Wtu W*.Terakreditasi No. 52lDIKTIIKEP 12002

t

KATAPENGAI{TAR

Assalamu'alaikr.rrn Wr. Wb

Syukur Alhamdulillah dengan Rahmat dan Karunia Allah SWT telah terbit Buletin Utama Teknik

FT-tllSU Vol. 8 No. 3 - September 2004. yang telah terakreditasi, baik menyangkut bidang science

dan ketiknikan/menrpakan tulisan hasil penelitian maLlpun Karya llmiah Populer yang dilakukan

oleh Staf Penga.iar'.

Kami mengharapkan r"rntuk terbitan bulan berikutnya Staf Pengajar dapat meningkatkan kualitas

lnaLrplul mutu dari tulisan, sehingga nremungkinkan sebagai bahan rujukan dalam melakukan

kegiatan penelitian.

Pada kesempatan ini Redaksi mengucaphan Selamat Menunaikan Ibadah Puasa Ramadhan 1425 H

dan terima kasih kepada Staf Perrgaiar/Dosen yang telah berpartisipasi menerbitkan Buliten Utama

'i'eknik FT-LrlSLl terutama pada Edisi Vol. 8 No. 3 - September 2004.

Sernoga FT-UISU sukses dan maiu.

Walbillahi Taufiq Walhidayah

Wassalamu'alaikum Wr. Wb,

Wassalam

Redaksi

Buletin IJtanrn Tahun 2004, Volunte I No. 3 : 177- /,83

Artificial Neural Networks for Peak Static Strength prediction

Nazaruddinl , Zu1rnjit, Adi Setiawan3

A bstralt

Peratnolun kekuatan statis ,sangallah penting untuk menyelesaikan ntasalah-masalah ergonomik terutama dal,"'nten?urangi teladinya musculoskeletal disrtrders ( MSD'S) dan cumulative trauma jisorder (CTD'S) da,'mmelakukan pekerjaan secara manual. Pada penelitian ini digunakan suatu leknik baru untuk peramalan Lekrs*slalis nranusia dalant kasus pengangkatan secara ntanual Sebelumnya, pendekotan statistik seperti regresi liruo-berganda telah digunakan. Di samping nrengusulkan suatu teknik baru, penelitian ini juga menguji kimampuzrneural nefwork's dalam perantalan kekuatan statis manu,sia dengan menggunakan ukuran-ukuro, tubrh. hieweneht'ork's clilatih oleh weight yang dibentttk dari data yapg diperoleh nelolui suatu survei. Didapati bahwa neuranelv'ctrks dopol dengan akurqt membqca datq dqn mqmpu dengan qkurol meramalkan dati tersebut. Deng*ntenggunakun pairutise t-test, didapati bqhwa perbedaan qntqra deta sebenarnya don data hasil ramalan aa;ottidak signi/ikan ( p-t,alue>0.05) clengan tingkat kepercayaan (CI) 9S%.

Kata kunci : Kekutan Statis, Peramalan, Artificial Neural Network, Musculoskeletal Disorders (MSD'sgPenanganan Manual

Abstract

S/utic"strength predictirtn i:s cruciql .fbr ergononric'.s.solution especially in reducing muscu[oskeletal disorden(MSD's) crnd ctrntulative trauma disorcler CfD',9 in manuel handling. ln this pqper propose a new techniquefwpretlictittg hunran stctlic strength for manual li/iing, Previously, slalistical approach such as multiple lineoregre'ssion.s has been used. Be,sides proposing a neu, technique, this paper is also examining the appticability qfartificial neural networks to predict u,orkers slatic strenglh. The neural networks was lrained by *"igl'ttt that ree;ecrealed./rom the data collected through o survey in Malaysia. Il u,as found that neural networks ian accuratehreading lhe data crnd /hus able to accurately predict the dala predicted. Using pairwise t-test, therewas insignificamtlifference betv'een predicting results and actuql data (p-value>0.05) with level of confident interval 95ok.

