16 h In e na ional Symposium on P ac ical Design o Ships and O he Floa ing S uc u es PRADS 2025
Ann A bo , MI, USA, Oc obe 19 h-23 d 2025
Towa d P ocess Imp o emen in Hull Cons uc ion: A Combined Wo k
Moni o ing and Simula ion F amewo k o Schedule De ia ion Analysis
Chenwei Gui1,*, Jaeyeon Nam1, Tomoyuki Taniguchi2,
Kohei Ma suo2and Kazuhi o Aoyama1
1The Uni e si y o Tokyo, Tokyo, Japan
2Na ional Ins i u e o Ma i ime, Po and A ia ion Technology, Japan
Abs ac . Mode n shipya ds ace pe sis en challenges in achie ing e icien , on- ime p oduc ion due
o he inhe en complexi y o hull cons uc ion. The high unce ain y in oduced by di e se p oduc
a ie ies and labou -in ensi e ope a ions makes i imp ac ical o ully p ede e mine and op imize p o-
duc ion wo k lows. In p ac ice, shipya ds ypically de elop only ough daily schedules a he shop le el,
lea ing many de ailed asks o he judgmen and expe ise o ield wo ke s. Howe e , he a iabili y in
wo ke beha io , d i en by di e en le els o expe ience, can in oduce addi ional non- alue-adding
ac i i ies ha may lead o signi ican de ia ions om he planned schedule. This pape in es iga es
he impac o wo ke beha io on schedule de ia ions in a subassembly line by in eg a ing wo k mon-
i o ing ools and agen -based simula ion. We employ a YOLO model o ex ac wo k- ela ed ea u es
om su eillance ideo eeds. These moni o ed da a o m a ime se ies o wo k ac i i ies, which we
hen segmen o pinpoin c i ical pe iods whe e delays a e mos likely o occu . By acking wo ke s’
beha io ajec o ies wi hin hese c i ical windows, we unco e pa e ns ha con ibu e o subop imal
pe o mance. In pa allel, we u ilize an agen -based hull cons uc ion simula o o gene a e nea -op imal
wo k ajec o ies, le e aging insigh s de i ed om he beha io pa e ns o e e an wo ke s. By high-
ligh ing gaps be ween simula ion-based op imal scena ios and ac ual wo k lows, ou indings e eal he
ole o nuanced human ac o s in delay occu ence. This in eg a ed analysis amewo k p o ides clea
isibili y in o how subop imal wo k ac i i ies a ise, enabling mo e a ge ed e alua ions ha can se e
as he basis o p ocess imp o emen s in hull cons uc ion.
Key wo ds: Wo k Moni o ing, Beha io Modeling, Shipbuilding P ocess Simu-
la ion, Schedule De ia ion Analysis
1. In oduc ion
Hull cons uc ion in mode n shipbuilding p esen s pe sis en challenges due o i s high p oduc a ie y,
eliance on manual labo , and dynamic shop loo condi ions. The complexi y o assembling highly cus-
omized componen s unde a ying spa ial and empo al cons ain s makes i di icul o de elop de ailed,
obus p oduc ion plans. As a esul , shipya ds ypically ely on coa se daily schedules a he shop le el,
wi h he ine g anula i y o ask execu ion le o he judgmen o indi idual wo ke s. This eliance on
human decision-making in oduces signi ican unce ain y in o he p oduc ion p ocess, pa icula ly when
di e ences in expe ience, skill, and si ua ional awa eness lead o non- alue-adding ac i i ies such as e-
pea ed adjus men s, ine icien mo emen s, o poo ly sequenced ac ions– all o which can ul ima ely esul
in p oduc ion delays.
To add ess he di icul ies in shipya d p oduc ion planning, a ious simula ion-based app oaches ha e
been de eloped o imp o e schedule accu acy and sys em pe o mance. K ause e al. [1] p oposed a dis-
c e e e en simula ion (DES) amewo k as a decision-suppo ool o e alua ing al e na i e in es men
and scheduling scena ios ac oss shipya d acili ies, emphasizing s a egic and ope a ional planning. Lee
e al. [2] de eloped a p ocess-cen ic simula ion pla o m aimed a imp o ing mas e scheduling accu acy
*Co espondence o: [email p o ec ed]
1
h ough o ecas ing and scena io e alua ion. Okubo and Mi suyuki [3] p oposed a me hod o au oma -
ically gene a ing ealis ic p oduc ion plans using s uc u ed sys em models o wo k low and eam allo-
ca ion. While hese wo ks demons a e he u ili y o simula ion in op imizing shipya d ope a ions, hey
p edominan ly ocus on mac o-le el o sys em-le el planning and ypically assume ideal ask execu ion
and o e look he a iabili y in oduced by wo ke -le el decisions—p ecisely whe e many delays o igina e.
