tree-structured knowledge in a distributed intelligent mems application
DESCRIPTION
Tree-structured knowledge in a distributed intelligent MEMS application. 1) Atsushi Sato, 2) Eugen Dedu , 2) Julien Bourgeois, 1) Runhe Huang 1) Hosei University, 2) UFC/FEMTO-ST. Table of Contents. Introduction Smart Surface Main issues Theory Tree-structured knowledge (TSK) - PowerPoint PPT PresentationTRANSCRIPT
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