Recei ed No embe 11, 2020, accep ed No embe 18, 2020, da e o publica ion No embe 24, 2020,
da e o cu en e sion Decembe 3, 2020.
Digi al Objec Iden i ie 10.1109/ACCESS.2020.3039885
A Machine Lea ning App oach o Pe o m
Physical Ac i i y Classi ica ion Using
a Senso ized C u ch Tip
ASIER BRULL MESANZA 1, SERGIO LUCAS1, ASIER ZUBIZARRETA 1, ITZIAR CABANES1,
EVA PORTILLO 1, AND ANA RODRIGUEZ-LARRAD2
1Depa men o Au oma ic Con ol and Sys ems Enginee ing, Facul y o Enginee ing o Bilbao, Uni e si y o he Basque Coun y (UPV/EHU),
48013 Bilbao, Spain
2Depa men o Physiology, Facul y o Medicine and Nu sing, Uni e si y o he Basque Coun y (UPV/EHU), 48940 Leioa, Spain
Co esponding au ho : Asie B ull Mesanza (asie [email p o ec ed])
This wo k was suppo ed in pa by he Uni e si y o he Basque Coun y [Uni e si y o he Basque Coun y (UPV/EHU)] unde G an
PIF18/067, in pa by he UPV/EHU unde P ojec GIU19/045 (GV/EJ IT1381-19), and in pa by he Minis e io de Ciencia e Inno ación
(MCI) unde G an DPI2017-82694-R (AEI/FEDER, UE).
ABSTRACT In ecen yea s, in e es in moni o ing Physical Ac i i y (PA) has inc eased due o i s posi i e
e ec on heal h. New echnological de ices ha e been p oposed o his pu pose, mainly ocused on spo s,
which include Machine Lea ning algo i hms o iden i y he ype o PA being pe o med. Howe e , PA
moni o ing can also p o ide da a use ul o assessing he eco e y p ocess o people wi h impai ed lowe -
limbs. In his wo k, a Machine-Lea ning based Physical Ac i i y classi ie design p ocedu e is p oposed,
which makes use o he da a p o ided by a Senso ized Tip ha can be adap ed o di e en Assis i e De ices
o Walking (ADW) such as canes o c u ches. The p ocedu e is based on h ee main s ages: 1) de ining
a wide se o po en ial ea u es o pe o m he classi ica ion; 2) op imizing he numbe o ea u es by a
Random-Fo es app oach, de ec ing he mos ele an ones o classi y i e ele an ac i i ies (walking a a
no mal pace, walking as , s anding s ill, going up s ai s and going down s ai s); 3) aining he ML-based
classi ie s conside ing he op imized ea u e se . A compa a i e analysis is ca ied ou o e alua e he
p oposed p ocedu e, using h ee ML-based classi ie (Suppo Vec o Machines, K-Nea es Neighbou and
A i icial Neu al Ne wo ks), demons a ing ha he p oposed app oach can p o ide e y high success a es
i p ope ea u e selec ion is ca ied ou . This wo k p esen s ou ele an con ibu ions o he PA moni o ing
a ea: 1) he app oach is ocused on people ha equi e ADW, which a e no conside ed in o he app oaches;
2) an analysis o he ea u es o cha ac e ize gai in people ha equi e ADW is ca ied ou ; 3) a design
p ocedu e o op imize he numbe o ea u es using a Random-Fo es app oach is used, a oiding a ypical
‘‘b u e o ce’’ p ocedu e; and 4) a compa a i e analysis is ca ied ou o demons a e he alidi y o he
app oach.
INDEX TERMS Ins umen ed c u ch, ehabili a ion, machine lea ning, physical ac i i y classi ica ion,
andom o es , a i icial neu al ne wo k, suppo ec o machine, K-nea es neighbo .
I. INTRODUCTION
Lowe -limb mobili y plays an impo an ole on au onomy
and quali y o li e. Neu ological diseases o auma inju ies
ha a ec he mobili y o he lowe -limb ha e a g ea impac
on he li es o people su e ing om hem. Hence, ying o
ully o pa ly eco e his unc ion is one o he main goals
when designing a ehabili a ion s a egy o hese pa ien s [1].
The associa e edi o coo dina ing he e iew o his manusc ip and
app o ing i o publica ion was Tyson B ooks .
In o de o be e ec i e, ehabili a ion in e en ions mus
be adap ed o he s a us o he pa ien du ing he whole
ehabili a ion p ocess [2]. This also includes he selec ion o
he assis i e de ice ha be e i s pa ien needs acco ding o
he /his unc ionali y. I he pa ien has los he abili y o walk
au onomously, he use o wheel chai s o scoo e s is he be e
op ion, while c u ches o canes a e ypically used when he
gai unc ion is main ained. Hence, he apis a e equi ed o
assess pa ien s a us pe iodically o moni o he e olu ion on
he s a us o he pa ien .
VOLUME 8, 2020 This wo k is licensed unde a C ea i e Commons A ibu ion 4.0 License. Fo mo e in o ma ion, see h ps://c ea i ecommons.o g/licenses/by/4.0/ 210023
A. B ull Mesanza e al.: ML App oach o Pe o m PA Classi ica ion Using a Senso ized C u ch Tip
Pa ien assessmen is ypically pe o med using he da a
collec ed h ough es s ca ied ou in clinical se ings.
