Luís Ca los Rod igues Mo ei a
Assis -As-Needed EMG-based Con ol S a egy o
Wea able Powe ed Assis i e De ices
Disse ação de Mes ado
Mes ado In eg ado em Engenha ia Biomédica
Ramo Ele ónica Médica
T abalho ealizado sob a o ien ação de
P o esso a Dou o a C is ina P. San os, Uni e sidade do Minho
Dou o a Joana So ia Campos Figuei edo, Uni e sidade do
Minho
Dou o a Elena Ga cía A mada, Consejo Supe io de
In es igaciones Cien í icas – Uni e sidad Poli écnica de Mad id
Dezemb o de 2019
Uni e sidade do Minho
Escola de Engenha ia
ii
DIREITOS DE AUTOR E CONDIÇÕES DE UTILIZAÇÃO DO TRABALHO POR TERCEIROS
Es e é um abalho académico que pode se u ilizado po e cei os desde que espei adas as eg as e
boas p á icas in e nacionalmen e acei es, no que conce ne aos di ei os de au o e di ei os conexos.
Assim, o p esen e abalho pode se u ilizado nos e mos p e is os na licença abaixo indicada.
Caso o u ilizado necessi e de pe missão pa a pode aze um uso do abalho em condições não
p e is as no licenciamen o indicado, de e á con ac a o au o , a a és do Reposi ó iUM da Uni e sidade
do Minho.
Licença concedida aos u ilizado es des e abalho
iii
ACKNOWLEDGMENTS
Fi s ly, I wan o hank my pa en s and b o he o all he suppo , condi ions, com o and
emo ional s eng h ha was gi en o me o e he las i e yea s. All he academic success was achie ed
because o you p esence.
I would like o hank my ad iso , P o esso C is ina P. San os, o he oppo uni y o wo k on his
p ojec , o all he guidance and mo i a ional con e sa ions owa ds he achie emen o he main goals.
To my co-wo ke , Doc o Joana Figuei edo, I wan o hank o all he a ailabili y, o gi e me he
bes ad ices, o being a cons an suppo along his disse a ion and o all he pa ience. You a e an
example o ollow.
Thanks o my lab colleagues o all he good daily disposi ion, co ees, game o ca ds and o all
he delicious Wednesday’ snacks.
To he bes iends ha B aga ga e me, João Mendes Lopes, C is iana Pinhei o, Ma ga ida
Machado, Ana Pe ei a, Ca la Pe ei a, Luciana Meneses and Ped o Mou a. We s a ed oge he and we
will con inue always oge he . When and whe e is he nex dinne ? You a e g ea people!
To all my musician iends o Banda Musical e Cul u al da Vila de Rio de Moinhos, hank you o
all he momen s ou o he wo k, o all he social mee ings, ba becues and good music ha we made
oge he .
Las bu no leas , I ha e o hank my gi l iend, Ma iana Cos a, o being p esen in all he mos
impo an momen s o my li e, o gi ing he suppo and he s eng h o ne e gi e up and o all he
imes ha you said “Fo ça Engenhei inho!”. You we e essen ial in he success o his jou ney.
Thank you e y much,
Luís Mo ei a
i
STATEMENT OF INTEGRITY
I he eby decla e ha ing conduc ed his academic wo k wi h in eg i y. I con i m ha I ha e no used
plagia ism o any o m o undue use o in o ma ion o alsi ica ion o esul s along he p ocess leading
o i s elabo a ion.
I u he decla e ha I ha e ully acknowledged he Code o E hical Conduc o he Uni e si y o Minho.
RESUMO
A assis ência e eabili ação obó ica usando disposi i os de assis ência a i os es í eis (WPADs),
como o ó eses e exosquele os, em c escido na á ea da eabili ação com o im de ecupe a e aumen a
a unção mo o a de sujei os com al e ações neu ológicas. Es es disposi i os de em o nece uma
assis ência pe sonalizada, uma ez que a condição ísica e a adiga muscula a iam de pacien e pa a
pacien e. Nes a á ea, sinais de ele omiog a ia (EMG) êm sido usados pa a con ola WPADs, dada a
sua capacidade de in e i a in enção de mo imen o do u ilizado . Con udo, em casos de de iciência
mo o a, os sinais de EMG ap esen am meno ampli ude quando compa ados com sinais de EMG em
condições saudá eis e, po an o, o uso de WPADs ge idos po sinais de EMG pode não o e ece a
assis ência que o pacien e necessi a.
O p incipal obje i o des a disse ação isa o desen ol imen o de uma es a égia de con olo
baseada em EMG capaz de o nece assis ência quando necessá io, pa a u u a in eg ação num sis ema
o ó ico a i o e in eligen e (Sma Os). Pa a a ingi es e obje i o o am desen ol idos e alidados os
seguin es elemen os: (i) sis ema de EMG pa a adqui i sinais de a i idade muscula dos músculos mais
ele an es no mo imen o da a iculação do o nozelo; (ii) e amen a de
machine lea ning
pa a es imação
do biná io da a iculação do o nozelo pa a se i como e e ência na es a égia de con olo; e (iii)
e amen a de es imação do biná io eal do o nozelo conside ando sinais de EMG dos músculos
Tibialis
An e io
(TA)
e
Gas ocnemius La e alis
(GASL) e ângulo eal do o nozelo.
O sis ema de EMG ap esen ou co elações sa is a ó ias com um sis ema come cial. O biná io de
e e ência pa a o o nozelo oi ge ado com base no ângulo de e e ência da mesma a iculação,
elocidade de ma cha (de 1 a é 4 km/h) e dados an opomé icos (al u as de 1.51 m a é 1.83 e massas
de 52.0 kg a é 83.7 kg), usando cinco algo i mos de
machine lea ning
:
Suppo Vec o Machine
,
Random
Fo es
,
Mul ilaye Pe cep on
,
Long-Sho Te m Memo y
e
Con olu ional Neu al Ne wo k
. CNN ap esen ou
a melho pe o mance, p e endo biná ios de e e ência do o nozelo com um i en e 74.7 e 89.8 % e
No malized Roo Mean Squa e E o s (NRMSE) en e 3.16 e 8.02 %. A es ima i a do o que com base
em sinais de EMG eque a inclusão de um maio núme o de músculos, uma ez que sinais de EMG dos
músculos TA e GASL não o am su icien es.
PALAVRAS-CHAVE
Disposi i os de Assis ência A i os Ves í eis, Es a égia de Con olo baseada em EMG, Modelos de
Reg essão, Ele omiog a ia, Reabili ação da Ma cha

i
ABSTRACT
Robo ic-based gai ehabili a ion and assis ance using Wea able Powe ed Assis i e De ices
(WPADs), such as o hosis and exoskele ons, has been g owing in he ehabili a ion a ea o eco e and
augmen he mo o unc ion o neu ologically impai ed subjec s. These WPADs should p o ide a
pe sonalized assis ance, since physical condi ion and muscula a igue modi y om pa ien o pa ien . In
his ield, elec omyog aphy (EMG) signals ha e been used o con ol WPADs gi en hei abili y o in e
he use ’s mo ion in en ion. Howe e , in cases o mo o disabili y condi ions, EMG signals p esen lowe
magni udes when compa ed o EMG signals unde heal hy condi ions. Thus, he use o WPADs managed
by EMG signals may no ha e po en ial o p o ide he assis ance ha he pa ien equi es.
The main goal o his disse a ion aims he de elopmen o an Assis ed-As-Needed (AAN) EMG-
based con ol s a egy o a u u e inse ion in a Sma Ac i e O ho ic Sys em (Sma Os). To achie e his
goal, he ollowing elemen s we e de eloped and alida ed: (i) an EMG sys em o acqui e muscle ac i i y
signals om he mos ele an muscles du ing he mo ion o he ankle join ; (ii) machine lea ning-based
ool o ankle join o que es ima ion o se e as e e ence in he AAN EMG-based con ol s a egy; and
(iii) a ool o eal EMG-based o que es ima ion using
Tibialis An e io
(TA) and
Gas ocnemius La e alis
(GASL) muscles and eal ankle join angles.
EMG sys em showed sa is ac o y pa e n co ela ions wi h a comme cial sys em. The e e ence
ankle join o que was gene a ed based on p edic ed e e ence ankle join kinema ics, walking speed
in o ma ion ( om 1 o 4 km/h) and an h opome ic da a (body heigh om 1.51 m o 1.83 m and body
mass om 52.0 kg o 83.7 kg), using i e machine lea ning algo i hms: Suppo Vec o Reg ession (SVR),
Random Fo es (RF), Mul ilaye Pe cep on (MLP), Long-Sho Te m Memo y (LSTM) and Con olu ional
Neu al Ne wo k (CNN). CNN p o ided he bes pe o mance, p edic ing he e e ence ankle join o que
wi h i ing cu es anging om 74.7 o 89.8 % and No malized Roo Mean Squa e E o s (NRMSEs)
be ween 3.16 and 8.02 %. EMG-based o que es ima ion bene icia es o a highe numbe o muscles,
since EMG da a om TA and GASL a e no enough o es ima e he eal ankle join o que.
KEYWORDS
Wea able Powe ed Assis i e De ices, EMG-based Con ol S a egy, Reg ession Models,
Elec omyog aphy, Gai Rehabili a ion
ii
CONTENTS
Acknowledgmen s ............................................................................................................................... iii
Resumo...............................................................................................................................................
Abs ac .............................................................................................................................................. i
Lis o Figu es ...................................................................................................................................... x
Lis o Tables ..................................................................................................................................... xii
Lis o Abb e ia ions and Ac onyms ................................................................................................... xi
Chap e 1 – In oduc ion ..................................................................................................................... 1
1.1. Mo i a ion ........................................................................................................................... 2
1.2. P oblem S a emen ............................................................................................................. 3
1.3. Goals and Resea ch Ques ions ............................................................................................ 3
1.4. Con ibu ions ...................................................................................................................... 5
1.5. Thesis Ou line ..................................................................................................................... 6
Chap e 2 – S a e o he A ................................................................................................................ 7
2.1. Physiological Aspec s .......................................................................................................... 7
2.2. Elec omyog aphic Signals ................................................................................................... 8
2.2.1. O e iew ..................................................................................................................... 8
2.2.2. Comme cial Sys ems ................................................................................................... 9
2.3. Assis i e Con ol S a egies ................................................................................................ 10
2.3.1. EMG-based Con ol S a egies .................................................................................... 10
iii
2.3.2. AAN EMG-based Con ol S a egies ............................................................................ 13
2.3.3. Discussion ................................................................................................................. 15
2.4. EMG-based To que Es ima ion ........................................................................................... 16
2.4.1. P opo ional Gain Me hods ......................................................................................... 16
2.4.2. Musculoskele al Models ............................................................................................. 17
2.4.3. Empi ical Me hods ..................................................................................................... 19
2.4.4. Discussion ................................................................................................................. 23
2.5. Gene al Conclusions .......................................................................................................... 24
Chap e 3 – Sys em O e iew ........................................................................................................... 25
3.1. Sma Os Desc ip ion ......................................................................................................... 25
3.2. P oposed AAN EMG-based Con ol S a egy ....................................................................... 27
3.3. Wi ed EMG Acquisi ion Sys em .......................................................................................... 28
3.3.1. Ha dwa e Speci ica ions ............................................................................................ 29
3.3.2. Expe imen al Valida ion P o ocol ................................................................................ 32
3.3.3. EMG Signal P ocessing .............................................................................................. 33
3.3.4. Resul s and Discussion .............................................................................................. 33
3.4. Gene al Conclusions .......................................................................................................... 35
Chap e 4 – Ankle Kinema ics T ajec o y Gene a ion ......................................................................... 36
4.1. In oduc ion....................................................................................................................... 36
4.2. Me hods ............................................................................................................................ 38
4.2.1. Da a Acquisi ion ........................................................................................................ 38
4.2.2. Reg ession Model Implemen a ion ............................................................................. 40
4.2.3. Reg ession Model E alua ion Me ics ......................................................................... 43
4.3. Resul s and Discussion ...................................................................................................... 44
4.4. Gene al Conclusions .......................................................................................................... 48
Chap e 5 – Ankle Kine ics T ajec o ies Gene a ion ........................................................................... 50
5.1. In oduc ion....................................................................................................................... 50
5.2. Me hods ............................................................................................................................ 51
5.2.1. Reg ession Models..................................................................................................... 51
ix
5.2.2. Da a P epa a ion ....................................................................................................... 57
5.2.3. Machine Lea ning E alua ion Me ics ......................................................................... 59
5.3. Resul s .............................................................................................................................. 59
5.3.1. Suppo Vec o Reg ession ......................................................................................... 59
5.3.2. Random Fo es .......................................................................................................... 60
5.3.3. Mul ilaye Pe cep on ................................................................................................. 61
5.3.4. Long Sho -Te m Memo y Neu al Ne wo k .................................................................. 63
5.3.5. Con olu ional Neu al Ne wo k .................................................................................... 65
5.4. Discussion and Gene al Conclusions.................................................................................. 67
Chap e 6 – AAN EMG-based Con ol S a egy ................................................................................... 70
6.1. In oduc ion....................................................................................................................... 70
6.2. Re e ence Ankle Join To que P edic ion ............................................................................ 71
6.3. EMG-based Real Join To que Es ima ion ........................................................................... 73
6.3.1. Model P esen a ion .................................................................................................... 73
6.3.2. Model Adap a ion ....................................................................................................... 74
6.3.3. Model Valida ion ........................................................................................................ 76
6.4. Gene al Conclusions .......................................................................................................... 78
Chap e 7 – Conclusions .................................................................................................................. 79
7.1. Fu u e Wo k ...................................................................................................................... 82
Re e ences ....................................................................................................................................... 84
x i
QF
Quad iceps Femo is
R Co ela ion Coe icien
R2
Coe icien o De e mina ion
RBFNN Radial Basis Func ion Neu al Ne wo k
RF Random Fo es
RF
Rec us Femo is
RMSE Roo Mean Squa e E o
RMSJ Roo Mean Squa e Je k
RQ Resea ch Ques ion
sEMG Su ace EMG
SM
Semimemb anosus
Sma OS Sma Ac i e O ho ic Sys em
SOL
Soleus
ST
Semi endinosus
SVR Suppo Vec o Reg ession
TA
Tibialis An e io
TF
Tenso Fasciae
TS Te minal S ance
TSw Te minal Swing
VI
Vas us In e medius
VL
Vas us La e alis
m Maximal speed o he muscle
VM
Vas us Medialis
WPAD Wea able Powe ed Assis i e De ice
γ
Op imal Fibe Leng h

1
CHAPTER 1 – INTRODUCTION
This disse a ion p esen s he wo k de eloped in he scope o he i h yea o he In eg a ed
Mas e ’s in Biomedical Enginee ing du ing he academic yea o 2018/19.
Du ing he i s semes e , skills in he ield o he human gai cycle we e achie ed a Ma si Bionics
S.L. in Mad id, Spain, in eg a ed in o an ERASMUS placemen p og am. Wi h his expe ience, i was
possible o lea n compe ences ega ding he human gai pa e n based on he analysis o i s main
cha ac e is ics and s a egies. As a esul , a hyb id dynamic and kinema ic model o simula e a heal hy
human gai cycle was cons uc ed, ed and alida ed wi h da a collec ed a he gai labo a o y o he
company.
The wo k p esen ed in his disse a ion was de eloped du ing he second semes e and he
concep s acqui ed in he i s semes e we e undamen al o achie e he main goals o his p ojec . This
disse a ion was de eloped a BiRD LAB (Biomedical Robo ic De ices Labo a o y) o he Cen e o
Mic oElec oMechanical Sys ems (CMEMs), a Uni e si y o Minho, B aga, Po ugal. This disse a ion
add esses he de elopmen o a con ol s a egy o pe sonalized human gai ehabili a ion wi h a
Wea able Powe ed Assis i e De ice (WPAD). To achie e his goal, an Assis ed-As-Needed (AAN) EMG-
based con ol s a egy was p ojec ed, combining concep s o machine lea ning, o p edic he e e ence
gai kinema ics and kine ics o ien ed o he use , wi h he de elopmen o musculoskele al models o
de e mine he eal gai kinema ics and kine ics o he use in eal- ime.
2
1.1. Mo i a ion
Acco ding o [1], s oke e en s co espond o he second leading cause o dea h and he hi d
leading cause o disabili y in he wo ld. The phenomenon behind hese episodes is ela ed o he absence
o oxygen in he b ain issue due o a up u e o an a e y in he b ain (hemo hagic s oke) o due o a
blocking in he blood low (ischemic s oke) [2]. Based on [3], 63% o he s oke su i o s canno walk
wi hou ex e nal suppo , no being able o pe o m hei daily li e ac i i ies. Consequen ly, he pa ien ’s
quali y li e is a ec ed due o social and wo k exclusion, cos ly medical assis ance and ea ly e i emen
[4].
The hemipa esis and hemiplegia a e he wo majo consequences de i ed om a s oke.
Hemipa e ic pa ien s exhibi weakness on one side o he body, whe eas hemiplegic pa ien s p esen
comple e pa alysis on one side o he body. Wi h his in o ma ion, he esidual lowe limb muscle o ce
o he hemipa e ic pa ien s can be conside ed in o hei gai ehabili a ion [5].
The lowe limbs ehabili a ion has been changed in he las yea s, o deal wi h (i) he disad an ages
associa ed wi h he in e - and in a- he apis a iances; (ii) he dependency o he malleabili y o he
pa ien ’s join (commonly a ec ed by spas ici y); and (iii) he absence o p ecise and epea able
mo emen s du ing he apy. Fo his pu pose, in he ehabili a ion a ea, he obo ic assis ance in eg a ing
WPADs, such as o hosis and exoskele ons, has s eadily gained impo ance [6].
