Measurement confidence in a R EAL S ENSOR N ETWORK Swarm - - PowerPoint PPT Presentation

Measurement confidence in a REAL SENSOR NETWORK

Swarm Intelligence project Authors: Sven Gowal, Maxime Augier

I say fifty, maybe a hundred horses... What do you say Red Eagle?

Abstract

Familiarize one self with TinyOS/nesC Implement some approach to filter raw light

measurements

Demonstrate the importance of collaboration

Background

TinyOS

OS designed for wireless embedded sensor networks.

nesC

Extension to the C programming language designed to

embody the structuring concepts and execution model of TinyOS.

TinyDB

query processing system for extracting information from a

network of TinyOS sensors.

Miscroscopic abstraction level

Although not required

Easier and faster algorithms implementation Easier testing

MoteSim

Our simulator Adapted to our task

MoteSim

Takes as input a network configuration file Outputs the filtered light as well as the raw light

sensed by each mote (node).

Able to plot

Realistic abstraction level

Once we decided on the algorithm according

to the results given by MoteSim.

Used TinyDB to help us get the results from

the MicaZ motes.

Add an attribute (f_light) Able to query with

« SELECT f_light FROM sensors »

Code still totally independent

Algorithm used (outline)

Weighted average of the raw light

measurements by a mote and its neighbors

Average given as input to a low pass filter

(leaky-integrator)

Algorithm used (average)

Averaging formula:

Raw light perceived by the neighbor n at time i Raw light perceived at time i by the mote Number of neighbors the mote has Personal contribution coefficient (k ∊ [0,1]) Averaging over space

Algorithm used (average)

Averaging already produces some kind of

filtering:

10 motes (9 neighbors). Raw light measured is N(10, 0.5). k = 0.1.

Mean error reduced from 0.5 to 0.17

300% improvement Hint: Collaboration is important.

Algorithm used (average)

Algorithm used (filtering)

Using only an average is not sufficient if the number

• f neighbors is limited.

Low pass filtering via a leaky-integrator (for

simplicity).

Leaky-integrator formula:

Filtered data at time i Data received by the filter at time i Leaky-integrator coefficient (l ∊ [0,1]) Averaging over time

Algorithm used (filtering)

Filtering is very efficient even with a lower

number of neighbors:

1 mote (0 neighbors). Raw light measured is N(10, 0.5). l = 0.3.

Mean error reduced from 0.5 to 0.19

267% improvement Hint: Filtering is important.

Algorithm used (filtering)

Results

Algorithm (average + filtering) was run on a

real 8 motes network

Number of neighbors between 0 and 3. High sampling rate to see how noise interfered. Whole day Radio strength reduced to the strict minimum

(approx. 3 meters radius).

Results

Noise (transmission errors mostly) Day Night Global pattern preserved even if the motes were in different light environment

Discussion (introduction of a metric)

Useful to measure the performance The metric has to take into account 2 aspects

Collaboration

• Different motes should follow the same pattern.

Personal data

• A mote should follow the pattern dicted by the light it

measures and not only rely on its neighbors.

Let’s use the statistical correlation

Discussion (metric)

Performance of a single mote: Performance of the whole network:

Raw light Filtered light Collaborative coefficient

Discussion (metric)

The metric yields a real number between 0

and 1.

1 is the better performance one can obtain. Let’s find a reference value on which we can

base our observations.

Will use a = 0.8

Discussion (comparison)

Used MoteSim to compare 3 basic algorithms

Nop: outputs as the filtered light the raw intensity

• f the light (no averaging, no filtering).

Weighted Mean: performs a weighted mean over

the raw intensities perceived by its neighbors and a small historic of the raw values perceived by the mote itself (averaging, little filtering).

Weight & Filter: performs a weighted average like

Weighted Mean but in addition it uses a low pass filter (averaging, filtering).

Discussion (comparison)

Simulated a network of 8 motes

Different neighborhood size (0, 2-3, 7) Raw light is a N(μ,0.5), μ is a U(9, 11). 1000 steps.

Discussion (comparison)

Reference value Filtering happens for bigger neighborhood size. Outperforms the other algorithms

Neighborhood size Performance

Discussion (comparison)

Performance obtained for the real networks is

unexpectedly very good: 0.9797

Very close to 1, probably due to:

Not as much noise. Maybe just lucky.

Conclusion

Filtering occurs with very simple algorithms

such as a simple average as long there are enough neighbors.

For unstable/small neighborhood, filtering

must be used in order to yield better results.

Thanks Any questions?

Discussion (metric - recall)

Correlation With

Covariance Standard deviation