Comparative Study of EMG based Joint Torque Estimation ANN Models for Arm Rehabilitation Device

Similar documents
Predicting EMG Based Elbow Joint Torque Model Using Multiple Input ANN Neurons for Arm Rehabilitation

Study of Stress Distribution in the Tibia During Stance Phase Running Using the Finite Element Method

An Energy Efficient Seizure Prediction Algorithm

SUPPLEMENTARY INFORMATION

Effects of physical exercise on working memory and prefrontal cortex function in post-stroke patients

Effect of linear and random non-linear programming on environmental pollution caused by broiler production

What is it? Flexibility Training involves stretching exercises to improve joint and muscle function.

Check your understanding 3

build Firm, sexy arms

Infrared Image Edge Detection based on Morphology- Canny Fusion Algorithm

ENERGY CONTENT OF BARLEY

Meat and Food Safety. B.A. Crow, M.E. Dikeman, L.C. Hollis, R.A. Phebus, A.N. Ray, T.A. Houser, and J.P. Grobbel

Using Paclobutrazol to Suppress Inflorescence Height of Potted Phalaenopsis Orchids

8/1/2017. Correlating Radiomics Information with Clinical Outcomes for Lung SBRT. Disclosure. Acknowledgements

Reducing the Risk. Logic Model

ET 100 EXTERIOR FRONT DOOR BLACK OUT TAPE INSTALLATION

Chapter 5: The peripheral nervous system Learning activity suggested answers

Optimisation of diets for Atlantic cod (Gadus morhua) broodstock: effect of arachidonic acid on egg & larval quality

Not for Citation or Publication Without Consent of the Author

Health-Related Quality of Life and Symptoms of Depression in Extremely Obese Persons Seeking Bariatric Surgery

THE EVALUATION OF DEHULLED CANOLA MEAL IN THE DIETS OF GROWING AND FINISHING PIGS

SUPPLEMENTARY INFORMATION

SUPPLEMENTARY INFORMATION

The potential future of targeted radionuclide therapy: implications for occupational exposure? P. Covens

Agilent G6825AA MassHunter Pathways to PCDL Software Quick Start Guide

The Effect of Substituting Sugar with Artificial. Sweeteners on the Texture and Palatability of Pancakes

Bioactive milk components to secure growth and gut development in preterm pigs ESTER ARÉVALO SUREDA PIGUTNET FA1401 STSM

The Acute Time Course of Concurrent Activation Potentiation

Analysis of Regulatory of Interrelated Activity of Hepatocyte and Hepatitis B Viruses

Input from external experts and manufacturer on the 2 nd draft project plan Stool DNA testing for early detection of colorectal cancer

MODERNIZATION of MAIN HELICOPTER GEARBOX with ASYMMETRIC TOOTH GEARS

EFFECTS OF AN ACUTE ENTERIC DISEASE CHALLENGE ON IGF-1 AND IGFBP-3 GENE EXPRESSION IN PORCINE SKELETAL MUSCLE

Trajectory of Contact Region On the Fingerpad Gives the Illusion of Haptic Shape

Extraction and Some Functional Properties of Protein Extract from Rice Bran

Finite Element Analysis of MOD Prosthetic Restored Premolars

Consumer perceptions of meat quality and shelf-life in commercially raised broilers compared to organic free range broilers

Non-invasive Diagnosis of Liver Clinical Condition by Real-time Tissue Elastography and Shear Wave Measurement : Get More Accessible by One Probe

Available online at ScienceDirect. Procedia Engineering 160 (2016 )

Single-Molecule Studies of Unlabelled Full-Length p53 Protein Binding to DNA

THE EFFECT OF DIFFERENT STIMULI ON MEAGRE (Argyrosomus regius) FEEDING BEHAVIOUR.