The i s gene a ion o WPADs o lowe limbs, in eg a ing ajec o y acking con ol s a egies, is
esponsible o conce n a cyclic pa e n o he use , based on p e-p og ammed ajec o ies [7]. Howe e ,
in hese cases, he pa ien pa icipa ion is educed, since he e o equi ed by he use o pe o m he
walking mo ion is small. Acco ding o [8], [9], he ehabili a ion p ocess is mo e e icien i he
encou agemen o he pa ien pa icipa ion is achie ed. On he o he side, i was al eady p o ed ha hese
ajec o y acking s a egies, when applied o hemipa e ic pa ien s, a e no e icien in he ehabili a ion
con ex [10], [11]. In such ci cums ances, he incapaci y le el a ies om pa ien o pa ien and i also
a ies du ing he ehabili a ion p ocess. F om his pe spec i e, since he physical condi ion and muscula
a igue comp ehend di e en aspec s ha modi y om pa ien o pa ien , a pe sonalized assis ance
should be p o ided [12].
A human-machine in e ac ion has been highligh ed by applying bioinspi ed con ol a chi ec u es
wi h use -o ien ed assis i e con ol s a egies in eg a ed in WPADs. Mo eo e , hese con ol s a egies a e
designed o conside in o ma ion abou he use ’s mo o condi ion and mo ion in en ion, by using bio-
signals acqui ed wi h biomedical senso s [13]. In his con ex , Elec omyog aphy (EMG) signals ha e been
3
widely used since hey can o ecas in o ma ion ela ed o he use ’s mo ion in en ion, namely 20 – 100
ms be o e he use ’s lowe limb mo ion [6], [14], [15]. In his sense, he p o ision o unc ional assis ance
is possible i EMG-based assis i e con ol s a egies a e implemen ed in o WPADs, a oiding he muscle
a ophy.
1.2. P oblem S a emen
No wi hs anding he EMG-based con ol assis i e s a egies help o a oid muscle a ophy, his
s a egy is des ina ed o ollow he in en ions o he use . Howe e , in cases o impai men s o he lowe
limbs, he muscle weakness esul s in lowe EMG signals when compa ed o EMG signals om heal hy
subjec s [16]. Consequen ly, he use o EMG-based con ol s a egies o manage he assis ance deli e ed
by WPADs may no p o ide he assis ance ha he pa ien needs o walk [17]. To comba his
phenomenon, while conside ing he pa ien mo ion in en ion, s udies ha e p oposed AAN con ol
s a egies o p o ide he assis ance needed by he pa ien o pe o m he walking mo ion [6].
The con en ional AAN s a egies conside he ajec o y o he use ’s lowe limbs and a p ede ined
e e ence ajec o y, no conside ing he human condi ion and mo ion in en ion [6]. Due o his, he e a e
di icul ies ela ed o (i) he synch onism be ween he e e ence ajec o y and he use ’s mo emen ; and
(ii) he adap a ion o he e e ence ajec o y acco ding o he use -speci ic needs and in en ions [18].
These d awbacks can be o e passed using EMG signals, aking ad an age o hei an icipa o y
pe o mance, by he cons uc ion o an AAN EMG-based con ol s a egy. Mo eo e , he EMG signals can
be use ul as a me ic o he muscle weakness and, hus, when in eg a ed in AAN con ol s a egies, a
pe sonalized assis ance could be p o ided [13].
This disse a ion explo es he po en ial o AAN s a egies based on EMG signals o a u u e
in eg a ion in o a WPAD, in o de o p o ide a pe sonalized assis ance in eal- ime, conside ing he
physical condi ion and he mo ion in en ion o each use .
1.3. Goals and Resea ch Ques ions
The ul ima e goal o his disse a ion aims he de elopmen o an AAN EMG-based con ol s a egy
owa ds he pe sonalized ehabili a ion o he ankle join using Sma Ac i e O ho ic Sys em (Sma Os).
The s a egy should be adap ed o each use , p o iding only he equi ed assis ance in eal- ime.
The de elopmen o EMG-based con ol s a egies equi es he con e sion o EMG signals in o join
o ques, pa icula ly ankle join o ques [14], [19]. On he o he side, he join o ques a e dependen on
4
he join kinema ics and, hus, o pe o m he ankle join o ques p edic ion o speci ic subjec s, he ankle
join kinema ics should be well adap ed. Fo his pu pose and conside ing he main goal o his p ojec ,
se e al objec i es we e es ablished:
• Objec i e 1: To pe o m a li e a u e sea ch o collec he main con ol s a egies al eady
de eloped o assis and o es o e he lowe limbs unc ions, using EMG signals. Pe o m
a li e a u e sea ch o collec he mos ele an wo ks o con e EMG signals in o join
o que alues. This objec i e is add essed in Chap e 2.
• Objec i e 2: To de elop an EMG sys em o moni o he muscula ac i i y o lowe limb
muscles, pa icula ly he mos impo an muscles esponsible o he ankle mo ion. The
sys em mus co ec ly de ec he muscle ac i a ions wi h low noise and wi h ew mo ion
a i ac s. Valida e he e ec i eness o he de eloped EMG sys em wi h a comme cial
solu ion. This objec i e is add essed in Chap e 3.
• Objec i e 3: To implemen and alida e an e ec i e me hod o es ima e a e e ence join
posi ion ajec o y, namely ankle join kinema ics, ackling he a iabili y o he walking
speed and subjec an h opome ic da a. This objec i e is add essed in Chap e 4.
• Objec i e 4: To de elop and alida e an au oma ic and accu a e machine lea ning-based
me hod o es ima e a e e ence join o que ajec o y, conside ing he use -o ien ed
e e ence join kinema ics and he a iabili y o he walking speed and subjec
an h opome ic da a. This objec i e is add essed in Chap e 5.
• Objec i e 5: To implemen and alida e an e icien me hod o es ima e he eal use ’s join
o que, based on EMG signals and join angles owa ds he implemen a ion o he AAN
EMG-based s a egy. This objec i e is add essed in Chap e 6.
Wi h his disse a ion, ou Resea ch Ques ions (RQs) we e iden i ied and answe ed, in o de o
comple e he main challenges o he p ojec :
• RQ1: Which a e he con ibu ions and he main di e ences o he EMG-based con ol and
he AAN EMG-based con ol s a egies?
• RQ2: Is i possible o ob ain join o que measu es only using EMG signals?
• RQ3: Is i possible o p edic e e ence walking kinema ics and kine ics ajec o ies elying
exclusi ely on he walking speed and an h opome ic da a?
5
• RQ4: Can EMG-based o que es ima ion s a egy p esen a good pe o mance?
1.4. Con ibu ions
The main con ibu ions o his disse a ion a e:
• A desc ip i e li e a u e e iew epo ing he assis i e con ol s a egies based on EMG signals and
in eg a ed in o WPADs;
• A desc ip i e li e a u e e iew epo ing he EMG-based o que es ima ion me hods;
• A ool o es ima e e e ence ankle join posi ion ajec o ies based on he walking speed and
an h opome ic da a;
• A machine lea ning-based me hod o e e ence ankle join o que es ima ion conside ing use -
o ien ed e e ence ankle join kinema ics ajec o ies, walking speed and an h opome ic da a;
• A me hod o pe o m an EMG-based o que es ima ion using EMG signals and eal ankle join
posi ion ajec o ies, owa ds he implemen a ion in o a WPAD des ina ed o ankle join
assis ance and ehabili a ion.
Fu he mo e, he de eloped wo k allowed he publica ion o ou con e ence pape s:
• Mo ei a, L., Pinhei o, C., Lopes, J. M., Sanz-Me odio, D., Figuei edo, J., San os, C. P., & Ga cia,
E. (2019). S udy o Gai Cycle Using a Fi e-Link In e ed Pendulum Model: Fi s De elopmen s.
In 2019 IEEE 6 h Po uguese Mee ing on Bioenginee ing (ENBENG) (pp. 1–4). IEEE.
h ps://doi.o g/10.1109/ENBENG.2019.8692451
• Lopes, J. M., Mo ei a, L. Pinhei o, C., Sanz-Me odio, D., Figuei edo, J., San os, C. P., & Ga cia,
E. (2019). Th ee-Link In e ed Pendulum o Human Balance Analysis: A P elimina y S udy. In
2019 IEEE 6 h Po uguese Mee ing on Bioenginee ing (ENBENG) (pp. 1–4). IEEE.
h ps://doi.o g/10.1109/ENBENG.2019.8692531
• Pinhei o, C., Lopes, J. M., Mo ei a, C., Sanz-Me odio, D., Figuei edo, J., San os, C. P., & Ga cia,
E. (2019). Kinema ic and kine ic s udy o si - o-s and and s and- o-si mo emen s owa ds a
human-like skele al model. In 2019 IEEE 6 h Po uguese Mee ing on Bioenginee ing (ENBENG)
(pp. 1–4). IEEE. h ps://doi.o g/10.1109/ENBENG.2019.8692569
• Fe nandes, P. N., Figuei edo, J., Mo ei a, L., Félix, P., Co eia, A., Mo eno, J. C., & San os, C. P.
(2019). EMG-based Mo ion In en ion Recogni ion o Con olling a Powe ed Knee O hosis. In
2019 IEEE In e na ional Con e ence on Au onomous Robo Sys ems and Compe i ions (ICARSC)
(pp. 1–6). IEEE. h ps://doi.o g/10.1109/ICARSC.2019.8733628

6
• Mo ei a, L., Figuei edo, J., Ga cia, E. & San os, C. P. (2020). Myoelec ic Con ol S a egies
Applied in Powe ed Lowe Limb Assis i e De ices: A Re iew. In Robo ics and Au onomous
Sys ems (Submi ed).
1.5. Thesis Ou line
This disse a ion is o ganized in he ollowing se en chap e s.
Chap e 2 p esen s he s a e o he a add essed o ou main poin s: (i) he main muscles
esponsible o he lowe limbs join mo ion; (ii) he mos used EMG sys ems o acqui e EMG signals; (iii)
he EMG-based con ol and AAN EMG-based con ol s a egies al eady implemen ed and in eg a ed in o
WPADs o es o e he lowe limbs unc ions; and (i ) he EMG-based o que es ima ion me hods al eady
de eloped.
Chap e 3 exhibi s an insigh abou Sma Os a chi ec u e, p esen ing in o ma ion abou i s
cons i uen ha dwa e and so wa e. I is also p esen ed he p oposed AAN EMG-based con ol s a egy
and he chap e ends wi h he ha dwa e o he EMG sys em de eloped in his disse a ion.
Chap e 4 p esen s a eg ession model o gene a e e e ence ankle join kinema ics, based on
walking speed and an h opome ic da a. Addi ionally, i desc ibes he da a acquisi ion p o ocol o alida e
all he algo i hms de eloped in his disse a ion.
In Chap e 5, machine lea ning algo i hms a e de eloped and alida ed o p edic he ankle join
o ques, ha will be used as a e e ence pa ame e on he con ol s a egy p oposed in Chap e 3.
Chap e 6 p esen s he alida ion o he esul s ob ained in Chap e 4 and Chap e 5, whe e he
ou pu esul s ob ained in Chap e 4 se e as inpu da a o he bes algo i hm de eloped in Chap e 5.
Mo eo e , he chap e desc ibes he implemen ed algo i hm esponsible o EMG-based eal join o que
es ima ion and p esen s i s alida ion.
Las ly, Chap e 7 p esen s he main conclusions o his mas e disse a ion, he esea ch ques ions
a e answe ed and opics o a u u e wo k a e p oposed.
7
CHAPTER 2 – STATE OF THE ART
This chap e begins wi h a b ie desc ip ion o he heal hy and he pa hological walking mo ion,
iden i ying he main muscles a ec ed by a s oke e en and he main muscles esponsible o he mo ion
o he knee and ankle join s. Subsequen ly, comme cial EMG sys ems a e p esen ed, as well as hei
main cha ac e is ics. I is ollowed by an exhaus i e e iew o he con ol s a egies al eady de eloped
and applied in o WPADs, using EMG signals o assis and es o e he lowe limbs unc ions. A las , s udies
ela ed o he con e sion o he EMG signals in o o que alues a e also p esen ed, since mos o he
EMG-based con ol s a egies pe o m his con e sion [14], [19].
2.1. Physiological Aspec s
S oke e en s cause unc ional o neu omuscula changes, depending on he loca ion o he
a ec ed a ea o he b ain. Acco ding o [16], [20], hese changes a e ela ed o a loss o s eng h in he
hemipa e ic leg, being e i ied weakness in he lexo muscles and spas ici y in he ex enso muscles,
causing an inc ease o he join s i ness. S udies concluded ha , gene ally, in s oke su i o s, he
Soleus
(SOL) and
Gas ocnemius
(GAS) muscles ( esponsible o he plan a lexion mo ion o he ankle join ) a e
con ac ed due o spas ici y, while he
Tibialis An e io
(TA) muscle ( esponsible o he do si lexion mo ion
o he same join ) emains weak [2], [16], [20], [21]. This is he main eason o he d op oo diso de in
pa ien s ha su e ed a s oke e en [22]. Rega ding he knee join , he knee ex enso s (
Vas us Medialis
(VM),
Vas us La e alis
(VL),
Vas us In e medius
(VI) and
Rec us Femo is
(RF)) exhibi spas ici y and he
knee lexo s (
Biceps Femo is
(BF),
Semi endinosus
(ST) and
Semimemb anosus
(SM)) a e weak, being
8
e i ied a knee hype ex ension and a dec ease in he lexion mo emen o his join du ing he gai cycle
[16], [20], [21]. In his sense, conside ing he muscles a ec ed by s oke e en s and acco ding o [23],
he muscles ha p o ide mo e in o ma ion abou he ankle join mo ion a e he TA muscle o do si lexion
and he SOL o he GAS muscles o plan a lexion. Conce ning he knee join , he muscles ha p esen
mo e in o ma ion a e he BF muscle o lexion and he VM and he RF muscles o ex ension.
2.2. Elec omyog aphic Signals
2.2.1. O e iew
The con ac ion and he elaxa ion episodes o he muscles a e con olled by he ne ous sys em,
h ough elec ic signals deli e ed by he neu ons o he muscles. This elec ical ac i i y can be eco ded
using EMG, p o iding a powe ul o ecas in o ma ion ela ed o he mo ion in en ion o he use [6], [14],
[15], [24].
EMG signals can be measu ed using needles inse ed di ec ly in he muscles (in amuscula EMG
– iEMG, also known as ine-wi e EMG – EMG) o using elec odes placed on he skin (su ace EMG -
sEMG). The choice o he me hod o measu e he elec ical ac i i y o he muscles depends on he
p ope ies o he muscles o s udy. Compa ing bo h, he i s enounced me hod p esen s less c oss- alk,
since i ecei es lowe muscle ac i i ies o he muscles a ound he desi ed muscle [25]. Howe e , due o
he di icul y o inse ion, g ea e in asi eness and highe cos , his me hod is no no mally used in EMG
s udies [25], [26]. In con as , sEMG echnique is commonly used, since i is non-in asi e, p ac ical,
inexpensi e and i may be used by non-clinical assesso s. In addi ion, some s udies epo ed ha his
me hod, when applied o he SOL, GAS and TA muscles, p esen s EMG ampli udes simila o iEMG and
a negligible alue o c oss- alk [27]–[29].
Rega ding he sEMG ( e e ed as “EMG” along his disse a ion), he e a e wo con igu a ions o
adop , namely he monopola and he bipola . The monopola con igu a ion uses wo elec odes:
de ec ion elec ode placed on he skin, abo e he muscle in s udy; and a e e ence elec ode ha should
be placed on an elec ically neu al issue o a bone [30]. Howe e , acco ding o [31], his con igu a ion
is no ecommended, since he signal-noise a io and he spa ial esolu ion o he EMG signals a e
educed, when compa ed o EMG acquisi ions pe o med wi h bipola con igu a ions. On he o he side,
he bipola con igu a ion uses h ee elec odes: wo de ec ion elec odes placed abo e he a ge muscle,
dis anced by 10 - 20 mm [32], [33] and a e e ence elec ode. In his sense, he bipola con igu a ion is
9
p e e able, since he common noise o bo h de ec ion elec odes is elimina ed, inc easing he signal-noise
a io and p oducing a cleane EMG signal [33], [34].
2.2.2. Comme cial Sys ems
The cha ac e is ics o he EMG signals a e well es ablished in he li e a u e and he mos ele an
a e he ampli ude ange be ween ±10 mV, he esponse equency be ween 0 and 500 Hz wi h he
dominan ene gy be ween 50 and 150 Hz and he s ochas ic na u e, cha ac e ized by a Gaussian
dis ibu ion unc ion [35].
Delsys, No axon, Mo ionLab, Plux, BTS Bioenginee ing, Con emplas, Come a, Biopac, Shimme ,
Cadwell, Myon ec and A hos a e examples o companies ha p oduce EMG acquisi ion sys ems able o
acqui e EMG signals wi h he a o emen ioned cha ac e is ics. An EMG acquisi ion sys em can each he
ens o housands o eu os and, hus, depending on he inali y o each p ojec , i is equi ed o analyze
he speci ica ions o each sys em. Table 1 summa izes he main cha ac e is ics o he 4 mo e used
comme cial EMG acquisi ion sys ems.