This gap highligh s he need o in eg a ing de ailed human beha io in o simula ion o cap u e he oo
causes o mic o-le el ine iciencies, which can help e ine he o e all accu acy o p oduc ion planning.
To complemen mac o-le el planning me hods and p o ide insigh s in o shop loo eali ies, many s ud-
ies ha e ocused on ex ac ing mic o-le el in o ma ion om uns uc u ed on-si e da a. In pa icula , su eil-
lance ideos accumula ed h ough pas shipbuilding p ojec s ha e been le e aged o cap u e de ailed wo ke
ac i i ies, wi h ecen e o s using compu e ision echniques o ans o m his da a in o s uc u ed ep e-
sen a ions o wo k beha io . Kim e al. [4] p oposed a ision-based sys em ha segmen s ideo s eams,
iden i ies block componen s, and compa es hem wi h CAD models o ack assembly p og ess. Shinoda e
al. [5] in oduced a deep neu al ne wo k amewo k ha classi ies wo k i ems om head-moun ed came a
oo age, enabling au oma ed obse a ion o welding ope a ions based on de ined ask ca ego ies. Gui e al.
[6] combined YOLO-based objec de ec ion wi h ime-se ies analysis o iden i y complex wo k ac i i ies
by ecognizing sequences o basic beha io s in ideo. While hese e o s success ully ex ac aluable
mic o-le el wo k in o ma ion, hey a ely close he loop om moni o ing o unde s anding he causes o
ine iciencies, no do hey in eg a e beha io al insigh s in o planning o simula ion models o p ocess im-
p o emen .
To help b idge his gap, his s udy p oposes an in eg a ed amewo k ha combines wo k moni o ing and
simula ion o sys ema ically analyze schedule de ia ions in hull cons uc ion. The co e idea is o cap u e
and model wo ke beha io du ing p oduc ion using compu e ision-based moni o ing, and hen compa e
hese obse ed pa e ns wi h s anda d wo k lows gene a ed h ough simula ion. This compa ison allows us
o quan i y de ia ions in ask execu ion and iden i y beha io -d i en ine iciencies ha con ibu e o delays.
The amewo k consis s o wo main componen s. Fi s , a wo k moni o ing sys em ex ac s ime-se ies
da a o wo ke ac i i ies om su eillance oo age, allowing us o de ec c i ical pe iods whe e schedule
de ia ions a e likely o occu . The cap u ed beha io is hen ep esen ed in a s uc u ed o ma using a Be-
ha io Pa e n Ma ix (BPM), which encodes he du a ion o asks a speci ic loca ions as well as ansi ions
be ween hem. Second, an agen -based simula ion gene a es nea -op imal wo k lows based on p ede ined
p ocess ules and physical cons ain s, se ing as a e e ence o compa ison. By analyzing he di e ences
be ween he moni o ed and simula ed beha io pa e ns, he amewo k enables a de ailed e alua ion o how
ac ual wo ke ac ions con ibu e o de ia ions om planned ope a ions.
This s udy makes se e al key con ibu ions o he ield o shipbuilding p oduc ion planning and con ol.
Fi s , i p oposes a sys ema ic me hod ha combines compu e ision echniques wi h ime-se ies analysis o
de ec ask execu ion pe iods whe e schedule delays a e likely o o igina e. Second, i in oduces a ma ix-
based ep esen a ion o mic o-le el wo k beha io , enabling quan i a i e e alua ion o wo ke pe o mance
and es ablishing a s uc u ed link be ween obse ed ac ions and simula ion-based planning—se ing as a
p oo o concep o in eg a ing human beha io in o digi al wins. Thi d, he amewo k o e s a wo k-
cen ic pe spec i e on schedule de ia ion analysis, shi ing ocus om sys em-le el p ocesses o indi idual
decision-making and i s impac on p oduc ion ou comes. Finally, he app oach is alida ed h ough a case
s udy in a eal shipya d subassembly p ocess, demons a ing i s easibili y and p ac ical ele ance.