INFLUENCE OF DIFFERENT STRAINS AND WAYS OF INOCULATION ON THE RABBIT S RESPONSE TO EXPERIMENTAL INFECTION WITH PASTEURELLA MULTOCIDA

A FACTORIAL STUDY ON THE EFFECTS OF β CYCLODEXTRIN AND POLOXAMER 407 ON THE SOLUBILITY AND DISSOLUTION RATE OF PIROXICAM

Copy Number ID2 MYCN ID2 MYCN. Copy Number MYCN DDX1 ID2 KIDINS220 MBOAT2 ID2

TEMPLATE SYNTHESIS OF Cu AND Cu-Sn NANOPARTICLES USING CARBON FOAM AS A SUPPORT. Ivania MARKOVA-DENEVA, Tihomir PETROV, Ivan DENEV

EMG Signal Features Extraction of Different Arm Movement for Rehabilitation Device

XII. HIV/AIDS. Knowledge about HIV Transmission and Misconceptions about HIV

USE OF SORGHUM-BASED DISTILLERS GRAINS IN DIETS FOR NURSERY AND FINISHING PIGS

Visualization of Stent Lumen in MR Imaging: Relationship with Stent Design and RF Direction

Clinical Study Report Synopsis Drug Substance Naloxegol Study Code D3820C00018 Edition Number 1 Date 01 February 2013 EudraCT Number

The step method: A new adaptive psychophysical procedure

PNEUMOVAX 23 is recommended by the CDC for all your appropriate adult patients at increased risk for pneumococcal disease 1,2 :

The study of Forage Quality of Smirnovia iranica In Different phonological stages in sandy areas-case-study: Band-e-Rig-Kashan

Severe Gummy Smile with Class II Malocclusion Treated with LeFort I Osteotomy Combined with Horseshoe Osteotomy and Intraoral Vertical Ramus

Invasive Pneumococcal Disease Quarterly Report July September 2018

2. Hubs and authorities, a more detailed evaluation of the importance of Web pages using a variant of

Detecting the Fetal Electrocardiogram by Wavelet Theory-Based Methods

ORIGINAL ARTICLE. Diagnostic Signs of Accommodative Insufficiency. PILAR CACHO, OD, ÁNGEL GARCÍA, OD, FRANCISCO LARA, OD, and M A MAR SEGUÍ, OD

Analytic hierarchy process-based recreational sports events development strategy research

AR Rice Performance Trials (ARPT) Color as a Quality Indicator. Functional Property Analyses. Cause of Chalkiness in Rice Kernels

N. J. Boddicker*, D. J. Garrick*, R. R. R. Rowland, J. K. Lunney, J. M. Reecy* and J. C. M. Dekkers* Summary. Introduction. doi: /age.

The Effects of Diet Particle Size on Animal Performance

Hamstrings stretch reflex in human spasticity

HEMOGLOBIN STANDARDS*

Automatic Detection of Malignant Melanoma using Macroscopic Images ABSTRACT

Summary. Effect evaluation of the Rehabilitation of Drug-Addicted Offenders Act (SOV)

Sleepy Stack Reduction of Leakage Power

Effect of supplemental fat from dried distillers grains with solubles or corn oil on cow performance, IGF-1, GH, and NEFA concentrations 1

A LAYOUT-AWARE APPROACH FOR IMPROVING LOCALIZED SWITCHING TO DETECT HARDWARE TROJANS IN INTEGRATED CIRCUITS

EFFECTS OF INGREDIENT AND WHOLE DIET IRRADIATION ON NURSERY PIG PERFORMANCE

Nicholas James Boddicker Iowa State University. Dorian J. Garrick Iowa State University, Raymond Rowland Kansas State University

P AND K IN POTATOES. Donald A Horneck Oregon State University Extension Service

Effect of 1-Methylcyclopropene on the Physiology and Yield of Cotton. Derrick Oosterhuis Eduardo Kawakami and Dimitra Loka University of Arkansas

EVALUATION OF DIFFERENT COPPER SOURCES AS A GROWTH PROMOTER IN SWINE FINISHING DIETS 1

3. DRINKING WATER INTAKE BACKGROUND KEY GENERAL POPULATION STUDIES ON DRINKING WATER INTAKE RELEVANT GENERAL POPULATION

WSU Tree Fruit Research and Extension Center, Wenatchee (509) ext. 265;

Combined Vision and Wearable Sensors-based System for Movement Analysis in Rehabilitation

The effect of encapsulated butyric acid and zinc on performance, gut integrity and meat quality in male broiler chickens 1

Geographical influence on digit ratio (2D:4D): a case study of Andoni and Ikwerre ethnic groups in Niger delta, Nigeria.