Table 1. EMG sys ems speci ica ions
Delsys [36], [37]
No axon [38]
Mo ion Lab [39]
MuscleBAN [40], [41]
EMG Signal Inpu
Range
± 11 mV
± 24 mV
± 250 mV
*
Resolu ion
16 bi s
16 bi s
16 bi s
16 bi s
EMG Signal
Bandwi h
20 – 450 Hz
Minimum:
20 – 500 Hz
10 – 2000 Hz
1 – 1000 Hz
Maximum:
5 – 1500 Hz
Sampling Ra e
2000 Hz
2000 Hz o 4000
Hz
4000 Hz
4000 Hz
T ansmission Range
40 m
40 m
18 m
10 m
Gain
1
1
10 – 500
1100
Senso Size
27 x 37 x 13 mm
24 x 37 x 16 mm
38 x 19 x 9 mm
28 x 70 x 12 mm
Mass
14 g
14 g
20 g
25 g
Au onomy
2 – 3 h
8 h
*
8 h
Tempe a u e Range
5 – 45 deg ees
Celsius
0 – 38 deg ees
Celsius
20 – 40
deg ees Celsius
*
*In o ma ion no speci ied
16
Fu he mo e, he wo ks de eloped by [17], [56]–[60] e eal o be a sui able app oach o assis
and ehabili a e he knee join , when he use equi es suppo . Based on he esul s, wo si ua ions we e
e i ied: (i) i he suppo a io p o ided by he WPAD inc eases, he use ’s o que dec eases, bu he knee
join angle emains equal; and (ii) i he suppo a io dec eases, he use ’s o que inc eases and he knee
join angle emains equal once again. This indica es ha he AAN s a egy p o ides he assis ance o
pe o m he mo emen s when equi ed, o e ing g ea po en ial o mo o ehabili a ion and use ’s mo o
au onomy. None heless, he muscle model should be imp o ed and sui able o each muscle, in o de o
conside di e en ac i a ion delays o each muscle.
Based on he AAN EMG-based model s a egies al eady de eloped, i is concluded ha i he use
does no pe o m any mo emen , he exoskele on p o ides he enough amoun o o que o comple e he
desi ed knee join ajec o y. On he o he hand, i he wea e has an elec ic signal associa ed, he
exoskele on only p o ides a o que upon he use ’s o que, in o de o achie e a co ec knee join
ajec o y.
2.4. EMG-based To que Es ima ion
Mos o he implemen ed con ol s a egies in ol ed he join o que es ima ion o he eal use ’s
join o que, based on he EMG signals acqui ed om speci ic muscles. Fo his pu pose, a li e a u e
sea ch was made o e iew he me hods applied o con e he EMG signals in o o que alues, namely
p opo ional gain me hods, musculoskele al models and empi ical models, as p esen ed in Table 4.
2.4.1. P opo ional Gain Me hods
The s udies de eloped by [13], [43], [44] ollowed he same me hodology o ob ain he join
o que, based on a p opo ional gain me hod. They ca ied ou a calib a ion s ep o ind p opo ional gains
pe join , o lexion and ex ension mo ions. Fo his pu pose, he pa icipan s had o main ain a ixed join
angle, while he o hosis/exoskele on ac ua o pe o med a ixed o que alue. This p ocess implied ha
he pa icipan needed o pe o m a join o que o main ain he equi ed join angle. A he same ime,
he EMG signals we e ead and, a p opo ional gain ha ela es he EMG signals wi h he join o que was
ound. In he h ee wo ks, he es ima ed join o que was simila o he eal one. In [13], a No malized
Roo Mean Squa e E o (NRMSE) and a phase delay be ween he measu ed o que and he es ima ed
o que was we e calcula ed, achie ing alues o 12 % and 22 ms, espec i ely. In [44], he RMSE and he
Roo Mean Squa e Je k (RMSJ) be ween he es ima ed and he measu ed o que was 3.56 ± 0.63 Nm
and 2.85 ± 0.78 Nm, espec i ely.

17
2.4.2. Musculoskele al Models
The Hill- ype model p oposed by [46] and de eloped in [47] was he i s musculoskele al model
p oposed o es ima e join o que using EMG signals and join angles. Fu he a ia ions o his model
ha e been de eloped, imp o ing he calib a ion p ocess, in o de o minimize he model unce ain y. In
his calib a ion p ocess, muscle pa ame e s a e iden i ied o cons uc he model, ollowing a ade-o
be ween he accu acy o he model and i s complexi y [51].
Based on [46], [47], he wo k o [48] p edic s he knee join o que, using EMG signals om
hi een muscles and he knee join angles. Eigh een pa ame e s we e ound in he calib a ion p ocess o
cons uc he model. Among hese pa ame e s, he o ce p oduced by he muscle- endon uni (
Fm
),
depends on he penna ion angle (ɸ) ha depends on
L0
m
. Consequen ly,
L0
m
is unc ion o he pe cen age
change in op imal ibe leng h, γ. This pe cen age was a ied in his s udy o in es iga e he impo ance
o a co ec calib a ion s ep in musculoskele al models. The easibili y o he con e sion was assessed by
six subjec s, who pe o med app oxima ely 204 asks, including dynamome e , unning and sides epping
ials. Coe icien o de e mina ion (
R2
) o 0.91 ± 0.04 was ob ained be ween he knee join o que
p edic ed by he model and by in e se dynamics. A mean esidual e o below 0.2 Nm/kg no malized o
body weigh was achie ed. These esul s we e ob ained wi h γ = 15 %. When γ was se o 0 %, a
R2
o
0.85 was ob ained, showing he sensi i i y o he model o he pa ame e s o he calib a ion s ep.
Based on he wo ks de eloped by [47], [48], o he six wo ks we e de eloped aiming he
simpli ica ion o he model, he educ ion o he con e sion ime and he educ ion o he con e sion
e o s. In [62] he use mo ion in en ions o he knee join we e mapped, using EMG signals (lis ed in
Table 4) and knee join angles. A he same ime, he join angles and he g ound eac ion o ces se ed
as inpu o he se en-link biomechanical model (composed by wo legs wi h ee , shanks, highs and he
o so) o ob ain he knee join o que h ough in e se dynamics. When compa ed o [48], only wo
pa ame e s we e calib a ed (p esen ed in Table 4). The e ec i eness o he me hodology was e alua ed
in one heal hy subjec du ing s epping up s ai s and wi h lexion and ex ension mo ions o he knee join .
A good co ela ion be ween bo h join o ques was achie ed, bu he ampli udes o he cu es we e
di e en in some cases.
The s udy ad anced by [49] conside ed ewe calib a ion pa ame e s han [48], aiming he
educ ion o he ime o he EMG-based o que con e sion. Fo his pu pose, he endon s ain pa ame e
( ound in he calib a ion p ocess o [48]) was neglec ed, since his pa ame e ep esen s only 3.3 % o
he endon leng h when a maximum isome ic o ce is gene a ed. This neglec ion allows he es ima ion
18
o he
L0
m
wi hou he ime-consuming Runge-Ku a-Fehle g in eg a ion, p oposed in [48]. Conside ing he
es ima ed and he expe imen al knee join o que, a co ela ion coe icien (
R
) o 0.892 ± 0.047 and a
RMSE o 8.10 ± 1.02 Nm, when a heal hy subjec pe o med 10 gai ials. Mo eo e , he calib a ion was
comple ed in 63.4 ± 1.20 s and he join o que es ima ion in his simpli ied model ook 0.0630 s,
whe eas he comple e model [48] ook 3 h and 0.691 ± 0.0146 s o accomplish he same he calib a ion
and he es ima ion asks, espec i ely.
Inspi ed in he model de eloped by [47], in [51] he knee join o ques we e es ima ed, using he
EMG signals o six muscles (lis ed in Table 4) and he knee join angles. Eigh pa ame e s we e
de e mined in he calib a ion p ocess (also p esen ed in Table 4), since he au ho s conside ed his
numbe as an accep able adeo be ween he complexi y o he model and i s unce ain y. Expe imen s
we e ca ied ou wi h a heal hy subjec pe o ming lexion and ex ension mo emen s o he knee join . A
NRMSE be ween he o que ob ained wi h he model and he measu ed o que om in e se dynamics
was 12.4 %.
Mo eo e , in [63], a musculoskele al model was de eloped, based on he Hill- ype muscle model
de eloped by [48]. EMG da a we e eco ded om h ee muscles o a heal hy pe son, along wi h he join
angles and hey ac ed as inpu in he model. In he calib a ion s ep, ou pa ame e s we e iden i ied, as
p esen ed in Table 4. The RMSE and he
R2
we e calcula ed o e alua e he model pe o mance, ob aining
alues below 1.99 Nm and an a e age o 0.89, espec i ely.
In [44], in addi ion o he de eloped linea p opo ional model, an EMG-d i en Hill- ype
neu omuscula model was also p oposed, based on [47]. Two muscles we e used o pe o m he
con e sion and only h ee pa ame e s we e de e mined in he calib a ion p ocess. The e ec i eness o
he con e sion was es ed in eigh heal hy subjec s, pe o ming maximum isome ic olun a y
con ac ions a di e en angles. RMSE o 3.49 ± 0.57 Nm and RMSJ o 1.21 ± 0.51 Nm we e ob ained
be ween he es ima ed and he measu ed o que.
An EMG-d i en musculoskele al model was p oposed in [64], in eal- ime, o p edic he use
in en ions o he ankle and knee join mo ions. This musculoskele al model was based on [48], whe e he
EMG signals and he ankle and knee join angles a e he model’s inpu s. This me hodology was
implemen ed in a Raspbe y Pi 2 o in es iga e he compu a ional cos . Th ee di e en asks we e
pe o med by i e di e en pa icipan s: (i) walking a sel -selec ed speed, (ii) knee squa , and (iii) cal ise
mo ion. The es ima ed o ques by he musculoskele al model we e compa ed wi h he calcula ed join
o ques using in e se dynamics. The esul s o he EMG-d i en model we e p esen ed in 2.7 ± 0.48 ms.
RMSE we e always below 0.37 ± 0.12 Nm/kg and he smalles alue obse ed was 0.01 ± 0.01 Nm/kg.
19
Pea son coe icien s we e also calcula ed and he alues we e always abo e 0.43 ± 0.36, and he highes
co ela ion alue ob ained was 0.9 ± 0.07.
2.4.3. Empi ical Me hods
F om he li e a u e esea ch, i was e i ied ha empi ical me hods, such as neu o- uzzy models
[53], [55] and neu al ne wo ks [11], [65], [66] ha e been applied o es ima e he join o que based on
EMG signals.
S udy [53] only used EMG da a o eigh muscles as inpu in a neu o- uzzy model o ob ain he hip
and knee join o ques. To cons uc his model, wen y uzzy IF-THEN con ol ules we e de ined and a
alue o join o que was ma ched. In [55], he EMG signals along wi h he join angles se ed as inpu o
ob ain he join o que. In bo h wo ks is no p esen ed he compa ison be ween he eal o que and he
es ima ed o que. When pe o ming di e en mo emen s, i he EMG signals we e educed, he esea ch
concluded ha he EMG-based con olle is adap ed o he use .
A Mul ilaye Pe cep on (MLP) neu al ne wo k was de eloped in [66] o es ima e he knee join
o que. EMG da a no malized o he maximal isome ic ac i a ion, along wi h he body mass, heigh , age,
gende , join eloci y, and join posi ion we e en e ed as inpu a iables in he h ee-laye neu al ne wo k.
The second laye was composed by a a iable numbe o hidden uni s. The hi d laye es ima ed join
o que o he knee join . The esul s (
R
= 0.96) demons a ed he e ec i eness o his neu al ne wo k o
es ima e he knee join o que. La e , in [65], he au ho s imp o ed he MLP p esen ed in [66]. Fi s , he
esea ch concluded ha how many mo e muscles we e inco po a ed, be e esul s a e achie ed. Second,
be e esul s we e ob ained when no malized da a a each isome ic angle was used in aining p ocess
o he neu al ne wo k, because i was p o ed ha he cu e o he no malized join o que a each angle
is simila o he cu e o he no malized EMG. Thi d, esul s poin ed ou ha MLP neu al ne wo k has a
be e join o que es ima ion when compa ed wi h o he neu al ne wo ks, such as Fully Connec ed
Cascade. A las , i e neu ons in he second laye we e ound as he bes numbe o hidden uni s.
A RBFNN wi h a wo-s ep lea ning s a egy was de eloped by [11] o con e he EMG signals in o
o que alues. The pe o mance o he me hodology was e alua ed h ough simula ions and expe imen s
wi h ou heal hy subjec s only o he swing phase. A RMSE o 2 Nm and a
R
highe han 0.8 we e ound,
be ween he measu ed and he es ima ed join o que.
20
Table 4. Me hods o con e EMG signals in o join o que alues
S udy
Con e sion
Me hod
Used in
Con ol?
Calib a ion Pa ame e s
Inpu s
Join
Muscles
Pa icipan s
(numbe )
Resul s
[43]
P opo ional Gain
Me hod
Yes
Two gain
pa ame e s/join
EMG da a
Hip
Knee
BF, VM, GM and RF
Heal hy (1)
*
[13]
P opo ional Gain
Me hod
Yes
Two gain
pa ame e s/join
EMG da a
Knee
VL, VM, SM and ST
Heal hy (2)
NRMSE = 12 %
Phase delay = 22 ms
[44]
P opo ional Gain
Me hod
Yes
One gain
pa ame e /join
Join angles and
EMG da a
Ankle
TA and GAS
Heal hy (8)
RMSE = 3.56 ± 0.63 Nm
RMSJ = 2.85 ± 0.78 Nm
[48]
Musculoskele al
Model
No
Eigh een pa ame e s
**
Join angles and
EMG da a
Knee
Thi een muscles **
Heal hy (6)
R2 = 0.91 ± 0.04
Mean Residual E o /body weigh <
0.2 Nm/kg
[62]
Musculoskele al
Model
No
F
max
,
F
m
, and
A
Join angles and
EMG da a
Knee
*
Heal hy (*)
Good co ela ion be ween he shape
o he knee join o que ob ained by
in e se dynamics and by he EMG-
based model
[49]
Musculoskele al
Model
No
F
max
,
Fm
,
L0
m
, ɸ and
maximal speed o he
muscle (
m
)
Join angles and
EMG da a
Knee
SM, ST, BF, Sa o ius (SAR), TF,
GRA, VL, VM, VI, RF,
Gas ocnemius Medialis (GASM)
and Gas ocnemius La e alis
(GASL)
Heal hy (1)
R = 0.892 ± 0.047
RMSE = 8.1 ± 1.02 Nm
Calib a ion ime = 63.4 ± 1.20 s
Join o que es ima ion ime =
0.0630 s
[51]
Musculoskele al
Model
Yes
F
max
,
F
m
,
L0
m
, ɸ,
A
and
h ee cons an s
Join angles and
EMG da a
Knee
RF, VM, VL, BF, SM and ST
Heal hy (1)
NRMSE = 12.4 %
21
[63]
Musculoskele al
Model
No
F
max
,
F
m
,
L0
m
, ɸ
Join angles and
EMG da a
Knee
RF, VM and VL
Heal hy (1)
RMSE = 1.99
R2 = 0.89
[44]
Musculoskele al
Model
Yes
F
max
,
F
m
,
L0
m
and
A
Join angles and
EMG da a
Ankle
TA and GAS
Heal hy (8)
RMSE = 3.49 ± 0.57 Nm
RMSJ = 1.21 ± 0.51 Nm
[64]
Musculoskele al
Model
No
F
max
,
F
m
,
L0
m
and
A
Join angles and
EMG da a
Ankle
and
Knee
BF, GASL, GASM, GRA, RF, SAR,
SOL, SM, VL, VM, TA, Pe oneus
Longus (PL) and Pe oneus
Te ius (PT)
Heal hy (5)
Join o que es ima ion ime = 2.7 ±
0.48 ms
RMSE < 0.37 ± 0.12 Nm/kg
Pea son co ela ions > 0.43 ± 0.36
[53]
Empi ical Model
Yes
***
EMG da a
GRF da a
Hip
Knee
TF, RF, VL, VM, AL, GRA, BF and
ST
Heal hy (1)
*
[55]
Empi ical Model
Yes
***
Join angles and
EMG da a
Hip
Knee
TF, RF, VL, AL, GRA, VM, BF and
ST
Heal hy (3)
*
[66]
Empi ical Model
No
***
Body mass, body
heigh , age, gende ,
join eloci y, join
posi ion and EMG
da a
Knee
VL and BF
Heal hy
(20)
R = 0.96
[65]
Empi ical Model
No
***
Body mass, body
heigh , age, gende ,
join eloci y, join
posi ion and EMG
da a
Knee
VL, BF, RF, VM and ST
Heal hy (1)
• Mo e muscles imply a be e
join o que es ima ion;
• The da a should be no malized
a each isome ic angle;
• The second laye should ha e
i e neu ons.

22
[11]
Empi ical Model
Yes
***
Join angles,
eloci ies and EMG
da a
Hip
QF and BF
Heal hy (4)
RMSE < 2 Nm
R > 0.8
* No speci ied
** Consul [48] o mo e in o ma ion
*** No Requi ed
23
2.4.4. Discussion
Among he di e en s a egies, h ee me hods we e iden i ied o con e he EMG signals in o join
o que alues: P opo ional Gain Me hods, Musculoskele al Models and Empi ical Me hods.
I was epo ed by [46], [67] ha he muscle ac i a ion, he a chi ec u al and biomechanical
p ope ies o he muscle ibe s and endons a e de e minan ac o s in he o ce o muscle- endon uni s.
Fo his pu pose, in a s udy de eloped by [68], epo ed ha only he use o EMG signals in p opo ional
gain me hod is no enough o p edic he o que o he ankle join . Mo eo e , [69] epo ed ha he
p opo ional o que con ol only using EMG alues is no eliable o wo- old easons: (i) he human body
muscles and coac i a ion e ec s a e exclusi e om pe son o pe son; and (ii) he EMG eadings a e
dependen on he elec odes placemen , loca ion and skin impedance. S udy [70] showed he di e ences
be ween using only EMG signals and mo ion da a usion (EMG and o he senso s da a), concluding ha
he usion o biomechanical da a wi h EMG signals imp o es join o que es ima ion. Addi ionally, s udy
[43] combined loo eac ion o ce wi h EMG signals o imp o e he hip o que a ex ension mo ion and
consequen ly, o e come he use discom o du ing s ance phase.
