Spec alp ocessing echniques o e icien 
moni o inginop icalne wo ks
FABIANOLOCATELLI,1,*KONSTANTINOSCHRISTODOULOPOULOS,2MICHELA
SVALUTOMOREOLO,1JOSEPM.FÀBREGA,1LAIANADAL,1SALVATORESPADARO3
1Cen eTecnològicdeTelecomunicacionsdeCa alunya(CTTC/CERCA),Cas ellde els(Ba celona),Spain
2NokiaBellLabs,S u ga ,Ge many
3Uni e si a Poli ècnicadeCa alunya,Ba celona,Spain
*Co espondingau ho : abiano.loca elli@c c.es
Recei edXXMon hXXXX; e isedXXMon h,XXXX;accep edXXMon hXXXX;pos edXXMon hXXXX(Doc.IDXXXXX);publishedXXMon hXXXX
Ha ingubiqui ousop icalmoni o sindensewa eleng h‐di isionmul iplexing(DWDM)o  lex‐g idne wo ksallows
hees ima ionin eal imeo c ucialpa ame e s.Suchmoni o ingwouldbee enmo eimpo an indisagg ega ed
op icalne wo ks, oinspec pe o manceissues ela ed oin e ‐ endo in e ope abili y.Se e alimpo an 
pa ame e scanbe e ie edusingop icalspec umanalyze s(OSAs).Howe e ,omnip esen OSAs ep esen san
in easiblesolu ion.Ne e heless, head en o new, ela i elycheap,compac andmedium‐ esolu ionop ical
channelmoni o s(OCMs)enableamo ein ensi edeploymen o  hesede ices.In hispape ,weiden i y womain
scena ios o  heplacemen o suchmoni o s:a  heing essanda  heeg esso  heop icalnodes.In heing ess
scena io,wecandi ec lyes ima e hepa ame e s ela ed o hesignals,bu no  hose ela ed o he il e s.On he
con a y,in heeg essscena io, he il e  ela edpa ame e scanbeeasilyde ec ed,bu no  hose ela ed oASE.
The e o e,wep esen  wome hods ha ,le e agingacu e i ingandamachinelea ning(ML) eg ession
algo i hm,allow ode ec  hemissingpa ame e s.We e i y hep oposedsolu ionswi hspec alda aacqui edin
simula ionandexpe imen alse ups.Weob ainedgoodes ima ionaccu acy o bo hse upsand o bo hs udied
placemen scena ios.No ewo hy,in heexpe imen alassessmen o  heing essscena io,weachie edamaximum
absolu ee o (MAE)lowe  han1GHzin il e bandwid hes ima ionandaMAElowe  han0.5GHzin il e 
equencyshi es ima ion.Inaddi ion,bycompa ing he ela i ee o so  heconside edpa ame e s,weiden i ied
heing essscena ioas hemo ebene icial.Inpa icula ,wees ima ed he il e cen al equencyshi wi h84%
and he il e 6‐dBbandwid hwi h75%highe accu acy,wi h espec  oda ashee / e e ence alues.This ansla es
in oa o al educ iono  hees ima edsignal‐ o‐noise a io(SNR)penal y,in oducedbyasingleop ical il e ,o 0.24
dB.© 2020 Op ical Socie y o Ame ica
h p://dx.doi.o g/10.1364/JOCN.99.099999
1. INTRODUCTION
Nowadays, disagg ega ion is an impo an end wi hin op ical
ne wo ks. Acco ding o his pa adigm, se e al elemen s o he ne wo k
may be p o ided by di e en endo s [1]. The e o e, in such si ua ion,
ad anced endo -independen moni o ing capabili ies would be
equi ed in o de o cope wi h speci ica ion misma ches and
pe o mance a ia ions o he disagg ega ed ne wo k elemen s [2].
Mo e in gene al, in an op ical ne wo k, an ideal scena io en isions
omnip esen and powe ul op ical pe o mance moni o ing (OPMs),
placed be o e and a e e e y ne wo k node. OPMs enable o moni o
di e en pa ame e s ac oss he ne wo k [3]. Among hem, he mos
impo an a e he ampli ied spon aneous emission (ASE) noise, di ec ly
ela ed o he op ical signal- o-noise a io (OSNR) and he il e ela ed
pa ame e s, such as he il e 3-dB o 6-dB bandwid h and he il e s
shi s. The OSNR is conside ed as one o he mos impo an pa ame e s
o be moni o ed, because i is di ec ly co ela ed o he bi e o a e
(BER) and i is anspa en o he modula ion o ma . Wi hin he
di e en ypes o OPMs, spec al moni o ing is he mos in e es ing
op ion, since i can iden i y all signal pa ame e s, excep o he
nonlinea in e e ence ela ed impai men s, which a e in gene al
ex emely ha d o moni o [3,4]. Nowadays, cos -e ec i e e sions o
he classic op ical spec um analyze s (OSAs), also known as op ical
channel moni o s (OCMs), a e a ailable on he ma ke [5]. These
de ices can be conside ed a cheap solu ion (o en in he o de o ew
hund eds o eu os), when compa ed wi h he cos o he es o he
ne wo k elemen s in dense wa eleng h-di ision mul iplexing (DWDM)
o lex-g id ne wo ks. Despi e hei low-cos , and excep o he
equency d i limi a ion which a ec s hem in he long e m, OCMs
show good pe o mance and esolu ions (up o sub-GHz o de ).
The e o e, we en ision he employmen o such moni o ing de ices,
conside ing hei e en ual eplacemen once he equency d i e ec s
© 2021 Op ical Socie y o Ame ica. Use s may use, euse, and build upon he a icle, o use he a icle o ex o da a mining, so
long as such uses a e o non-comme cial pu poses and app op ia e a ibu ion is main ained. All o he igh s a e ese ed.


