Event Processing Middleware
for Wireless Sensor
Networks
Selvakennedy Selvadurai
School of Information Technologies University of Sydney
Outline
• Introduction
• System Assumptions
• EPM Architecture
• Group Management and Centre
Localisation Components
• Results
WSN Background
• This technology is evolving rather rapidly
– E.g., the Mica2 sensor has roughly 8x the memory and bandwidth as its predecessor, the Rene sensor for the same power budget
• Some apps: earthquake monitoring, target
tracking and surveillance, structural monitoring
and precision agriculture
• Typically, these apps entail simple sensing and
relaying to the sink(s) for processing and storage
In-Network Processing
• Now, need more intelligence into the network
itself
– inline with the hardware advance
• Moreover, in many control apps, latency
constraints call for more in-network processing
• A number of efforts being undertaken
– These efforts may be classified either as statistical
processing or semantics processing
• Most fall under the former class
– data processing purely involves statistical treatments informed by any temporal/spatial correlation structure in the data
Semantic Processing
• In semantics processing, data processing is informed by the application semantics
• Sensor readings of high significance are tracked reflecting the readings of phenomena of interest
• Sensor nodes will need to collaboratively process their
data
– Enables rapid reactive mechanism towards events through appropriate actuators.
– For instance, some nodes may detect a fire event, and then
collaboratively process and activate the relevant water sprinklers – Would complete the sense-control-actuate loop.
Collaborative Processing
• Phenomena may span an area
– Collaborative processing is needed
• We might be interested to capture its area
coverage, its boundary and its centre
• Various algorithms have been proposed
independently to process specific event properties
• It may be more beneficial to provide event
Event Processing M’ware
(EPM)
• Advantageous to provide them as middleware services • Middleware has often been useful for bridging the gap
between the lower-level components and the application
– Eases development of distributed applications
• A WSN middleware, termed MiLAN [3], allows apps to specify their quality needs and proactively adjusts the network characteristics
• DSWare middleware [4] provides real-time event
detection services that sits between the application layer and the network layer.
• Our proposed middleware could logically complement
EPM
• It masks the collaborative processing required for some services
• As concrete examples, we implement the event boundary, event area and event centre processing services
– The edge detection component will form the core component – Other services uses the above to estimate both event centre and
area.
• This is the first such attempt.
• The middleware also exploits the existence of
clusterheads in certain data gathering protocols, like LEACH, HEED and T-ANT
• Using this approach, significant overhead reduction is achieved while maintaining up to 95% accuracy of event information
System Assumptions
• Each node knows its location
• Nodes know interesting sensor readings
(app-defined)
• Nodes within event are termed
affected
nodes
– Otherwise, unaffected nodes.
• The
edge
of a phenomenon is the imaginary
boundary between these sets
EPM Architecture
Event Subscription Higher-Level Processing Event Notification Event Data Storage Group Management Edge DetectionEPM Components
• Event Notification allows a node to discover its neighbourhood related to certain events
• Event Data Storage: Some information are kept at the nodes and cluster info at CHs
• Group Management: This component controls the collaboration among sensor nodes
• Edge Detection: It tests whether the current node is on the event edge (based on T-Fit [9])
• Higher-Level Processing: event centre localization and area computation
Group Management
• An
overlay tree
is formed among CHs
• When an event transpires, sensors in disjoint
clusters might detect it
– CHs collect from their clusters
• To choose the root, a natural choice would be a
CH close to the event region
• Data are aggregated and finally processed at the
root
• Any subscribed higher-level services is pushed to
the sink from root
Event Centre Localisation
• A x-coord sorted list of edge points becomes available at
the tree root
• For center localization, we need to find the furthest
distance between any pair of edge nodes • Since the diameter of a circle can
be approximated by the longest possible pair-wise distance, the
centre is the midpoint
• An exhaustive search over the edge nodes can be
performed but results in O(n2)
C A B D I II III IV
Event Centre Localisation
• The nodes are organised into
k subgroups
• Two node pointers are used for traversing, and
they are initialized at opposite ends of the edge
node list
• Then, line segments involving nodes in subgroup 1
and
k are checked, followed by nodes in
subgroups 2 and
k-1 and so on.
1x_start x_end
2 k-1 k
C
A B
Simulation Scenario
• Sensor nodes are randomly distributed in a
square
MxM region with M = 100 m
• The data message size is fixed at 30B.
• Node radio range is fixed at 20m
CL Estimation Error
0 5 10 15 20 25 30 35 40 45 50 60 110 160 210 260 310 360Num ber of Nodes
E s ti m a ti o n A ccu ra cy ( % )
CL Estimation Error
0 5 10 15 20 25 30 35 40 45 50 15 20 25 30 35 40 45 50 55 Event Radius (m ) E s ti m a ti o n A c c u ra c y (% )Communications Overhead
0 5000 10000 15000 20000 25000 15 20 25 30 35 40 45 50 55 60 65Event Area Radius (m)
To ta l Co mm un ic at io n O v e r h e a d ( B y t e )
Mean Value with EPM Mean Value without EPM
0 5 10 15 20 25 30 35 40 45 50 15 20 25 30 35 40 45 50 Radio Range (m) Total Energ y Consumption
Mean Value with EPM Mean Value without EPM
Conclusions
• EPM is proposed
• The edge detection module forms the core
component.
– it supports higher-level event processing services.