IMechatronic Systems, Applications Edited by Annalisa Part 3 docx

Due to this fact the neural network is used for force vector identification based on measured deformations of sensor body.. a contact sensor an ing principle sensor is based strain gaug

Zhou, Y., Liu, W., & Huang, P (2007) Laser-activated RFID-based Indoor Localization

System for Mobile Robots, Proceedings of IEEE International Conference on Robotics

and Automation, pp 4600 – 4605, ISBN: 1-4244-0601-3, Roma, Italy, April 2007

Petr Krejci

X

Contact sensor for robotic application

Petr Krejci

Brno University of Technology

Czech Republic

Abstract

The chapter deals with design of contact force vector sensor The information about

interaction between robotic parts and surroundings is necessary for intelligent control of

robot behavior The simplest example of such interaction is mechanical contact between

working part of robot and surroundings Than the knowledge of contact characteristic is

important for robot control This mechanical contact could be described by vector of contact

force which includes information about force magnitude as well as information about

orientation and contact point The information about contact force vector will allow to

predict the geometry of object which is in the contact with robots parts and modify robots

behaviour This kind of sensor can be used for instance for control of robotic hand gripping

force as well as for detection of collision between robot and surrounding

1 Introduction

The design of contact force sensor was published by Schwarzinger, 1992 This design

requires application of 24 strain gauges on active part of sensor The quantity of strain

gauges is sufficient for analytical determination of contact force vector

Demand on small size of sensor for a lot of robotic applications (Grepl, R., Bezdicek, M.,

Chmelicek, J., Svehlak, M., 2004) disable application of a large number of strain gauges

Quantity of applied strain gauges and their size is limiting factor for using such design in

our applications

Our design of contact sensor supposes to use only three strain gauges on active part of

sensor However three strain gauges are not enough for the analytical expression of contact

force vector Due to this fact the neural network is used for force vector identification based

on measured deformations of sensor body The application of three strain gauges and new

design will reduce size of sensor but requires a lot of numerical simulations for correct and

accurate sensor behaviour

The main advantage of using neural network is in low computational requirements for

vector determination It means fast response of sensor to contact load The neural network is

able to process measured data faster than nonlinear equations for force vector expression in

analytical way

The other advantage of our design is in reduced requirements for strain measurement by

strain gauges Generally, the Wheatstone bridge has to be used for strain measurement

3

com

ele

Fig

2.

Ba

loc

ide

of

for

sen

nu

wo

3.

Th

sen

he

cor

Th

FE

2

thi

Hoffman, 1989) fo

mplexity of elect

ectronics into sens

g 1 Geometry of

Sensor worki

asic principle of

cations by three s

entified by neura

neural network

rce vectors corre

nsor was used fo

umerical simulati

ork of neural netw

FE model of s

he FE model of se

nsor flange The

ad It is necessa

rrect work of sen

hese positions cor

E model the force

The total strains

is numerical simu

or each strain g trical measuring sor body

a) contact sensor an

ing principle

sensor is based strain gauges on

al network Huge The training mat esponding to sen

or creating of trai ions with varied work

sensor

ensor (see Fig 1b) contact force is

ry to hold geom nsor Deformation rrespond to posit

of 20N was appl

in z-direction ar ulation

gauge It means unit This reduc

nd FE model of se

on measuring o Based on these matrix of trainin trix contains pair nsor body defor ining matrix It i magnitude and

) consists among simulated as app metry and dimen

n of sensor durin tions of strain ga ied on sensor hea

e S1= 2.9um/m,

that our design ction will allow

ensor

f deformation of deformations the

ng pairs is necess

s of deformations rmation Finite e

is necessary to m position of con

others of sensor plied loads in sel nsions of model w

ng load is calcula auges on real sen

ad Results of the S2=12.6um/m an

n significantly re

us to build in c

b)

f sensor body in

e contact force ve ary for proper fu

s in three location element (FE) mo make a large num ntact force for pr

head, sensor bod lected nodes on with real structu ated in three pos nsor For verificat

e verification are o

nd S3= -22.97um/

educes control

n three ector is unction

ns and odel of mber of roperly

dy and sensor ure for sitions

tion of

on Fig

/m for

Fig

4.