Howe e , moni o ing he ypes o Physical Ac i i y (PA) he
pa ien pe o ms h oughou he day is becoming inc easingly
impo an in he unc ional assessmen o pa ien s, due o
he well-known bene i s i has on hei heal h and i s con-
ibu ion o he p e en ion o non-communicable diseases
[3], [4]. I also allows in e p e ing he esul s o he pe iodic
clinical es s and gi ing indi idualized ecommenda ions and
eedback on how much and how o pe o m hese ac i i ies in
o de o aid in he eco e y p ocess.
In o de o pe o m PA iden i ica ion, h ee main s eps
a e ypically ca ied ou : 1) da a ela ed o he pa ien is
cap u ed by a moni o ing de ice; 2) a se o ea u es ha
allow cha ac e izing PA is ex ac ed om he aw senso da a;
and, 3) he se o ea u es a e p ocessed by a classi ie , which
de ec s he pa icula PA being execu ed. The main wo ks
ela ed o each s ep will be summa ized nex .
Rega ding he da a cap u e and moni o ing, di e en ech-
nological solu ions ha e been p oposed [5]. The mos pop-
ula ones a e wea able de ices, which ha e o be a ached
o speci ic places o he lowe -limb o he pa ien , and
ypically cap u e mo ion da a using Ine ial Measu emen
Uni s (IMUs) [6]–[10] o biomedical signals such as EMG
[11], [12]. A numbe o comme cial de ices exis on he ma -
ke , such as XSens [13], BioS ampRC [14], T acmo [15]),
FlexiFo ce [16], BioCap u e [12]). These solu ions equi e
o be p ope ly placed and a ached o he limbs, and may
gene a e ejec ion on pa ien s. In o de o educe he impac
o moni o ing de ices he use o he in eg a ed senso s
o sma wa ches and mobile phones has been p oposed
[17]–[20]. These la e de ices do no ha e a speci ic place-
men in he body, bu , on he o he side, his posi ioning
lexibili y and he a iabili y in oduced by pa asi ic mo ions
a e issues o be conside ed when p ocessing he da a.
Once he aw da a has been cap u ed by he moni o ing
de ice, he second s ep is o ex ac a se o ea u es ha
will allow o cha ac e ize he di e en PA. Fo his pu pose,
he use o ime-se ies segmen a ion using ime-windows is
a common app oach, as i allows o educe he numbe o
da a o be p ocessed [21], [22]. In he case o gai moni o ing,
he selec ed window ypically ma ches a s ep. Hence, ea u es
o di e en na u e ha allow o cha ac e ize each s ep can
be ex ac ed om hese windows. S a is ic (mean, s anda d
de ia ion,...) [10], [15], [23]–[25], equency [19], [26] o
phase [9] ope a ions a e ypically applied o he cap u ed
a iables o his pu pose. In addi ion, in he pa icula case
o gai , ea u es such as he a e age speed, ime be ween
s eps o he numbe o s eps [27] ha e also been p oposed.
I is o be no ed ha he e is no s anda ized app oach o de ine
hese ea u es, and ha in gene al, a b u e o ce app oach is
used in which a wide se o ea u es is de ined so ha he
classi ie o be designed has enough inpu da a o pe o m
i s job.
Finally, in he hi d s ep, using he se o selec ed ea-
u es, he PA iden i ica ion o classi ica ion is pe o med.
Machine Lea ning (ML) echniques such as K-Nea es
Neighbou (K-NN) [10], [24], [28], Suppo Vec o
Machine (SVM) [29]–[31] and A i icial Neu al Ne wo ks
(ANN) [15], [32]–[34] a e he p e e ed solu ion o gai -
ela ed PA classi ica ion due o hei lexibili y and capabili y
o gene aliza ion, which p o ide accep able esul s wi h a
success a e up o 91% [15]. No e ha all hese app oaches a e
o supe ised na u e, and equi e a se o p ope ly designed
aining da a in which he selec ed ea u es a e he inpu , and
he ype o PA o be iden i ied a e he ou pu s. In he case o
gai - ela ed PA he p oposed classi ie s ypically iden i y i
he pa ien is walking (a an habi ual o no mal speed o as e ,
i.e. unning), going up and down s ai s o s anding
s ill [9], [15], [28], [32], [35].
The h ee-s ep p ocedu e de ailed p e iously p o ides a
gene al me hodology o PA classi ica ion. Howe e , i is o
be no ed ha he e is no s anda ized app oach o be ollowed
in each s ep, and ha di e en open esea ch a eas s ill exis .
In pa icula , he ea u e selec ion p ocedu e is ypically ca -
ied ou using a b u e o ce app oach, in which a wide se
o possible ea u es a e p oposed as inpu s o he ML-based
Physical Ac i i y classi ie , so ha i can ha e enough da a
o pe o m he classi ica ion. This app oach, howe e , leads
o non-op imal classi ie s, which ypically use mo e ea u es
han equi ed leading o o e sized solu ions, as he ela i e
impo ance o each ea u e is no usually analyzed.