make hem imp ac ical. All hese ea u es could enable a wide use o
such moni o s h oughou he op ical ne wo k. The mos app op ia e
posi ion o place hem is close o he econ igu able op ical add/d op
mul iplexe s (ROADMs), which a e key enable s o op ical co e/me o
ne wo ks [6].
In deployed wa eleng h swi ched op ical ne wo ks employing
DWDM o lexg id channels, a ROADM allows indi idual channels (i.e.
indi idual wa eleng hs) o pass- h ough he node, o o be added o
d opped. So, i allows o he e mina ion and en y o se ices bu also
he anspa en bypass o he node, a oiding cos ly op ical-elec ical-
op ical con e sions [7,8]. In cu en gene a ion ROADMs, hese
unc ions a e implemen ed by means o wa eleng h selec i e swi ches
(WSSs). A WSS is an 1xN op ical de ice, which allows any en e ing
wa eleng h on he common inpu po o be swi ched o any o he N
a ailable ou pu po s. The WSS also wo ks in he opposi e di ec ion,
selec ing ou o he N inpu po s he channels o be o wa ded on he
common ou pu po . To do so, he WSSs include op ical il e s ha
in oduce di e en il e ing penal ies on he op ical signals [9].
Fu he mo e, he il e s in oduce a sha p powe d op a he signal
sides, which makes he measu emen o he ASE noise challenging. In
ac , one o he classical ways o measu e i a e de ined as ou -o -band
me hods, since hey ely on noise measu emen s aken ou side he
signal bandwid h. One example is he in e pola ion me hod [10], whe e
he noise le els a he sides o he conside ed channel spec um a e
in e pola ed o gi e an es ima ion o he ASE noise alue inside he
channel. These kinds o app oaches a e e en mo e complex o be
applied in mode n DWDM and lex-g id ne wo ks, whe e he spacing
be ween channels is educed o he minimum. Indeed, he challenges o
ace a e no only he al eady men ioned s ong il e ing, bu also he ac
ha in such ne wo ks, each channel exhibi s a di e en noise le el,
acco ding o he ou e i has aken. The e o e, new OSNR moni o ing
echniques which ope a e in-band, a e needed.
In he pas ew yea s, ROADM a chi ec u e e ol ed om a
“B oadcas and Selec ” (B&S) app oach o a mo e lexible and be e
pe o ming pa adigm, called “Rou e and Selec ” (R&S) [11]. The o me
employs a b oadcas ing powe spli e and only a single WSS pe deg ee
(i.e. pe di ec ion) a he eg ess ibe , yielding educed il e ing penal ies
a low cos [12]. Con a ily, he R&S app oach uses wo independen
WSSs pe deg ee, bo h a he ing ess and eg ess ibe s and yields be e
isola ion o connec ions and lowe inse ion losses. O cou se, since R&S
has wice he numbe o WSSs o B&S, i in oduces a la ge passband
na owing e ec and i is a mo e expensi e solu ion [13].
Se e al moni o ing s a egies ha use a ious amoun o moni o s
can be implemen ed in he ne wo k. O cou se, he highe he numbe
and he mo e ad anced speci ica ions he deployed moni o s ha e, he
mo e expensi e he solu ion is. In he ideal case, depic ed in Fig. 1,
powe ul spec al moni o s a e a ailable be o e, a e and inside e e y
node o he ne wo k (no e ha , Fig. 1 does no display he in e nal
moni o s). This solu ion in eali y is no easible because o i s cos . Thus,
we iden i ied 2 al e na i e scena ios o he placemen o he moni o s,
limi ing/selec ing hei posi ioning, in o de o educe he o e all cos .
The i s scena io equi es he moni o s o be placed be o e he ing ess
po s o each ROADM, i.e. be o e hei ing ess WSSs, as shown in Fig. 2.
In he second scena io he op ical moni o s a e placed a e he eg ess
po s o he ROADMs nodes, i.e. a e hei eg ess WSSs, as depic ed in
Fig. 3.
As we p e iously men ioned, pa ame e s such as he ou -o -band
OSNR, he o al op ical powe o he wa eleng h d i can be easily
es ima ed h ough op ical spec um-based echniques [3]. Fo example,
he al eady men ioned in e pola ion me hod, in which he in-band noise
le el is es ima ed in e pola ing he noise a he wo sides o he signal
Fig. 1. Ideal scena io whe e powe ul moni o s a e a ailable a he ing ess and eg ess po s o e e y node o he ne wo k.
Fig. 2. Ing ess scena io. WSS: wa eleng h selec i e swi ch; OCM:
op ical channel moni o .