Rega ding he musculoskele al models [44], [48], [49], [51], [62]–[64], he e a e se e al ac o s
o conside o ob ain a good EMG- o que con e sion, as ollows: he numbe o calib a ion pa ame e s o
ind, he numbe o muscles o acqui e he EMG signals, he complexi y and he accu acy o he model,
and he ime o pe o m he con e sion. The complexi y o he model inc eases wi h he inc emen o (i)
he numbe o calib a ion pa ame e s; and (ii) he numbe o muscles. Consequen ly, he compu a ional
bu den will inc ease. The wo ks de eloped by [48], [62] a e no capable o pe o m he con e sion in
eal- ime, in con as wi h he wo ks de eloped la e by [44], [49], [51], [63], [64]. The eal- ime
con e sion is an impo an aspec o conside when i is desi ed o assis pa ien s using EMG signals in
o que con olle s.
As al eady e e ed, he p opo ional gain and he musculoskele al me hods equi e calib a ion
s eps ha may no be ealized by pa hological indi iduals. Fo his pu pose, he empi ical models, as he
ones p esen ed in [11], [53], [55], [65], [66], can be a sui able s a egy o con e he EMG signals in o
o que alues, a oiding calib a ion s eps.
In he s udy de eloped by [65], some imp o emen s in a MLP neu al ne wo k we e done. I was
epo ed ha he numbe o muscles, he inpu s no maliza ion and he numbe o hidden uni s had a
s ong impac in he pe o mance o he con e sion. Howe e , he da ase used o cons uc he neu al
ne wo k p esen ed wen y subjec s whe e he age ange o he pa icipan s, as well as he ange o
physical ac i i y o he pa icipan s was small. Mo eo e , he EMG- o que con e sion implemen ed in bo h
24
wo ks [65], [66] is no comple ed in eal- ime. In [11], he join o que is ob ained h ough a RBFNN,
whe e he inc emen o he numbe o muscles leads good esul s wi hou calib a ion s eps. Howe e , he
e o o he o que es ima ion inc eases when he hip mo emen o he o hosis’s join changed i s
di ec ion. A las , his me hodology was only de eloped o he swing phase, equi ing mo e ad ances o
co e all he gai cycle.
2.5. Gene al Conclusions
The e is in e es in using EMG signals o con ol WPADs, since hey a e di ec ly ela ed o use
mo ion in en ion. Two ypes o con ol s a egies in eg a ing EMG da a we e iden i ied, he EMG-based
con ol and he AAN EMG-based con ol. Bo h s a egies con ibu ed o muscula ac i i y imp o emen
and endu ance. EMG-based con ol s a egies can be used o e- ain he lowe limbs o pa ien s wi h
impai men s. Howe e , i is no expec able ha hese s a egies ha e he po en ial o assis hei mo o
condi ion. AAN EMG-based con ol s a egies ha e he po en ial o assis and ehabili a e he lowe limbs,
such ha he use can achie e au onomy o pe o m hei daily li e ac i i ies, in con as o EMG-based
con ol s a egies. I is expec able ha he in eg a ion o AAN EMG-based con ol s a egies in o WPADs
may os e mo o assis ance, e- ain, ehabili a ion and au onomy when applied o disabled use s.
Howe e , alida ion es s wi h pa hological subjec s a e s ill missing o alida e his phenomenon.
Fu he mo e, un il he momen , he e is no explana ion by he exis ence o AAN EMG-based con ol
s a egies only des ina ed o he knee join .
Fu he , his e iew concluded ha mos o he con ol me hods which use EMG signals o con ol
WPADs con e hese EMG signals in o o que alues. The e is e idence ha only EMG eadings a e no
enough o ob ain he accu a e alues o join o ques, being necessa y o use EMG da a wi h
biomechanical da a, such as, join angles, eloci ies, accele a ions. Mo eo e , o pe o m he EMG-based
o que es ima ion in AAN EMG-based con ol s a egies, i was no iced ha only musculoskele al models
we e used.
O e all, his disse a ion aims he de elopmen o an AAN EMG-based con ol s a egy des ina ed
o he ankle join assis ance, using a musculoskele al model ed wi h EMG signals, join angles and
eloci ies o pe o m he EMG-based o que es ima ion, since he e we e no ound li e a u e e idences
o he use o hese s a egies o assis he e e ed join .
25
CHAPTER 3 – SYSTEM OVERVIEW
In his chap e , i is p esen ed an o e iew ela i e o he o ho ic sys em used in his disse a ion
o assis indi iduals wi h disabili ies in he ankle join . The con ol a chi ec u es al eady implemen ed in
his o ho ic sys em a e b ie ly explained and he AAN EMG-based con ol s a egy is p oposed. A las ,
he chap e ends wi h he desc ip ion o he cons uc ed EMG acquisi ion sys em, ha ep esen s an
essen ial elemen o he p oposed con ol s a egy.
3.1. Sma Os Desc ip ion
Sma Os is a WPAD embedded wi h wea able senso s des ina ed o he analysis and con ol o he
gai mo ion. Sma Os was de eloped in [71] and i p esen s he capaci y o p o iding a pe sonalized and
epe i i e gai aining, such as join ajec o y acking con ol, o assis he use acco ding wi h i s needs
and o enable he abno mal gai pa e n co ec ion, such as d op oo gai in s oke su i o s. Besides
ha , his WPAD in okes he use pa icipa ion, in o de o ehabili a e he ankle join , by conside ing he
use ’s mo ion in en ion based on muscula in o ma ion (an example o an EMG-based con ol s a egy)
and i was designed o achie e walking speeds anging om 0.5 o 1.6 km/h [71]. Mo eo e , Sma Os
also p esen a bio eedback sys em in eg a ed in i s a chi ec u e o accele a e he amilia iza ion o he
use o i sel and o accele a e he gai eco e y.
The concep ual design o Sma Os is p esen ed in Figu e 1 and a b ie explana ion o each block
is p o ided. None heless, he scope o his disse a ion is he p ojec ion o an AAN EMG-based con ol
s a egy ha has no ye been de eloped.
32
Analyzing Figu e 5, di e ences wi h espec o he a enua ion le els be ween he expe imen al
and heo e ical implemen a ions a e isible. Based on he heo e ical FR, i would be expec able o
achie e highe le els o a enua ion in he desi able ange o equencies. Howe e , despi e o lowe
le els o a enua ion ha e been e i ied expe imen ally, his phenomenon does no comp omise he
p ope unc ioning o he p ojec ed sys em. Mo eo e , he expe imen al a enua ion le els in oduced
by he il e s a equencies anging om 80 o 300 Hz a e be e han he expec ed ones, since a
hese equencies, he signal is no a enua ed.
The inal EMG sys em can be seen in Figu e 6. Since his sys em was designed o acqui e EMG
signals om only one muscle, o he boa ds mus be used o acqui e signals om o he muscles.
Thus, an inpu /ou pu supply was p ojec ed o supply ene gy be ween consecu i e boa ds.
Figu e 6. Final EMG sys em.
3.3.2. Expe imen al Valida ion P o ocol
To analyze he easibili y o he de eloped EMG sys em, a alida ion ial was pe o med wi h a
heal hy male subjec (wi h 23 yea s old, heigh o 1.70 m and weigh o 78.1 kg) walking in a eadmill
a 1 km/h, o 1 minu e. The choice o he walking speed was based on he minimum (0.5 km/h) and
maximum (1.6 km/h) ange o speeds allowed in Sma Os [71]. The MuscleBAN EMG sys em o [40]

33
was used as g ound u h. The de ec ion elec odes we e placed on he TA muscle and he e e ence
elec ode was placed on he knee, as exhibi ed in Figu e 7.
Figu e 7. Elec ode con igu a ion adop ed o a alida ion es o measu e he elec ical ac i i y o he TA muscle.
3.3.3. EMG Signal P ocessing
Du ing he da a collec ion, EMG da a om en gai cycles we e acqui ed. To his da a, he DC
o se componen o 1.65 V in oduced in ha dwa e was emo ed o achie e a signal wi h null mean.
Then, he ec i ica ion o he signal was pe o med, aking he absolu e alue o he signal. The en elope
o he EMG signal was de e mined using he Roo Mean Squa e (RMS) alue o he signal wi h a 300 ms
mo able window. The RMS ea u e was used o c ea e he EMG en elop since i is he mos sui able
me hod o ep esen a physical meaning o muscle o ce [74]. To con i m his ac , a ecen s udy showed
ha he RMS me hod is be e co ela ed wi h he muscle con ac ion o ce when compa ed wi h o he
me hods, such as mean absolu e alue, median equency and mean powe equency, since an
inc ease/dec ease o he muscle o ce con ac ion is mo e easily de ec ed wi h RMS alue [75]. The
choice o he window leng h (below 300 ms) was based on [76]. Highe alues o window leng h do no
cause an inc ease o he in o ma ion. In some cases, i was epo ed ha an ex eme inc ease o he
window leng h could p oduce a loss o in o ma ion [76].
3.3.4. Resul s and Discussion
Resul s o he expe imen al alida ion ial a e p esen ed in Figu e 8, whe e i is possible o e i y
ha he signals acqui ed wi h he p ojec ed EMG boa ds p esen a conside able co ela ion le el wi h he
signals acqui ed wi h he sys em o [40]. No wi hs anding he g ound u h sys em enable a be e
34
muscula ac i a ion de ec ion ( e i ied by he highe ampli ude ol age), compa ing bo h sys ems, he
muscula ac i a ions a e de ec ed a he same ins an , con e ing obus ness o he cons uc ed EMG
sys em. Besides he compa ison o he EMG signals acqui ed wi h di e en sys ems, ano he compa ison
was pe o med, in o de o analyze he easibili y o he measu e based on li e a u e e idences. In his
sense, he black pe ime e delimi ed in Figu e 8 was zoomed and he EMG signals o bo h sys ems we e
illed. The eason why his pe ime e was chosen is because he ange ime selec ed co esponds o a
single gai cycle.
The EMG signals inside he zoomed pe ime e we e compa ed wi h li e a u e EMG signals o a
subjec walking eely and, based on he esul s, i is possible o in e ha he signals acqui ed wi h he
p oposed EMG sys em a e simila o he expec ed ones.
Figu e 8. EMG signals acqui ed wi h he p ojec ed EMG sys em (blue line) and wi h a sys em om [40] ( ed line). The black
signal ep esen s he Li e a u e da a, whe e HS, LR, MS , TS , PSw, ISw, MSw and TSw mean Heel S ike, Load Response,
Mid S ance, Te minal S ance, P e-Swing, Ini ial Swing, Mid Swing and Te minal Swing, espec i ely.
35
3.4. Gene al Conclusions
In his chap e , he main cha ac e is ics and s a egies p esen ed in Sma Os we e p esen ed.
Since i was iden i ied he necessi y o cons uc an use -o ien ed s a egy, an AAN EMG-based con ol
app oach was p oposed, aiming i s u u e inse ion in o he p esen ed WPAD.
Fu he mo e, an EMG acquisi ion sys em was cons uc ed o acqui e signals om he main
muscles e iewed in Chap e 2. A good co ela ion be ween he signals acqui ed wi h he cons uc ed
EMG sys em and he acquisi ion sys em o [40] was ound. Mo eo e , he signals acqui ed wi h he
de eloped EMG sys em we e also compa ed wi h li e a u e e idences wi h espec o he ac i a ion o he
TA along a single s ide. Wi h he achie ed esul s, i was concluded ha he EMG sys em was co ec ly
p ojec ed and i can be used and in eg a ed in he p oposed AAN EMG-based con ol s a egy.
Once cons uc ed he EMG sys em o collec EMG signals om he muscles o in e es , ano he
equi emen o he p oposed AAN EMG-based con ol s a egy is he p edic ion o use -o ien ed e e ence
walking kinema ics and kine ic ajec o ies, ha will be he ocus o he Chap e 4 and Chap e 5,
espec i ely.
36
CHAPTER 4 – ANKLE KINEMATICS TRAJECTORY GENERATION
To implemen he AAN EMG-based con ol s a egy, i is equi ed o p edic he ankle join o ques,
o se e as he e e ence in he p oposed a chi ec u e. Since he join o que is associa ed wi h he join
kinema ics [77], a undamen al s ep o p edic he ankle join o que consis s o he gene a ion o ankle
join kinema ics o each subjec . Mo eo e , in his ield, some wo ks al eady epo ed he possibili y o
model ankle join angles based on he walking speed [78], [79].
Fo his pu pose, i is necessa y an ankle kinema ics p edic ion model based on well-known da a
om he subjec . In his sense, body heigh and body mass, as well as he walking speed a e possibili ies
o accomplish his issue. F om his pe spec i e, his chap e is ocused on he gene a ion o e e ence
ankle join kinema ics along he gai cycle, o ien ed o he subjec .
4.1. In oduc ion
Usually, p e- eco ded join ajec o ies o heal hy indi iduals ep esen he mos used me hod o
c ea e join kinema ics in he sagi al plane o se e as e e ence in assis ance applica ions o he human
gai . These ajec o ies a e no mally eco ded a slow (3.2 – 5.0 km/h), no mal (5.0 – 6.5 km/h) and
as (6.5 – 7.5 km/h) speeds [80]. Howe e , he walking speed o s oke, o ISCI, o Pa kinson su i o
is, app oxima ely, om 1.8 km/h o 2.5 km/h [81]. A possible solu ion o o e come his issue could be
he eco ding o join ajec o ies using nume ous walking speeds, in o de o cons uc a huge da abase
o be used as e e ence in gai assis an applica ions. Howe e , his solu ion is no sui able due o he
high ime consump ion o he da a collec ion and la ge amoun o da a, becoming imp ac icable o
37
ad ance wi h his app oach. In his di ec ion, eg ession models ha e been p oposed o enable he
p edic ion o join kinema ics ajec o ies du ing gai cycle [78], [79], [82].
Acco ding o [78], he ankle angle wa e o m du ing gai cycle p esen s a linea and quad a ic
ela ionship wi h he walking speed. In his sense, i join ajec o ies o heal hy indi iduals a slow walking
speeds a e used as e e ence in con ol s a egies in eg a ed in o WPADs, he assis ance will no be
e icien , because people wi h neu ological inju ies p e e lowe walking speeds and, consequen ly,
di e en join ajec o ies. Fo his eason, in [78], ou eg ession equa ions based on he walking speed
we e de eloped o p edic he peaks o he ankle join do si lexion and plan a lexion du ing he s ance
and swing phases. The
R2
was used as me ic o e alua e he accu acy o he model. Howe e , he highes
alues ob ained we e 0.1110, demons a ing unsa is ac o y esul s. The equa ions and he esul s
ob ained by [78] can be consul ed in Table 7, whe e 𝑣 ep esen s he walking speed.
Table 7. Reg ession equa ions and he co esponden esul s, achie ed in [78], o each peak o he ankle join angle
Pa ame e
Equa ion
𝑹𝟐
Peak Ankle Plan a Flexion (S ance Phase)
−1.7583𝑣 + 9.190961
0.0496
Peak Ankle Do si lexion (S ance Phase)
−2.4𝑣 + 13.62415
0.105
Peak Ankle Plan a Flexion (Swing Phase)
3.7834𝑣 + 12.88073
0.0870
Peak Ankle Do si lexion (Swing Phase)
4.16𝑣2 – 10.7498𝑣 + 10.03869
0.111
Con a y o he ou equa ions p esen ed by [78], in [79], only one eg ession equa ion was
de eloped conside ing he linea and quad a ic ela ionship be ween he walking speed and he ankle
join ankle. Be e esul s we e achie ed using he equa ion exhibi ed in Table 8.
Table 8. Reg ession equa ion and he co esponden esul s, achie ed in [79], whe e
a
,
b
and
c
ep esen eg ession
coe icien s o he model
Pa ame e
Equa ion
RMSE
θ
𝑎𝑣2+𝑏𝑣+𝑐
2.79 ± 2.05º
In acco dance wi h o he s udies, he e ec o he body heigh in he cons uc ion o he ankle
ajec o y is no ele an [80]. Howe e , in [82], i was epo ed ha i he a iabili y o he body heigh is
highe , his pa ame e may p esen a la ge e ec on he ankle ajec o y. Based on his sugges ion, in
his disse a ion, he body heigh was added o he equa ion p esen ed in Table 8, as p oposed by [82].
Thus, he ankle join ajec o ies a e dependen on he walking speed and body heigh o each subjec .
Wi h his app oach, he necessi y o eco ding join ajec o ies a many di e en body heigh s and

38
di e en walking speeds is a oided. The esul s o his app oach a e compa ed and alida ed wi h da a
acqui ed wi h a mo ion-cap u e sys em (Oqus; Qualysis – Mo ion Cap u e Sys em, Gö ebo g, Sweden)
unde speci ic expe imen al condi ions.
4.2. Me hods
4.2.1. Da a Acquisi ion
The g ound u h da a we e collec ed du ing locomo ion es s. Despi e only lowe limbs join
kinema ic and an h opome ic da a a e equi ed, join kine ic and EMG da a we e also collec ed, being
use ul da a o he nex chap e s.
A.
Volun ee s
The s udy was pe o med wi h six een adul subjec s (8 males and 8 emales wi h mean age o
23.8 ± 2.02 yea s, mean weigh o 67.5 ± 10.8 kg and mean heigh o 1.69 ± 0.109 m), wi h no e idence
o any ype o physical and physiological diso de ha could in e e e wi h hei walking pa e n, pe o med
walking es s. The minimum and maximum body heigh egis e ed in he da a collec ion was 1.51 and
1.83 m, espec i ely and, consequen ly, he econs uc ion o he ankle join ajec o ies can only be
pe o med in his ange.