The emainde o his pape is o ganized as ollows. Sec ion 2 desc ibes he shipya d con ex and he
moni o ing amewo k used o de ec de ia ion-p one pe iods based on isual da a. Sec ion 3 in oduces
he agen -based simula ion en i onmen and explains how s anda d wo k lows a e gene a ed o compa -
ison. Sec ion 4 p esen s he esul s o he case s udy, including he ex ac ion o beha io pa e ns om
moni o ing da a, he gene a ion o simula ed wo k lows, and a quan i a i e compa ison be ween he wo
o analyze beha io al de ia ions. Finally, Sec ion 5 summa izes he indings and discusses di ec ions o
u u e esea ch and p ac ical implemen a ion.
2
2. Wo k Moni o ing o Schedule De ia ion De ec ion and Beha io al
Modeling
2.1. O e iew o he Subassembly Line and Fi ing Tasks
In hull cons uc ion, he subassembly p ocess se es as a ounda ional s ep in o ming he s uc u al
blocks ha make up he essel’s hull. I in ol es he ab ica ion and assembly o smalle componen s —such
as s aigh o cu ed pla es and hei co esponding s eng hening s i ene s— in o la ge sub-blocks, called
sub-assemblies, which a e la e joined du ing block assembly s ages [7]. The accu acy and e iciency o
he subassembly p ocess di ec ly impac he quali y and imeliness o downs eam ope a ions, making i a
c i ical phase in he shipbuilding wo k low.
F1 F2 RB W1 W2
Subassembly Line
Figu e 1.: Layou o he sequen ial wo ks a ions in he subassembly line.
The subassembly line examined in his s udy comp ises a se ies o sequen ial wo ks a ions, as illus a ed
in Figu e 1.. Pla es a e anspo ed along a con eyo h ough he ollowing s a ions:
•F1: Pa eeding and alignmen - Base pla es a e ca ied in o he wo ks a ion, and s i ene s a e
hen deli e ed and aligned nea he base pla e o ini ial posi ioning.
•F2: Fi ing - Wo ke s adjus and empo a ily ix s i ene s on o he base pla e.
•RB: Robo ic Main Welding - Au oma ed welding o mos o he weldable sec ions.
•W1: Manual Welding - Wo ke s pe o m manual welding on a eas inaccessible o obo s, comple -
ing he s uc u al connec ions.
•W2: Finishing and Inspec ion Final quali y checks and adjus men s a e made be o e he subassem-
blies p og ess o he nex s age.
Among he a ious asks pe o med on he subassembly line, he i ing ope a ion a he F2 wo ks a ion
is o pa icula in e es due o i s manual and decision-in ensi e na u e. In his ask, wo ke s a e esponsible
o accu a ely posi ioning s i ene s on he base pla e, making eal- ime adjus men s, and de e mining he
sequence and loca ion o ini ial ack welds o empo a ily secu e componen s. Al hough each i ing job is
guided by a shop-le el p oduc ion plan ha de ines he spa ial layou and expec ed comple ion imeline, he
de ailed execu ion—especially he assembly sequence—is le o he disc e ion o indi idual wo ke s.
This eliance on human judgmen in oduces a iabili y in o he p ocess. Mic o-le el decisions, in lu-
enced by expe ience, local cons ain s, o wo king condi ions, can lead o subop imal wo k lows ha include
epea ed adjus men s, unnecessa y mo emen , o ewo k. Such de ia ions om he ideal sequence no only
educe e iciency a he ask le el bu can also con ibu e o b oade delays in he p oduc ion schedule. In
his s udy, we ocus on he i ing ask as a ep esen a i e example o in es iga e how indi idual beha io
and on-si e decision-making impac schedule adhe ence in hull cons uc ion.
3
2.2. Visual Moni o ing F amewo k o De ia ion-P one Task De ec ion
To iden i y ine iciencies ha may con ibu e o schedule de ia ions, a isual moni o ing amewo k
was de eloped o cap u e and analyze wo ke ac i i ies and wo ks a ion s a us h oughou he p oduc ion
p ocess. As shown in Figu e 2., he amewo k combines objec de ec ion wi h ime se ies analysis o
p o ide a s uc u ed iew o p oduc ion beha io .
(a) Time Se ies Analysis o Job Segmen a ion (b) Time Se ies Analysis o Local Fea u e Visualiza ion
Wo ke Pos u es
S and
Squa
Weld
Video F ames
YOLO11
Image
Classi ica ion
Objec
De ec ion
Wo ks a ion S a us
G ound Base Pla e Deli e ed WIP Subassembly
Figu e 2.: Visual moni o ing amewo k o he iden i ica ion o po en ial sou ces o schedule de ia ion. (a)
Segmen ed wo ks a ion s a us imeline showing de ec ed job bounda ies alongside he planned job schedule
o compa ison. (b) Wo ke pos u e ime se ies wi h co esponding local wo k ea u e cu es, illus a ing
a ia ions in wo k pa e ns ac oss di e en jobs.
