Title: Predictive QoS Routing to Mobile Sinks in Wireless Sensor Networks
1Predictive QoS Routing to Mobile Sinks in
Wireless Sensor Networks
Brano Kusy, HJ. Lee, M. Wicke, N.
Milosavljevic, and L. Guibas Computer
Science Stanford University
2Complex Urban Environments
- People have multiple everyday tasks work,
dining, entertainment, navigation, - Personal mobile devices now offer applications
that help us to sift through the information
3New Mobile Applications
Novel Applications Activity
recognition Automated rating Environmental
sensing Collaborative rating Complex
gestures Photo tourism
4Sensors and Mobile Devices
- iPhone 1.0
- ambient light sensor
- proximity sensor
- accelerometer
- microphone camera
- iPhone 2.0
- assisted GPS
- 2nd proximity sensor
- iPhone 3.0
- GPS
- magnetometer
- 3rd proximity sensor?
5Problem Statement
- Deliver sensor data to a mobile user, from any
sensor node that might have such data
6Talk Overview
- Data Collection in Sensor Networks
- Gradient based data collection
- Problems with mobility
- Formalizing User Motion the Mobility Graph
- Prediction can we estimate where users are going
ahead of time? - Data Delivery To Mobile Users
- Routing optimization using mobility constraints
and prediction
7Information Gradients For Routing
H. Lin, M. Lu, N. Milosavljevic, J. Gao, and L.
Guibas, Composable Information Gradients in
Wireless Sensor Networks, IPSN'08
- Diffuse information about user location away from
the user in the form of a scalar field - gt create implicit gradients towards the user
- Use function f as an information gradient for
routing it implicitly defines a routing tree
7
8Data Collection for a Mobile Sink
- Static sink ? periodically update gradient, due
to link dynamics - Mobile sink ? increase frequency of the gradient
updates - In uniform networks, information gradient changes
little as the sink moves and gradient can be
adapted quickly
8
9Non-Uniform Networks
- Networks with large holes (rivers, lakes, etc)
- Problem is the non-local motion the whole
gradient field changes
Blue nodes the sink can be reached using current
gradient field Red nodes the sink cannot be
reached
9
10Talk Overview
- Data Delivery (Routing) in Sensor Networks
- Gradient based data collection
- Problems with mobility
- Formalizing User Mobility Patterns
- Prediction can we estimate where users are going
ahead of time? - Data Delivery To Mobile Users
- Routing optimization using mobility constraints
and prediction
11Mobility vs. Connectivity Graph
- Mobility graph (MG) data structure that
represents possible user movement patterns within
the network - Mobility graph is different
- from connectivity graph
- People can take shortcuts outside coverage area
- Radio signals can pass through walls
Clark Center _at_Stanford
Network Connectivity Graph
12Mobility Graph Construction
- Observe connectivity patterns of mobile users in
the network
- Signature of path in RSSI space
- The best connected nodes define the mobility graph
8
9
7
6
10
1
- RSSI signature is stored with each edge
2
4
5
3
12
13Mobility Prediction
- Prediction Algorithm
- Training phase observe and learn mobility
patterns of users - Testing phase predict
- Technique
- Match local signatures as the user travels along
mobility graph - Dynamic Time Warping
- Estimate progress, predict the next node and
transition time
13
14Prediction Results
- Experimental evaluation (750 m2 lab space, based
on 25 training, 20 test trajectories, collected
over multiple weeks) - Mean prediction accuracy of the next node and
time to transition
14
15Talk Overview
- Data Delivery (Routing) in Sensor Networks
- Gradient based data collection
- Problems with mobility
- Formalizing User Mobility Patterns
- Prediction can we estimate where users are going
ahead of time? - Data Delivery To Mobile Users
- Routing optimization using mobility constraints
and prediction
16Using Prediction to Improve Routing
- Optimization I
- calculate gradients both for predicted and
current location - faster update, the same cost
- Optimization II
- precompute gradients at non-local transition
edges - no need for iteration
- Optimization III
- new gradient spreads from the root, slowly
propagates to leafs - switch to gradient at predicted transition time
(synchronously)
16
17Optimization II Non-Local Transitions
- Need an automated way to determine at which nodes
the pre-computed gradients are stored - Do we have enough storage for these gradients?
Edge stretch
17
18Routing Results, Simulation
- TOSSIM, 9000m2, 100 nodes, 20 random nodes
transmit - nodes send data to the sink _at_4pps
- Path2 gt86 for all speeds with no optimization
(4 improvement) - Path1 more significant improvement
PredCT full prediction and pre-computation
NoPred no prediction
18
19Summarizing
- Formalization of User Mobility
- the mobility graph
- characterization of non-local transitions
- Short Term Prediction
- based on signal strength measurements
- environments with constrained mobility
- Routing Improvements
- use the mobility graph to improve reliability,
latency - Questions?
20Backup Slides
21Mobility Prediction Assumptions
- Stability of the RSSI signal
- we dont measure user location directly (no
inverse distance law) - we assume that observation sequences will be
similar, if user follows the same trajectory at
different times
21
22Signal Processing
filtering
RSSI (dbm)
RSSI (dbm)
RSSI signal before processing.
RSSI signal after processing.
22
23Non-Local Updates, Continued
- Do we have enough storage for several gradients?
- limit number of saved gradients
- clustering of the saved gradients
23
24Training and Testing Sets
25Study of Prediction Errors
Prediction Error vs Delivery Rate
- Low effect of prediction error on delivery rate
- 9 different trajectories, each repeated 5 times
- training set 5 trajectories testing set 4
trajectories
Training Set Size vs Prediction Error
Training Set Variability vs Prediction Error
25
26Robustness of Information Gradients
Path robustness random walk view
is the probability of hitting the
source before hitting the boundary in a random
walk starting from
Harmonic paths tend to avoid narrow passages and
use non-critical edges
26
27How Can we Utilize Prediction?
- Optimization I use predicted location
- Optimization II store gradients in-network
- Optimization III use predicted transition time
27