This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and

Evaluation of a task-oriented client-centered upper extremity skilled performance training module in persons with tetraplegia

A reservoir of time constants for memory traces in cortical neurons

INFLUENCE OF JUMPING STRATEGY ON KINETIC AND KINEMATIC VARIABLES

Review article. Key words non specific low back pain, exercise therapy, stretching, motor control. Grade D

SYNOPSIS Final Abbreviated Clinical Study Report for Study CA ABBREVIATED REPORT

Am J Nephrol 2013;38: DOI: /

EFFECT OF DIETARY ENZYME ON PERFORMANCE OF WEANLING PIGS

Cushion. Ride Designs JAVA

Supplementary Figure 1

arxiv: v2 [cs.ne] 15 Sep 2017

SEIZURES AND EPILEPSY

Author's personal copy

Life satisfaction 6 15 years after a traumatic brain injury

Evaluation of the Masticatory Part and the Habitual Chewing Side by Wax Cube and Bite Force Measuring System (Dental Prescale )

Review TEACHING FOR GENERALIZATION & MAINTENANCE

Particle-size distribution of very low density plasma lipoproteins during fat absorption in man

Arthroscopic Anatomy of Shoulder

JOB DESCRIPTION. Volunteer Student Teacher. Warwick in Africa Programme. Warwick in Africa Programme Director

Invasive Pneumococcal Disease Quarterly Report. July September 2017

Optimizing Metam Sodium Fumigation in Fine-Textured Soils

Transcription:

Interntionl Journl of Applied Engineering Reserch ISSN 0973-4562 Volume 9, Numer 10 (2014) pp. 1289-1301 Reserch Indi Pulictions http://www.ripuliction.com Comprtive Study of EMG sed Joint Torque Estimtion ANN Models for Arm Rehilittion Device M. H. Jli 1,, M. F. Sulim 1,, T. A. Izzuddin 1, c, W. M. Bukhri 1, d nd M. F. Bhrom 1, e 1 Fculty of Electricl Engineering, Universiti Teknikl Mlysi Melk Hng Tuh Jy, 76100 Mlcc, Mlysi mohd.hfiz@utem.edu.my, fni@utem.edu.my, c trmizi@utem.edu.my, d ukhri@utem.edu.my, e mohmd.fizl@utem.edu.my Astrct Rehilittion device is used s n exoskeleton for people who hd filure of their lim. Arm rehilittion device my help the reh progrm whom suffered with rm disility. The device used to fcilitte the tsks of the progrm should improve the electricl ctivity in the motor unit nd minimize the mentl effort of the user. Electromyogrphy (EMG) is the techniques to nlyze the presence of electricl ctivity in musculoskeletl systems. The electricl ctivity in muscles of disle person is filed to contrct the muscle for movements. In order to minimize the used of mentl forced for disle ptients, the rehilittion device cn e utilize y nlyzing the surfce EMG signl of norml people tht cn e implemented to the device. The ojective of this work is to compre the performnce of the joint torque estimtion model from the muscle EMG signl to torque for motor control of the rm rehilittion device using Artificil Neurl Network (ANN) technique. The EMG signl is collected from Biceps Brchii muscles to estimte the elow joint torque. A two lyer feed-forwrd network is trined using Bck Propgtion Neurl Network (BPNN) to model the EMG signl to torque vlue. The comprison etween two ANN models is mde to oserve the performnce difference etween these models. The experimentl results show tht ANN model with doule input nodes hs etter performnce result in term of Men Squred Error (MSE) nd Regression (R) which is crucilly importnt to represent EMG-torque reltionship for rm rehilittion device control.