Mo eo e , all he a o emen ioned wo ks a e designed o
people ha do no equi e Assis i e De ices o Walking
(ADW) such as c u ches o canes. Howe e , se e al pa ame-
e s change signi ican ly in he case o people ha use ADW,
as hey p esen non-symme ical gai and pa ame e s such
as he load applied o he ADW migh be ele an . These
di e ences ha e o be conside ed in he h ee-s ep p ocedu e.
Recen wo ks ha e demons a ed ha pa ien s ha equi e
ADW in hei ehabili a ion p ocess equi e speci ic moni-
o ing app oaches [36], being senso ized ADW de ices he
bes op ion o his popula ion [37]–[41]. Hence, he se o
ea u es o be de ined also has o conside ADW da a.
Based on he p e ious analysis, in his wo k, a no el
app oach o he de elopmen o Physical Ac i i y classi ie s
o pa ien s ha equi e ADW is p oposed. The p oposed
app oach aims o gi e some insigh in o he p e iously ci ed
issues, wi h ou ele an con ibu ions: 1) The app oach is
ocused on people ha equi e ADW, which a e no consid-
e ed in o he wo ks; 2) A comp ehensi e se o ea u es o
classi y i e ele an ypes o Physical Ac i i y is p oposed
and analyzed ; 3) A Fea u e Selec ion me hodology based
on a Random-Fo es app oach is p oposed; and, 4) A ho -
ough compa a i e analysis using h ee ML app oaches
(K-NN, SVN and ANN) is ca ied ou o alida e he p o-
posed app oach.
The es o he wo k is s uc u ed as ollows. Sec ion II
p esen s he Senso ized Tip and i s senso iza ion capabili ies.
Sec ion III de ails he se o es s ca ied ou o gene a e he
da abase used o de elop he ML-based PA classi ica o s.
In Sec ion IV a ho ough analysis o he po en ial ea u es
210024 VOLUME 8, 2020
A. B ull Mesanza e al.: ML App oach o Pe o m PA Classi ica ion Using a Senso ized C u ch Tip
p oposed in he li e a u e o gai moni o ing is ca ied ou ,
and he p oposed me hodology o selec he mos ele an
ea u es and ain h ee di e en ML-based PA classi ica o s
using K-NN, SVN and ANN app oaches is de ailed. Finally,
in Sec ion V, a compa a i e analysis is ca ied ou o e alua e
he app oach. Finally, he mos impo an ideas a e summa-
ized in Sec ion VI.
II. SENSORIZED TIP FOR GAIT MONITORING
In o de o moni o he pe o mance o people ha equi e
ADW, di e en app oaches can be used, as analyzed in
Sec ion I. Wea able de ices, al hough widely used, p esen
some d awbacks o his popula ion, as hey may gene a e
ejec ion due o he need o a aching he senso s o he limbs,
and do no conside he in e ac ion o ce be ween he ADW
and he pa ien , which p o ides ele an moni o ing da a.
Sma phones and wa ches, on he o he hand, p esen pa asi ic
mo ions ha ha e o be conside ed.
Hence, se e al wo ks ha e p oposed o senso ize ADW,
p o iding a nonin asi e app oach ha p o ides accu a e
measu emen s o bo h ADW mo ion and in e ac ion o ce
[37]–[41]. In pa icula , in his wo k he Senso ized Tip p o-
posed in [41] (Figu e 1) is used o cap u e gai da a. Di e -
en om he o he ci ed app oaches in which a senso ized
c u ch o cane is designed, he p oposed Senso ized Tip can
be a ached o he pe sonal c u ch o cane used by he pa ien ,
which is ypically adap ed o his/he needs.
FIGURE 1. Senso ized Tip o Cap u e Gai Da a.
The Senso ized Tip in eg a es h ee senso s in i s alu-
minum enclosu e. A 9 deg ees-o - eedom Ine ial Measu e-
men Uni MTi-3 by XSens p o ides linea accele a ion da a,
angula speed and magne ic ield in he local (x,y,z) axes.
FIGURE 2. La e omedial and An e opos e io angles in he ADW.
In addi ion, his de ice in eg a es a p op ie a y algo i hm
based on a Kalman il e ha allows o es ima e he oll-
pi ch-yaw Eule angles in he global e e ence ame (Roll
and Pi ch dynamic e o o 0.5◦, and Yaw dynamic e o 1◦).
The a o emen ioned da a can also be used o es ima e he
an e opos e io and la e omedial c u ch angles (see Figu e 2).
A BMP280 ba ome e p o ides in o ma ion on a mosphe ic
p essu e, which allows o es ima e he ela i e heigh o he
de ice ( ela i e p ecision o 0.12hPa). Finally, a C9C piezo-
elec ic o ce senso by HBM, wi h 1 kN ange, p o ides
in o ma ion on he axial load exe ed by he pa ien . The
o e all weigh o he Tip is 160g.
