Tree-structured knowledge in a distributed intelligent MEMS

application

1)Atsushi Sato, 2)Eugen Dedu, 2)Julien Bourgeois, 1)Runhe Huang1)Hosei University, 2)UFC/FEMTO-ST

Table of Contents

• Introduction– Smart Surface– Main issues

• Theory– Tree-structured knowledge (TSK)– Reconstruction of the object with TSK– Differentiation of the object with TSK

• Analyses• Conclusions and future work

2/30

Smart Surface

MEMS-arrayed manipulation surface– Recognition– Conveyance– Positioning

Air-Flow Pressure

35mm35mm

3/30

Distributed control

Sensor

Processing unit

Actuator

MEMS

MEMS– Sense– Act– Decide– Communicate

4/30

Recognition

• Offline stage– Create database of shapes of models

• Online stage– Reconstruction– Differentiation

?

5/30

Offline stage

0010011111100100

CriteriaMatrixA: 10P: 16S: 8

Database

Model data…

Database is uploaded to every cell

Rotate and translate the object on the Smart Surface

6/30

( the previous approach )

Repeat

Online stage

• Reconstruction phase00000000000000000100000000000000000000000000000000000000

• Differentiation phase– Calculate criteria– Compare with database

failure

00000000000000000110000001000000000000000000000000000000

00000000001000000111000011100000010000000000000000000000

success

7/30

( the previous approach )Repeat

Main issues

• Message size is the same as the Smart Surface -> redundant

• excessively comparison -> there is no trigger

Relative position based representation

8/30

Tree-structured knowledge

N ESW

E’W’N’

Smart Surface

root node

N

W E

S

E’

N’

W’

9/30

( our current approach )

Tree-structured array

1

1

010

000 10 0

000

1 1 0 0 0 0 1 0 0 1 0 0 0 01 1 1 1Smart Surface

10/30

Matrix00000000000000000110000001000000000000000000000000000000

00000000001000000111000011100000010000000000000000000000

10010001000

100110011000100010001000

64 bits

64 bits

11 bits

24 bits

Tree-Structured Array

Matrix VS. TSK

11/30

Reconstruction

① Initialize its array

② Generate and send messages

③ Receive and merge messages

④ Check duplication

Differentiation phase12/30

Repeat

Generate messages

0 0 1 0 0 1 0 0 01 111 0 0 01

1 1 0 0 0 0 1 0 0 1 0 0 0

1 1 0 0 0 0 1 0 0 1 0 0 0

1 1 0 0 0 0 0 13/30

Merge messages(1/3)

Message size : 1 bit

14/30

Merge messages(2/3)

Message size : 4 bits

15/30

Merge messages(3/3)

All leaf values are 0

Go todifferentiation phase

16/30

Duplication check

Smart Surface

1 0 0 0 0

1 0 0 1 1

1 0 0 1 0 0 1 1 0 1 0

1 0 0 1 0 0 1 1 0 0 0

17/30

Differentiation

① Transform its tree to the regular form

• Change the root to the north

• Change the root to the west

② Compare the array with all the shapes in database

Repeat

if (discover the same array) Send the result to the motion controllerelse Restart the online stage

Until the root is most northern and western

18/30

Transformation (1 / 2)

0 1 0 0 1 0 0 0 001 0 01

Change the root to the north cell

19/30

Transformation (2 / 2)

0 01 1 1 0 0 0 0 0 0 01 0

Change the root to the west cell

20/30

Comparison

shape model10001001001000 210001010000 010001110000000 110010001000 210010100100000 210010101000000 1

10010100100000

Compare

Database

0 1 2

Models

L21/30

Performance analyses

• Number of communication iterations

• Communication traffic

• Computation time

• Memory footprint

22/30

The number of communication iterations :

Edge cells need morecommunications

Central cells need fewer communications

Iteration Iteration 0 0 1 1 2 2 3 3 4 4 5 6 7

h h𝑒𝑖𝑔 𝑡+ h𝑤𝑖𝑑𝑡2 +1≤𝑵 ≤h h𝑒𝑖𝑔 𝑡+ h𝑤𝑖𝑑𝑡 −1

8 5

23/30

Communication traffic

Smart Surface 10 x 10

Matrixbits

Tree-structured arraybits

The number ofmessages at a time

The number ofactive cells

× 118The number of

communication iterations

×

24/30

Computation time

• Reconstruction time (TR)• Transformation time (TT)• Comparison time (TC)• Computation time (TA)

𝑇= ∑𝑖=1

𝑁 𝑅𝑒𝑐𝑜𝑛𝑠𝑡𝑟𝑢𝑐𝑡𝑖𝑜𝑛

𝑇𝑅𝑖+ ∑𝑗=1

𝑁 𝑇𝑟𝑎𝑛𝑠𝑓𝑜𝑟𝑚

𝑇 𝑇 𝑗+ ∑𝑘=1

𝑁 𝐶𝑜𝑚𝑝𝑎𝑟𝑖𝑠𝑜𝑛

𝑇 𝐶𝑘

𝑇 ′= ∑𝑖=1

𝑁 ′𝑅𝑒𝑐𝑜𝑛𝑠𝑡𝑟𝑢𝑐𝑡𝑖𝑜𝑛 (𝑇𝑅 ′𝑖+𝑇𝐴𝑖+ ∑𝑗=1

𝑁 ′𝐶𝑜𝑚𝑝𝑎𝑟𝑖𝑠𝑜𝑛

𝑇𝐶 ′ 𝑖𝑗)

Proposed approach

Previous approach

25/30

Memory footprint(1/2)

Previous approach• One shape needs 29 bytes • bytes for matrices• One criterion needs 4 bytes• bytes for criteria• Total Memory for models is :

26/30

Proposed approach• One shape needs: bits• is the number of cells covered by the object• Total Memory for models is :

Memory footprint(2/2)

is 22, bits

𝑴= ∑𝒊=𝟏

𝑵 𝑺𝒉𝒂𝒑𝒆𝒔

𝑴𝑺𝒊

27/30

Simulation of the offline stage(1/2)

31 mm 29 mm

40 mm39 mm

33 mm

13 mm

8 mm

● 8 mm

1.6 mm

ModelsMEMS

Circle : 48 shapesRectangle : 248 shapesH : 428 shapes

Criteria : 58

724 shapes

Every model covers less than 25 cells 28/30

Simulation of the offline stage(2/2)

𝑴=𝟐𝟗×𝑵𝑺𝒉𝒂𝒑𝒆+𝟒×𝑵𝑪𝒓𝒊𝒕𝒆𝒓𝒊𝒂

bits bytes

𝑴= ∑𝒊=𝟏

𝑵 𝑺𝒉𝒂𝒑𝒆𝒔

𝑴𝑺𝒊

Previous approach

Proposed approach bytes

bytes 29/30

fewer The number of shapes many

reduction of the memory footprint

724 shapes are too many to store in every cell

the probability of matchinglow high

Include the shape appearing rarely

Reduce the stored shapes

30/30

Conclusions and future work

• Representing the shapes as tree-structured array reduces their memory footprint and redundant information in messages.

• The number of shapes can be reduced, but it trades off with the probability of the successful differentiation.

• Reduction of the number of shapes to be stored in every cell.

31/30