Th con con Th pro sug tra

Fig

g 2 Sensor defor

Neural netwo

he architecture of ntains deformati ntains informatio

he training matrix oject This amou ggested sensor F aining matrix

g 3 Architecture

rmation (applied

ork

f artificial neural ions of sensor b

on about contact f

x of 1000 training unt of training p For better accura

of used neural ne

load of 20 N) (un

l network (ANN) body measured force and position

g pairs was used pairs was used j acy of ANN as w

etwork

nit of results are m

) is shown on fig

by strain gauge

n of contact force for training of A just for verificati well as sensor can

m/m)

g 3 The input

es The output

on sensor head ANN in first step ion of functiona

n by use much g

vector vector

of this ality of greater

com

ele

Fig

2.

Ba

loc

ide

of

for

sen

nu

wo

3.

Th

sen

he

cor

Th

FE

2

thi

Hoffman, 1989) fo

mplexity of elect

ectronics into sens

g 1 Geometry of

Sensor worki

asic principle of

cations by three s

entified by neura

neural network

rce vectors corre

nsor was used fo

umerical simulati

ork of neural netw

FE model of s

he FE model of se

nsor flange The

ad It is necessa

rrect work of sen

hese positions cor

E model the force

The total strains

is numerical simu

or each strain g trical measuring

sor body

a) contact sensor an

ing principle

sensor is based strain gauges on

al network Huge The training mat esponding to sen

or creating of trai ions with varied

work

sensor

ensor (see Fig 1b) contact force is

ry to hold geom nsor Deformation

rrespond to posit

of 20N was appl

in z-direction ar ulation

gauge It means unit This reduc

nd FE model of se

on measuring o Based on these matrix of trainin

trix contains pair nsor body defor ining matrix It i magnitude and

) consists among simulated as app metry and dimen

n of sensor durin tions of strain ga

ied on sensor hea

e S1= 2.9um/m,

that our design ction will allow

ensor

f deformation of deformations the

ng pairs is necess

s of deformations rmation Finite e

is necessary to m position of con

others of sensor plied loads in sel nsions of model w

ng load is calcula auges on real sen

ad Results of the S2=12.6um/m an

n significantly re

us to build in c

b)

f sensor body in

e contact force ve ary for proper fu

s in three location element (FE) mo

make a large num ntact force for pr

head, sensor bod lected nodes on

with real structu ated in three pos nsor For verificat

e verification are o

nd S3= -22.97um/

educes control

n three ector is unction

ns and odel of mber of roperly

dy and sensor ure for sitions

tion of

on Fig

/m for

Fig

4.

Th con con Th pro sug tra

Fig

g 2 Sensor defor

Neural netwo

he architecture of ntains deformati ntains informatio

he training matrix oject This amou ggested sensor F aining matrix

g 3 Architecture

rmation (applied

ork

f artificial neural ions of sensor b

on about contact f

x of 1000 training unt of training p For better accura

of used neural ne

load of 20 N) (un

l network (ANN) body measured force and position

g pairs was used pairs was used j acy of ANN as w

etwork

nit of results are m

) is shown on fig

by strain gauge

n of contact force for training of A just for verificati well as sensor can

m/m)

g 3 The input

es The output

on sensor head ANN in first step ion of functiona

n by use much g

vector vector

of this ality of greater

Fig

5.