The moni o ing begins wi h objec de ec ion using a ine- uned YOLO11n model [8]. Fo wo ke pos-
u e de ec ion, he model was ained on 975 anno a ed images collec ed om h ee di e en shipya ds,
augmen ed o o e 3,700 images. I de ec s h ee pos u es: s anding, squa ing, and welding, e lec ing
di e en s ages o ask execu ion, including inciden al ac ions such as posi ioning o adjus men s. Wo ks a-
ion s a us is iden i ied using YOLO11n wi h a classi ica ion head YOLO11n-cls, ine- uned on 475 images
o classi y i e o e all s ages: g ound, base pla e, deli e ed, WIP, and subassembly. Due o he complex
bo de s and o e lapping elemen s in shipya d en i onmen s, con en ional objec de ec ion s uggled wi h
accu a ely de ec ing he s a us o indi idual componen s. Image classi ica ion was he e o e used o as-
sess he o e all s a us o he wo ks a ion, whe e g ound indica es no pla e p esen , base pla e signals pla e
placemen , deli e ed ma ks s i ene deli e y, WIP shows ongoing i ing wo k, and subassembly deno es
comple ion.
These de ec ion esul s a e p ocessed h ough ime se ies analysis. Wo ks a ion s a us changes a e used
o segmen he job in o dis inc phases, while sequences o wo ke pos u es help isualize local wo k pa -
e ns. Toge he , hese elemen s suppo he iden i ica ion o a ia ions in ask execu ion ha may signal
ine iciencies. As shown in Figu e 2.a, job segmen a ion is pe o med based on he ime se ies o wo k-
s a ion s a us. The aw s a us da a is i s smoo hed using a mo ing a e age, a e which Change Poin
De ec ion (CPD) is applied o iden i y po en ial ansi ions be ween jobs. The CPD algo i hm o mula es
he segmen a ion ask as an op imiza ion p oblem, seeking b eakpoin s ha maximize he selec ed good-
ness o i measu e o he segmen s [9]. To cap u e non-linea changes ypical in p oduc ion p ocesses, we
use ke nel CPD wi h a adial basis unc ion (RBF) ke nel. This allows de ec ion o a ious changes in he
dis ibu ion, no limi ed o shi s in mean o a iance, and is pa icula ly e ec i e o complex s uc u ed
4
da a [10]. A low penal y se ing in CPD algo i hm enables he de ec ion o mul iple candida e b eakpoin s,
as shown in he second ow o Figu e 2.a. These candida es a e u he e ined using ule-based il e ing,
whe e pa e ns aligned wi h con eyo mo emen s se e as indica o s o ue job bounda ies ( hi d ow).
Compa ing he il e ed segmen a ion ( hi d ow) wi h he planned job imeline ( ou h ow), i is e iden
ha he hi d and ou h jobs ook longe han expec ed, wi h he hi d job showing a pa icula ly signi ican
delay.
To p o ide mo e speci ic insigh s o schedule de ia ion analysis, local wo k ea u es we e analyzed
wi hin each segmen ed job, as shown in Figu e 2.b. The i s h ee ows display he YOLO-based pos u e
de ec ions o e ime (coun o s anding, squa ing, and welding), ollowed by CPD esul s applied wi hin
each job segmen , and he compu ed local ea u e alues o hese segmen s. Two key ea u es we e used:
local main wo k a e, de ined as welding ime di ided by o al wo k ime (in man-hou s), and p og ess
inc emen , calcula ed as he welding ime ela i e o he o al welding ime o he job. Abno mal pa e ns
in hese ea u es, such as low main wo k a es o s agna ion in p og ess, sugges ine iciencies. In Job 3,
wo dis inc pla e pe iods we e obse ed, along wi h no ably low main wo k a es, indica ing his job may
be a majo con ibu o o he o e all delay. This p o ides supplemen a y e idence linking speci ic wo k
beha io o schedule de ia ions.