Keyword(s) : Static Strength, Prediction, Artificial Neural Network. Musculoskeletal Disorders (MSD,s), ManualHandlino

1. IntroductionErgonourics today is growing and changing.Developrnent stelns frorn increasing and irnprovingl<nowledge about the hurnan, and is driven by newapplications and new technological developments. Theword of ergononrics was derived from the Greek termsergon, indicating work and effort, and nornos, law orrules (Murrel, 1994). From basic sciences such associoloey. psyclrology, physiology, medicine,nrathernatics. etc., a group of more applied disciplinesdeveloped into the core of ergonornics such asanthropornetry, biomechanics, industrial hygiene, etc.This paper is corrcerned about biomechanics especiallyin using artificial neural networks as new model to solvein predicting the workers static strengths.The strength prediction is an important consideration.'ri hen solving ergononric problems. Many of researches

use regression method (Potvin et.al, 1992 and K.M.Bolte et.al, 1998) and others use polynomial method(McGill SM et.al, 1996) for predicting the humanstrength. As we know, regression method does not giveaccurate prediction for non linear modeling or complexmodeling. Previously, most of lifting strength values iscalculated based on the NIOSH lifting equation. TheNIOSH lifting equation is still being revised until rodaybecause it does not accurately calculate or predicthuman strength.Now days, artificial neural networks is very popular forpredicting. This is mainly due to their successes inrnodeling complex domain tasks as opposed totechniques such as linear regression and Kalman filters.Besides, artificial neural network modeling techniqueare very flbxible to apply, there are g"nerally no needsto make presumptions abbut the characteristic of static

|J : (-nrnntelce Dcpartmtnt, State Polytechrric l,hohsettmawe'' \lcchanical Dcpartrnent, Faculty of Enginccring, l-lNlNlA' Lhokseumawe

Artificial I\ieural Nelworks for Peak Static Strengtlr Prediction

strengtlr. The tlse of artificial neural network for

strength prediction has not been used recently'

Artiliciai neural networks have been seen involve in

various applicarions dotrlain. such as tnotion analysis

(Taha et,al, 1996), tl-ie prediction of stock trrarket

trading volrttrte (Weigencl, A.S. and B' Lebaron'1994),

currency exchange (f{el-enes et al'' 1993)' ellergy

consuuiption lMaiKay, I993). For ergonotnics solution'

an artificial neural network hits been used in reach

posture prediction (.lurrg ancl Park, 1994) and humatr

nrotion analysis such as liand grasping, walking'

lirnning (l-aha et.al. 1995) tsksioglu, et al (1996) used

rotir aitiircial neural netrvorl<s and statistical model for

ersonomic solution about prediction of pealt pinch

str-ength from a variety of factors, such as elbow and

shoulder flexion, age, weight' grip strength, and various

:r-m and hand dimensions. They found that the ANN

r.odel better at predicting peal< pinch strength than

.:itistical rnoclel. In this paper we study a type of static

r::ength whiclt has different flavors fi'onr those in the

.:rVe appl ications"-r: data come fi'orn the direct lreasurement through the

,,:re1,' of indr"rstrial workers static strength' The data

:,: set of anthropometric, age, and static strength which

r:re collectecl from industrial workers Static strerrgtlt

:-:liction is very usefitl il1 the design rvork task for the

::iels especially irr rrranual rlaterial handling'

l. Back Propagntion 'fechnique

s(net,):I + d-ntt'

In our study the number of inputs into the network are

variety, we try tel find the best combination of physical

dirnensions to fit the static strength. The output is

standing lifting strength for position with two-handed

pull 100 crn height level of the handle.

(5)

N+

Figure l. Topology of a

network

N+

fully connected neural

(

The network weights W;, ar€ updated using the back

propagation rnethod by minimising square error by the

trairring set:

d A tt 1,,\.E= ZE(k) = III/z)(Y -Y)' (6)

k-t k--l t=l

d"

ttrrtr€€{

jt3ry'

.r[b.

--: ar-tificial rreural networks are typically organized in

:.::s. Layers are rnade up 'bf a nurnber of K is the number of training sets. In this study the

r-.:::onnected ,nodes' rvhich contain an 'activation training sets are the inputs and outputs at the

..-::iou', These data will be used to find a correlation anthropometrics data and static strength data from

:,:-.i,3er.t the irrputs and outputs r.rsing a fully connected measurements. Yi are outputs of training sets

". , -. net.,vork. corresponding to lifting strength and grasp.-". ',Lllr' cotrnected network irsed to sitnr-rlate the data

-r- :: desct'ibed by the follorving set of equations:

=\ l<i<m, (l)3. Proposed Approach

The proposed approach is based on survey that

measurements of human static strength for example

using sorne of equiprnent for measure static strength do

direct measurements of the workers static strength.