[10]. Mo eo e , nowadays, a la ge amoun o ad anced
analy ics/machine lea ning (ML) me hods ha e eme ged, enabling he
enhancemen o he a ailable spec al da a. Thus, om he op ical
spec um one can e ie e in o ma ion no only abou he signal i sel ,
bu also ela ed o he condi ion o he ne wo k o he elemen s ha
compose i , like he op ical ampli ie s and il e s. In [14], o example, he
au ho s p oposed h ee ML-based me hods o de ec and iden i y he
il e ela ed so ailu es, such as il e shi /lase d i and il e
igh ening. These p oposed app oaches, ely on a se ies o equency-
powe pai s, e ie ed di ec ly om he op ical spec a cap u ed wi h
OSAs placed a he eg ess po o e e y node o he ne wo k. On he
o he hand, in [15], an al e na i e solu ion o OSAs o il e impai men s
moni o ing was p esen ed. The e, he au ho s p oposed a cen e o
mass-based app oach employed inside he cohe en ecei e , which
essen ially ope a ed as an OSA moni o a he end o he connec ion.
Ano he solu ion based on moni o ing in o ma ion o exis ing
connec ions, o es ima e il e ing unce ain ies and he e o e imp o e
he quali y o ansmission (QoT) es ima ion o u u e connec ions, was
p oposed in [16]. In ha pape , he au ho s, le e aging a ML eg ession
model and p ocessing spec al da a acqui ed wi h moni o s such as
OCMs, we e able o show, in simula ions, an 80% educ ion o he
ma gin o a new connec ion. In [17], op ical spec al da a analysis was
applied o il e less op ical ne wo ks. The au ho s, p oposed a me hod
o moni o he powe luc ua ions and lase d i s o he ansponde s,
exploi ing op ical spec a collec ed by a single OCM pe il e less
segmen . In [18], he au ho s compa ed he pe o mance o 4 di e en
ML algo i hms, in pa icula , suppo ec o machine (SVM), a i icial
neu al ne wo k (ANN), k-nea es neighbo s (KNN) and decision ee,
o es ima ing pa ame e s, such as cen al wa eleng h, OSNR and signal
bandwid h, by p ocessing he spec al da a. Howe e , in [18], wide
op ical spec a we e conside ed, which a e no a ailable in deployed
il e ed ne wo ks, whe e only in-band op ical spec a can be e ie ed
om he moni o s (i.e. inside he channel c ea ed by he il e s). Finally,
op ical spec a p ocessing has also been employed o ne wo k secu i y
pu poses, as in [19], whe e he au ho s implemen ed a ML-based
app oach able o de ec unau ho ized signals in he ne wo k.
Selec ing one o he placemen s a egies ou lined abo e esul s in
di e en moni o ed pa ame e s and in he lack o some o he s. In he
ing ess scena io, we can di ec ly es ima e he pa ame e s ela ed o he
signal such as he ASE noise, bu no hose ela ed o he il e . On he
con a y, in he eg ess scena io, he il e ela ed pa ame e s can be
easily de ec ed, bu hose ela ed o he ASE noise canno . To cope wi h
he missing in o ma ion, we p opose o enhance he collec ed spec a
wi h adequa e da a analy ics and ML me hods. Th oughou his pape ,
when we men ion he spec al moni o ed da a, we e e o he
moni o ed powe spec al densi y (PSD); we will use hese wo e ms
in e changeably.
In ou p e ious wo ks, we p oposed a ML-based solu ion o in-
band OSNR moni o ing [20,21] in he eg ess scena io, he e o e
exploi ing op ical spec a cap u ed a he node’s ou pu . We e i ied he
alidi y o he p oposed solu ion h ough simula ed and expe imen al
se ups. In pa icula , in [20], we compa ed he pe o mance o wo ML
algo i hms, SVM and Gaussian p ocess eg ession (GPR) using op ical
spec a collec ed a wo di e en esolu ions: a e y-high esolu ion,
namely 12.5 MHz, and a a medium esolu ion, namely 1.25 GHz. We
obse ed good es ima ion accu acy, while no subs an ial imp o emen s
eme ged using he e y-high esolu ion spec a. In addi ion, in [21], we
u he e i ied he app oach p esen ed in [20], implemen ing a new
expe imen al se up, which conside ed s ong il e ing condi ions and
achie ed good es ima ion accu acy. In [22], we ocused on he ing ess
scena io. The e, we p oposed a me hod o e ie e he ans e unc ion
(TF) o a bandpass op ical il e , e.g. o a WSS, exploi ing he op ical
spec a cap u ed a he node’s inpu po . Ob aining he il e TF,
allowed us o unde s and he quali y o he il e i sel . As o he
p e ious wo ks, we e i ied he p oposed solu ion wi h spec al da a
se s collec ed h ough bo h simula ion and expe imen al se ups.
In his pape , we compa e he wo iden i ied placemen scena ios
and he ela ed pa ame e s es ima ion me hods. To do so, we p o ide a
ho ough o e iew o he p oposed moni o ing scena ios. We p esen
in de ail he iden i ied da a analy ics and ML app oaches o p ocessing
he a ailable spec a and we imp o e hem wi h espec o ou p e ious
wo ks. We also assess hem in uni ied simula ion and expe imen al
se ups. Finally, a e compa ing hem, we p o ide guidelines o he
p e e ed one in e ms o es ima ion accu acy imp o emen s and
signal- o-noise (SNR) penal y educ ion.
The es o his pape is s uc u ed as ollows. In Sec ion 2, we
p esen a de ailed o e iew o he wo main iden i ied placemen
scena ios and he de eloped co esponding p ocessing me hods. We
hen p esen in Sec ion 3 he implemen ed simula ion se up along wi h
he esul s. In Sec ion 4, we desc ibe he expe imen al se up and he
ela ed esul s. In Sec ion 5, we compa e he esul s ob ained wi hin he
di e en placemen scena ios and we p o ide some guidelines on hei
employmen . Finally, in Sec ion 6, we summa ize he achie ed esul s
and conclude he pape .
2.OPTICALSPECTRALMONITORSPLACEMENT
SCENARIOSANDSPECTRALPROCESSINGMETHODS
We ocus ou a en ion on he wo mos impo an ca ego ies o
pa ame e s ha can be moni o ed h ough OPMs, especially in
disagg ega ed op ical ne wo k scena ios: signal- ela ed and il e -
ela ed pa ame e s. The i s g oup includes he signal cen al
equency, he signal 3/6-dB bandwid h and he ASE noise added o he
signal ( he e o e, i s OSNR). The second g oup is composed o he il e
cen al equency and he il e 3/6-dB bandwid h. We p opose o
e ie e such in o ma ion h ough spec al moni o ing, employing
medium-accu acy OCMs. In bo h he scena ios iden i ied in Fig. 2 and
Fig. 3, we we e able o neglec he e ec s o he po ion o he ne wo k
p eceding he conside ed OCM. To do his, a e iden i ying he channel
which we a e in e es ed in, we would collec he ela ed spec a om
he OCM placed a he WSS a e sed by ha channel. In pa icula , i we
de ine wi h OCM
n+1
he moni o loca ed a he poin n+1 o he ne wo k,
we di ided, in he linea domain, he OCM
n+1
cap u ed PSDs by hose
cap u ed wi h OCM
n
. This ope a ion allows us o ocus only on wha
happens be ween he moni o loca ed in posi ion n and he moni o
loca ed in he conside ed posi ion n+1, wi hou any knowledge and any
e ec o wha comes be o e posi ion n. Pe o ming such di isions along
he pa h o an op ical connec ion (ligh pa h), enables us o see he e ec
o each link/node on ha pa h, nulli ying hose o he p e ious ne wo k
elemen s along ha same ligh pa h. Assuming ha no equency d i
a ec s he employed OCMs, he only equi emen o such ope a ion is
he collec ion o bo h spec a ( om OCM
n+1
and OCM
n
) a a single poin ,
which would ypically be he cen al con olle . In addi ion, again
assuming no equency e o on he OCM, a ia ions o he spec a
esolu ions can be easily accoun ed wi h some simple p ocessing, e.g.
Fig. 3. Eg ess scena io. WSS: wa eleng h selec i e swi ch; OCM: op ical
channel moni o .