1290 M. H. Jli et l Keywords - Electromyogrphy, Artificil Neurl Network, Arm Rehilittion Device, Joint Torque Estimtion, Exoskeleton 1. Introduction Humn support system is endoskeleton. Endoskeleton plys role s frmework of the ody which is one. Our dily movements re fully depends on the functionlity of our complex systems in the ody. The disility one or more of the systems in our ody will reduce our physicl movements. The ssistive device is need for reh s n exoskeleton. The functionlity of the rehilittion device hs to smooth s the physicl movement of norml humn. The rehilittion progrms provide the suitle progrm for conducting the nerve nd stimulte the muscles. People who hve temporry physicl disility hve the chnces to recover. Nowdys, rehilittion progrm re using exoskeleton device in their tsks. The functionlity of exoskeleton depends on muscle contrction. Electromyogrm studies help to fcilitte the effectiveness of the rehilittion device y nlysing the signl trnsmitted from the muscle. The technique of mesuring electricl ctivity tht produced from the muscles during rest or contrctions known s electromyogrphy (EMG). The electric signl genertes from the rin nd sends to the muscles vi motor neuron. The EMG my detect the dysfunctionl of the muscles or filure in signl trnsmission from nerve to muscle. The filure of sending the electricl signl from the rin requires electricl stimultion from the externl source to muscles. Electrodes re used for signl detection of electricl ctivity in muscles. The study of this electricl ctivity is importnt for comintion of electromyogrm nd rehilittion device. The rehilittion device is tool tht used to help the movements for dily life ctivities of the ptients who suffer from the filure of muscle contrctions, due to the filure of the muscles contrctions the movements is limited. The ility of the ptients to do the tsks in the rehilittion progrms need to e mesured. The rehilittion progrms hve to ssure whether the tsks will cuse effective or ring hrm to the ptients [1]. Historiclly, the rehilittion tsks hve een voided due to elief tht it would increse spsticity [2]. In this reserch, the nlysis of the dt will e focusing on upper lim muscles contrction consisting of iceps muscles only. The experiment is limited to the certin of upper lim movements tht use in trining. EMG is division of io signl; the io signl nlysis is the most complex nlysis. Thus, the signl nlysis is complicted process tht hs to e through mny phses of nlysis [3]. EMG is used s control signl for the rm rehilittion device. A system needs model to estimte reltionship etween EMG nd torque [4]. EMG signl sed control could increse the socil cceptnce of the disled nd ged people y

Comprtive Study of EMG sed Joint Torque Estimtion ANN Models 1291 improving their qulity of life [5]. The joint torque is estimted from EMG signls using Artificil Neurl Network [6]. The BPNN is used to find solution for EMGjoint torque mpping. The EMG signl of the iceps rchii muscle ct s the input of the ANN model whiles the desired torque ct s the idel output of the model. Hence the EMG signls considered the intent of the system while the joint torque is the controlled vrile for the rm rehilittion device [7]. There re severl work tht hs pplied BPNN for modeling the muscle ctivity of nkle to joint torque reltionship. It is done y pproximting the force from the EMG signl under sttic conditions [7]. The network is evluted sed on the est liner regression etween the ctul joint torque nd the estimted joint torque [4]. 2. Methods Figure 1 shows lock digrm of our reserch tht consists of two mjor phses. First phse is EMG dt processing nd desired torque determintion. Second phse is the ANN construction nd testing. The dt collection from the first phse is used to vlidte nd tech the ANN lgorithm in second phse. Figure 1. Reserch methods 2.1 Experimentl Setup Implementtion of rm rehilittion device sed on movement is recorded from the EMG signl of helthy sujects. From the humn ntomy studies, different ngle movements of upper lim with elow s the reference is depends on reltion of gonist nd ntgonist. In this study is focusing on the ehviour of iceps muscle s gonist nd the triceps s the ntgonist respectively. Muscle tht involved in this movement is iceps nd triceps, however in this study to understnd the electricl ctivity during muscle contrction, the iceps is the only muscle tht tking into ccount. The movements rnges in etween position of rm flexion until rm fully extend. The environment is in room with low lighting especilly the fluorescent light, ny electromgnetic devices is wy from the experiment equipment nd the