The 16 sou ces o da a p o ided by he a o emen ioned
senso s a e cap u ed by a nRF52832 mic op ocesso , which
adds a imes amp and sends he p ocessed da a wi h a 20ms
pe iod o a mobile phone de ice using he Blue oo h Low
Ene gy (BLE) p o ocol. The da a is s o ed in he phone using
a sel -de eloped app, so ha i can be p ocessed la e . The
cap u ing sys em is powe ed by a s anda d 5V powe bank,
which is placed ex e nally o he Tip in o de o minimize he
weigh o he de ice (Figu e 1).
The ull cha ac e iza ion o he measu emen e o s and
he in eg a ed algo i hms o he Senso ized Tip can be
ound a [41].
III. DATA BASE FOR CLASSIFIER DESIGN
In o de o de elop a Physical Ac i i y classi ie using
Machine Lea ning app oaches, a p ope da a base is equi ed,
in which he selec ion o he ypes o PA o be iden i ied is a
key issue.
As analyzed in Sec ion I, when conside ing gai - ela ed PA
classi ie i e ypes o PA a e ypically conside ed [9], [15],
[28], [32], [35]: walking a a no mal pace; walking a a as
pace (app oxima ely 30% as e han no mal pace); going up
s ai s; going down s ai s; and s anding s ill. The iden i ica ion
o hese ypes o PA will allow o moni o he ac i i y o
a pa ien h ough i s daily li e, de ining pa e ns o ac i i y,
VOLUME 8, 2020 210025
A. B ull Mesanza e al.: ML App oach o Pe o m PA Classi ica ion Using a Senso ized C u ch Tip
seden a iness, e c,...This da a can be used by he he apis o
p o ide indi idualized ecommenda ions o o de ec possible
modi ica ions in he pa ien unc ional s a us [26], [42].
In o de o cap u e ele an da a o he classi ie design,
a o al o i e es s ha e been ca ied ou using a c u ch in
which he Senso ized Tip de ailed in Sec ion II was a ached:
•Walking 30m in a s aigh line a he no mal speed.
•Walking 30m in a s aigh line a a speed highe han
no mal (app oxima ely 30% as e ).
•S anding s ill o app oxima ely 10 seconds.
•Going up an 11-s ep ligh o s ai s.
•Going down an 11-s ep ligh o s ai s.
The es s we e ca ied ou by 11 heal hy olun ee s om
he esea ch g oup o he au ho s (4 women and 7 men,
anging be ween 24-48 yea s), a he acili ies o he Facul y
o Enginee ing o Bilbao UPV/EHU. Each es was epea ed
h ee imes o each olun ee .
In o de o gene a e he da abase, a segmen a ion p ocedu e
was ollowed [28]. This p ocedu e is ca ied ou by conside -
ing each cycle o use o he c u ch, which is composed by
a s ance phase (in which he c u ch is in con ac wi h he
g ound), and he swing phase (in which he c u ch is li ed
hough he ai and no con ac exis s). This way, he aw da a
p o ided by each senso is di ided in sequen ial windows,
each associa ed o a c u ch cycle. The ini ial poin o each
window is de ined a he e y i s s a o he s ance phase,
in which he c u ch ip con ac s he g ound. This can be easily
de ec ed by conside ing he o ce senso signal, as seen in
Figu e 3, as no o ce exis in he swing phase. The o al num-
be o segmen ed windows gene a ed in he a o emen ioned
es s a e summa ized in Table 1.
TABLE 1. Numbe o Windows pe Physical Ac i i y (PA). Tes and T aining
Se s.
No e ha in he case o S anding S ill, he a o emen ioned
app oach is no longe alid, as no c u ch cycles exis . In hese
scena ios a i ual s ep is conside ed as a ixed segmen a ion
window o 1.8s, which is sligh ly longe han he a e age
cycle ime o he cycles conside ed in he walking a no mal
pace scena io.
Once he da abase is de ined, i will be di ided in o wo
balanced se s (T aining and Tes ), as equi ed by he design
p ocedu e o supe ised ML-based app oaches [43]. The
T aining se will be used o ain he p oposed ML-based
PA classi ica o s. Fo ha pu pose, a balanced se has been
de ined, wi h app oxima ely he same numbe o windows
conside ed o he di e en iden i ied ypes o PA. This
allows o ain he classi ie wi h he same ela i e impo ance
o each ype o PA. The Tes se , in he o he hand, will be
FIGURE 3. Cycle o use o an ADW and i s phases. Da a Segmen a ion in
windows by using he da a acqui ed om he o ce senso .
used o es he designed classi ie s. Hence, Tes da a will
no be used in he PA classi ie design p ocedu e, bu o he
alida ion analysis ca ied ou Sec ion V. No e ha in his
la e case, a balanced window selec ion has also been ca ied,
so ha he es ed classi ica ion success a es can be simila in
na u e o each ype o PA o be iden i ied [44], [45].
IV. MACHINE LEARNING-BASED PA CLASSIFIER
DESIGN METHODOLOGY
The use o segmen a ion allows o de ine disc e e uni s o
da a, one o each c u ch cycle, om which a se ies o ea u es
can be ex ac ed. These ea u es, which may be di e se in
na u e (s a is ical, equency based,...) can be used o cha -
ac e ize each cycle, and be used as inpu s o he PA classi i-
ca ion sys em o be de eloped. In his sec ion, a me hodology
is de ailed o selec he mos app op ia e ea u es and design
he ML-based PA classi ica o .