upsampling he lowes esolu ion. Mo eo e , we a e cons aining he
equency ange ela ed o he channel o in e es , elying on he
spec al g id de ined by he ITU-T s anda ds [23]. I is wo h no ing ha
in his pape we ocus on he ex ac ion o he pe link/node pa ame e s
o a single connec ion. Co ela ing in o ma ion among connec ions ha
c oss he same link/node, can imp o e u he he unde s anding o he
link/node ea u es. This would equi e a u he p ocessing/analysis o
he ou pu s o ou p oposed solu ion a some cen alized poin (e.g. he
so wa e-de ined ne wo king (SDN) con olle ). Such p ocessing
me hods a e ou side he scope o his pape , he e o e he eade is
e e ed o [24,25], o a ious solu ions a ha p ocessing le el. In
addi ion, in his pape we ocused ou in es iga ion on di e en
ansmission aspec s, wi hou conside ing he consequences o he
nonlinea in e e ence (NLI) caused by he Ke e ec . In ac , all he
spec a we collec ed and p ocessed o his wo k e e ed o a single
channel con igu a ion. Howe e , o almos all he in es iga ed cases we
will p esen in he nex sec ions, we conside ed s ong il e ing
scena ios, hus neglec ing he e ec s o he linea c oss alk coming om
any e en ual adjacen channels, on he op ical spec a. Mo eo e , since
sel and c oss-channel NLIs only ha e a mino impac on he spec a, we
claim ou solu ions o be una ec ed by such e ec s.
The wo main OCM moni o s placemen scena ios ha we
iden i ied, along wi h he wo solu ions we p opose o e ie ing he
ela ed missing pa ame e s, a e discussed in de ails in he nex
subsec ions.
A.Ing essMoni o ingScena io
In he ing ess moni o ing scena io, depic ed in Fig. 2, he OCMs a e
placed be o e he ing ess WSS o each ROADM nodes o he ne wo k.
This ansla es in o a ela i e ease in e ie ing in o ma ion abou he
signals en e ing he node, since hey a e moni o ed be o e hey a e
il e ed by he WSSs o he node i sel . The e o e, pa ame e s such as he
signal cen al equency, he signal 3/6-dB bandwid h and he OSNR,
can be measu ed di ec ly o h ough (mild) p ocessing o he op ical
spec a collec ed by he OCMs. On he con a y, he op ical spec a
collec ed h ough such con igu a ion will no gi e any di ec
in o ma ion abou he il e s, hus some ad anced p ocessing is
equi ed. In he ideal scena io o Fig. 1 (OCMs e e ywhe e), we would
simply di ide in he linea domain he PSD cap u ed a he eg ess po
o he node o he one cap u ed a i s ing ess. As discussed in Sec ion I,
depending on he conside ed ROADM a chi ec u e, he numbe o WSSs
con ained in a ROADM node, can a y be ween one (B&S) and wo
(R&S) pe deg ee. We ea bo h ROADM a chi ec u e he same: i he
node con ains wo il e s, we model an equi alen il e o he speci ic
ing ess-eg ess di ec ion. So, by doing he abo e di ision, we would
ob ain he TF o he single il e (B&S) o he equi alen o bo h il e s
(R&S) in ol ed in he speci ic ing ess-eg ess di ec ion. Howe e , since
he moni o ing placemen scena io we a e now conside ing does no
ha e op ical moni o s a he eg ess po s, we ha e o ind ano he way
o ob ain he eg ess ela ed PSD. To cope wi h his lack, we eplace he
PSD moni o ed a he eg ess po , wi h he one moni o ed a he ing ess
o he ollowing node. Re e ing o Fig. 2, and ollowing he app oach
p oposed in [22], we di ide, in he linea domain, he op ical spec a
cap u ed wi h OCMn+1,i by he one cap u ed wi h OCMn,i, whe e n
ep esen s a gene ic node o he op ical ne wo k, n+1 he node ollowing
n in he di ec ion o ou in e es , and OCMn,i and OCMn+1,i, ep esen he
OCMs a he ing ess po o node nand node n+1, espec i ely. O cou se,
he esul o his di ision ep esen s no only he il e TF o node n, bu
also he noise accumula ed o e he link connec ing he wo nodes,
e e ed o as link (n,n+1) in Fig. 2. As discussed abo e, his ope a ion
emo es he e ec s accumula ed by he connec ion o e i s pa h be o e
eaching he moni o o node n. So we can only ocus on he las pa , he
il e o node n and he link (n,n+1). I is wo h no ing ha when wo o
mo e equency adjacen channels come om he same link and
con inue a he same link, hey can also sha e he same il e and hus
he in e nal il e edge(s) would no show up in he TF. Fo such
channels ou me hod will no wo k. On he con a y, when wo adjacen
channels come om di e en inpu po s, o go owa ds di e en
ou pu po s (including d op po s), he il e edges a e isible, and he
solu ion we p oposed can be employed.
To iden i y he noiseless TF o he il e , we can es ima e he ASE
noise h ough he OCMn+1,i cap u ed spec um and emo e i , in he
linea domain, om he PSD ob ained wi h he a o emen ioned di ision.
Due o il e ing and/o he esponse o he OCM, measu ing he noise a
he sides o he spec a migh no be e y easy. Thus, enhancing wha
we p oposed in [22], we implemen ed a p ocess ha wo ks on he
moni o ed spec a and iden i ies he noise con ibu ion. The unc ion
sea ches o e a se o noise alues. They can come om he moni o ed
spec a o om he basic link/span knowledge wi h he addi ion o an
accu acy co ec ion ac o . Then, he unc ion selec s he noise amoun
ha esul ed in he bes (lowes ) i ing e o . In a sense, i elies on how
good he shape o he e e ence il e ma ches he shape o he eal one.
In ac , emo ing a lowe o highe amoun o noise would e u n a
w ong il e shape, yielding o a wo se ma ch wi h espec o he case
whe e a co ec noise amoun is sub ac ed. The assump ion behind he
ASE es ima ion is ha he il e a node n supp esses he noise
in oduced by he ampli ie s/links cascade up o ha poin and
he e o e, he noise a he sides o he OCMn+1,i collec ed spec a is mainly
due o he link (n,n+1) con ibu ion. This noise iden i ica ion p ocess
allowed us o be e es ima e he ASE noise, wi h espec o wha we
epo ed in [22]. An example o a TF ob ained h ough he noise
emo al, s a ing om expe imen ally collec ed spec a, is shown in Fig.
4. Obse ing he il e TF ob ained h ough he abo e s eps, we see ha
we a e s ill no able o measu e i s 6-dB bandwid h, due o he ac ha
i s wo sides do no each deep alues. This happens in mos cases ha
we obse ed wi h simula ion and expe imen al spec al da a and occu s
because a he spec a edges we a e p ocessing he noise ins ead o he
signal con ained in he channel.
To o e come his obs acle, we p opose o econs uc he ull TF, by
i ing he ob ained po ion o spec um wi h a unc ion ha
co esponds o he ‘ideal’ shape o he il e . Typically, op ical il e s a e
Fig. 4. Example o an expe imen ally ob ained il e TF spec um
be o e (blue) and a e (o ange) he noise emo al p ocess. The solid
yellow plo is he i ed cu e, while he do ed yellow line ep esen s
he iden i ied cen al equency o he i ed cu e, which is shi ed wi h
espec o he o iginal cen e (do ed black line).


conside ed o ha e a high-o de Gaussian shape. In [26], he au ho s,
le e aging he e o unc ion e 󰇛𝑥󰇜, p oposed a model o he
cha ac e iza ion o he op ical ield spec um 𝑆󰇛𝑓󰇜 o a bandpass il e
c ea ed by a WSS. The unc ion modeling an ideal il e , which is
symme ic and cen e ed a 0 equency, is he ollowing:
𝑆󰇛𝑓󰇜