1292 M. H. Jli et l environment is in silent room. Then, the experimentl is set up with the suject sit on the chir while the hnd is on the tle. The suject hs to complete the tsk of lift up the dumell with 2.268 kg of weight in Figure 2() for 5 times. Normlly, the ppernce of EMG signl is chos nd noisy depends on the type of electrodes lso the noise fctor. To simplify the difference of mplitude response for the motion, the dumell is functioned to mplify the mplitude in nlysing the electricl ctivity during rest nd contrct. The rehilittion devices (white in color on Figure 2()) helps to keep the position of the elow joint nd the wrist joint in line. Mostly, the EMG signl is otined fter severl trils of the movements. These movements re specified from ngle of 0 (rm in rest position), up to 120 (rm is fully flexion). Dt ws collected from two sujects y 5 repetitions of ech flexion movements [8]. Figure 2. Suject is set-up with rm rehilittion ssistive device for experiment (), Simultion of suject s to lift up the dumell 2..268kg of weight (). Prior of dt collection process, the skin needs preprtion. The preprtion of skin is ruled y the Surfce Electromyogrphy for the Non-Invsive Assessment of Muscles (SENIAM) procedure for non-invsive methods. The suject s skin hs to e shved y using smll electricl shver nd clened with sterile lcohol sws sturted with 70% Isopropyl Alcohol. This step is to e tken for minimizing the noise nd to hve good contct with the electrodes of the skin y decresing the impednce of the skin. The skin hs to e clen from ny contmintion of ody oil, ody slt, hir nd the ded cells. The preprtion of skin cn e done y wiping the lcohol sw into the re of skin tht electrode plcement to e pplied. The plcements of the electrode hve to e t the elly of the muscles not in the tendon or

Comprtive Study of EMG sed Joint Torque Estimtion ANN Models 1293 motor unit. This ensured the detecting surfce intersects most of the sme muscle on suject s in Figure 3() t the iceps rchii, nd s result, n improved superimposed signl is oserved. Reference electrode hs to e t the one s the ground, for this experiment it plced t elow joint s shown in Figure 3(). These electrodes re connected to the comintion of hrdwre Olimex EKG-EMG-PA nd Arduino Meg for dt collection. Figure 3. The iceps rchii muscles for electrode positions (), The electrode plcements on suject skin () 2.2 EMG dt processing Figure 4 shows the EMG dt processing lock digrm. After otined stisfctory EMG signl s shown in Figure 5(), Fst Fourier Trnsform (FFT) is performed to the signl to nlyses the frequency content of the signl. The EMG signl is rek into its frequency component nd it is presented s function of proility of their occurrence. In order to oserve the vrition of signl in different frequency components, the FFT signl is represented y Power Spectrl Density (PSD). From the PSD we cn descries how the signl energy or power is distriuted cross frequency. Figure 5() shows tht most of the power is in the rnge of elow 10Hz, therefore the EMG signl should e filtered in the rnge of ove 10Hz s cut off frequency for low pss filtered. After decide the cut off frequency for filtering, the DC offset of the EMG signl is removed nd is rectified to otin its solute vlue s shown in Figure 5(c). Finlly, the signl ws smoothed nd normlized pssing it through 5 th order Butterworth type low-pss filter with cut off frequency 10Hz nd the smooth signl is illustrte in Figure 5(d) [8][9].