Th

Th

mo

ap

S1=

vec

(co

g 4 Points of mo

g 5 Sensor coord

Verification o

he force of 20N a

he position of the

odel load used fo

plied force are 6

=10.81um/m, S

ctor of trained A

oordinate system

del load

dinate system

of ANN function

applied on senso

e force was differ

or training matrix 1mm, -3.31mm, 1 S2=-23.57um/m a ANN The resul

of sensor is show

nality

r head was used rent than forces creation are show 15.95mm respecti and S3=10.04um

lt of contact for

wn on .)

d for verification applied for train

wn on Fig 4) The ively The total st /m These strain

ce vector determ

of ANN functio ning of ANN (po

e x, y and z direc trains in z-directi

ns were used as mination is in T

onality

ints of tion of ion are input Table 1

[%]

Point of FE model load determined by Simulated by ANN Position of contact force

Table 1 Result of verification

6 Experimental verification of sensor functionality

The sensor functionality was verified by experimental simulation in laboratory of Mechatronics During experiment the loads of sensor was applied in several positions of sensor head Gauging fixture (Fig 6) was used for sensor positioning Load was applied by materials testing machine Zwick Z 020-TND (Fig 7, Fig 8) where the real load force was measured The deformation of sensor body was measured by strain gauges through HBM Spider 8 unit which is among other things designed for measuring of deformation by strain gauges

Fig 6 Gauging fixture Measured deformations was transferred to information about contact force position and magnitude by neural network implemented in Matlab software The results of experimental verification for selected points are shown in Table 2 for four positions of load force and shows really good accuracy of designed sensor

Fig

5.

Th

Th

mo

ap

S1=

vec

(co

g 4 Points of mo

g 5 Sensor coord

Verification o

he force of 20N a

he position of the

odel load used fo

plied force are 6

=10.81um/m, S

ctor of trained A

oordinate system

del load

dinate system

of ANN function

applied on senso

e force was differ

or training matrix 1mm, -3.31mm, 1 S2=-23.57um/m a ANN The resul

of sensor is show

nality

r head was used rent than forces

creation are show 15.95mm respecti

and S3=10.04um

lt of contact for

wn on .)

d for verification applied for train

wn on Fig 4) The ively The total st

/m These strain

ce vector determ

of ANN functio ning of ANN (po

e x, y and z direc trains in z-directi

ns were used as mination is in T

onality

ints of tion of ion are input Table 1

[%]

Point of FE model load determined by Simulated by ANN Position of contact force

Table 1 Result of verification

6 Experimental verification of sensor functionality

The sensor functionality was verified by experimental simulation in laboratory of Mechatronics During experiment the loads of sensor was applied in several positions of sensor head Gauging fixture (Fig 6) was used for sensor positioning Load was applied by materials testing machine Zwick Z 020-TND (Fig 7, Fig 8) where the real load force was measured The deformation of sensor body was measured by strain gauges through HBM Spider 8 unit which is among other things designed for measuring of deformation by strain gauges

Fig 6 Gauging fixture Measured deformations was transferred to information about contact force position and magnitude by neural network implemented in Matlab software The results of experimental verification for selected points are shown in Table 2 for four positions of load force and shows really good accuracy of designed sensor

Fig

Load p

1

2

3

4 ble 2 Results of v

g 7 Testing mach

point Direction [mm]

x

y

z

x

y

z

x

y

z

x

y

z verification

hine Zwick Z 020

Contact force co Position of force during experiment 0.8 -2.5 20.0 -1.0 -1.9 22.0 -1.0 0.1 22.0 2.0 -5.0 16.0

-TND

oordinates Simulated by ANN 0.81 -2.75 20.62 -0.97 -2.03 24.1 -1.02 0.11 22.91 2.07 -4.52 16.38

Accuracy [%]

98.8 91.0 96.8 97.0 93.6 90.9 98.0 90.9 96.0 96.6 90.4

7.

Th tes inc Th top

"sh top cri red Th loc op Th dir wh red Th

g 8 Loaded senso

Optimization

he low sensitivity sting Therefore creasing of sensiti

he Finite elemen pological optimiz hape" optimizati pological optimiz iterion takes on a duction)

he sensor body i cated under sup ptimization

his optimization rections The first hile second step w duction of 80% in

he boundary cond

or during experim

of sensor des

of sensor was ob the topological ivity for loads ap

nt model of sen zation procedure

on, sometimes r zation is to find t maximum/mini

is the volume w pposed location was done for

t step of optimiz was done for rad

n Fig 10 The figu ditions used durin

mental verificatio

ign

bserved in axial d optimization o pplied to sensor in sor in finite ele

e Topological op referred to as "

he best use of ma imum value subje which was subjec

ns of strain gau two load steps zation procedure dial load The resu ure shows distrib

ng optimization p

on of functionality

direction during

of sensor geome

n axial direction

ement software ptimization (ref

layout" optimiza aterial for a body ect to given const cted to optimiza uges was exclu