𝛼√2𝜋󰇣𝑒𝑟𝑓󰇡

⁄
√
󰇢𝑒𝑟𝑓󰇡
 
⁄
√
󰇢󰇤
󰇛1󰇜
whe e 𝛼 is he pa ame e ela ed o s eepness o he il e edges (i.e. he
il e o de ), 𝛽 is he 6-dB bandwid h o he il e and 𝑓 ep esen s he
equency.
Due o misalignmen o il e o he signal ( ansmi e lase and/o
il e shi s) and o powe le eling issues, he po ions o il e TFs ha
we wan o i a e o en no cen e ed. The e o e, o conside he wo
e en ual shi s in he di ec ions o x and y-axis, we ex ended he Eq. 1
based model wi h wo new pa ame e s: 𝛿 and 𝛾, espec i ely. The
equa ion ep esen ing he new model is he ollowing:
𝑆󰇛𝑓󰇜


𝛼√2𝜋󰇣𝑒𝑟𝑓󰇡

⁄
√
󰇢𝑒𝑟𝑓󰇡
 
⁄
√
󰇢󰇤𝛾
󰇛2󰇜
whe e 𝛿 ep esen s he shi o he il e cen al equency and 𝛾 is a
no maliza ion ac o o he y-axis shi s. By uning he pa ame e s 𝛼,𝛽,
𝛾 and 𝛿, we can now i he ob ained po ions o il e s TFs spec a wi h
he loga i hmic squa ed e sion o Eq. 2. The ange in which he
pa ame e s a e uned, can be na owed down based on he il e
speci ica ions. This allows a be e i ing and he e o e a be e
es ima ion o he il e ela ed ea u es. In ac , once he i ing p ocess is
o e , he alues assumed by he pa ame e s 𝛽 and 𝛿 ep esen he
es ima ed il e 6-dB bandwid h and he es ima ed il e cen al
equency shi , espec i ely. We ocused ou analysis on he wo il e
pa ame e s ha we conside ed as he mos ep esen a i e, bu
essen ially we es ima e he ull shape o he il e TF, hus making any
o he pa ame e o in e es e ie able. No e ha he abo e model can
be also ex ended o cap u e il e shape asymme ies, which we e no
hea ily p esen in ou simula ed and expe imen al da a. So, we do no
epo he e alua ion o such e ec s he e.
B.Eg essMoni o ingScena io
In he eg ess moni o ing scena io, he moni o s a e placed
exclusi ely a e he eg ess WSS o each op ical node o he ne wo k, as
shown in Fig. 3. F om he op ical spec a collec ed wi h such
con igu a ion, we can easily e ie e in o ma ion ega ding he op ical
il e s o he node, bu no abou he signals en e ing he node. Fil e
ela ed pa ame e s can be e ie ed because he il e di ec ly a ec s
he signal a e sing he node. In ac , by di iding (in he linea domain)
he PSD cap u ed wi h OCM
n+1,e
by he one cap u ed wi h OCM
n,e
, we
ob ain he con ibu ions o link (n,n+1) and o he il e s o node n+1.
This ope a ion emo es he e ec s o he cascade up o he loca ion o
OCM
n,e
, lea ing he il e TF clea ly isible, while hiding he signal ela ed
pa ame e s, such as he ASE noise. Once he TF is iden i ied, he il e
ela ed pa ame e s can be e ie ed applying a i ing app oach simila
o he one p oposed o he ing ess scena io, wi h he main di e ence
ha in his case he i ing e o would be negligible. In addi ion, an
e en ual il e cen al equency shi can be de ec ed om he TF, o
example employing he cen e o mass app oach, p esen ed in [15], a
e e y node.
On he o he hand, as we men ioned in Sec ion 1, signal ela ed
pa ame e s a e no di ec ly moni o able, since he il e s end o hide
se e al o he signals o iginal cha ac e is ics, making adi ional
moni o ing echniques ine ec i e. Thus, pa ame e s such as he signal
3/6-dB bandwid h and he associa ed ASE noise (co ela ed wi h he
signal in OSNR me ic) a e no di ec ly measu able om he op ical
spec a.
We p opose a supe ised ML-based me hod o es ima e in-band he
ASE noise om op ical spec a collec ed in he eg ess moni o ing
scena ios [20,21]. Wi h espec o ou p e ious wo ks [20,21], we
enhanced he p oposed me hod di iding he OCM
n+1,e
cap u ed PSD by
he one cap u ed wi h OCM
n,e
(i.e. neglec ing he e ec s o he cascade up
o node n) and imp o ing/ uning he employed ML model coe icien s.
The p oposed solu ion, as a i s s ep, equi es he collec ion o di e en
se s o op ical spec a which a e hen classi ied acco ding o some
pa icula signal pa ame e s and labeled wi h hei co esponding ASE
noise/OSNR alues. Le e aging he labeled spec al da a, we hen ain
sepa a e ML eg ession models o he di e en classes, o p edic he
OSNR. Ou p e ious esul s indica ed e y good p edic ion
pe o mance, o new op ical spec a ha we e no pa o he aining.
Indeed, in [20,21] we al eady p o ided a solid pe o mance e alua ion
o he p oposed ML-based solu ion, using ex ensi e simula ion and
expe imen al da ase s. Thus, he main goal o his cu en wo k is he
compa ison o he wo placemen scena ios and hei assessmen unde
common simula ion and expe imen al da ase s.
To o mally p esen he me hod, we ep esen an acqui ed op ical
spec um ins ance wi h a ec o 𝒔 o leng h 𝑙, and we name i s
co esponding OSNR alue as 𝑦. The goal is o ind he mapping 𝑓,
be ween he spec um 𝒔 and i s OSNR alue 𝑦, ha is 𝑦𝑓󰇛𝒔󰇜. To do
so, we implemen a ML model 𝑄

, whe e 𝑐 ep esen s he g oup o
pa ame e s o which he model is alid (e.g. he oll-o ac o , he baud
a e o he connec ion and he nominal il e bandwid h). We hen ain
he ML model wi h a se o moni o ed and labeled spec a 󰇛𝑺

,𝒚

󰇜. 𝑺

is a ma ix o dimension 𝑙𝑚, which ep esen s he se o op ical
spec a wi h he same pa ame e s 𝑐, while 𝒚

is he ec o o leng h 𝑚,
o hei co esponding OSNR alues. No e ha , he a o emen ioned
pa ame e s 𝑐 co esponds o nominal ansponde s and il e alues, so
we could c ea e he aining se in a calib a ion phase in he lab o in he
ield, be o e commissioning a connec ion. We also deno e by 𝒚