1294 M. H. Jli et l Figure 4. EMG dt processing c d Figure 5. Rw EMG signl (), Power Spectrl Density (), Rectified signl (c), Smooth signl (d) 2.3 Desired Torque Desired torque of the elow joint is used s trget dt for our ANN techniques s well s ct s output signl for muscle. The dt is collected throughout the ngle from 0 0 to 120 0 ngle with increment of 0.0619 0 ech step to lign with the smple numer of EMG signl. Figure 6() shows the rm rehilittion device position for torque clcultion. The desired torque for elow joint is determined y pplying stndrd torque eqution: -

Comprtive Study of EMG sed Joint Torque Estimtion ANN Models 1295 τ = (r lod F lo od + r rm F r rm) cos θ (1) Where r lod is distnce from the elow joint to the lod, F lod is force due to lod, F rm is the force due to the mss of the lever rm nd r rm is evenly distriuted distnce of mss of rm distnce which is hlf of r lod. The ngle θ etween r nd F is drwn from the sme origin. A pplied lod is 5 pounds (2.268kg) dumell nd the distnce from the elow joint is 0..25m while the mss of the lever rm is 0.1kg. Figure 6() shows the desired torque chrcteristic for the elow joint. Figure 6. Arm rehilittion device position (), Desired torque chrcteristic () 2.4 Artificil Neurl Network (ANN) Model ANNN is computing prdigm tht is loosely modeled fter corticl structures of the rin. It consists of interconnected processing elements clled nodes or neuron tht work together to produce n output function. It cple to mp dt set of numeric inputs with set of numeric outputs. It is lso the most widely pplied trining network whichh hs input lyer, hidden lyer nd output lyer. The neurons on ech lyer need to e considered crefully to produce high ccurcy network. The numer of hidden neurons could ffect the performnce of the network. The network performnce not lwys een improved if the hidden lyer nd its neurons is incresed [4]. Therefore the numer of hidden neurons is tested to chievee the optimized network. However theree is constrint in determining the numer of neurons. If the numers of hidden neurons is too lrge, the network requires more memory nd the network ecome more complicted while if the numer of hidden neurons is too smll, the network would fce difficulty to djust the weigh properly nd could cuse over fitting which is prolem wheree the network cnnot e generlized with slightly different inputs [10]

1296 M. H. Jli et l The input signl is propgted forwrd through network lyer using ck propgtion lgorithm. An rry of predetermined input is compred with the desired output response to compute the vlue of error function. This error is propgted ck through the network in opposite direction of synptic connections. This will djust the synptic weight so tht the ctul response vlue of the network moved closer to the desired response [11]. BPNN hs two-lyer feed-forwrd network with hidden neurons nd liner output neurons. The function used in the hidden lyer of network is sigmoid function tht genertes vlues in rnge of -1 to 1 [8]. There re lyers of hidden processing units in etween the input nd output neurons. For ech epoch of dt presented to the neurl network, the weights (connections etween the neurons) nd ises re updted in the connections to the output, nd the lerned error etween the predicted nd expected output, the delts, is propgted ck through the network [12]. A Lvenerg-Mrqurdt trining ck propgtion lgorithm is implemented for this work to model the EMG to torque signl. Two ANN models is considered to compre the performnce of ech model. ANN model 1 consist of input lyer with single node which is denoted s tht represent EMG signl from iceps while ANN model 2 hs dditionl node t the input lyer which denoted s t tht represent the movement time which ct s trining dt. The output lyer hs only one node which is denoted s τ tht represents the desired torque ct s trget dt. Tle I descrie the difference etween the models nd the network structure of ech model is shown in Figure 7[4]. The network ws trined using 1839 sets of EMG dt for rm flexion motion from 0 0 ngles to 120 0 ngle. It lso hs output dt which is torque of correspondent rm motion. The trining process ws itertively djusted to minimize the error nd incresed the rte of network performnce [10]. MATLAB softwre is used to construct the BPNN network. The network requires dt for lerning nd testing in order to determine the weights ech node uses. The trining hs een done y dividing the input dt of 70% for trining, 15% for vlidtion nd 15% for testing [9]. The network will e trined until the following condition fulfilled efore it stop [10]: - rech mximum numer of epochs grdient performnce ecme less thn the minimum grdient vlidtion performnce incresed more thn the mximum fil times since the lst decresed one