2.3. Beha io Pa e n Ma ix Cons uc ion om Moni o ing Da a
To enable de ailed analysis o wo ke beha io du ing he i ing ask, a quan i a i e ep esen a ion was
de eloped in he o m o a Beha io Pa e n Ma ix (BPM), which cap u es bo h he du a ion o wo k a each
wo kplace ( ypically a s i ene ) and he ansi ions be ween wo kplaces. The BPM is de ined as a squa e
ma ix, whe e diagonal en ies indica e he o al ime spen a each wo kplace, and o -diagonal en ies
ep esen he numbe o ansi ions om one wo kplace o ano he . This s uc u e allows o a compac ye
exp essi e iew o bo h ask ocus and mo emen pa e ns, making i sui able o compa ison, isualiza ion,
and u he analysis.
To cons uc he BPM om moni o ing da a, DBSCAN clus e ing [11] was applied o he spa ial-
empo al dis ibu ion o de ec ed welding poin s, enabling he iden i ica ion o disc e e welding wo kplaces.
These clus e s we e hen aligned wi h spa ial a angemen planning da a o map hem o speci ic s i ene
loca ions on he pla e. Once wo kplaces we e es ablished, he du a ions o welding and inciden al wo k
we e accumula ed o compu e he diagonal elemen s. Fo he o -diagonal elemen s, he welding ime se-
ies was comp essed in o a one-dimensional sequence o wo kplace isi s, om which ansi ions we e
coun ed. Sho -du a ion, noisy mo emen s we e il e ed ou o a oid o e es ima ing ansi ions caused by
mino hesi a ions o de ec ion e o s.
This ma ix o ma no only enables a s uc u ed compa ison o ac ual and simula ed wo k lows bu also
p o ides a basis o downs eam ea u e ex ac ion and ne wo k-based analysis. Me ics such as modula i y,
ansi ion en opy, o cen ali y can be de i ed om he BPM o assess beha io al pa e ns and e iciency.
By p ese ing bo h loca ion-speci ic engagemen and ask low, he BPM se es as a ounda ional ep esen-
a ion o iden i ying de ia ions, e alua ing pe o mance, and unde s anding how execu ion un olds ac oss
space and ime.
3. Simula ion-Based Baseline Modeling and Compa ison Me hod
3.1. Simula ion Mechanism o S anda d Wo k low Gene a ion
To be e suppo p ecise p oduc ion planning, especially in es ima ing wo kloads and de ining bench-
ma k wo k lows, a mul i-agen -based simula o was used o ealis ically model he shipbuilding p ocess, in-
cluding inciden al wo k. This simula o was de eloped by he Na ional Ma i ime Resea ch Ins i u e (NMRI)
and is implemen ed on he Uni y pla o m [12]. I cap u es bo h he beha io o indi idual wo ke s and he
physical layou o he shipya d in a i ual en i onmen .
As shown in Figu e 3., he simula o is composed o h ee main elemen s: A i ual shipya d, which
models bo h he acili ies and he p oduc s being assembled. Wo ke agen s, which ac au onomously,
5
Facili ies
P oduc
Wo k
Task 1
Task 2
Basic
Task 1
Basic
Task 2
Basic
Task 3
Vi ual Shipya d (En i onmen )
Wo ke Agen
Obse a ion Ac ion
(Task Selec ion
& Execu ion)
Pa e n A
Decision-Making
Model
Pa e n B
Pa e n C
Wo k low
A
Wo k low
B
Wo k low
C
Figu e 3.: O e iew o he agen -based shipya d simula o and i s mechanism.
making decisions based on hei su oundings. A hie a chical wo k model whe e asks a e s uc u ed in o
laye s, b eaking down complex ac i i ies in o simple , mo e manageable basic asks.
Be o e unning he simula ion, a Bill o P ocess (BOP) needs o be de ined o each p oduc , ou lining
he equi ed s eps o assembly. Once he simula ion s a s, each wo ke agen e alua es a ailable asks and
selec s wha o do nex based on hei pe cep ion o he en i onmen and p ede ined decision-making ules.
This selec ion p ocess is go e ned by a p io i y unc ion 1:
ask =1
k(Wp+b)×Y
pc∈p
sgn(pc)(1)
whe e:
•p ep esen s he ea u e ec o o he en i onmen , including he cu en ask and i s con ex .
•Wis a weigh ec o ha de e mines he impo ance o each ea u e.
•bis a bias e m ela ed o he ype o wo k.
•kis he numbe o ea u es in p, used o no maliza ion.
•pc e e s o speci ic cons ain - ela ed ea u es, such as ask p ecedence, and sgn(pc)adjus s p io i y
based on whe he cons ain s a e sa is ied.