The p oposed me hodology is summa ized in Figu e 4:
Fi s , a se o po en ial ea u es based on he ones p oposed in
he li e a u e is p oposed (Sec ion IV-A). This se is de ined
wi h a high numbe and a ie y o ea u es, so ha he
maximum amoun o in o ma ion can be conside ed. Then,
in a second s ep, a Random-Fo es app oach is used o de e -
mine he ela i e impo ance o each ea u e, allowing o
o de he po en ial ea u e se conside ing he ele ance o
each ea u e (Sec ion IV-B).This o de ed se will be used
210026 VOLUME 8, 2020
A. B ull Mesanza e al.: ML App oach o Pe o m PA Classi ica ion Using a Senso ized C u ch Tip
FIGURE 4. Fea u e selec ion me hodology.
o design he ML-based classi ie . In a hi d s ep, he se
o op imal hype pa ame e s will be calcula ed o each se
o n ea u es o be conside ed as inpu s. Finally, using he
selec ed hype pa ame e s and he se o n ea u es selec ed
(based on hei ele ance), he ML app oach will be ained
(Sec ion IV-C). An analysis and e alua ion o he p ocedu e
will be ca ied ou in Sec ion V.
A. POTENTIAL FEATURES SET GENERATION
Fea u es a e ela ed o he da a sou ces a ailable, as hey a e
used o ex ac , using a simple me ic, a pa icula cha ac-
e is ic o he signal con ained in he segmen ed window. Fo
he pa icula case de ailed in his wo k, 17 sou ces o da a a e
conside ed based on he da a p o ided by he Senso ized Tip
(Sec ion II): 9 associa ed o he aw IMU da a (x,y,zcompo-
nen s o accele a ion, angula speed and magne ic ield in he
local axes); 5 ela ed o he p ocessed IMU da a (RPY Eule
Angles, and c u ch an e opos e io and la e omedial angles);
1 ela ed o he o ce senso alue, which is il e ed; and,
2 associa ed o he ba ome e signal ( il e ed and un il e ed).
Fo each segmen ed window, he ime e olu ion o hese
17 da a sou ces can be p ocessed o ex ac a ea u e. This is
ca ied ou by applying an ope a o , which may be o di e en
na u e (s a is ical, ime-based,...). Al hough he pa icula
case o people ha equi e ADW has no been analyzed in
he li e a u e, based on he ope a o s p oposed in he ela ed
wo ks and he clinical expe ience o he au ho s, he ollowing
se o ope a o s a e p oposed:
•S a is ic-based ope a o s: They a e widely used in gai
cha ac e iza ion wo ks, as hey a e easily applied o
any da a sou ce. Mean alue,s anda d de ia ion, a i-
ance,ku osis,co ela ion coe icien s XY (i.e., be ween
X and Y signals), pe cen iles,a ea unde each cu e and
in e qua ile anges [15], [23]–[25] ha e been selec ed
o be applied o he da a p o ided by all senso s. In he
pa icula case o co ela ion coe icien s, he co ela ion
be ween he di e en angles/axes alues p o ided by
a senso a e conside ed, i.e. co ela ion be ween he
accele ome e xand ysignals, co ela ion be ween oll
and pi ch Eule Angles, e c.
•Mo ion-based ope a o s: The alues o mo ion- ela ed
sou ces o da a in speci ic e en s allows o de ine ea-
u es ela ed o he use o he ADW. In pa icula , he
alues associa ed o he s a o he s ance phase (S ance
S a Value), he end o he s ance phase (S ance End
Value) and he alue associa ed o he maximum suppo
(Value a Max. Fo ce) a e o pa icula in e es . The
Ampli ude, de ined as he absolu e di e ence be ween
he maximum and minimum alues o a mo ion a iable
is also de ined.
•Time-based ope a o s: Measu ing he ime be ween spe-
ci ic e en s allows o ob ain spa io- empo al ea u es.
In he case o ADW, cycle ime, his is, he ime be ween
consecu i e s a s o he s ance phase, allows o de ine
speed- ela ed ea u es [27]. The use o he ADW can also
be de ined by compa ing he ela i e pe cen age o he
cycle ime he pa ien uses he de ice o suppo , his
is, he ime o he s ance phase wi h espec o he cycle
ime (S ance Phase %) [11].
By combining he se o da a sou ces and he de ined ope -
a o s, a ull se o 176 ea u es can be de ined. All a e sum-
ma ized in Table 2, whe e an Xde ines a ea u e (o ea u es)
ha has been ob ained by applying a pa icula ope a o ( ow)
o a da a sou ce (column). No e ha his se o 176 ea u es
is ex ac ed o each ADW cycle, ollowing he segmen a ion
p ocedu e de ailed in Sec ion III.