Comprtive Study of EMG sed Joint Torque Estimtion ANN Models 1297 Tle 1. ANN model description ANN model Model 1 Model 2 No. of input node 1 2 Node 1 EMG dt, EMG dt, Node 2 Movement time, t - Hidden neurons Output node 10 Desired torque, τ 10 Desired torque, τ The performnce evlution of the network is sed on the Men Squred Error (MSE) of the trining dt nd Regression (R) etween the trget outputs nd the network outputs s well s the chrcteristics of the trining, vlidtion, nd testing errors. The network is considered hs the est performnce if it hs lowest MSE nd highest R while exhiit similr error chrcteristics mong the trining, vlidtion nd testing. However even if the MSE shows very good result ut the vlidtion nd testing vry gretly during the trining process, the network structure is still considered unstisfctory ecuse the network is not generlized. Therefore further tuning nd trining need to e conducted in order to improve the network performnce [10]. Figure 7. () Model 1, () Model 2 3. Experimentl Results & Discussion In order to optimize the network performnce, different numer of hidden neurons is simulted for severl times until chieved the stisfctory results [10].The network is trined using Lvemerg-Mrqurdt lgorithm nd the performnce of the network is mesured using MSE nd R. The est vlidtion performnce of Model 1 is 0.17928 t epoch 70 s shown in Figure 8() while 2.6521e-10 t epoch 1000 for Model 2 in Figure 8(). Thus it is cler tht Model 2 provide etter result in term of MSE compre to Model 1. However the stop epoch of Model 2 is higher thn Model 1.

1298 M. H. Jli et l Both results show tht the MSE hs decresed rpidly long the epochs during trining. The regression for Model 1 is 0.989 while 1.000 for Model 2. Figure 9() shows tht Model 2 produces etter curve fitness for trining, test nd vlidtion dt round compre to Model 1 s shown in Figure 9() [9][13]. Figure 8. () Best vlidtion performnce Model 1, () Best vlidtion performnce Model 2 Error sizes re well distriuted if most error pproching zero vlues tht mke the trined model performs etter. Figure 10() shows tht Model 1 hs mximum instnce round 450 of MSE distriuted round the zero line of the error histogrm while Figure 10() shows tht Model 2 hs mximum instnce round 700 of MSE distriuted round zero line [13]. This histogrm could e interpreted with error fluctution digrm long the zero vlues s shown in Figure 11(). It descried tht Model 1 exhiit lrge vlue of fluctution from the zero vlues compred to Model 2. Fig. 11() shows comprison etween prediction output from the trined network nd trget torque for Model 1 nd Model 2. The prediction output for Model 2 hs solutely spectculr greement with the trget output compre to Model 1tht hs only firly good greement with the chrcteristics with the trget dt [4]. Tle 2 summrize the performnce result of oth model.

Comprtive Study of EMG sed Joint Torque Estimtion ANN Models 1299 Figure 9. () Regression of the trined Model 1, () Regression of the trined Model 2 Figure 10. () Error histogrm of the trined Model 1, () Error histogrm of the trined Model

1300 M. H. Jli et l Figure.11 () Error fluctution long zero line, () Trget vs trined network output Tle 2. Summry of the result BPNN Structure Model 1 Model 2 Stop Epochs 70 1000 Regression 0.9891 1 MSE 0.17928 2.6521e-10 Time Elpsed (Seconds) 0.04 1.00 Conclusions Bsed on the result, it cn e concluded tht the ANN Model 2 hs etter performnce in term of MSE nd R compre to ANN Model 1. Model 2 hs super low MSE vlue s well s perfect result for R. However the stop epochs nd time elpsed for Model 2 is worse thn Model 1. The computtionl time for Model 2 consume longer time which is 1.00 seconds compre to Model 1 tht required only 0.04 seconds. Therefore there is trdeoff etween these models. In order to improve the Model 2 stop epochs nd time elpsed, evolution neurl network trining lgorithm such s genetic lgorithm nd prticle swrm optimiztion cn e implemented to produce good men squred error nd regression performnce result while reducing the computtionl time. Model 2 is considering good performnce s it shows tht this ANN model cn well represent the reltionship etween EMG signls nd elow joint torque. Hence this model cn e used for motor torque control of the rm rehilittion devices. Acknowledgements The uthors would like to thnks Universiti Teknikl Mlysi Melk (UTeM) nd Ministry of Eduction, Mlysi for the finncil supports given through Reserch Grnt.