A e selec ing a ask, i is b oken down in o basic asks, whose du a ions a e calcula ed using simula ion
pa ame e s. Fo example, he ime o a i ing ask is de e mined by he leng h o he welding line di ided
by he wo ke ’s i ing speed, plus addi ional ime o adjus ing posi ions based on he s i ene ’s condi ion.
This se up allows he simula o o accoun o bo h main and inciden al wo k, p oducing ealis ic wo k lows
ha can la e be compa ed o ac ual obse ed beha io .
While i is di icul o de ine a single op imal wo k low in p ac ice, he simula o gene a es baseline
wo k lows by implemen ing h ee di e en ask selec ion s a egies. These s a egies ep esen di e en
app oaches o sequencing and decision-making, p o iding a ied beha io pa e ns o analysis:
6
1. Au onomous S a egy: Agen s p io i ize minimizing mo emen dis ance, selec ing he nea es a ail-
able wo kplace a each decision poin . This models an e iciency-d i en app oach wi hou adhe ence
o a p ede ined sequence.
2. S anda d S a egy: Agen s ollow a ecommended sequence based on s anda d ope a ing p oce-
du es, ep esen ing s ic compliance wi h p ede ined wo k ins uc ions.
3. Heu is ic S a egy: Agen s assign highe p io i y o ho izon al s i ene s while main aining lexi-
bili y in choosing subsequen wo kplaces. This hyb id s a egy balances e iciency wi h s uc u al
conside a ions, modeling expe -like beha io .
In ac ual p oduc ion, some p ecedence cons ain s go e n he sequence o assembly. I hese cons ain s
a e no ollowed, addi ional ime and ewo k isk can a ise due o di icul ies such as poo welding pos u e
o limi ed access. In he simula o , when a gene a ed wo k low iola es such cons ain s, an addi ional ime
penal y is applied based on he speci ic condi ion o he p oduc .
3.2. Beha io Pa e n Compa ison Me hod o Schedule De ia ion Analysis
To enable s uc u ed compa ison be ween ac ual and simula ed wo k lows, bo h we e ans o med in o
Beha io Pa e n Ma ices (BPMs) wi h a uni ied o ma . Each BPM was in e p e ed as a di ec ed weigh ed
g aph (DWG), whe e nodes co esponded o wo kplaces and edge weigh s ep esen ed he equency o
ansi ions be ween hem. Based on his ep esen a ion, clus e ing was pe o med using g eedy modula i y
maximiza ion [13] o iden i y cohesi e subs uc u es wi hin he wo k low, e lec ing he unde lying s uc u e
o ask execu ion.
In addi ion o isual compa isons, a se o quan i a i e ea u es was ex ac ed. F om he diagonal en-
ies o he BPM, ime-based s a is ics—including a e age, s anda d de ia ion, and skewness o wo kplace
du a ions— we e compu ed o desc ibe he ypical wo kload dis ibu ion, a iabili y, and e o balance.
Fo s uc u al analysis, he o -diagonal en ies we e ea ed as an adjacency ma ix, and se e al g aph-
heo e ic ea u es we e de i ed. Modula i y was used o quan i y he s eng h o communi y s uc u e in
he ansi ion ne wo k, indica ing how well he wo k low di ided in o dis inc phases. The a e age deg ee
was calcula ed o ep esen he numbe o ansi ions pe wo kplace, e lec ing o e all wo k low complex-
i y. The clus e ing coe icien [14] calcula ed using he weigh ed di ec ed o mula ion based on geome ic
means, indica ing he p esence o igh ansi ion loops possibly associa ed wi h ewo k o localized loops.
Eigen ec o cen ali y [15] was also calcula ed o cap u e he quali y o connec ions based on he idea ha a
node is impo an i i is connec ed o o he impo an nodes. Las ly, g aph densi y was measu ed o assess
he o e all le el o in e connec i i y ac oss he ne wo k.
To complemen hese s uc u al ea u es, he BPM was also no malized in o a ansi ion p obabili y
ma ix, allowing he en opy o mo emen s o be calcula ed. This alue e lec ed he deg ee o andomness
in a wo ke ’s ansi ions, wi h highe en opy sugges ing less p edic able o less s uc u ed wo k lows.
These ime- and g aph-based me ics oge he o med a ea u e ec o ha enabled quan i a i e simila i y
analysis be ween ac ual and simula ed beha io , suppo ing he iden i ica ion o beha io -d i en de ia ions
and guiding imp o emen e o s.