𝑄

󰇛𝑺

󰇜 he ec o ep esen ing he es ima ed OSNR alues and by 𝜺


𝒚


𝒚

he es ima ion e o s. The goal o he aining p ocess is o
iden i y he model 𝑄

, which minimizes some unc ion ela ed o he
es ima ion e o s 𝜺

, such as he mean squa ed e o (MSE) unc ion.
Once he ML algo i hm is ained wi h he spec a 𝑺

and hei e e ence
OSNR alues 𝒚

, i will be able o e u n he es ima ed OSNR alue 𝑦


o an ope a ing channel ℎ wi h he same pa ame e s 𝑐, om i s op ical
spec um 𝒔.
3.SIMULATIONSANDRESULTS
A.Simula ionse up
To simula e bo h he ing ess and he eg ess moni o ing placemen
scena ios and o also e alua e he e ec o moni o s esolu ion, we
implemen ed he VPIpho onics [27] se up depic ed in Fig. 5. We
Fig. 5. Scheme o he implemen ed VPIpho onics simula ion se up. TX:
ansmi e ; OCM: op ical channel moni o .



gene a ed a 32 GBd pola iza ion mul iplexed-quad a u e phase shi
keying (PM-QPSK) modula ed signal, wi h 0.1 oll-o ac o , cen e ed a
193.4 THz (1550.116 nm). In o de o simula e he op ical links, a e he
ansmi e (TX) we cascaded a numbe o spans which included 80 km
leng h s anda d single mode ibe s (SSMFs) and e bium-doped ibe
ampli ie s (EDFAs) wi h 5 dB noise igu e (NF). We se he ou pu powe
o he TX o 0 dBm as well as he ou pu powe o all he EDFAs. Taking
in o accoun ha he op imum launch powe a ies depending on
se e al ac o s (i.e. he amoun o ASE noise, he a enua ion o he ibe ,
he ampli ie ’s NF and he NLIs), we chose 0 dBm bea ing in mind he
speci ic single channel con igu a ion we conside ed. Va ying he
numbe o ibe spans and EDFAs, we we e able o simula e di e en
ASE noise le els and he e o e di e en OSNR alues. Each link was
ollowed by an op ical node, which we implemen ed as a cascade o 2
op ical il e s, wi h 2nd o de Gaussian TFs. We assumed ha he 2
op ical il e s o e e y node had he same cha ac e is ics, in ac ou
me hod conside s he wo il e s as an equi alen one, as discussed in
Sec ion 2.A. Fo each equi alen il e , we conside ed di e en 6-dB
bandwid h alues. In addi ion, in o de o emula e he impai men due
o he lase d i o il e shi , we also shi ed he cen al equency o
each il e wi h espec o he TX lase ’s equency. Finally, wi h he aim
o co e ing bo h he ing ess and he eg ess scena ios, we placed an OCM
a he inpu and a he ou pu o e e y node o he se up. The spec al
esolu ion o all he employed OCMs was 1 GHz, while he spec al
sampling esolu ion o he collec ed spec a was 15.625 MHz. This alue
ep esen ed he ecip ocal o he ime window se in VPI, which was
di ec ly co ela ed o he bi a e alue (i.e. 128 Gb/s).
In o al, ou simula ion se up included 3 nodes, 4 links and 7 OCMs:
4 OCMs we e used o collec he op ical spec a o he ing ess scena io
model ( he blue blocks in Fig. 5), while he emaining 3 o he eg ess
model ( he ed blocks in Fig. 5). Table 1 summa izes he 16 cases
conside ed: we lis he e he numbe o ibe spans and EDFA pe link,
and he 6-dB bandwid h and cen al equency shi o each il e . The
numbe o ibe spans and EDFAs o each link a ied and we e chosen
wi h he ollowing alues: 2, 3 and 5. Va ying his alue, allowed us o
ha e di e en ASE noise le els a he inpu po s o he il e s.
Addi ionally, o eplica e he na owing o he il e bandwid h
in oduced by he il e cascading e ec (FCE) and also o ake in o
accoun impe ec ions in he p oduc ion and a ia ion in he ageing
condi ions, we assigned he ollowing alues o he il e s 6-dB
bandwid hs: 36.5 GHz, 37.5 GHz and 38.5 GHz. Finally, o eplica e he
misalignmen be ween he lase and he il e cen al equencies, due
o impe ec ions and ageing, we assumed o each il e a shi which
anged be ween -2 GHz and +2 GHz.
B.Ing essscena io esul s
We conside ed he se up depic ed in Fig. 5 o he ing ess scena io.
The e, he only a ailable moni o s we e he blue ones, i.e. OCMn,i, whe e
n ∈ [1,4] ep esen s he node be o e which he moni o was placed.
Following he me hod p esen ed in Sec ion 2.A o he il e pa ame e s
de ec ion in he ing ess scena io, we i s e ie ed he noisy TFs o he
il e s co esponding o each node o he se up, which a e Node 1, Node
2 and Node 3 o he 16 cases lis ed in Table 1. Then, a e he noise
iden i ica ion and emo al p ocess, we i he esul ing po ions o
spec a wi h he loga i hmic squa ed e sion o Eq. 2. Obse ing he
alues assumed by he pa ame e s o Eq. 2, in pa icula he pa ame e s
𝛽 and 𝛿, we we e able o es ima e he il e 6-dB bandwid h and he
il e cen al equency shi , espec i ely. Finally, we compa ed he
alues e u ned by he algo i hm wi h hose epo ed in Table 1 and
calcula ed he es ima ion e o s. In Table 2 we lis he s anda d
de ia ion (σ), he MSE, he minimum (MIN) and he maximum (MAX)
e o s o he es ima ion o he 6-dB il e bandwid h and he il e
cen al equency shi o he h ee nodes. Di iding (in he linea
domain) he PSD acqui ed a OCM3,i by he one acqui ed a OCM2,i,
should nulli y he e ec o he pa h up o OCM2,i loca ion. Ne e heless,
he es ima ion e o s epo ed in Table 2 indica e a small e ec o he
cascade. In ac , he es ima ion e o o Node 3 was sligh ly highe han
hose co esponding o he o he wo nodes. In gene al, a Node 1 and
Node 2, he spec a a e almos no a ec ed by he cascade e ec and he
es ima ion e o s a e also e y low. The e o e, keeping he same
accu acy o he i s wo nodes, as he cascade o links inc eases, is e y
ha d.
C.Eg essscena io esul s
To s udy he eg ess scena io, we conside ed a pa icula case o he
simula ion se up shown in Fig. 5. The moni o s a ailable a e only he ed
ones, i.e. OCMn,e, whe e n ∈ [1,4] ep esen s he node whose eg ess po
is moni o ed by he OCM. On he spec a cap u ed wi h hese moni o s,
we applied he ML-based me hod desc ibed in Sec ion 2.B o he in-
band ASE noise es ima ion in he eg ess scena io. To do so, we le e aged
he SVM eg ession algo i hm, a ke nel-based nonpa ame ic ML
echnique. In his s udy, we o mula ed he es ima ion as a eg ession
p oblem. T aining he SVM model using a linea ke nel unc ion
e u ned be e pe o mance han aining i wi h a Gaussian one,
he e o e o he aining p ocess, we conside ed he o me one. In
Table1.Thespansnumbe and he il e pa ame e s o  he
16conside edcases
Links (# o spans) Fil e s (6-dB bandwid h [GHz]
+ cen al eq. shi [GHz])
0-1 1-2 2-3 3-4 1 2 3
5 2 2 3
37.5+0 37.5+2 37.5+1
37.5+1 37.5+2 37.5+2
37.5+2 37.5+1 37.5+0
37.5+2 37.5-1 37.5+2
2 5 2 3
37.5+0 37.5+2 37.5+1
37.5+1 37.5+2 37.5+2
37.5+2 37.5+1 37.5+0
37.5+2 37.5-1 37.5+2
2 5 2 3
36.5+0 36.5+2 36.5+1
36.5+1 36.5+2 36.5+2
36.5+2 36.5+1 36.5+0
36.5+2 36.5-1 36.5+2
2 5 2 3
38.5+0 38.5+2 38.5+1
38.5+1 38.5+2 38.5+2
38.5+2 38.5+1 38.5+0
38.5+2 38.5-1 38.5+2