Comprtive Study of EMG sed Joint Torque Estimtion ANN Models 1301 References [1] S. D. C. G. C. Louise Ad.: Strengthening Interventions Increse Strength nd Improve Activity After Stroke: A Systemtic Review. Austrlin Journl of Physiotherpy, vol. 52, pp. 241-248, 2006. [2] B. B.: Adult Hemiplegi: Evlution nd tretment..oxford, Butterworth- Heinemnn., 1990. [3] J. Muthuswmy.: Biomedicl Signl Anlysis in Stndrd Hndook Of Biomedicl Engineering And Design. Mc-Grw Hill, 2004, pp. 18.1-18.30. [4] Li Dpeng, Zhng Yxiong.: Artificil Neurl Network Prediction of Angle Bsed on Surfce Electromyogrphy. Interntionl Conference on Control, Automtion nd Systems Engineering (CASE).pp. 1-3, 2011. [5] M. B. I. Rez, M. S. Hussin, nd F. Mohd-Ysin.: Techniques of EMG Signl Anlysis: Detection, Processing, Clssifiction nd Applictions. Biologicl Procedures Online, vol. 8, pp. 11-35, 2006. [6] Morit, S. Kondo, T. Ito, K.: Estimtion of Forerm Movement from EMG Signl nd Appliction to Prosthetic Hnd Control, IEEE Interntionl Conference on Rootics nd Automtion(ICRA).vol.4.pp.3692-36972. 2001. [7] Kent, L.M. Siegler, S. Guez, A. ; Freedmn, W..: Modelling Of Muscle EMG To Torque By The Neurl Network Model Of Bckpropgtion. IEEE Interntionl Conference of the Engineering in Medicine nd Biology Society, Proceedings of the Twelfth Annul. Pp 1477-1478.1990. [8] Fvieiro, G.W. Blinot, A. Brreto, M.M.G.: Decoding Arm Movements y Myoeletric Signls nd Artificil Neurl Networks. Conference of Biosignls nd Biorootics(BRC). pp 1-6. 2011. [9] Neem, U.J. Adullh, A.A. ; Cihu Xiong.: Estimting humn rm's muscle force using Artificil Neurl Network. IEEE Interntionl Symposium on Medicl Mesurements nd Applictions Proceedings (MeMeA), pp 1-6.2012 [10] Ahsn, M.R. ; Irhimy, M.I. ; Khlif, O.O. EMG Motion Pttern Clssifiction through Design nd Optimiztion of Neurl Network. Interntionl Conference on Biomedicl Engineering (ICoBE), pp 175-179.2012 [11] B. Hudgins, P. Prker, nd R. Scott.: A new strtegy for multifunction myoelectric control. IEEE Trns. Biomed. Eng., vol. 40, no. 1, pp. 82 94, Jn. 1993. [12] Mrs, P, Chen JR, Nmir, R Lerning Algorithms.: Theory nd pplictions in signl processing, control nd communictions. CRC 1996. [13] Supeni, E. E. nd Eprchchi, J. A. nd Islm, M. M. nd Lu, K. T. (2013) Development of rtificil neurl network model in predicting performnce of the smrt wind turine lde. In: 3rd Mlysin Postgrdute Conference (MPC 2013), 4-5 Jul 2013, Sydney, Austrli.

1302 M. H. Jli et l