Besides, the anthropometrics data also was collected to

know whether any correlation between anthropometrics

and static strengths. The anthropometrics data can be as

input and the output is only static strength data for

different size of anthropometrics data. Supposing the

output parameters are defined by the static strength

-i,'' '-tt"l

- :tet,)

,r,o1<N-l-r?. (2)

m<i <Nrn, (3)

l<i<n, (4)T[r!m

m

4rra

imG

', r:-= \ is the input and Y; is the out put' rr is tlre*!iT-r:=- of inputs. ir is ntltnber of outputs and N is

r .i-: integer less than rrr. The value of N I n decides

' :r '- -rr3r-o?n.uront in the netr.vork. The function 's in

r:' -: 1,1) is the signroid ftrnction:

Input

Nuzorudtlitr, Zuarni,,4tli Setiowan

oredicted (Yi) and the input as anthropometry and static

str.ngtl' (Xi) to find a function:Yi:.fiXi) 0)

The fr.rnctiorr / is rnapping the outputs and the inputs'

'iir. o,'tp",o 'can be static strengtlr fronr different

nositio'r such as knee bend, knee straight' sit etc'

il";; .ut., *. only take one output that is static

;;;.;il' oit*o-t'und"i pull 100-ln'rf:u'l of handle on

standfng lifting posture (Figure 3)' The inputs data X;

will consists of:

X'= Ittutut., lveight, artn reach, circunrference' hand

erc) (8)

The input components are not linrited to those

pr."l"r'iy describe bttt can be extended to irrcltrde other

pal'ameters such as age and other physical dimension

data as need arises. The inputs parameter will be created

for different nurnber of inputs and different types of

physical dlmension. Even though these data can be

predicte{ using linear regression method' but in the case

of predicted problem are more reliably using artificial

neural networks. The approach used here is illustrpte/ in

Figure 2.

4. Case StudY

Data was taken from 146 industrial workers across

Malaysia. The sample consists of 104 male (32 heavy

indusiries and 72 light industries) and 42 female (light

industries) workers. Subject was chosen randomly that

was involved in manual material handling'

Figure 2. Basic back-propagation

Age. weight. grasp strength and 9-different types

of arrthrJporretrics meaiLrrelnents frorn 52 types arrd 2

,fi'oui l2 type's lneasurelnents of static strength are

taken.

Table l. Physical dimensions measllrement

The anthropometry measurements are based on

literature by Kroemer (1999) and White (1964), whilst

the static sirengths based on literature by Chaffin and

Anderson (1991) and the guidelines by FAA. (Section

14, 1996). There are S [ina of physical dimension

measurement were taken, namely stature, knuckle

height, arm reach, knee height, wrist c^ircumference'

l,pi.t'un" circumference, crotch circumference' ankle

lir.u,rlf.r.n.., and lower thigh circumference (see

iable l). The other hand the strength data. were taken

onfV t,unaing strength for position with two-handed pull

100 cm height.levelof the handle

. Statltre

. l(nuckle height(NH)

. Arrn reach (ARF)

. Knee height (KH)o Ankle (AC)

o Wrist (WC)o Upper arm (UAC). Crotclr (CTC)o Lower Thiglr

(Lrc)

4rtificinl Neurd Networhs.for Penk Stnlic Slrenglh.Predicliott

Figure 3. Illustration of the power for 100 crn high level of the handle with twohanded pLrll in standing posture.

Tt-

_3

I

fr 5. Learning from Dnta and Simulating

ihe first step in this approaeh is retrieve physical

irmensiou data to be sirnulated, Whiehever methods are

:sed. the output data of training value for ANN is in:::rns of static strength data for standing lifting with

:.,.o-handed pr"rll on 100 cm height level of the handle

. ruure 3). Training is learning process by which inputs

.-.J oLrtpLrt data are repeatedly presented to tlre network'

.:rs process is rrsed to deterrnine the best set of weights

' : the network. r,vhich allerw the ANN to classify the

input vectors with a satisfactory level of accuracy. Inthis study we want to see wh€ther any correlationbetween types of physical dirnensions and staticstrength for standing posture with two-handed pull on100 crn height levelof the handle.From the result (Table 2) shows the static strength errorof male workers (heavy industry) for different number

of inputs of different physical dimension. It also showwhich part of the physical dimension have related to the

static strength,

T'able 2. Root Mean Square Error (RMSE) for difference cornbination of parameters input.ilr

r-i5.d-L

::.rililI

rrn.ici*-ref;[s,.

ag-e

:fNsrd pnl

\ I easured

\\ eielrt

S tatu le

\RFl .\(l

r lt

\t. lra

* tl'. Selectetl l'ttriable,s

': rnportant restrlt from this study is the artificial:--:. netwol'k cart trairl the physical dinrension data

. : -:;nlan static strength data altrost no significarrtly

different compare to the output. Besides, ANN also can

prove there are any relationship between physical

dirnension and human static strength.