Table2.Es ima ionaccu acyo  he il e  ela ed ea u esin
hesimula ioncase o  heing essplacemen scena io
Node Es ima ed
Fea u e MSE σ
[GHz]
MIN
[GHz]
MAX
[GHZ]
1 Cen . eq. shi 0.0019 0.0334 -0.0391 0.0655
6-dB BW 0.0183 0.0807 -0.0249 0.1937
2 Cen . eq. shi 0.0008 0.0178 -0.0147 0.0454
6-dB BW 0.0024 0.0479 -0.1057 0.0672
3 Cen . eq. shi 0.0026 0.0482 -0.0702 0.0997
6-dB BW 0.0163 0.1247 -0.1470 0.2962



o de o label he aining spec a and o e alua e he accu acy o he
es ima ion, we measu ed he ASE noise alues ha we e used as
e e ence, di ec ly on he spec a collec ed h ough he OCM
n,i
, a he
ing ess po o e e y node. Fo he calcula ion o he noise spec al
densi y in eg al, we conside ed a e e ence noise bandwid h equal o
12.5 GHz (0.1 nm). The o al numbe o spec a ha we used was 48 (16
o each one o he nodes): we used he 80% o hese spec al da a o
ain he model, he 10% o c oss- alida e i and he emaining 10% o
es i . The c oss- alida ion was used o une he pa ame e ε o he SVM
eg ession model. ε ep esen s hal he wid h o he insensi i e band, i.e.
ha ole ance a ea whe e no-penal y is assigned o he e o s. In
addi ion, o p ecisely assess he es ima ion accu acy o he model, we
also andomly shu led he aining and he es ing da a se s 4000 imes,
ained a di e en model each ime and es ed i wi h i s co esponding
es ing se . The MSE, he MIN and he MAX es ima ion e o s a e
summa ized in Table 3. We achie ed a maximum absolu e e o (MAE)
lowe han 0.61 dB and a MSE o 0.0018 o he es ima ion o he ASE
noise in he eg ess scena io. These esul s e lec he goodness o ou
app oach, also conside ing ha he spec al se we used e e ed o a
se up whe e we uned he il e 6-dB bandwid h, while in he pas we
always conside ed he 3-dB bandwid h o he il e s.
4.EXPERIMENTALSETUPANDRESULTS
A.Expe imen alse up
Fo a u he alida ion o he p oposed app oaches, we also
implemen ed he expe imen al se up shown in Fig. 6. By means o a
unable lase wo king a 193.4 THz (1550.116 nm), we gene a ed a 64
GBd PM-QPSK modula ed signal wi h wo di e en oll-o ac o s: 0.1
and 0.2. We se he TX ou pu powe o -11 dBm. Wi h he aim o
emula ing he FCE ha can occu in a ne wo k, igh a e he TX we
placed a i s op ical il e , namely Fil e 1, h ough which we simula ed
his kind o beha io na owing i s 6-dB bandwid h. A e his i s il e ,
in o de o emula e op ical links o di e en leng hs and he e o e
di e en ASE noise con ibu ions, we cascaded a a iable op ical
a enua o (VOA) and an EDFA ope a ing in powe con ol mode wi h
ou pu powe se o 0 dBm and NF o 5 dB. These 2 blocks oge he
ep esen ed link (1,2) in Fig. 6. As pe he simula ion se up, we chose
such ou pu powe conside ing he single channel con igu a ion we
planned. Following he i s link, we placed a second op ical il e ,
namely Fil e 2, he il e on which we es ed he p oposed models. Thus,
we a ied i s 6-dB bandwid h and i s cen al equency o gene a e a
numbe o di e en possible impai men cases. A e Fil e 2, we
cascaded a second link, composed by a VOA and an EDFA, o simula e
he link a e which we place he OCM in he ing ess scena io. Finally,
h ee OCMs we e placed in he se up: he wo ela ed wi h he ing ess
scena io (namely OCM
2,i
and OCM
3,i
), a he end o he wo links, and he
one o he eg ess scena io, a Fil e 2 ou pu , namely OCM
2,e
. The
moni o we used, was he Finisa Wa eAnalyze 1500S, a high-
esolu ion cohe en OSA able o each esolu ion up o 150 MHz [28]. In
o de o simula e he pe o mance o an OCM, we collec ed all he op ical
spec a a 2 di e en esolu ions: 600 MHz and 1 GHz. The spec al
sampling esolu ion o he collec ed spec a was equal o he in insic
alue o he employed Finisa OSA, i.e. 20 MHz [28]. Table 4 shows how
we clus e ed he collec ed spec a in o 7 di e en cases. Each case
consis s o 9 sub-cases wi h di e en 6-dB il e s bandwid hs and
di e en link a enua ion alues. While Fil e 1 bandwid h and he
a enua ions o he wo links a y o each case, Fil e 2 bandwid h
assumes he same alues in e e y clus e . Case 1 ep esen s he
“de aul ” si ua ion, wi h Fil e 1 6-dB bandwid h se a 74 GHz, no
(addi ional) a enua ion se in link (1,2) and 10 dB a enua ion se in link
(2,3). All he o he cases cons i u e a wo sening o case 1: a leas one o
he 3 a ying pa ame e s (Fil e 1 bandwid h, VOA
1
and VOA
2
) assumes
a alue wo se han in he de aul case. In o al, o each o he wo oll-o
alues and esolu ions, we collec ed 189 op ical spec a, 63 o each
OCM in he se up.
B.Ing essscena io esul s
As pe he simula ion scena io, also o he expe imen al ing ess
scena io, we e ie ed he il e TF ollowing he s eps desc ibed in
Sec ion 2.A. We conside ed he se up depic ed in Fig. 6 wi h he
moni o s associa ed o he ing ess scena io, ha a e OCM
2,i
and OCM
3,i
.
We also used OCM
2,e
o e ie e he TF o be used as e e ence o he
e alua ion o he es ima ion accu acy. The es ima ion e o s o he 6-
dB il e bandwid h and he il e cen al equency shi a e shown in
Fig. 7 and Fig. 8 o oll-o ac o o 0.1 and in Fig. 9 and Fig. 10 o oll-
o ac o o 0.2. The box and whiske s plo s g aphically ep esen he
mean e o s, he s anda d de ia ions and he MIN and MAX es ima ion
e o s, o each one o he 7 conside ed cases. F om he p esen ed
esul s, i is clea how he de aul case (i.e. case 1), showed he bes
pe o mance, especially o he il e bandwid h es ima ion. On he
o he hand, when conside ing he cases wi h na owe Fil e 1
Table3.Es ima ionaccu acyo  heASEnoisein heeg ess
placemen scena io
MSE MIN [dB] MAX [dB]
Simula ion 0.0018 -0.3134 0.6013
Expe imen 0.0136 -0.3911 0.3866