B.
Ma e ials
The da a acquisi ion was pe o med using a mo ion-cap u e sys em wi h 12 came as (Oqus;
Qualysis – Mo ion Cap u e Sys em, Gö ebo g, Sweden), acqui ing a 200 Hz, i e o ce pla o ms
embedded in he loo (wi h cha ac e is ics p esen ed in Table 9), acqui ing a 200 Hz and an 8-channel
wi eless elec omyog aph (Delsys, Massachusse s, Uni ed S a es o Ame ica) [37], acqui ing a 2000
Hz.
Table 9. Fo ce pla es used in he da a acquisi ion
Fo ce Pla e Sys em
Model
Quan i y
Be ec, Ohio, Uni ed S a es o Ame ica
FP4060
2
Be ec, Ohio, Uni ed S a es o Ame ica
FP6090
2
Kis le , Win e hu , Swi ze land
9281 EA – FP4060
1
39
C.
Subjec P epa a ion
The Newing on-Helen Hayes model was adop ed as ma ke se , in eg a ing ou mo e ma ke s
placed in ochan e , medial ube osi y o he emu , medial malleolus and in he i s me a a sal head o
ob ain esul s mo e accu a e [83]. In his sense, a o al o 24 e lec i e ma ke s we e used. In ela ion o
he EMG signals, hey we e ex ac ed om TA, GASL, BF and VL o bo h legs. Bo h ma ke s and EMG
senso s a e p esen ed in Figu e 9.
Figu e 9. EMG senso s and ma ke s placemen on he use ’s body.
Fi s , he use was ins umen ed wi h EMG senso s. Then, wo maximum olun a y con ac ions
we e pe o med o each muscle, in o de o no malize all he EMG da a o he maximum isome ic
con ac ion egis e ed. A e his s ep, he 24 e lec i e ma ke s we e placed on he body and hen, he
subjec s we e asked o pe o m a s anding s a ic calib a ion wi h hei a ms c ossed in on o he ches ,
looking o wa d and wi h hei ee in a com o able posi ion, aligned wi h he shoulde s. This s ep was
use ul o i he an h opome ic da a o he body model o he acquisi ion sys em.
D.
Walking Expe imen s
All subjec s we e ins uc ed o pe o m 10 walking ials on a 10-me e la su ace wi h 5
embedded o ce pla o ms, a se en di e en speeds (1.0, 1.5, 2.0, 2.5, 3.0, 3.5 and 4.0 km/h),
con olled wi h a me onome. Be ween each speed change, he subjec s es ed o one minu e and hey
pe o med a habi ua ion ial o become acquain ed wi h he new walking speed.
E.
Da a Collec ion and P ocessing
The da a collec ed we e he join angles, angula eloci y and angula accele a ion o ankle, knee
and hip join s, he g ound eac ion o ce and he EMG da a om TA, GASL, BF and VL.
40
Following he da a collec ion o join angles, he kinema ic and kine ic da a we e il e ed wi h a low-
pass Bu e wo h il e . The cu o equency chosen was 6 Hz, since in [84], a equency a ound 6 and
7 Hz was de e mined as he op imal cu o equency o smoo h ajec o ies du ing he walking mo ion.
Rega ding he EMG da a, a band-pass il e wi h 20 and 450 Hz as cu o equencies was applied o he
aw da a. Fu he , he EMG en elop was de e mined using he RMS alue o he signal wi h a 300 ms
mo able window.
In his chap e , only ankle join posi ion ajec o ies will be used. The kine ic and EMG da a will be
use ul o he Chap e s 5 and 6.
4.2.2. Reg ession Model Implemen a ion
The implemen a ion o he eg ession model o es ima e ankle join ajec o ies du ing walking
mo ion was pe o med using
MATLAB
® and i was inspi ed in he wo k de eloped in [82]. Based on he
collec ed kinema ic da a, he ankle join angles we e spli in o indi idual gai cycles, whe e he i s sample
co esponds o he heel-s ike e en . Fo each ial ( ha ep esen s a single gai cycle), he join angle
alues and he ins an ame a se en key-e en s (lis ed in Table 10) we e ex ac ed.
Table 10. Key-e en s conside ed o cons uc he ankle join angle
Key-e en
Ins an o he s ide
1
Heel-S ike
2
Minimum Angle o he S ance Phase
3
Minimum Angula Veloci y o he S ance Phase
4
Maximum Angle o he S ance Phase
5
Minimum Angle o he Swing Phase
6
Maximum Angle o he Swing Phase
7
Heel-S ike
To de ec all hese e en s, wo s a egies we e adop ed: (i) c ea ion o a de ec ion algo i hm based
on he second de i a i e o he ankle join angle (angula eloci y); (ii) c ea ion o a de ec ion algo i hm
wi h e e ence o he maximums and minimums peaks o he ankle join angle.
In bo h s a egies, he 1s and he 7 h e en co espond o he i s and inal samples o each ial,
espec i ely. In case (i), he e en s numbe 2, 4, 5 and 6 we e ex ac ed based on a ze o-c oss de ec ion.
E e y ime ha he angula eloci y o he ankle join c ossed he ze o alue, an e en was ma ked. To
dis inguish be ween a do si lexion o a plan a lexion e en , he p e ious and he nex samples in ela ion
41
o he ze o-c oss alue we e also conside ed. I he p e ious and he nex samples co esponded o
posi i e and nega i e alues, espec i ely, a do si lexion e en was conside ed, whe eas i he p e ious
and he nex samples co espond o nega i e and posi i e alues, espec i ely, a plan a lexion e en was
ma ked. Only wo plan a lexion and wo do si lexion e en s mus be de ec ed. A las , he e en numbe
3 co esponds o he minimum angula eloci y be ween he minimum (second e en ) and he maximum
angle o he s ance phase ( o h e en ). Figu e 10 p esen s he de ec ion o all key-e en s using he
s a egy (i).
Figu e 10. Iden i ica ion o he se en key-e en s enounced on Table 10, o a walking speed o 4 km/h.
Howe e , mainly a slow walking speeds, he e en de ec ion p ocedu e does no wo k e icien ly due
o i egula i ies in he pa e n o he ankle join angles and angula eloci ies collec ed, as p esen ed in
Figu e 11.
Figu e 11. Poo iden i ica ion o he se en key-e en s enounced on Table 10, o a walking speed o 1 km/h.
48
Despi e he bes esul has been achie ed o a walking speed o 4 km/h, based on Figu e 13, he
walking speed ha is able o p oduce esul s wi h a smalle e o dis ibu ion is 3 km/h, because a his
eloci y, 99.3 % o he p edic ions p esen a RMSE below 5 º, a R be ween 0.871 and 0.981, a GOF om
30 o 70 % and a NRMSE below 20 %.
Addi ionally, he e is ano he limi a ion in he cu en eg ession model. The p oposed me hodology
will p o ide he same ankle join ajec o ies o indi iduals wi h he same body heigh , walking a he
same walking speed. Howe e , based on he da a collec ed, i is seen ha he in e -subjec a ia ion o
indi iduals in he same condi ions is conside able. This phenomenon can be isualized in Figu e 16. In
his sense, as u u e wo k, a use -o ien ed eg ession me hod should be de eloped, in o de o conside
mo e cha ac e is ics o he subjec s (e.g. body mass), cons uc ing di e en join ajec o ies o subjec s
in he same condi ions.
Figu e 16: Subjec 6 and 10 wi h a body heigh o 1.80 m, walking wi h a speed o 4 km/h.
4.4. Gene al Conclusions
In his chap e , a eg ession model was implemen ed o gene a e e e ence ankle join ajec o ies
in he sagi al plane o subjec s wi h body heigh s om 1.51 m o 1.83 m and walking speeds om 1 o
4 km/h, acco ding o he walking speed and body heigh o each subjec . The esul s we e sa is ac o y
and simila o he esul s achie ed in [82]. Howe e , imp o emen s a e needed. The easons ha jus i y
imp o emen s can be ela ed wi h wo ac s: he da a collec ion and he eg ession model. I he da a
collec ion had been pe o med in a eadmill, ce ainly, he s abili y o he subjec s walking a slow speeds
would no ha e been comp omised. The e o e, he ankle join da a de i ed om he walking mo ion would
no ha e been so i egula . Fu he mo e, using he p oposed eg ession model, subjec s in he same
condi ions o speed and body heigh will acqui e he same walking ajec o y. Howe e , i was seen ha

49
he in e -subjec a ia ion is signi ican and, o his eason, i is ecommended o de elop a model
o ien ed o he use , conside ing mo e da a, such as body mass.
50
CHAPTER 5 – ANKLE KINETICS TRAJECTORIES GENERATION
This chap e ocuses on modeling o e e ence ankle join o ques based on i e inpu s, namely he
e e ence ankle join angles de e mined in he Chap e 4 and hei de i a i es (angula eloci y and
accele a ion), body heigh o he subjec and walking speed. The chap e s a s wi h a b ie explana ion
o he dependency o he join o que on he body mass and hen, machine lea ning me hods used o
c ea e he e e ence join o ques a e p esen ed. A las , he esul s o each model a e exhibi ed,
compa ed and discussed, ending he chap e wi h he bes app oach o model he e e ence ankle join
o que. In his con ex , wi h he cu en and he p e ious chap e , he c ea ion o e e ence ankle join
o ques based on he walking speed, body heigh and body mass o each subjec can be achie ed.
5.1. In oduc ion
In he Chap e 4, based on [78], [82] and acco ding o he achie ed esul s, i was seen ha he
ankle join kinema ics du ing he gai cycle depends on he walking speed and he body heigh . Besides
ha , du ing he walking mo ion, he ankle plan a lexo s muscles a e esponsible o p o iding he suppo
equi ed o boos he body o wa d [86]. In his sense, wi h an inc ease o he body mass, i is expec able
ha he suppo gi en by he ankle join should be highe , inc easing he muscle unc ion and he join
o que and powe . Some s udies al eady e i ied ha he ankle join kine ics depends no only on he
walking speed, bu also on he body mass [87]–[89]. Thus, in his disse a ion, all he join o ques we e
no malized by he body mass.
51
In he las yea s, s udies ha e e ealed ha machine lea ning algo i hms ha e capaci y o model
nonlinea ela ionships o da a om he walking mo ion. In his ield, mos o hese machine lea ning
algo i hms a e selec ed o classi y and o ecognize locomo ion modes [90]–[92]. The e is no e idence
o he use o machine lea ning we e ound o p edic join kine ics o ien ed o he subjec du ing gai cycle.
In his connec ion, his chap e ocuses on he use o di e en machine lea ning algo i hms o model he
ankle join kine ics.
To explo e he ela ion be ween a iables, wo app oaches o machine lea ning can be used:
supe ised and unsupe ised lea ning echniques. A supe ised lea ning echnique de elops a p edic ion
model, using a aining da ase wi h a ela ion be ween he inpu and he ou pu da a. Then, his de eloped
model can p edic he ou pu o a new da ase and i s gene aliza ion capaci y wi h high p edic i e
accu acy is dependen o he a iance o he aining da ase . On he o he hand, in unsupe ised lea ning
echniques, he lea ning p ocess is based on lea ning pa e ns h ough g ouping ins ances. While
supe ised lea ning echniques comp ehend eg ession and classi ica ion algo i hms, unsupe ised
lea ning echniques co espond o dimensionali y educ ion o clus e ing [93]. Thus, supe ised lea ning
echniques a e used in his chap e o p edic he e e ence ankle join o que.
5.2. Me hods
In his sec ion, di e en eg ession models a e implemen ed o achie e he mos p oximal ankle
join o que o each subjec . Based on a li e a u e sea ch, he mos in es iga ed machine lea ning
app oaches used in eg ession p oblems include Suppo Vec o Machine (SVR), Random Fo es (RF) and
A i icial Neu al Ne wo ks (ANN). In he ield o ANN, MLP neu al ne wo ks and Deep Lea ning
a chi ec u es, including Long-Sho Te m Memo y (LSTM) and Con olu ional Neu al Ne wo ks (CNN),
ha e been widely used o sol e eg ession p oblems [65], [66], [94]–[96]. In his sense, hese eg ession
models (SVR, RF, MLP, LSTM and CNN) we e implemen ed, ained and op imized in
MATLAB
® and he
bes me hod was chosen o p edic he e e ence ankle join o que.
5.2.1. Reg ession Models
A.
Suppo Vec o Reg ession
As in Suppo Vec o Machine, de eloped by [97], SVR is widely used in eg ession p oblems due
o i s g ea e capaci y o gene aliza ion [93], [98]. This echnic is cha ac e ized o map a lowe -
dimensional da a in o a high-dimensional ea u e space using ke nel me hods o achie e highe
accu acies. The mos commonly used ke nels o his me hod a e linea , polynomial and gaussian o adial
52
basis unc ion. Howe e , he e is no e idence abou he bes ke nel o use. In SVR, he ain is pe o med
using a symme ical loss unc ion. Du ing his p ocess, a lexible ube (ԑ- ube) is o med a ound he
es ima ed unc ion and he main objec i e o his me hod is o ind he ube wi h he minimal wid h (ԑ -
epsilon) ha app oxima es he con inuous- alued unc ion, penalizing poin s ou side he ube and
p o iding no penaliza ion o he poin s inside. A he same ime, he model complexi y and he p edic ion
o he e o a e balanced. In addi ion o he g ea gene aliza ion capaci y wi h high accu acy, he
dimensionali y o he inpu da a does no a ec he complexi y o he model and, since i is less sensi i e
o he noise o he inpu s, he model is mo e obus [93].
B.
Random Fo es
Be o e p o iding a b ie explana ion abou RF applied in eg ession p oblems, i is equi ed o know
some ope a ion concep s behind Decision T ees (DT). When using his me hod (DT), a single p edic ion
is made as a esul o ques ions ha a e p oduced based on a spli ing me hod. The decision s a s a
he oo node, on he op o he ee, and i p og esses h ough he ee conside ing he answe s o he
ques ions a each decision node. This p ocess uns un il eaching he bes p edic ion, ha is ound a he
e minal node, also known as lea node, as p esen ed by Figu e 17.
Figu e 17. Example o an a chi ec u e o a DT.
Du ing he aining p ocess, he DT lea ns he bes ques ions/decisions o ask/pe o m, as well as
he o de o ask, wi h he inal pu pose o achie ing he mos accu a e es ima ion. The decision c i e ia
ela ed o he spli ing me hod is no mally based on he mean squa ed e o and, hus, acco ding o his
alue, du ing he aining p ocess, he model decides i i is equi ed o spli a node [99]. Howe e , he
53
majo p oblem ela ed wi h DT is o e i ing, because when aining, he model ends o adap i sel oo
much o he aining da a. Ano he nega i e aspec o his me hod is he ins abili y, p oducing a model
wi h high a iance. In hese cases, small a ia ions in he da a imply a gene a ion o a di e en DT. These
issues can be o e passed using RF. This eg ession me hod agg ega es many ees ha a e ained wi h
di e en andom pa s o he same aining da ase , in o de o educe he high a iance, using a
echnique called bagging. This echnique enables be e p edic ions, since he di e en ees a e no
co ela ed. In his sense, his means ha he a e age pe o mance o many ees is less sensi i e o
noise, while he p edic ion o a single ee is conside ably sensi i e [99].
C.
A i icial Neu al Ne wo ks: Mul ilaye Pe cep on
ANN a e composed by h ee laye s: inpu , hidden and he ou pu laye , con aining independen
a iables, ac i a ion unc ions and dependen a iables, espec i ely. The inpu laye is connec ed o he
hidden laye ha is composed by hidden neu ons. These neu ons ecei e he in o ma ion om he inpu
laye , hey calcula e he weigh s o he a iables using ac i a ion unc ions (such as, hype bolic angen ,
sinusoid, logis ic, bina y s ep o iden i y unc ions) and p edic he inal ou pu [100]. I he ou pu o one
laye se es as inpu in he nex laye and all nodes a e ully connec ed, he ne wo k is called eed o wa d
neu al ne wo k and he in o ma ion ne e ed back in he neu al ne wo k because he e a e no loops in
i s a chi ec u e [101]. Figu e 18 ep esen s an example o an a chi ec u e o a eed o wa d
neu al ne wo k
wi h wo hidden laye s.
Figu e 18. Feed o wa d neu al ne wo k a chi ec u e wi h wo hidden laye s.
Based on hese cha ac e is ics, and acco ding o [102]–[104], MLP is he mos used ANN in he
majo i y o he ields due o i s simplici y. Thus, a MLP neu al ne wo k was applied in his disse a ion.
The aining p ocess o MLPs has been widely used along wi h backp opaga ed algo i hms, since
he con e gence o he
neu al ne wo k is imp o ed [105]. The p inciple behind MLPs wi h
backp opaga ion algo i hms is based on he e o g adien compu a ion wi h espec o he weigh s and

54
i comp ehends wo phases: (i) o p ocess he da a om he inpu laye o he ou pu laye , passing
h ough he hidden laye (s); (ii) ha ing he ou pu s, a compa ison be ween he es ima ed and he eal
signal is done, compu ing an e o . This e o is backp opaga ed h ough he neu al ne wo k based on
he g adien descenden echniques. Thus, he weigh s a e upda ed acco ding wi h he e o be ween he
di e ence o he p edic ed ou pu and he eal signal. This can enable he achie emen o p edic i e
alues close o he a ge [101], [106]. Ma hema ically, he e o /cos unc ion in backp opaga ion
algo i hms can be calcula ed as he sum o squa es, using Equa ion (4).