Male (Lieht lndustriesFernale (Lieht Industries)

N azsru ddirr, Zu utti, Atli S et iaw an

M ale - Heavy lndustries

60

50

40

1 3 5 7 I l',l 13 ls 17 19 2'l 21 25 27 29 31

Num bcr ol gubjsct

+Msar'ur'd +-- Predicte d

o'-6

gF;Jo ttt

6Oot6E:

30

2C

10

0

Fieure 4. Cornparisons between measured and predicted for standing posture on l00cm

height level of the handle (Male in Heavy lndustries)

The neural networks also able to produced model from

human static strength in different position and different

dimension of arrthropometrics data' The tnodel is stored

as a set of neural tretwork weights generating during the

learning phase.

fn ng;re 3 above, shows that for male in heavy

indusiries, the input parameters with the smallest RMSE

(4.002) were age, weight, KH, UAC, -CTC,,and

AC'

Wittl ti. pairwise t-test with 95% confident level was

applied, tirere was no significant difference, between

pr.ai.t.a values and lneasurelnent values whereas p-'valtte

calculated was 0'456 (p<0'05)' Figure 4 shows

that results was approxitnately no difference cornpared

to actual measurements' As a result that the ANNs

model has capability with higher accuracy for predicting

the standing strength with l00cm height level of male in

light indusries.Figure 5 shows that the number of passes also affects to

thi result, There is any certain number of passes that

can nrake the smallest error' For the human static

strength of two-handed pull 100 cm levels in standing

postwe for heavy industries workers have the smallest

error when the number of passes is 20000.

6r

,l

Standing Strongth on 100cm Helght Favol

1 0o0o 1 6oOo 20000 25000 50000 1 00000 200000 500000

Number of Passes

fio6

goGit

=oot

5000

+Female(Lightlndustries) s--Msl6(Lightlndustri6s) Msl€(H€aWlndustries)

Figure 5. RMSE with Difference Nurnber of Passes (standing Lifting on l00cm)

The preceding nutnbers and types of. parameters

assurned for stirrding strength on 100cm height level for

male in light industiies was iderrtical with 50crn height

.level, however, tlre combination of parameters were not

exactly same. For male in light industries, the input

parameters with the smallest RMSE (3'848) were

Ailirtciol Nearsl Networks for Peok Stotic Strength prediction

exactly same with female in light industries such as

Grasp. By using the pairwise t-test with 9570 corrfidentlevel, it was found that there was no significantdifference between predicted valrtes and tneasurementvalues rvhereas p- value calculated was 0.508 Qt<0.05).

weight, stature, KH, NH, UAC, WC, CTC, AC, andThe Figure 6 shows that the accuracy of ANNs modelfor predicting wap high, and slightly differencecompared to actual measurements for male in Iightindustries.

M ale " Lig ht lnd ustries

- 50c45U';g|40

! JJ

5E eog6 zs

i*'oa x 15

I ro

r015 13 17 21 25 29 33 37 41 45 49 53 57

Num ber ot SubJ6c!

-+- M ea 6 ured --+.-... P redicted

FigLrre 6. Comparisons between measured and predicted for standing posture on l00cm heightlevel of the handle (Male in Light tndustries).

The earlier nurnbers and types of parameters assutned

ior standing strength on l00cm height level was age,,'reight, stature. KH, NH,, ARF, UAC, WC, AC, LTC,CTC, and Grasp, hor.vever, the combination offararneters were not exactly same, For fernale in lightrndustries. the input parameters with the smallest errorRMSE : 2.108) were weight, stature, KH, NH, UAC,

rI'C, CTC, AC, and Grasp.

The pairwise t-test with 95o/o confident level wereapplied, and found that there was significant differencebetween predicted values and measurement valueswhereas p-valtre calculated was 0.024 (p<0.05). Eventlrough it was significant difference using pairwise t-test, but in the Figure 7 shows the predicted resultsapproximately no difference compared to actualmeasurements.