- for load forces was done for ax ult of optimizatio bution of pseudod procedure are sho

y

simulations and etry was require ANSYS was use ANSYS) is a fo ation The purp

y such that an ob traints (such as v ation process Vo ded from proce

s oriented in dif ial load of senso

on is shown for v density in sensor own in Fig 9

sensor

ed for

ed for orm of ose of bjective volume olumes ess of fferent

r head volume

r body

Fig

Load p

1

2

3

4 ble 2 Results of v

g 7 Testing mach

point Direction [mm]

x

y

z

x

y

z

x

y

z

x

y

z verification

hine Zwick Z 020

Contact force co Position of force during

experiment 0.8

-2.5 20.0 -1.0 -1.9 22.0 -1.0 0.1

22.0 2.0

-5.0 16.0

-TND

oordinates Simulated by ANN

0.81 -2.75 20.62 -0.97 -2.03 24.1 -1.02 0.11 22.91

2.07 -4.52 16.38

Accuracy [%]

98.8 91.0 96.8 97.0 93.6 90.9 98.0 90.9 96.0 96.6 90.4

7.

Th tes inc Th top

"sh top cri red Th loc op Th dir wh red Th

g 8 Loaded senso

Optimization

he low sensitivity sting Therefore creasing of sensiti

he Finite elemen pological optimiz hape" optimizati pological optimiz iterion takes on a duction)

he sensor body i cated under sup ptimization

his optimization rections The first hile second step w duction of 80% in

he boundary cond

or during experim

of sensor des

of sensor was ob the topological ivity for loads ap

nt model of sen zation procedure

on, sometimes r zation is to find t maximum/mini

is the volume w pposed location was done for

t step of optimiz was done for rad

n Fig 10 The figu ditions used durin

mental verificatio

ign

bserved in axial d optimization o pplied to sensor in sor in finite ele

e Topological op referred to as "

he best use of ma imum value subje which was subjec

ns of strain gau two load steps zation procedure dial load The resu ure shows distrib

ng optimization p

on of functionality

direction during

of sensor geome

n axial direction

ement software ptimization (ref

layout" optimiza aterial for a body ect to given const cted to optimiza uges was exclu

- for load forces was done for ax ult of optimizatio bution of pseudod procedure are sho

y

simulations and etry was require ANSYS was use ANSYS) is a fo ation The purp

y such that an ob traints (such as v ation process Vo ded from proce

s oriented in dif ial load of senso

on is shown for v density in sensor own in Fig 9

sensor

ed for

ed for orm of ose of bjective volume olumes ess of fferent

r head volume

r body

Fig 9 Loads of sensor used in topological optimization procedure

Fig 10 Results of optimization (pseudodensity - red color means that volume will be

included in final design, blue color mean that volume will be excluded from final design)

Optimized shape of sensor body need to by simplified by reason of good manufacturing

Due to this fact few shapes of cutting was designed with consideration of optimized shape

(Fig 10) and machining Based on results of structural analysis rectangular shape of cutting

with 1 mm hole (Fig 11b) ) produces the best results in terms of sensitivity This shape is

also suitable for simple machining Fig 12 shows prototype of optimized and

non-optimized sensor which is made from aluminium alloy

Fig

Fig

7.1

Str lin Fig de

g 11 Optimized s

g 12 Optimized a

1 Structural ana

ructural analysis near behaviour of

g 13 for load forc formation of sens

a) sensors with diffe

and non-optimiz

alysis of sensor

of optimized sen structure occurs

ce of 140N This v sor body can occu

b) erent shapes of cu

ed sensor prototy

prototype

sor was done in o Results of this si value defines upp urs

uttings

ype

order to find out imulation are sho per bound of sens

c)

load limits where own in

sor limits where

e the plastic