Whe e
C
ep esen s he cos unc ion be ween he a ge (
T
) and he p edic ed ou pu (
O
) o each
ou pu neu on,
N
. As epo ed by [107], o a ne wo k wi h only one laye wi h one neu on and a ec i ied
linea uni ac i a ion, he p edic ion/ a ge o he model co esponds o he ac i a ion o he neu on. In
his case, he ac i a ion unc ion is desc ibed by Equa ion (5).
Whe e,
w
and
b
ep esen he pa ame e s o he model (weigh and bias, espec i ely) and
X
ep esen s he model inpu s. As a esul , he cos unc ion can be e o mula ed by Equa ion 6 and 7:
Di e en ia ing Equa ion (7) wi h espec o he weigh s, i is possible o upda e he cu en weigh s,
a ins an
, conside ing he e o g adien descen and he weigh s o he p e ious ins an
– 1
, ob aining
Equa ion (8). The de i a ion s eps o achie e Equa ion 8 can be consul ed in [107].
No wi hs anding he p edic ion can achie e be e pe o mances, he g adien descen algo i hms
used o backp opaga e he e o can s uck in a local minimum, as p esen ed in Figu e 19.
𝐶(𝑇,𝑂)= 1
𝑁෍(𝑇𝑘−𝑂𝑘)
2
𝑁
𝑘=1
(4)
𝑎𝑐𝑡𝑖𝑣𝑎𝑡𝑖𝑜𝑛(𝑋)=𝑚𝑎𝑥 (0,𝑤∙𝑋+𝑏)
(5)
𝐶(𝑇,𝑂,𝑤,𝑏)= 1
𝑁෍(𝑇𝑘−𝑎𝑐𝑡𝑖𝑣𝑎𝑡𝑖𝑜𝑛(𝑋𝑘))
2
𝑁
𝑘=1
(6)
C(T,O,w,b)= 1
N෍(Tk−max (0,w ∙ Xk+b))2
N
k=1
(7)
𝑤𝑡=𝑤𝑡−1 −
𝛿𝐶
𝛿𝑤𝑡−1
(8)
55
Figu e 19. G adien descen algo i hm ge ing s uck in a local minimum.
To sol e his p oblem, wo new pa ame e s can be conside ed, namely
lea ning a e
,
𝜼
and
momen um
,
𝜶
, as ep esen ed by Equa ion (9).
Wi h hese pa ame e s, he e ec o he e o g adien on he weigh s upda e is con olled.
Mo eo e , he pe o mance o he neu al ne wo k and i s ime o con e gence a e also dependen on
hese pa ame e s. Rega ding he lea ning a e, i is equi ed o conside ha i his pa ame e is se oo
small, he algo i hm will ake oo much ime o con e ge, because he g adien descen algo i hm will
wo k slowly. In hese cases, he con e gence o he model is comp omised and i is possible o s op again
in a local minimum. On he o he hand, i he lea ning a e is se oo high, he algo i hm will con e ge
apidly bu wi h no s abili y. Thus, low alues o lea ning a e a e p e e able, also conside ing he
momen um pa ame e o escape om local minimums, since his pa ame e adds a pe cen age o he
las upda e, imp o ing he cu en upda e [108].
Howe e , he e o on he aining se can be small, bu when he neu al ne wo k p edic s
esponses o unknown da a, he e o is la ge. This means ha he neu al ne wo k is no gene alized and
i canno p o ide good esponses o new si ua ions. Acco ding wi h [109], in MLP, he gene aliza ion
pe o mance is be e using Bayesian egula iza ion. Ano he egula iza ion echnique ha p o ides a
gene alized neu al ne wo k is he Le enbe g-Ma qua d [100].
In his sense, using he neu al ne wo k oolbox o
MATLAB®
, in his disse a ion, a MLP wi h
backp opaga ion algo i hm is applied, in es iga ing wo di e en aining algo i hms:
ainb
and
ainlm
,
co esponding o he Bayesian egula iza ion me hod and Le enbe g-Ma qua d me hod, espec i ely. In
bo h cases, he lea ning a e and momen um pa ame e s we e used as de aul . Mo eo e , based on
w = w −1− η
δC
δw −1
−αw −1
(9)
56
[109], he pe o mance o he neu al ne wo k can inc ease, ob aining a s able lea ning i he lea ning a e
is upda ed du ing he aining p ocess. In his sense, ano he aining algo i hm, namely,
aingdx
was
implemen ed o upda e he lea ning a e pa ame e .
D.
A i icial Neu al Ne wo ks: Deep Lea ning
Ankle join o que is a con inuous signal ha is dependen on he ime and, hus, i can be
ecognized as a Time-Se ies signal. In his ield, s udies epo ed ha he use o Deep Lea ning me hods,
such as LSTM and CNN, can p o ide sa is ac o y esul s [110], [111].
LSTM consis s o a ype o ecu en neu al ne wo ks. As desc ibed in he p e ious sub-sec ion, he
beha io o he ac i a ion unc ions exis ing on he neu ons o he hidden laye s o a eed o wa d neu al
ne wo ks depends on he beha io o he ac i a ion unc ions o neu ons o he p e ious hidden laye s. In
con as , wi h he a chi ec u e o ecu en neu al ne wo ks, he beha io o he ac i a ion unc ions o he
hidden neu ons depends no only on he beha io o he p e ious ac i a ion unc ions, bu also on he
beha io a an ea lie ime. Fo his eason, neu al ne wo ks whe e he beha io o he ac i a ion unc ions
is dependen o he ime a e called by ecu en neu al ne wo ks and, hus, Time-Se ies da a is lea ned
in an e icien way [101]. No mally, hese a chi ec u es a e used o p edic he u u e using pas
in o ma ion. Howe e , i he in e al in ime be ween he cu en p edic ion and he p e ious in o ma ion
is oo la ge, he p edic ions o he ecu en neu al ne wo k a e no p omising. Acco ding o [101], i he
ime o un o a ecu en neu al ne wo k is oo long, he g adien will become emendously uns able and
he capaci y o lea n will be minimal. To sol e his p oblem, LSTM neu al ne wo ks we e p oposed. These
kind o neu al ne wo ks a e a conjuga ion o ecu en
neu al ne wo ks wi h g adien descenden lea ning
algo i hms, making easie o ob ain good esul s wi h uns able g adien p oblems unde con ol [112].
In o ma ion ela i e o he ain o he LSTM neu al ne wo k can be ound in [112].
CNN a e ano he ype o neu al ne wo ks ha is ained based on backp opaga ion algo i hms. The
a chi ec u e o hese neu al ne wo ks was inspi ed in he mammalian isual co ex, since acco ding o
[113], he i s p ocessing s ages in he isual co ex consis in he de ec ion o simple ea u es, such as
edges and ba s. In he ollowing p ocessing s ages, mo e complex associa ions a e done and he objec
ha is being seen is ecognized. The beha io o CNN is iden ical, since hese neu al ne wo ks a e
composed by h ee laye s (con olu ional, pooling and ully-connec ed laye s) and du ing he aining
p ocess, he objec i e is o educe he dimensionali y o he ep esen a ion based on ea u e ex ac ion
me hods, in o de o cap u e he mos ele an spa ial and empo al dependencies o an image [113],
[114]. Thus, his neu al ne wo k was used o ex ac he mos ele an ea u es o achie e he mos
p oximal ankle join o que.
57
The con e gence and gene aliza ion pe o mance o neu al ne wo ks can be a di icul ask o
accomplish due o he di icul y in o ind he bes lea ning a e and momen um alue. Mo eo e , i only
one lea ning a e and momen um alue is applied o all weigh s, he weigh upda e is he same o all
neu al ne wo k, causing he con e gence pe o mance less e icien . To imp o e he con e gence
pe o mance, some g adien descenden algo i hms ha e been eme ged, such as, he Adap i e Momen
Es ima ion (ADAM). Wi h his algo i hm, he weigh s o he neu al ne wo k a e upda ed using adap i e
lea ning a es, a oiding he exis ence o a global one. Thus, he con e gence pe o mance is inc eased,
while he lea ning e o is educed, ob aining a e sa ile neu al ne wo k [115]. In Deep Lea ning
algo i hms explo ed in his disse a ion, ADAM was used. In he neu al ne wo k oolbox o
MATLAB®
is
no possible o apply his algo i hm when using MLP neu al ne wo k and hus, he mos p oximal algo i hm
ound was
aingdx
.
5.2.2. Da a P epa a ion
The da a collec ed in he Chap e 4 was used o ain he implemen ed machine lea ning models.
To ain all he models, he da a we e andomly di ided in blocks, whe e 60% o he da a we e used o
aining, 20% o alida ion and 20% o es ing, co esponding o 10, 3 and 3 subjec s, espec i ely. The
inpu s o he models we e he ankle join angles, angula eloci y, angula accele a ion, body heigh and
walking speed. The ou pu o he models was he ankle join o que no malized by he body mass. A
k
-
old c oss- alida ion algo i hm was implemen ed o analyze he obus ness o he models. The numbe o
olds was se o 4.
Table 13 p esen s he maximum ange and he uni s e ealed by he inpu and ou pu ea u es. In
acco dance o hese alues, i is possible o in e ha he e is a disc epancy in he magni ude o de o
each ea u e.
Table 13. Uni s and Maximum Va ia ion Range o each model ea u e
Fea u e
Uni
Maximum
Range
Ankle Angle
deg ees
76.0
Ankle Angula Veloci y
deg ees/s
3.75×102
Ankle Angula
Accele a ion
deg ees/s2
3.68×103
Body Heigh
m
0.320
Walking Speed
km/h
3.00
Ankle To que
N·m/kg
2.18
64
Table 21. Key pa ame e s o LSTM neu al ne wo k
Pa ame e
Value
Maximum numbe o epochs
10000
Maximum alida ion ailu es
10
G adien Descenden Algo i hm
ADAM
Ini ial Lea ning Ra e
0.01
Lea ning Ra e D op Pe iod
50
Lea ning Ra e D op Fac o
0.2
Ba ch Size
64
The esul s o he LSTM neu al ne wo k wi h he pa ame e s de ined in Table 21 a e p esen ed in
Table 22.
Table 22. LSTM esul s wi h de aul pa ame e s, whe e he wo s and bes esul s a e colo ed in ed and g een, espec i ely
Ba ch Size
Hidden
Neu ons
GOF (%)
NRMSE (%)
Valida ion mean
(s d)
Tes
Valida ion mean
(s d)
Tes
64
10
78.1
(0.400)
77.7
4.76
(0.121)
4.70
70
77.2
(2.71)
78.4
4.96
(0.634)
4.55
110
79.4
(0.195)
79.6
4.92
(0.117)
4.31
Conside ing he eg ession pe o mance achie ed wi h LSTM, an inc emen o he numbe o
neu ons does no o e signi ica i e imp o emen s. The bes esul s we e achie ed o a LSTM neu al
ne wo k wi h 110 hidden neu ons.
The e ec o he ba ch size was s udied in his neu al ne wo k, inc easing and he dec easing i s
alue, in o de o ind he a chi ec u e ha p o ided he bes esul s. Due o he high ime consump ion
encoun e ed du ing he ini ial de elopmen s o his neu al ne wo k, he ba ch size was only a ied o he
LSTM wi h 110 hidden neu ons, since ha was he a chi ec u e ha p o ided he bes pe o mance in
de aul pa ame e s. The esul s achie ed a e p esen ed in Table 23, whe e he wo s and he bes
pe o mances a e colo ed in ed and g een, espec i ely.

65
Table 23. LSTM esul s wi h ba ch size a ia ion
Ba ch Size
Hidden
Neu ons
GOF (%)
NRMSE (%)
Valida ion mean
(s d)
Tes
Valida ion mean
(s d)
Tes
128
110
79.3
(0.200)
79.5
4.95
(0.125)
4.57
64
110
79.4
(0.195)
79.6
4.92
(0.117)
4.31
32
110
77.4
(0.352)
76.8
4.92
(0.155)
4.87
Acco ding wi h Table 23, wi h a dec easing o he ba ch size, he pe o mance o he neu al ne wo k
s a ed o dec ease. On he o he side, conside ing a la ge ba ch size, he pe o mance o he neu al
ne wo k emains simila . Thus, an LSTM wi h 110 hidden neu ons and ained wi h a ba ch size o 64
can p o ide good pe o mances.
5.3.5. Con olu ional Neu al Ne wo k
Gene ally, CNN a e used o image p ocessing and hus, his neu al ne wo k ope a es wi h 3 –
dimensional ma ices. In his sense, he inpu da a we e ea ed as images, ha ing su e ed a
ans o ma ion. Once he e a e i e inpu ea u es and se e al gai cycles, he da a was o ganized in 3 –
dimensional ma ices, as ollows: wid h – numbe o each gai cycle samples; heigh – numbe o inpu
ea u es; dep h – numbe o gai cycles.
Conce ning he aining p ocess, his neu al ne wo k also ained while he numbe o maximum
epochs o he numbe o maximum alida ion ailu es we e no eached. A he beginning, he ini ial
lea ning a e alue, he d op pe iod, d op ac o and he ba ch size we e se acco ding wi h he de aul
alues. The de ined pa ame e s a e p esen ed in Table 24.
66
Table 24. Key pa ame e s o CNN
Pa ame e
Value
Maximum numbe o epochs
10000
Maximum alida ion ailu es
10
G adien Descenden Algo i hm
ADAM
Ini ial Lea ning Ra e
0.01
Lea ning Ra e D op Pe iod
50
Lea ning Ra e D op Fac o
0.2
Ba ch Size
64
Numbe o Con olu ional Laye s
3 (wi h 8, 16 and 32
il e s applied)
Ke nel Size
5×5
Pooling (Pool Size)
A e age (2)
S ide
2
Fully Connec ed Laye s
(Numbe o ou pu s)
1 (1)
Acco ding o [118], wi h smalle ke nel sizes, he compu a ional pe o mance is mo e e icien han
wi h la ge ke nel sizes and he neu al ne wo k is able o lea n mo e complex non-linea ea u es. In his
sense a ke nel size o 3 was also expe ienced and he esul s a e p esen ed in Table 25.
Table 25. CNN esul s wi h ke nel size a ia ion
Ke nel Size
GOF (%)
NRMSE (%)
Valida ion mean
(s d)
Tes
Valida ion mean
(s d)
Tes
5×5
81.1
(0.721)
80.5
4.10
(0.122)
4.10
3×3
81.0
(1.08)
80.7
4.13
(0.188)
4.06
Based on he achie ed esul s, he pe o mance o he neu al ne wo k wi h bo h ke nel sizes was
iden ical, achie ing smalle be e esul s in he es da ase wi h a ke nel size o 3.
Since he las laye o CNN is a ully connec ed laye iden ical o a MLP and conside ing ha an
inc emen in he numbe o hidden neu ons in MLP caused an imp o emen in he eg ession
pe o mance, ano he ully connec ed laye was added o CNN. Since a he end o he las con olu ional
laye , a ec o wi h a leng h o 32×3 is c ea ed and conside ing he desi ed ou pu o que alue, he
67
numbe o neu ons o he ully connec ed laye added was se o 32. The esul s achie ed wi h his new
a chi ec u e a e p esen ed in Table 26.
Table 26. CNN esul s wi h wo ully connec ed laye s
Ke nel Size
GOF (%)
NRMSE (%)
Valida ion mean
(s d)
Tes
Valida ion mean
(s d)
Tes
3×3
80.6
(0.596)
80.0
4.20
(0.0814)
4.22
In con as o he expec able, he pe o mance o CNN dec eases wi h he addi ion o ano he ully
connec ed laye . Thus, only one ully connec ed laye should be conside ed. In his neu al ne wo k, he
ba ch size was also a ied. The pe o mances achie ed a e exhibi ed in Table 27, whe e he wo s and
bes pe o mances a e colo ed in ed and g een, espec i ely.
Table 27. CNN esul s wi h ba ch size a ia ion
Ke nel Size
Ba ch Size
GOF (%)
NRMSE (%)
Valida ion mean
(s d)
Tes
Valida ion mean
(s d)
Tes
3×3
128
81.6
(0.533)
81.2
3.99
(0.137)
3.97
64
81.0
(1.08)
80.7
4.13
(0.188)
4.06
32
79.9
(0.548)
79.7
4.36
(0.137)
4.27
By analyzing Table 27, a educ ion o he ba ch size is no a o able, since he eg ession
pe o mances s a ed o dec ease. On he o he side, wi h an inc emen o he ba ch size, be e esul s
we e achie ed. Thus, a CNN wi h 3 con olu ional and 1 ully connec ed laye , a ke nel size o 3×3 and a
ba ch size o 128 was ound as he a chi ec u e ha p o ides be e esul s.
5.4. Discussion and Gene al Conclusions
Machine lea ning algo i hms ha e been used in he las yea s o model nonlinea ela ionships o
he walking. Howe e , o he bes o he au ho ’s knowledge, he e a e no li e a u e e idences on he
using o machine lea ning o p edic he ankle join o que o ien ed o he subjec . In his sense, he mos
68
used eg ession models applied o model nonlinea ela ionships, namely SVR, RF, MLP, LSTM and CNN,
we e used o p edic he e e ence ankle join o que based on i e inpu s: ankle angle, angula eloci y,
angula accele a ion, body heigh and walking speed.
Table 28 p esen s he bes esul s ob ained pe each machine lea ning algo i hm.