Female - Light Industries

7 9 11 13 15 17 t92t 23 2527 2931 33 35 37 39 41

Numbe. of Subiccl

' +Measured , e-Predlcled,

Figure 7. Comparisott between Actual Measuretnent and Prediction for StandingLifting on Fleight l00cm (Female in Light Industries)

€.9-30

E lcEc;@

^ tco Y ,^

:0

Ir

ItvI

rurl

:r1{r*€ill[

N azr rw tldin, Z u orn i, A tli Sritinwn

6. Discussions

The artificial nettral network can be used to predict

lrurnan static strength for different subject's physical

dirnensions. fhese can be proved on the above graph

and table that there are not significantly different

between data tneasurement and ANN.At tlre human static strengtlr prediction for both male

and fetnale workers in light industries, age can affect

the human static strength. J'lre other hand the human

static stretrsth for rnale workers in heavy industries in

the sanre position, age is affect to the hurnan static

st rett gtlr.

From the results was shown that for the human static

strength prediction of rnale workers in heavy industries

need only 6 irrputs paranleter, but for trrale and female

workers in light indtrstries only 9 input to get the

snrallest error. Even between light and heavy industries

had difference rrtttrber of input, but there wgre no

significant different especially in the shape of graph'

Hswever, tlrere was still having any differences between

data outputs frorn ANN and data lneasurement' These

may caused there was not strong correlatiorr between

part of plr,vsical ditnension and hutnan static strength oftu'o-handed pull 100 clr level.

Generally, the simulation results using ANN have at

least two satrre values rvith the l'lleastlrelrent data for

each oLrtput paralneter. All the inputs have been passed

iteration 20000 times.

z. Conclusions

We conclLrde tlrat artificial neural network can form the

basis of physical ditnensiorr and weight for modeling

lrunran static strength and predict tlre human static

stren gtlr fbr d iff'erent phys ical d ilnerrsion.

The strength predicting rnodel that is developed willcreate into software, so that the users are easy to predict

the strength. Next research, we will test the data and the

resLrlt statistically and develop for predigting the other

positiorr of lil'ting strength and also dynarnic strength

nrodel of'u'orkers.

Refcrences

Chaffin, D.B. and Andersson, G'B.J', (1991)'

Occupational Biomephanics, ,lohn lVilley & Son

I tt,'

ELrr S. Jung ancl Suttgioon Park. (1994), Prediction ofIrurnan reach posture using neitral netwot'k for

et'gonotrr ics lllan urodels, Cornputer and

lndustrial Engineering, Yol. 27, (l-a) pp. 369-372, Elsevier Science.

Eksioglu, M., Fernandez, l. E., and Twomey, J. M.' (1996), Predicting peak pinch strength:

Artificial neural networks vs. regression,

lnternational Journal of lndustrial Ergonomics,

18,431-441.

Human Factors Design Guidelines, Section 14, (1996),FAA Willian J. Hughes Technical Center.

K.M. Bolte, M.H, Pope, V.K. Goel, K. Spratt, (1998),

Regression equations to predict spinal loadings,

North America Congress in Biomechanics.

Kroemer, Karl H, 8., (1999), EngineeringAnthropometry, In Karwowski, W., Marras, W.S'(Eds) The Occupational Ergonomics Handbook,l 39- I 62.

Kroemer, K. H. E., Kroemer, H. B., "Ergonomics: Howto Design for Ease and Efficiency", l'r ed, New

Jersey: Prentice HalL 1993..

MacKay, D.J.C., (1993), Bayesian non-linear modellingfor the energy prediction competition. Technical

Repart, Canvendish Laboratory, Cambridge

University,

McGill SM, Norman RW, Cholewicki J., (1996), Asirnple polynomial that' predicts low-back

compression during complex 3-D tasks,

Ergonomics, 39: I 107-l I 18.

Murrell, K. F. H., (1969), Ergonomics - Man and

His Working Environment, London: Chapman

and Hall

Potvin JR, Norman RW, Eckenrath ME, McGill SM'Bennett CW, (1992), Regression models for the

prediction of dynamic L4-5 compression forces

during lifting, Ergonomics, 35: 187'201 .

Refenes, A.N., M, Azemar-Barac, et ?1., (1993),

Currency rate exchange prediction and neural

network design strategies, Neural Computing and

Applications, Vol. l, 46-58.

Weigend, A.S. and B. Lebaron, (1994), Evaluating

neural network predictors by bootstrapping' Proc'

Q/'ICONIP '94, Seoul.

Z. Taha, R, Brown and D. Wright, (1996), Realistic

animatiott of human figures using artificial neural

networks, Journal of Medical Engineering

Physic, Vol 18. No. 8, pp. 662-669. Great Britain.