LSM
Sense2Web
Semantic sensor metadata Sensor APIs
ETALIS
Semantic Sensor Network ontology Videk
SwissEx
BOTTARI,
UrbanMatch
AEMET
transporte.
linkeddata.es
SSN +CEP
…many many more working on this
* Sheth et al. 2008, Semantic Sensor Web
6

7. SSN Ontology with other ontologies
tool for modeling our sensor data
~what we are observing
7

8. Sensor Metadata
swissex:Sensor1
rdf:type ssn:Sensor;
ssn:onPlatform swissex:Station1;
ssn:observes cf-property:wind_speed.
swissex:Sensor2
rdf:type ssn:Sensor;
ssn:onPlatform swissex:Station1;
ssn:observes cf-property:air_temperature.
swissex:Station1
:hasGeometry [ rdf:type wgs84:Point;
wgs84:lat "46.8037166";
wgs84:long "9.7780305"].
8

9. Sensor Observations
swissex:WindSpeedObservation1
rdf:type ssn:Observation;
ssn:featureOfInterest cf-feature:wind;
ssn:observedProperty cf-property:wind_speed;
ssn:observationResult [rdf:type ssn:SensorOutput;
ssn:hasValue [qudt:numericValue "6.245"^^xsd:double]];
ssn:observationResultTime [time:inXSDDatatime "2011-10-26T21:32:52"];
ssn:observedBy swissex:Sensor1 ;
9

10. Sensor Observations
Heterogeneity
Reuse this data?
Publish as Linked Sensor Data?
Query with SPARQL-Stream?
10

11. Cosm/Pachube Datastreams
not enough reliable metadata about
the observations
11

12. Looking into the Data
Air Pressure?
Air Temperature?
12

13. Classifying Sensor Data
Unclassified input series
Already classified time series
Representation
Classification
Metadata
13

14. Related Tasks
• Querying time series
• e.g. find a sub-sequence in a time series database
• Measuring time series similarity
• e.g. are these time series the same?
• Time series classification
• e.g. classify heart beat series: normal, murmur, et
14

15. Deriving Semantic Metadata
15

16. Data Representation
Represent the data approximating
with fewer linear segments
Accuracy vs Numerosity
16

17. Constant Approximation
• Use constant segments for a subset of points
4.5
4
3.5
3
2.5
2
1.5
1
0 10 20 30 40 50 60 70 80 90 100
17

18. Piecewise Linear Approximation
4.1
4.05
3.95
4
Reflect data trends
3.9
3.85
3.8
Apply with different resolutions
3.75
3.7
0 1 2 3 4 5 6 7 8 9 10
Applicable for different rates
4.1
4.05
4
3.95
3.9
3.85
3.8
3.75
3.7
3.65
0 1 2 3 4 5 6 7 8 9 10
18

19. Piecewise Linear Approximation
ri hi+1
hi
li
begi endi
Constructing segments
19

20. Linear Approximations
We care about the angles π/2
a π/4
a c b
d
0
a
c
-π/4
d
Divide the angle space in sectors
Distribution of angles in training set
20

21. Slope Distributions
ts1 adacdaaad [5a,0b,1c,3d]
ts2 adabbaaad [5a,2b,0c,2d]
ts3 adccdaaad [4a,0b,2c,3d]
Distance measure
Classification
21

22. Use the representation for Classifying
Linear approximation
Compute distribution of the slopes
K-nearest neighbor classification
Training-Test datasets:
SwissExperiment
AEMET
22

23. Experiments SwissEx
Confusion matrix

24. Experiments AEMET
Confusion matrix

25. Evaluation AEMET
25

26. Evaluation AEMET
cf-property:wind_speed rdf:type dim:VelocityOrSpeed;
rdfs:label "wind speed";
ssn:isPropertyOf cf-feature:wind;
qu:propertyType qu:scalar;
qu:generalQuantityKind qu:speed.
26

27. Metadata Properties
cf-property:air_temperature rdf:type dim:Temperature;
ssn:isPropertyOf cf-feature:air;
qu:propertyType qu:scalar;
qu:generalQuantityKind qu:temperature.
cf-property:soil_temperature rdf:type dim:Temperature;
ssn:isPropertyOf cf-feature:soil;
qu:propertyType qu:scalar;
qu:generalQuantityKind qu:temperature.

28. Querying Metadata
SELECT ?sensor
WHERE {
?sensor a ssn:Sensor ;
ssn:observes cf-property:air_temperature .}
SELECT ?stream ?observedProperty
WHERE {
?sensor a ssn:Sensor ;
ssn:observes ?observedProperty .
?stream ssn:isProducedBy ?sensor .
?observedProperty qu:generalQuantityKind qu:temperature .}
28

29. Evaluation vs SAX
29

30. Evaluation vs SAX
30

31. How much data do we need?
31

32. Conclusions
Classify Sensor Data
• Piecewise Linear Representation
• Segment slope distributions
• kNN classification
Generate Metadata
• Observed properties
• Potentially unknown metadata
Future work
• Combine with tag disambiguation?
• Use pattern mining for online queries
• Other techniques, shapelets, use
other parameters
32

33. …Thanks
Questions, please.
jp.calbimonte@upm.es
33