B. FEATURE SELECTION USING
RANDOM-FOREST APPROACH
The a o emen ioned se o 176 ea u es can be used o de elop
ML-based PA classi ie s. This way, he se o ea u es will be
conside ed as he inpu o he classi ie , which will iden i y a
ype o PA o each ADW cycle as seen in Figu e 4.
Howe e , his b u e o ce app oach, which is ypical in he
wo ks ci ed in he in oduc ion is no an e icien one. Fi s ,
a high numbe o ea u es inc eases he compu a ional cos
o he classi ie . Second, he ea u e selec ion impac s he
pe o mance o he classi ie , as some ea u es may be no
be ela ed o he ypes o PA conside ed, o e en some a e
co ela ed one wi h he o he . Hence, in o de o op imize he
PA classi ie design, p ope ea u e selec ion app oaches mus
be used.
De ec ing he bes ea u e se o design an PA classi-
ie is no a i ial ask. In ecen yea s, Machine Lea n-
ing app oaches ha e demons a ed hei abili y o analyze
he ela i e impo ance o di e en ea u es when analyzing
VOLUME 8, 2020 210027
A. B ull Mesanza e al.: ML App oach o Pe o m PA Classi ica ion Using a Senso ized C u ch Tip
TABLE 2. Fea u es gene a ed om he da a p o ided by he senso ized ip (R=Roll, P=Pi ch, Y=Yaw, A=An e opos e io , L=La e omedial).
a classi ica ion o eg ession p oblem. One o he mos in e -
es ing app oach in his ield is he Random Fo es (RF)
[46], [47] app oach, which consis s on he gene a ion o a
wide se ( o es ) o di e en decision ees o classi ica ion
pu poses. The ees a e gene a ed using a se o andom
samples and ea u es, so ha in he aining p ocedu e, di -
e en ea u es can be es ed. This echnique has been used
in di e en applica ion ields such as diagnosis [48], mine al
p ocess indus ies [49] o DNA analysis [50], o es ima e
he ela i e impo ance o each ea u e. This way he mos
ele an ones can be iden i ied, and he ones ha a e edun-
dan o unimpo an elimina ed.
Hence, in his wo k, a Random-Fo es app oach is p o-
posed o analyze he ela i e ea u e signi icance o he PA
classi ica ion. Fo ha pu pose, only he samples con ained
in o he T aining Se de ined in Sec ion III ha e been used.
The p oposed RF has been implemen ed using Ma lab’s
S a is ics and Machine Lea ning Toolbox [51] and expe i-
men ally uned conside ing he ollowing se o hype pa am-
e e s: he numbe o ees in he o es has been uned o
5000; a sample wi h eplacemen s a egy has been selec ed;
a node size o 1 was de ined; he numbe o a iables an-
domly chosen a each spli (m y) has been uned o √M,
whe e Mis he o al numbe a iables; and he p edic o used
has been he in e ac ion-cu a u e o a oid he dis u bances
caused by co ela ed ea u es.
The ob ained esul s om his p ocedu e a e summa ized
in Table 3, in which all he po en ial ea u es ha e been
so ed in dec easing o de o dec easing ela i e signi icance
acco ding o he RF app oach. No e ha he RF app oach
o de s he ea u es by conside ing hei ela i e weigh o con-
ibu ion o he desi ed classi ica ion p ocess, being he A ea
Unde he Cu e o he Yaw angle and he Cycle Time some
o he mos ele an ea u es o he p oposed s udy-case.
I is o be no ed ha i all weigh s o he 176 ea u es
a e analyzed, all p esen a posi i e weigh wi h he excep ion
o he las wo, ela ed o he Ba ome e In e qua ile Range.
This means ha ollowing he RF analysis, he ea u es wi h
posi i e weigh con ibu e (o add in o ma ion) o he PA
classi ica ion. Howe e , he ela i e impo ance o he mos
signi ican one A ea Unde he Cu e Yaw is mo e han
50 imes highe wi h espec o he less signi ican ones.
Hence, designing an PA classi ie using only some o he mos
ele an ones should p o ide be e esul s han he use o he
less ele an ones. In he nex sec ion, a compa a i e analysis
will be ca ied ou o analyze he e ec o he p oposed ea u e
selec ion.
C. CLASSIFIER HYPERPARAMETER SELECTION
AND TRAINING
Once he po en ial se o ea u es has been o de ed acco ding
o i s ele ance, a subse o n ea u es can be selec ed o design
a PA classi ie . The goal o he classi ie s is o be able o de ec
i e ele an PA: Walking no mal, Walking as , Going up and
down s ai s and s anding s ill. Hence, all classi ie s will be
implemen ed wi h 5 ou pu s/classes, one associa ed o each
PA ype.
In his wo k, he h ee mos commonly used app oaches
in ela ed wo ks ha e been selec ed, so ha a compa a i e
analysis can be ca ied ou in he nex sec ion: Suppo Vec o
Machine (SVM), K-Nea es Neighbo (K-NN) and A i icial
Neu al Ne wo k (ANN).