Table 28. Bes esul s o he implemen ed eg ession models whe e he bes model pe o mances achie ed a e colo ed in
g een, while he wo s a e colo ed in ed
Me hod
GOF (%)
NRMSE (%)
Valida ion mean
(s d)
Tes
Valida ion mean
(s d)
Tes
SVR
69.4
(9.52)
68.7
5.97
(2.01)
6.59
RF
73.9
(1.95)
73.6
5.68
(0.470)
5.56
MLP
67.3
(1.57)
70.9
7.10
(0.380)
6.13
LSTM
79.4
(0.195)
79.6
4.92
(0.117)
4.31
CNN
81.6
(0.533)
81.2
3.99
(0.137)
3.97
In addi ion o he bes esul s summa ized and p esen ed by Table 28, Figu e 20 exhibi s he
g ound u h ankle join o que and he p edic ions achie ed wi h each one o he models o a andom
selec ed ange o he es da ase .
Figu e 20. P edic ions o he SVR, RF, MLP, LSTM and CNN machine lea ning models in compa ison wi h he eal ankle join
o que.
69
In acco dance o Figu e 20, i is possible o in e ha he p edic ions made by LSTM and CNN
models p oduced an ankle join o que close o he expec ed one, when compa ed o he emaining
eg ession models. This ac is con i med by he highe GOF and lowe NRMSE alue p esen ed by hese
models, in Table 28. None heless, pe o ming a deepe analysis cen e ed on he gai cycle, du ing he
s ance phase, i was e i ied ha all he models p esen capaci y o model he ankle join o que. Howe e ,
in he swing phase, i egula i ies and sudden peaks we e e i ied o he p edic ions achie ed by he SVR
and MLP. This phenomenon can be con i med by he lowes e alua ion me ics p esen ed by hese wo
models. Thus, conside ing he ob ained esul s, he CNN was he model chosen o p edic he e e ence
ankle join o que o ien ed o he subjec .

70
CHAPTER 6 – AAN EMG-BASED CONTROL STRATEGY
This chap e desc ibes he p oposed AAN EMG-based con ol s a egy p esen ed on Chap e 2. The
impo ance o he eg ession models de eloped in Chap e 4 and Chap e 5 is in es iga ed. Addi ionally,
i is de eloped a me hod o es ima e he eal ankle join o ques based on EMG signals and join angles.
The pe o mance o he ankle join o que es ima ion is e alua ed and he chap e ends wi h a discussion
o he achie ed esul s.
6.1. In oduc ion
The li e a u e analysis in Chap e 2 demons a ed ha he EMG-based con ol s a egies
in eg a ed in o WPADs only ollow he in en ions o he use and, consequen ly, i may be possible ha
indi iduals wi h impai men s o he lowe limbs and whose EMG signals a e weak do no ecei e he
equi ed suppo o walk. AAN con ol s a egies in eg a ing EMG signals ep esen a ecen al e na i e
o o e come hese limi a ions, aking ad an age o he an icipa o y pe o mance o hese signals. Thus,
his disse a ion aims o cons uc an AAN EMG-based con ol s a egy o a u u e in eg a ion a
Sma Os.
To ad ance wi h he p oposed s a egy, i is equi ed o de e mine wo pa ame e s: a e e ence
and a eal join o que, as ep esen ed in Figu e 4. Rega ding he e e ence ankle join angle de e mined
in Chap e 4, i s impac is de e minan on he p edic ion o he e e ence ankle join o que. Acco ding
o [77], he join o que abou he ankle join is dependen on he angula accele a ion. Since his
pa ame e co esponds o he second de i a i e o he join angle, he p edic ion o he e e ence ankle
71
join o que is dependen on he e e ence ankle join angle p edic ion. In his sense, h ough his
chap e , he connec ion be ween he Chap e 4 and Chap e 5 is e alua ed: i.e., he p edic ion o he
e e ence ankle join o ques wi h he bes model ound in Chap e 5 ed wi h he p edic ed ankle join
kinema ics de e mined a Chap e 4.
Conside ing he collec ed wo ks esponsible o con e ing he EMG signals in o join o que
alues, he s udy de eloped by [64] was chosen o pe o m his con e sion due o p esen he sho es
con e sion ime, o e ing p omising esul s. Fu he mo e, he oolbox o pe o m his es ima ion is open
sou ce, p og ammed in
C++
language and i is a ailable on [119]. Thus, in his chap e , he code o
he oolbox is adap ed o he a chi ec u e o Sma Os, in o de o es ima e he eal ankle join o que in
eal- ime.
6.2. Re e ence Ankle Join To que P edic ion
The p edic ion o he e e ence ankle join o que (
τ e ( )
) was based on he p edic ion o kinema ic
da a o each subjec . Fi e inpu s we e in ol ed, namely,
BH, WS,
𝜃
e ( ), ѡ e ( )
and
α e ( )
, as p esen ed
in Figu e 2.
To e alua e he pe o mance o he e e ence ankle join o que p edic ion, wo cases a e
analyzed: (i) he wo s ; and (ii) he bes ankle join angle p edic ion. Based on he esul s achie ed in
Chap e 4, he wo s p edic ion was e i ied o a subjec wi h a body heigh o 1.79 m, walking a 1
km/h, while he bes esul s we e achie ed o a subjec wi h a body heigh o 1.65 m, walking a 4 km/h.
Fo each case, he body mass was 81.9 and 60.0 kg, espec i ely. Thus, o bo h cases i was p edic ed
he e e ence ankle join o que using he bes eg ession model ound in Chap e 5, namely, CNN wi h
a ke nel size o 3×3, 3 con olu ional and 1 ully connec ed laye s, ained wi h a ba ch size o 128. The
esul s o he p edic ions a e ep esen ed in Figu es 21.
72
Figu e 21. Re e ence ankle join o que p edic ion based on he wo s (a)) and bes ((b)) e e ence ankle join angle
p edic ion.
Al hough he ankle join angle p edic ion o he i s case ((i)) has been less a o able, he
p edic ion o he o que o his join is accep able and compa able wi h he eal join o que, p esen ing
a GOF o 74.7 % and a NRMSE o 8.02 %. Based on Figu e 15 – a), ( ha ep esen s he wo s ankle join
angle p edic ion), he ankle join angle p edic ed p esen s a wide ange o mo ion (ROM) when compa ed
wi h he eal ROM. Hence, he angula accele a ion also p esen s a wide ange and, since he join o que
is di ec ly dependen on his pa ame e , i would be expec able ha he p edic ed join o que could
p esen highe alues han he eal one. In ac , his phenomenon occu ed, as i is p o en by Figu e 21
– a). Conce ning he second case ((ii)), he p edic ion o he join o que is qui e simila o he eal join
o que, p esen ing a GOF o 89.8% and a NRMSE o 3.16%.
73
O e all, e en he p edic ed e e ence ankle join kinema ics p esen less a o able i s, he
implemen ed machine lea ning algo i hm (CNN) can model he e e ence ankle join o ques wi h high
accu acy.
6.3. EMG-based Real Join To que Es ima ion
Once he p edic ion o he e e ence ankle join o que has been achie ed, i was equi ed o
es ima e he eal ankle join o que o a subjec , in eal- ime. Along his sec ion, all he algo i hms
implemen ed a e exhibi ed, ending he chap e wi h he alida ion o he p oposed es ima ion.
6.3.1. Model P esen a ion
Figu e 22 p esen s he block diag am o es ima e he eal ankle join o que, based on eal EMG
signals and eal ankle join angles.
Figu e 22. Block diag am o es ima e he eal ankle join o que,
τ eal
.
The oolbox a ailable on [119] pe o ms he con e sions illus a ed in he Blocks 1, 3, 4 and 5 o
Figu e 23. Howe e , i does no calcula e he musculo endon leng hs (LMTs), nei he he momen a ms
(MAs) illus a ed in Block 2, bo h equi ed o es ima e he eal ankle join o que. To sol e his p oblem,
an algo i hm p esen ed by [120] and adap ed by [121] was used o ob ain he LMTs and MAs based on
eal ankle join angles. To de e mine hese wo pa ame e s, a calib a ion s ep is equi ed, in o de o
iden i y six calib a ion pa ame e s, namely,
Fmax
,
max
,
L0
m
, op imal leng h o he endon (
Lslack
), ɸ and ype
I
ibe s pe cen age. In he wo k ad anced by [121], hese six pa ame e s (p esen ed in Table 29) we e
ound o a heal hy subjec wi h 1.80 m and 80 kg.
80
pe o med o iden i y he EMG-based and AAN EMG-based con ol s a egies al eady de eloped, as well
as he mos ele an me hodologies o es ima e he join kine ics based on EMG signals, since his
con e sion is pe o med in he majo i y o he con ol s a egies using EMG signals. P opo ional
myoelec ic, musculoskele al and empi ical models we e in es iga ed as me hods o con e he EMG
signals in o o que alues. Empi ical models may be mo e app op ia ed o pe o m he desi ed con e sion
o disable pe sons, since calib a ion s eps a e no equi ed.
An EMG sys em was p ojec ed, implemen ed and alida ed wi h a heal hy subjec walking a 1
km/h. The esul s achie ed when compa ed wi h a comme cial sys em and wi h li e a u e indings,
e ealed a good pe o mance.
In his disse a ion, a eg ession model was explo ed o p edic he e e ence ankle join kinema ics
ajec o y o a speci ic subjec in he sagi al plane, based on he walking speed and body heigh . The
esul s achie ed we e sa is ac o y and simila o he esul s ob ained in [82]. Howe e , imp o emen s
ela ed o (i) he da a collec ion; and (ii) he eg ession model mus be done. Rega ding he i s poin ,
du ing he da a collec ion, i was e i ied ha a slow walking speeds, he pos u al s abili y o mos o he
subjec s was comp omised. Du ing he eg ession model aining, in he p esence o slow walking speeds,
he eg ession model lea ned i egula ankle join ajec o ies. Rega ding he second poin , i was e i ied
ha he p oposed eg ession model a ibu es he same e e ence ankle join angle o subjec s wi h he
same body heigh and walking a he same walking speed. Howe e , he in e -subjec a ia ion is
conside able and, o his eason, i is ecommended o de elop a model o ien ed o he use , conside ing
mo e da a, such as body mass.
Based on ankle join kinema ics, body heigh o he subjec and walking speed, machine lea ning-
based me hods we e explo ed o gene a e he e e ence ankle join kine ics. The pe o mance o i e
machine lea ning models was explo ed, namely: SVR, RF, MLP, LSTM and CNN, whe e he bes esul s
we e achie ed o a CNN.
The esul s demons a ed ha CNN can accu a ely p edic he e e ence ankle o que ajec o ies,
e en conside ing e e ence ankle join kinema ics wi h less a o able p edic ions. The co ec p edic ion
o he e e ence ankle join o que o ien ed o he use based on p edic ed e e ence ankle join
kinema ics, body heigh and walking speed is u mos impo ance in he AAN EMG-based con ol s a egy,
in o de o compa e he e e ence ankle join o que wi h he eal one. In his ield, conside ing he bes
acknowledge o he au ho , he e is no e idence on he li e a u e ega ding he ankle join o que
p edic ion o ien ed o he subjec .

81
To es ima e he eal ankle join o que, an EMG-based o que es ima ion was pe o med, based on
a musculoskele al model p esen ed in [64]. Only wo muscles we e used o pe o m he es ima ion, in
con as o he six muscles used in [64]. Thus, EMG da a om TA and GASL we e used wi h eal ankle
join angles o es ima e he eal ankle join o que. Since he model depends on calib a ion pa ame e s
ha we e ound o a subjec wi h 1.80 m and 80 kg, he alida ion o his me hodology was pe o med
only conside ing one subjec wi h he e e ed physical cha ac e is ics. Based on he ob ained esul s, i
was e i ied ha he magni ude o he o que p o ided by he model was in e io o he eal magni ude,
indica ing ha only wo muscles a e no enough o es ima e he ankle join o que. Thus, i is necessa y
o include mo e EMG con ibu ions om GASM, SOL, PT and PL muscles.
Aiming he u u e in eg a ion o he p oposed AAN EMG-based con ol s a egy in o Sma Os and
conside ing ha he EMG-based o que es ima ion was de eloped as a s andalone amewo k, a TCP/IP
communica ion was c ea ed o simula e he p o ocol be ween he EMG-based o que es ima ion model
and he CCU o Sma Os.
Fu he mo e, wi h he de eloped mas e hesis, he goals ha we e es ablished in Chap e 1 we e
all achie ed and he RQ can be answe ed:
• RQ1:
Which a e he con ibu ions and he main di e ences o he EMG-based con ol and he
AAN EMG-based con ol s a egies?
This RQ was add essed in Chap e 2. Bo h s a egies con ibu ed o muscula ac i i y
imp o emen and endu ance, a oiding muscle a ophy. Mos o he EMG-based con ol
s a egies e i ied a dec eased muscula ac i i y while he join angles we e con olled. Howe e ,
hese e iewed s a egies we e only es ed in heal hy subjec s. Thus, he join angles we e
con olled gi en he mo o abili y o he heal hy subjec s. None heless, i is no expec able ha
hese s a egies ha e he po en ial o assis he lowe limbs o disabled subjec s. In con as ,
he AAN EMG-based con ol s a egies compa e, in eal- ime, he use ’s mo o pe o mance
wi h he desi ed use ’s mo ion o ensu e a su icien le el o assis ance o ien ed o he use ’s
needs. Thus, AAN EMG-based con ol s a egies ha e he po en ial o assis and ehabili a e he
lowe limbs, such ha he use can achie e au onomy o pe o m hei daily li e ac i i ies.
• RQ2:
Is i possible o ob ain join o que measu es only using EMG signals?
This RQ was add essed in Chap e 2. Based on he e iewed in o ma ion, i was
concluded ha mos o he con ol me hods which use EMG signals o con ol WPADs con e
hese EMG signals in o o que alues. None heless, he e a e li e a u e e idences ha only EMG
82
eadings a e no enough o ob ain he accu a e alues o join o ques, being necessa y o use
EMG da a wi h biomechanical da a, such as, join angles, eloci ies, accele a ions.
• RQ3:
Is i possible o p edic e e ence walking kinema ics and kine ics ajec o ies
elying
exclusi ely on he walking speed and an h opome ic da a?
This RQ was app oached in Chap e s 4, 5 and 6. Based on he esul s achie ed in
Chap e 4, i is possible o p edic e e ence walking kinema ics using exclusi ely walking speed
and an h opome ic da a. The esul s o his disse a ion show he needed o adding mo e
an h opome ic da a (such as, he body mass) o a oid he a ibu ion o he same ankle join
angle o subjec s wi h he same body heigh , walking a he same walking speed. Howe e ,
based on he esul s achie ed in Chap e 5 and Chap e 6, his is no a limi a ion o p edic he
e e ence walking kine ics, since good pe o mances we e ob ained.
• RQ4:
Can EMG-based o que es ima ion s a egy p esen a good pe o mance?
This RQ was app oached in Chap e 6. Based on he esul s achie ed in he li e a u e
[64], he EMG-based o que es ima ion wi h join angles can be pe o med in eal- ime wi h
good pe o mances, when six muscles a e conside ed. In his disse a ion, only EMG da a om
wo muscles (TA and GASL) and ankle join angles we e conside ed. Based on he achie ed
esul s, he magni ude o he es ima ed ankle join o que was in e io o he expec able. Wi h
he addi ion o one mo e muscle (GASM), highe magni udes we e achie ed. Thus, mo e
muscles mus be conside ed o es ima e he eal ankle join o ques, conside ing EMG signals
and eal ankle join angles.
7.1. Fu u e Wo k
Fu u e wo k in he scope o his disse a ion include: (1) he imp o emen o he
eg ession model o p edic e e ence ankle join angles, in o de o a oid he a ibu ion o he
same ajec o y o subjec s wi h he same body heigh and walking speed; (2) he imp o emen
o he quali y o he da a acquisi ion by pe o ming walking es in a eadmill wi h o ce
pla o ms o a oid loss o he balance, specially a slow walking speeds; (3) he inclusion o
mo e muscles o es ima e he eal ankle join o que based on EMG signals and join angles;
(4) he explo a ion o machine lea ning algo i hms ope a ing in eal- ime o es ima e he eal
ankle join o que based on EMG signals and join angles, since hese empi ical me hods may
83
be mo e app op ia ed o pe o m he con e sion in mo o impai ed pe sons, because calib a ion
s eps a e no equi ed.
84
REFERENCES
[1] W. Johnson, O. Onuma, M. Owolabi, and S. Sachde , “Bulle in o Wo ld Heal h O ganiza ion,”
S oke: a global esponse is needed
, 2016. [Online]. A ailable:
h ps://www.who.in /bulle in/ olumes/94/9/16-181636/en/. [Accessed: 05-Jun-2019].
[2] F. A. Lo e io, “Análise do Pad ão de A i ação Muscula de Indi íduos Hemipa é icos Pós-AVC em
Ma cha Assis ida po Andado Robó ico,” Uni e sidade Fede al do Espí i o San o, 2015.
[3] S. H. Jang, “The eco e y o walking in s oke pa ien s: a e iew,”
In . J. Rehabil. Res.
, ol. 33,
no. 4, pp. 285–289, Dec. 2010.
[4] D. G. da Saúde, “P og ama Nacional de P e enção e Con olo das Doenças Ca dio ascula es,”
Dgs
, p. 28, 2006.
[5] A. H. Neu oHeal h, “A Pa en ’s Quick Guide o Hemiplegia and Hemipa esis,” 2017. [Online].
A ailable: h ps://neu oheal hah.com/guide- o-hemiplegia-hemipa esis/. [Accessed: 05-Jun-
2019].
[6] J. Cao, S. Q. Xie, R. Das, and G. L. Zhu, “Con ol s a egies o e ec i e obo assis ed gai
ehabili a ion: The s a e o a and u u e p ospec s,”
Med. Eng. Phys.
, ol. 36, no. 12, pp. 1555–
1566, Dec. 2014.