4. Resul s and Discussion
4.1. Beha io Pa e n Ex ac ion and Analysis om Moni o ing Da a
As illus a ed in Figu e 2., Job 3 was iden i ied as a po en ial sou ce o schedule de ia ion. This job
in ol ed wo pla es occupying he en i e wo kshop a ea, placed symme ically on he le and igh sides,
as shown in Figu e 4.a. Bo h pla es had iden ical designs bu we e assigned o wo di e en wo ke s—
Wo ke A on he le and Wo ke B on he igh —who pe o med he asks independen ly. Acco ding o
expe e alua ion based on epea ed ideo e iews, Wo ke B appea ed o exhibi highe e iciency and skill
in ask execu ion, al hough his sequence did no always align wi h he p ede ined s anda d wo k low. In
con as , Wo ke A gene ally adhe ed o he sugges ed sequence bu exhibi ed equen local mo emen s,
pa icula ly wi hin con ined a eas, sugges ing possible hesi a ion o epea ed adjus men s.
7
0
1 2
3
45
6
7
8
9
10
11
0
1
2
3
4
5
6
7
8
9
10
11
(a) Layou o wo kplaces comple ed in
Job3 a he F2 s a ion.
400 200 0 200 400
Time (s)
WP 0
WP 1
WP 2
WP 3
WP 4
WP 5
WP 6
WP 7
WP 8
WP 9
WP 10
WP 11
0.070.06
0.110.04
0.370.03
0.080.31
0.130.33
0.280.48
0.200.27
0.190.07
0.350.77
0.090.03
0.030.07
0.080.12
Welding (A)
Inciden al Wo k (A)
Welding (B)
Inciden al Wo k (B)
(b) Compa ison o welding and inciden al wo k imes o Wo ke A and
Wo ke B ac oss all wo kplaces.
Figu e 4.: Wo kplace dis ibu ion and wo ke e iciency compa ison o Job 3.
Figu e 4.a shows he wo kplaces (s i ene s) comple ed a he F2 s a ion, while Figu e 4.b compa es he
welding and inciden al wo k imes pe wo kplace o bo h wo ke s. The s acked ba s ep esen he ime
spen on welding and inciden al asks, and he alue a he end o each ba ep esen s he main wo k a e a
ha loca ion. O e all, Wo ke B main ained highe main wo k a es ac oss mos wo kplaces, suppo ing he
expe judgmen ega ding supe io ask execu ion skills. In con as , Wo ke A eco ded highe inciden al
wo k ime a se e al loca ions. The signi ican ime spen a WP 0 may indica e an ex e nal dis u bance o
abno mal di icul y a ha loca ion.
To u he in es iga e he di e ences in wo k pa e ns, Beha io Pa e n Ma ices (BPMs) we e con-
s uc ed o bo h wo ke s, as shown in Figu e 5.. These ma ices cap u e bo h he ime spen a each wo k-
place (diagonal elemen s) and he equency o ansi ions be ween wo kplaces (o -diagonal elemen s).
Figu e 5.a and 5.b display he BPMs o Wo ke A and Wo ke B, espec i ely. Wo ke A’s BPM e eals
mo e sca e ed ansi ions and highe a iabili y in wo k ime ac oss wo kplaces, while Wo ke B’s BPM
shows mo e ocused ac i i y, pa icula ly a ound WP 3, WP 9, and WP 10.
To isualize hese dynamics mo e in ui i ely, he BPMs we e also ep esen ed as di ec ed weigh ed
g aphs, as shown in Figu e 5.c and 5.d. In hese g aphs, node posi ions co espond o he ac ual spa ial
layou o he wo kplaces. Node sizes e lec he o al ime spen a each wo kplace, and edge wid hs ep-
esen he equency o ansi ions. Addi ionally, node colo s indica e clus e s ob ained ia g aph-based
communi y de ec ion, which co espond o he modula blocks isible in he ma ix ep esen a ion.
Figu e 5.c shows ha Wo ke A ollowed a highly in e connec ed mo emen pa e n, wi h dense an-
si ions among WP 4, WP 6, WP 7, and WP 9, indica ing complex and po en ially ine icien ask ou ing.
In con as , Figu e 5.d shows Wo ke B’s wo k low as mo e modula and spa ially cohesi e. F om an in u-
i i e pe spec i e, B’s g aph appea s mo e egula , wi h clea ly de ined ask phases and ewe c oss- egion
ansi ions. This isualiza ion ein o ces he ma ix-based analysis, illus a ing how indi idual beha io
di e ences—especially in mo emen s uc u e and ask alloca ion—can signi ican ly in luence execu ion
e iciency and highligh a eas whe e a ge ed guidance o in e en ion may imp o e pe o mance.