As de ailed in Figu e 4, o a gi en se o n
ele ance-o de ed inpu ea u es, i s he op imal subse o
210028 VOLUME 8, 2020
A. B ull Mesanza e al.: ML App oach o Pe o m PA Classi ica ion Using a Senso ized C u ch Tip
TABLE 3. Fea u e signi icance and hei ela i e weigh acco ding o Random-Fo es p ocedu e. (Magne=Magne ome e , Accel=Accele ome e ,
AU C=A ea Unde he Cu e, 25P=25 h Pe cen ile, 50P=50 h Pe cen ile, 75P=75 h Pe cen ile, IR=In e cua ile Range, SD=S anda d De ia ion, Co .
Coe .=Co ela ion Coe icien , An e o=An e omedial Angle, La e o=La e omedial Angle, WoF=Wi hou Fil e , WF =Wi h Fil e , n=Posi ion).
hype pa ame e s o each ML-based classi ie is o be cal-
cula ed. Fo ha pu pose a K- old c oss- alida ion app oach
is p oposed wi h K=5 [52]. This app oach allows o
e ec i ely e alua e di e en ML-based models. No e ha
o his pu pose only he da a om he T aining Se de ined
in Sec ion III. is used. Once he bes hype pa ame e s ha e
been chosen, hese a e used o ain he ML-app oach using
supe ised me hods and he T aining Se da a.
I is o be no ed ha in he case o he SVM and K-NN,
Ma lab’s S a is ic and Machine Lea ning Toolbox in eg a es
VOLUME 8, 2020 210029
A. B ull Mesanza e al.: ML App oach o Pe o m PA Classi ica ion Using a Senso ized C u ch Tip
he a o emen ioned s eps, op imizing he ela ed hype pa am-
e e s (Ke nel unc ions, numbe o neighbou s,...) [51]. Fo
he case o he ANN, he au ho s ha e ad hoc p og ammed he
hype pa ame e selec ion. In his la e case, a single hidden
laye Mul i Laye Pe cep on (MLP) ANN has been selec ed,
wi h 5 ou pu neu ons (one o each PA), a numbe o inpu s
equal o he n ea u e se o be p ocessed and mhidden laye
neu ons wi h hype bolic angen sigmoid ac i a ion unc ion.
The numbe o hidden laye neu ons mhas been conside ed
as he hype pa ame e o be uned using he a o emen ioned
p ocedu e, wi h m anging om 1 o 10 neu ons, since expe -
imen al es s ha e de e mined ha ANN wi h 10 o lowe
neu ons p o ide good esul s. Once he bes (highe success
a e) alue o mhas been selec ed, a Bayesian egula iza ion-
based aining algo i hm is used o ain he ANN.
V. COMPARATIVE ANALYSIS
In his sec ion, a compa a i e analysis is ca ied ou consid-
e ing he ea u es selec ed in Sec ion IV-B. The aim is o:
1) analyze he bes app oach o he p oposed PA classi ica o
applica ion; and 2) analyze he alidi y o he ea u e selec ion
app oach in di e en ML-based classi ica ion app oaches.
No e ha all he ML-based classi ie s analyzed in his
sec ion ha e been ained ollowing he me hodology p o-
posed in he p e ious sec ion.
A. ANALYSIS OF THE EFFECT OF THE NUMBER OF
FEATURES CONSIDERED FOR CLASSIFICATION
In o de o analyze he e ec o he numbe o ea u es
conside ed, a compa a i e analysis is ca ied ou conside ing
he ea u es de ined in Table 3. This way, each ML-based
classi ica ion app oach p oposed p e iously is ained wi h
176 di e en ea u e se s ollowing he p ocedu e de ailed in
Sec ion IV-C. These ea u e se s a e de ined inc emen ally
conside ing he nmos ele an ea u es. This is: in he i s
se , only he mos ele an ea u e is conside ed; in he second
one, he wo mos ele an ea u es a e conside ed; while in
he las one, all 176 po en ial ea u es a e conside ed.
FIGURE 5. Success a e o he classi ie s based on K-NN, SVM and ANN,
wi h espec o he numbe o he nmos ele an ea u es o de ed
acco ding o he RF.
Figu e 5shows he o al classi ica ion success a e pe -
cen age o he p oposed app oaches wi h espec o he
numbe no he mos ele an ea u es acco ding o he
RF app oach. This success a e is de ined as he pe cen age o
PA samples o he Tes Se whose ype he classi ie iden i ies
co ec ly wi h espec o he o al numbe o PA samples in
he se . No e ha he samples in his la e se ha e no been
conside ed in he aining p ocedu e, so ha he esul s can
be used o analyze also he gene aliza ion capabili y o he
app oaches.
As i can be seen, i he se en mos signi ican ea u es
a e conside ed, a success a e pe cen age o o e 90% can be
achie ed in all cases (92.8% o he K-NN, 97% o he SVM
and 96.8% o he ANN). This alue inc eases up o 97% i
he nine mos ele an ea u es a e selec ed o all app oaches.
The gene al endency is ha a highe numbe o ea-
u es conside ed allows be e classi ica ion. A maximum
success a e o 98.4% (66 ea u es) o he K-NN, 99.1%
(87 ea u es) o he SVM and 99.6% (174 ea u es) o he
ANN is ob ained. No e ha he small oscilla ions a e due o
he andomized na u e o he ML app oaches aining, wi h a
success a e a ia ion in he ange om 7 o 176 mos ele an
ea u es o 2.8% in he case o he ANN and 4.6% o he
SVM.