[7] G. Colombo, M. Joe g, R. Sch eie , and V. Die z, “T eadmill aining o pa aplegic pa ien s using
a obo ic o hosis.,”
J. Rehabil. Res. De .
, ol. 37, no. 6, pp. 693–700, 2000.
[8] A. Kaelin-Lang, L. Sawaki, and L. G. Cohen, “Role o Volun a y D i e in Encoding an Elemen a y
Mo o Memo y,”
J. Neu ophysiol.
, ol. 93, no. 2, pp. 1099–1103, 2004.
85
[9] M. Lo ze, “Mo o lea ning elici ed by olun a y d i e,”
B ain
, ol. 126, no. 4, pp. 866–872, Ap .
2003.
[10] L. Ma chal-C espo and D. Reinkensmeye , “Re iew o con ol s a egies o obo ic mo emen
aining a e neu ologic inju y,”
J. Neu oeng. Rehabil.
, 2009.
[11] K. Gui, H. Liu, and D. Zhang, “A P ac ical and Adap i e Me hod o Achie e EMG-Based To que
Es ima ion o a Robo ic Exoskele on,”
IEEE/ASME T ans. Mecha onics
, ol. 24, no. 2, pp. 483–
494, Ap . 2019.
[12] C. Ma, N. Chen, Y. Mao, D. Huang, R. Song, and L. Li, “Al e a ions o Muscle Ac i a ion Pa e n
in S oke Su i o s du ing Obs acle C ossing,”
F on . Neu ol.
, ol. 8, no. MAR, Ma . 2017.
[13] P. N. Fe nandes
e al.
, “EMG-based Mo ion In en ion Recogni ion o Con olling a Powe ed Knee
O hosis,” in
2019 IEEE In e na ional Con e ence on Au onomous Robo Sys ems and
Compe i ions (ICARSC)
, 2019, pp. 1–6.
[14] P. A emiadis, “EMG-based Robo Con ol In e aces: Pas , P esen and Fu u e,”
Ad . Robo .
Au om.
, ol. 1, no. 2, pp. 1–3, 2012.
[15] P. R. Ca anagh and P. V. Komi, “Elec omechanical delay in human skele al muscle unde
concen ic and eccen ic con ac ions,”
Eu . J. Appl. Physiol. Occup. Physiol.
, ol. 42, no. 3, pp.
159–163, No . 1979.
[16] S. Olney and C. Richa ds, “Hemipa e ic gai ollowing s oke. Pa I: Cha ac e is ics,”
Gai Pos u e
,
ol. 4, pp. 136–148, 1996.
[17] C. Fleische and G. Hommel, “A Human-Exoskele on In e ace U ilizing Elec omyog aphy,”
IEEE
T ans. Robo .
, ol. 24, no. 4, pp. 872–882, Aug. 2008.
[18] S. Jeze nik, G. Colombo, and M. Mo a i, “Au oma ic Gai -Pa e n Adap a ion Algo i hms o
Rehabili a ion Wi h a 4-DOF Robo ic O hosis,”
IEEE T ans. Robo . Au om.
, ol. 20, no. 3, pp.
574–582, Jun. 2004.
[19] R. M. Singh, S. Cha e ji, and A. Kuma , “T ends and challenges in EMG based con ol scheme
o exoskele on obo s—a e iew,”
In . J. Sci. Eng. Res
, ol. 3, no. 8, pp. 933–940, 2012.
[20] L. L. M. e Sil a, C. E. M. de Mou a, and J. R. P. de Godoy, “A ma cha no pacien e hemipa é ico,”
Uni . Ciências da Saúde
, ol. 3, no. 2, pp. 261–273, Ap . 2008.
[21] B. Balaban and F. Tok, “Gai Dis u bances in Pa ien s Wi h S oke,”
PM&R
, ol. 6, no. 7, pp. 635–
642, Jul. 2014.
[22] C. A. Johnson, J. H. Bu idge, P. W. S ike, D. E. Wood, and I. D. Swain, “The e ec o combined
use o bo ulinum oxin ype A and unc ional elec ic s imula ion in he ea men o spas ic d op

86
oo a e s oke: a p elimina y in es iga ion,”
A ch. Phys. Med. Rehabil.
, ol. 85, no. 6, pp. 902–
909, Jun. 2004.
[23] J. Pe y,
Gai Analysis: No mal and Pa hological Func ion
. Tho o a e: SLACK Inco po a ed, 1992.
[24] K. P. Pa ekh and J. B. Vyas, “Myoelec ic Leg o T ans emo al Ampu ee,”
In . J. Sci. Res. De .
,
ol. 1, no. 5, pp. 12–15, 2013.
[25] B. S Raja a nam, “A Compa ison o EMG Signals om Su ace and Fine-Wi e Elec odes Du ing
Shoulde Abduc ion,”
In . J. Phys. Med. Rehabil.
, ol. 02, no. 04, 2014.
[26] D. L. Wai e, R. L. B ookham, and C. R. Dicke son, “On he sui abili y o using su ace elec ode
placemen s o es ima e muscle ac i i y o he o a o cu as eco ded by in amuscula
elec odes,”
J. Elec omyog . Kinesiol.
, ol. 20, no. 5, pp. 903–911, Oc . 2010.
[27] M. Solomonow
e al.
, “Su ace and wi e EMG c oss alk in neighbou ing muscles,”
J. Elec omyog .
Kinesiol.
, ol. 4, no. 3, pp. 131–142, Jan. 1994.
[28] R. Bogey, K. Ce ny, and O. Mohammed, “Repea abili y o Wi e and Su ace Elec odes in Gai ,”
Am. J. Phys. Med. Rehabil.
, ol. 82, no. 5, pp. 338–344, May 2003.
[29] R. A. Bogey, J. Pe y, E. L. Bon age , and J. K. G onley, “Compa ison o ac oss-subjec EMG
p o iles using su ace and mul iple indwelling wi e elec odes du ing gai ,”
J. Elec omyog .
Kinesiol.
, ol. 10, no. 4, pp. 255–259, Aug. 2000.
[30] C. J. De Luca, “Su ace elec omyog aphy: De ec ion and eco ding,”
DelSys Inc.
, ol. 10, no. 2,
pp. 1–10, 2002.
[31] B. Ge dle, S. Ka lsson, S. Day, and M. Djupsjöbacka, “Acquisi ion, P ocessing and Analysis o he
Su ace Elec omyog am,” in
Mode n Techniques in Neu oscience Resea ch
, Be lin, Heidelbe g:
Sp inge Be lin Heidelbe g, 1999, pp. 705–755.
[32] D. A. Gab iel, “E ec s o monopola and bipola elec ode con igu a ions on su ace EMG spike
analysis,”
Med. Eng. Phys.
, ol. 33, no. 9, pp. 1079–1085, No . 2011.
[33] R. Me le i and P. Pa ke ,
Elec omyog aphy: Physiology, Enginee ing and Nonin asi e
Applica ions
. New Je sey: John Wiley & Sons, Inc, 2004.
[34] A. Su esh, “Body swa m in e ace ( BOSI ): Con olling Robo ic Swa ms Using Human Bio-
Signals,” Bos on Uni e si y, 2016.
[35] L. Shaw and S. Bagha, “Online EMG Signal Analysis o diagnosis o Neu omuscula diseases by
using PCA and PNN,”
In . J. Eng. Sci. Technol.
, ol. 4, pp. 4453–4459, 2012.
[36] C. ADIns umen s, “T igno EMG Senso s.” [Online]. A ailable:
h ps://www.adins umen s.com/p oduc s/ igno-emg-senso s. [Accessed: 02-Sep-2019].
87
[37] Delsys Inco po a ed,
T igno TM Wi eless Bio eedback Sys em
. 2018.
[38] No axon,
Ul iumTM EMG Sys em: Senso and Recei e Use Manual
. .
[39] I. Mo ion Lab Sys ems,
Mo ion Lab Sys ems: Use Guide
. 2004.
[40] BITalino: Biomedical Equipmen ,
MuscleBAN BE Da a Shee
. 2015.
[41] BITalino: Biomedical Equipmen ,
OpenSignals: Use Manual
. 2012.
[42] D. P. Fe is and C. L. Lewis, “Robo ic lowe limb exoskele ons using p opo ional myoelec ic
con ol,” in
2009 Annual In e na ional Con e ence o he IEEE Enginee ing in Medicine and Biology
Socie y
, 2009, pp. 2119–2124.
[43] H. Kawamo o and Y. Sankai, “Com o able powe assis con ol me hod o walking aid by HAL-
3,” in
IEEE In e na ional Con e ence on Sys ems, Man and Cybe ne ics
, 2003.
[44] D. Ao, R. Song, and J. Gao, “Mo emen Pe o mance o Human–Robo Coope a ion Con ol Based
on EMG-D i en Hill-Type and P opo ional Models o an Ankle Powe -Assis Exoskele on Robo ,”
IEEE T ans. Neu al Sys . Rehabil. Eng.
, ol. 25, no. 8, pp. 1125–1134, Aug. 2017.
[45] Rong Song, Kai-yu Tong, Xiaoling Hu, and Le Li, “Assis i e Con ol Sys em Using Con inuous
Myoelec ic Signal in Robo -Aided A m T aining o Pa ien s A e S oke,”
IEEE T ans. Neu al Sys .
Rehabil. Eng.
, ol. 16, no. 4, pp. 371–379, Aug. 2008.
[46] A. V. Hill, “The hea o sho ening and he dynamic cons an s o muscle,”
P oc. R. Soc. London.
Se . B - Biol. Sci.
, ol. 126, no. 843, pp. 136–195, Oc . 1938.
[47] M. G. Hoy, F. E. Zajac, and M. E. Go don, “A musculoskele al model o he human lowe ex emi y:
The e ec o muscle, endon, and momen a m on he momen -angle ela ionship o
musculo endon ac ua o s a he hip, knee, and ankle,”
J. Biomech.
, ol. 23, no. 2, pp. 157–169,
Jan. 1990.
[48] D. G. Lloyd and T. F. Besie , “An EMG-d i en musculoskele al model o es ima e muscle o ces
and knee join momen s in i o,”
J. Biomech.
, ol. 36, no. 6, pp. 765–776, Jun. 2003.
[49] M. Sa o i, D. G. Lloyd, M. Reggiani, and E. Pagello, “A s i endon neu omusculoskele al model
o he knee,” in
2009 IEEE Wo kshop on Ad anced Robo ics and i s Social Impac s
, 2009, pp.
132–138.
[50] M. Sa o i, M. Reggiani, A. J. an den Boge , and D. G. Lloyd, “Es ima ion o musculo endon
kinema ics in la ge musculoskele al models using mul idimensional B-splines,”
J. Biomech.
, ol.
45, no. 3, pp. 595–601, Feb. 2012.
[51] N. Ka a as, A. Ajoudani, N. Tsaga akis, J. Saglia, A. Bicchi, and D. Caldwell, “Tele-impedance
based assis i e con ol o a complian knee exoskele on,”
Rob. Au on. Sys .
, ol. 73, pp. 78–90,
88
No . 2015.
[52] Suncheol Kwon, Yunjoo Kim, and Jung Kim, “Mo emen S abili y Analysis o Su ace
Elec omyog aphy-Based Elbow Powe Assis ance,”
IEEE T ans. Biomed. Eng.
, ol. 61, no. 4, pp.
1134–1142, Ap . 2014.
[53] H. He and K. Kiguchi, “A S udy on EMG-Based Con ol o Exoskele on Robo s o Human Lowe -
limb Mo ion Assis ,” in
2007 6 h In e na ional Special Topic Con e ence on In o ma ion
Technology Applica ions in Biomedicine
, 2007, pp. 292–295.
[54] T. Yan, M. Cempini, C. M. Oddo, and N. Vi iello, “Re iew o assis i e s a egies in powe ed lowe -
limb o hoses and exoskele ons,”
Rob. Au on. Sys .
, ol. 64, pp. 120–136, Feb. 2015.
[55] K. Kiguchi and Y. Imada, “EMG-based con ol o lowe -limb powe -assis exoskele ons,” in
2009
IEEE Wo kshop on Robo ic In elligence in In o ma ionally S uc u ed Space
, 2009, pp. 19–24.
[56] W. Hassani, S. Mohammed, H. Ri aï, and Y. Ami a , “Powe ed o hosis o lowe limb mo emen s
assis ance and ehabili a ion,”
Con ol Eng. P ac .
, ol. 26, pp. 245–253, May 2014.
[57] C. Fleische and G. Hommel, “To que Con ol o an exoskele al knee wi h EMG signals,” in
P oceedings o he join con e ence on obo ics
, 2006.
[58] C. Fleische and G. Hommel, “Calib a ion o an EMG-Based Body Model wi h six Muscles o con ol
a Leg Exoskele on,” in
IEEE In e na ional Con e ence on Robo ics and Au oma ion
, 2007, pp.
2514–2519.
[59] W. Hassani, S. Mohammed, and Y. Ami a , “Real-Time EMG D i en Lowe Limb Ac ua ed O hosis
o Assis ance As Needed Mo emen S a egy,” in
Robo ics: Science and Sys ems IX
, 2013.
[60] W. Hassani, S. Mohammed, H. Ri ai, and Y. Ami a , “EMG based app oach o wea e -cen e ed
con ol o a knee join ac ua ed o hosis,” in
2013 IEEE/RSJ In e na ional Con e ence on
In elligen Robo s and Sys ems
, 2013, pp. 990–995.
[61] M. Mol edo, T. Baček, T. Ve s a en, C. Rod iguez-Gue e o, B. Vande bo gh , and D. Le ebe ,
“Powe ed ankle- oo o hoses: he e ec s o he assis ance on heal hy and impai ed use s while
walking,”
J. Neu oeng. Rehabil.
, ol. 15, no. 1, p. 86, Dec. 2018.
[62] C. Fleische , C. Reinicke, and G. Hommel, “P edic ing he in ended mo ion wi h EMG signals o
an exoskele on o hosis con olle ,” in
2005 IEEE/RSJ In e na ional Con e ence on In elligen
Robo s and Sys ems
, 2005, pp. 2029–2034.
[63] L. Liu, M. Luken, S. Leonha d , and B. J. E. Misgeld, “EMG-d i en model-based knee o que
es ima ion on a a iable impedance ac ua o o hosis,” in
IEEE In e na ional Con e ence on
Cybo g and Bionic Sys ems (CBS)
, 2017, ol. 2018–Janua, pp. 262–267.
89
[64] G. Du andau, D. Fa ina, and M. Sa o i, “Robus Real-Time Musculoskele al Modeling D i en by
Elec omyog ams,”
IEEE T ans. Biomed. Eng.
, ol. 65, no. 3, pp. 556–564, Ma . 2018.
[65] M. Chand apal, X. Chen, W. Wang, B. S anke, and N. Le Pape, “In es iga ing imp o emen s o
neu al ne wo k based EMG o join o que es ima ion,”
Paladyn, J. Beha . Robo .
, ol. 2, no. 4,
pp. 185–192, Jan. 2011.
[66] M. E. Hahn, “Feasibili y o es ima ing isokine ic knee o que using a neu al ne wo k model,”
J.
Biomech.
, ol. 40, no. 5, pp. 1107–1114, Jan. 2007.
[67] P. A. Huijing and G. C. Baan, “S imula ion le el-dependen leng h- o ce and a chi ec u al
cha ac e is ics o a gas ocnemius muscle,”
J. Elec omyog . Kinesiol.
, ol. 2, no. 2, pp. 112–
120, Jan. 1992.
[68] K. N. An, K. Takahashi, T. P. Ha igan, and E. Y. Chao, “De e mina ion o Muscle O ien a ions
and Momen A ms,”
J. Biomech. Eng.
, ol. 106, no. 3, pp. 280–282, 1984.
[69] W. Meng, Q. Liu, Z. Zhou, Q. Ai, B. Sheng, and S. S. Xie, “Recen de elopmen o mechanisms
and con ol s a egies o obo -assis ed lowe limb ehabili a ion,”
Mecha onics
, ol. 31, pp.
132–145, Oc . 2015.
[70] Y. H. Yin, Y. J. Fan, and L. D. Xu, “EMG and EPP-In eg a ed Human–Machine In e ace Be ween
he Pa alyzed and Rehabili a ion Exoskele on,”
IEEE T ans. In . Technol. Biomed.
, ol. 16, no. 4,
pp. 542–549, Jul. 2012.
[71] J. S. Campos Figuei edo, “Sma Wea able O hosis o Assis Impai ed Human Walking,”
Uni e si y o Minho, 2019.
[72] STMic oelec onics, “UM1472: STM32F407VG MCU - Use Manual.” 2017.
[73] J. Wang, L. Tang, and J. E B onlund, “Su ace EMG Signal Ampli ica ion and Fil e ing,”
In . J.
Compu . Appl.
, ol. 82, no. 1, pp. 15–22, No . 2013.
[74] C. J. De Luca, “The Use o Su ace Elec omyog aphy in Biomechanics,”
J. Appl. Biomech.
, ol.
13, no. 2, pp. 135–163, May 1997.
[75] A. Abdelouahad, A. Belkhou, A. Jba i, and L. Bella bi, “Time and equency pa ame e s o sEMG
signal — Fo ce ela ionship,” in
2018 4 h In e na ional Con e ence on Op imiza ion and
Applica ions (ICOA)
, 2018, no. 1, pp. 1–5.
[76] F. D. Fa án, J. C. Poli i, and C. J. Felice, “E alua ion o EMG p ocessing echniques using
In o ma ion Theo y,”
Biomed. Eng. Online
, ol. 9, no. 1, p. 72, 2010.
[77] D. A. Win e ,
Biomechanics and Mo o Con ol o Human Mo emen
. Hoboken, NJ, USA: John
Wiley & Sons, Inc., 2009.