4.2. S anda d Wo k low Gene a ion ia Simula ion
To es ablish baseline wo k lows o compa ison, simula ions we e conduc ed using he h ee ask se-
lec ion s a egies in oduced ea lie : au onomous, s anda d, and heu is ic. The esul ing Beha io Pa e n
Ma ices (BPMs) and hei co esponding DWG a e p esen ed in Figu e 6..
The au onomous s a egy minimizes o al mo emen by allowing he agen o selec he nea es a ail-
able wo kplace a each s ep. Howe e , his o en leads o iola ions o p ecedence cons ain s— pa icula ly
hose equi ing ho izon ally placed s i ene s ( ypically cen al nodes) o be ack welded be o e hei e -
ically placed neighbo s. When such cons ain s a e igno ed, ime penal ies a e imposed o simula e he
8
01234567891011
Wo kplace
0
1
2
3
4
5
6
7
8
9
10
11
Wo kplace
484 22000000000
21341000000000
0160100100000
10018701200000
00001560200100
0000058200000
000431 214 50000
000000495 1 1 0 1
0000001046000
00000002020701
00000000011781
00000010002158
0
1
2
3
4
5
T ansi ions (O -Diagonal)
0
100
200
300
400
Time (Diagonal)
(a) Beha io pa e n ma ix o Wo ke A (le pla e).
01234567891011
Wo kplace
0
1
2
3
4
5
6
7
8
9
10
11
Wo kplace
190 10000000000
1 120 0 200000000
0197100000000
012131 1 3010000
1002960000000
0002 2 65000000
0001018102000
0000000961000
00000030 52 0 0 0
000000000297 12
0000001001208 0
0000000001152
0.0
0.5
1.0
1.5
2.0
2.5
3.0
T ansi ions (O -Diagonal)
0
50
100
150
200
250
Time (Diagonal)
(b) Beha io pa e n ma ix o Wo ke B ( igh pla e).
0
1 2
3
4 5
6
7 8
9
10 11
(c) BPM ne wo k o Wo ke A (le pla e).
0
1 2
3
4 5
6
7 8
9
10 11
(d) BPM ne wo k o Wo ke B ( igh pla e).
Figu e 5.: Compa ison o BPM and co esponding ne wo k g aphs o Wo ke A and Wo ke B.
01234567891011
Wo kplace
0
1
2
3
4
5
6
7
8
9
10
11
Wo kplace
59 10000000000
0117 1 010000000
1 1 194 000000000
0005800100000
0000122 1 000000
00010124 000000
0000005610000
0000000114 1 000
00000000111 001
0000000005500
0000000001 122 0
00000000001 122
0.0
0.2
0.4
0.6
0.8
1.0
T ansi ions (O -Diagonal)
0
25
50
75
100
125
150
175
Time (Diagonal)
(a) BPM o simula ed au onomous
s a egy.
01234567891011
Wo kplace
0
1
2
3
4
5
6
7
8
9
10
11
Wo kplace
124 00100000000
0106 1 000100000
01 111 000000000
0105600000000
0000104 000010 0
00001 101 000000
0000015800000
0000000102 0010
00000001 101 000
00000000161 0 0
0000000000112 1
00000000000112
0.0
0.2
0.4
0.6
0.8
1.0
T ansi ions (O -Diagonal)
0
20
40
60
80
100
120
Time (Diagonal)
(b) BPM o simula ed s anda d s a -
egy.
01234567891011
Wo kplace
0
1
2
3
4
5
6
7
8
9
10
11
Wo kplace
124 00100000000
0113 0010000000
01 113 000000000
0005600100000
0000102 0010000
0010 0 104 000000
000000560010 0
0100000107 0000
0000010 0 102 000
0000000005801
0000000010102 0
00000000001 101
0.0
0.2
0.4
0.6
0.8
1.0
T ansi ions (O -Diagonal)
0
20
40
60
80
100
120
Time (Diagonal)
(c) BPM o simula ed heu is ic s a -
egy.
0
1 2
3
4 5
6
7 8
9
10 11
(d) BPM ne wo k g aph o au o-
ma ed s a egy.
0
1 2
3
4 5
6
7 8
9
10 11
(e) BPM ne wo k g aph o s anda d
s a egy.
0
1 2
3
4 5
6
7 8
9
10 11
( ) BPM ne wo k g aph o heu is ic
s a egy.
Figu e 6.: BPM and ne wo k g aphs o he simula ed baseline wo k lows.
9