The e is an excep ion in he case o he K-NN app oach,
as he pe cen age o success dec eases sligh ly when he
numbe o ea u es is highe han 119, eaching a alue lowe
han 96% (92.8% wi h 147 and 160 mos ele an ea u es).
The esul s con i m ha i a p ope ea u e selec ion is
ca ied ou , a small se o ea u es can be used o design he
ML-based PA classi ica o , as he e ec o inc easing he
numbe o ea u es is small in he o al success a e o he clas-
si ie . Mo eo e , his has an impac on compu a ional cos .
As p e iously s a ed, a K-Fold c oss- alida ion p ocedu e has
been used o calcula e he bes con igu a ion o hype pa ame-
e s o each ea u e se . Fo ins ance, in he pa icula case o
he ANN he ob ained op imal numbe o hidden laye neu-
ons is summa ized in Figu e 6 o each ea u e se . I can be
seen ha al hough a lowe numbe o neu ons (5-6) is equi ed
o small alues o n, he numbe o neu ons s abilizes wi h
a mean o 9 neu ons. Hence, selec ing a mode a e numbe o
ea u es ( o example he 7 mos ele an ones) also leads o
smalle ANN and lowe compu a ional cos .
Finally, in o de o illus a e he classi ica ion capabili-
ies o he ML-based PA classi ie s, a pa icula example
o he classi ie s pe o mance is shown in Table 4, whe e
he Con usion Ma ices o all classi ie s when all ea u es
a e conside ed a e shown. In his pa icula case, he o e all
pe o mance o he K-NN is 96.1%, SVM pe o mance is
96.8% and in ANN 99.6%. Howe e , i can be seen ha he
K-NN has a p oblem classi ying Walking No mal case, as up
o 20 samples a e iden i ied e oneously as Walking Fas and
Going Up S ai s. The same e ec is seen in he SVM’s Con u-
sion Ma ix. The ANN ou pe o ms he p e ious app oaches,
ob aining be e esul s.
B. ANALYSIS OF THE EFFECT OF FEATURE SELECTION
In o de o emphasize he impo ance o he ea u e selec ion
p ocedu e, he p ocedu e de ined in he p e ious sec ion has
210030 VOLUME 8, 2020
A. B ull Mesanza e al.: ML App oach o Pe o m PA Classi ica ion Using a Senso ized C u ch Tip
TABLE 4. Con usion ma ix o K-NN, SVM and ANN wi h all ea u es.
FIGURE 6. Op imal numbe o neu ons in he hidden laye o he ANN
wi h espec o he numbe o he nmos ele an ea u es o de ed
acco ding o he RF.
been epea ed om he less signi ican ea u e o he mos
signi ican one. This is, 176 se s o ea u es ha e been ana-
lyzed: he i s se has conside ed only he less signi ican
ea u e; he second, he wo less signi ican ea u es; and so
on. As p e iously de ailed, o each se o ea u es, he op i-
mum hype pa ame e s ha e been uned, by he use o a K- old
p ocedu e. Fo he pa icula case o he ANN, an a e age
o 7.99 neu ons wi h a s anda d de ia ion o 1.75 neu ons
ha e been ob ained.
Resul s a e summa ized in Figu e 7 o all p oposed
app oaches. As i can be seen, he success a es e olu ion
p esen s an inc easing endency. This is, as mo e signi ican
ea u es a e added, he classi ie quali y inc eases. Hence,
he success a e inc eases when adding mo e and mo e ea-
u es, om app oxima ely 22% o 99%.
No e ha his is a e y di e en e olu ion compa ed wi h
he one analyzed in he p e ious sec ion (Fig. 5). In he p e-
ious case, wi h ew o he mos signi ican , success a es up
o 95% could be achie ed, while in his la e case, a g ea e
numbe o ea u es a e equi ed o achie e he same pe o -
mance: 86 o SVM, 68 o ANN, and almos all ea u es o
K-NN. This emphasizes he need o co ec ly selec ing he
ea u es o designing PA classi ie s.
FIGURE 7. Success a e o he K-NN, SVM and ANN based classi ie s, wi h
espec o he numbe o he nless ele an ea u es o de ed acco ding o
he RF app oach.
FIGURE 8. Success a e o he K-NN, SVM and ANN based classi ie s, wi h
7 indica o s as inpu selec ed acco ding o he ela i e signi icance
p o ided by he RF app oach.
The ele ance o co ec ly selec ing he ea u es is also
demons a ed in Figu e 8. As analyzed in he p e ious sub-
sec ion, he se en mos signi ican ea u es p o ide accep -
able success a es o he classi ie (o e 92%). Hence, all
p oposed app oaches ha e been e alua ed by conside ing
se s o 7 ea u es. This is, he i s 7 ea u es ha e been
e alua ed i s , hen he nex 7 and so on, o de ed om he
VOLUME 8, 2020 210031