Fig. 6. Scheme o he expe imen al se up. TX: ansmi e ; VOA:
a iable op ical a enua o ; OCM: op ical channel moni o .
Table4.The7conside edexpe imen alcases
Fil e 1 6-dB
BW (shi )
[GHz]
Link(1,2)
VOA
1
[dB]
Fil e 2 6-dB
BW (shi )
[GHz]
Link(2,3)
VOA
2
[dB]
1 74 (-2) 0 73, 75, 77
(-1, 0, +1) 10
2 74 (-2) 0 73, 75, 77
(-1, 0, +1) 20
3 74 (-2) 10 73, 75, 77
(-1, 0, +1) 10
4 69 (-2) 10 73, 75, 77
(-1, 0, +1) 10
5 74 (-2) 5 73, 75, 77
(-1, 0, +1) 10
6 74 (-2) 2.5 73, 75, 77
(-1, 0, +1) 15
7 69 (-2) 7.5 73, 75, 77
(-1, 0, +1) 20


bandwid h (i.e. cases 4 and 7) o highe links a enua ion alues (i.e.
cases 2 and 7), he es ima ion accu acy ends o deg ade. In gene al, he
accu acy o he 6-dB il e bandwid h es ima ion was lowe compa ed
o he accu acy o he il e cen al equency shi , o bo h he oll-o
alues. All he esul s we p esen ed he e we e ob ained wi h spec a
collec ed a 600 MHz esolu ion. We did no obse e any subs an ial
di e ence in he es ima ion accu acy using spec a a 1 GHz esolu ion.
C.Eg essscena io esul s
As we did o he simula ion case, we exploi he op ical spec a
collec ed h ough he expe imen al se up shown in Fig. 6, o es he
eg ess scena io me hod o he ASE noise es ima ion p esen ed in
Sec ion 2.B. The op ical spec a we conside ed o he es s we e hose
ela ed o 0.1 oll-o ac o , 74 GHz Fil e 1 6-dB bandwid h and OCM
spec al esolu ion o 600 MHz, esul ing in a o al o 45 op ical spec a.
As pe he simula ion case, we e ie ed he noise alues o be used as
e e ence, calcula ing he noise spec al densi y in eg als on he op ical
spec a collec ed h ough he ing ess placed moni o s. In pa icula , we
conside ed a e e ence noise bandwid h equal o 12.5 GHz (0.1 nm).
Again, we employed SVM eg ession algo i hm, using he 80% o he
o al spec al da a o aining he model, he 10% o c oss- alida e i
and he emaining 10% o es ing i , andomly shu ling he spec a
4000 imes. Also o he expe imen al case, we used he c oss- alida ion
o une ε, he pa ame e ep esen ing hal he wid h o he insensi i e
band. The esul s o he es ima ion a e epo ed in Table 3. We achie ed
a MAE lowe han 0.4 dB and a MSE o 0.0136. Again, as pe he
simula ed case, i is impo an o s ess ha he il e bandwid h we
conside ed in he expe imen al se up we e e e ing o he 6-dB
measu ed alues. The e o e, wi h espec o he wo ks we ca ied ou in
he pas [20,21], he e ec s o he il e s on he noise we e way mo e
isible his ime, since hei 3-dB bandwid hs we e na owe .
Ne e heless, compa ing ou cu en esul s wi h a simila case we had
in [21] (i.e. PM-QPSK signal wi h 0.1 oll-o , 64 GBd baud a e and 72
GHz 3-dB bandwid h), we imp o ed he accu acy o he ASE noise
es ima ion. In ac , in he conside ed case o [21], he ASE noise
es ima ion MAE was almos 1 dB. In addi ion, since in [21] we used
op ical spec a collec ed wi h a spec al esolu ion o 150 MHz, we again
did no obse e any dependency o he p oposed me hod on he OSA
spec al esolu ion.
5.OCMPLACEMENTSCENARIOSCOMPARISON
In he wo s udied moni o placemen scena ios, he es ima ed
pa ame e s ha e di e en uni s o measu emen . The e o e, in o de o
compa e hem, we con e ed each es ima ion e o in o a pe cen age
wi h espec o i s nominal alue, ansla ing he MAEs in o ela i e
e o s. To do so, he i s s ep is o iden i y he alues o be used as
e e ences. Fo he ing ess scena io, he da ashee o he Finisa il e we
used in he lab epo ed a cen al equency se ing accu acy o ±2.5 GHz
and a bandwid h se ing accu acy o ±5 GHz [29]. WSSs deployed in eal
Fig. 7. 6-dB il e bandwid h es ima ion e o s o he 7 expe imen al
cases, wi h ollo ac o = 0.1.
Fig. 8. Fil e cen al equency shi es ima ion e o s o he 7
expe imen al cases, wi h ollo ac o = 0.1.
Fig. 10. Fil e cen al equency shi es ima ion e o s o he 7
expe imen al cases, wi h ollo ac o = 0.2.
Fig. 9. 6-dB il e bandwid h es ima ion e o s o he 7 expe imen al
cases, wi h ollo ac o = 0.2.


ne wo ks could ha e be e cha ac e is ics, he e o e we decided o
adop o bo h he il e - ela ed pa ame e s a e e ence accu acy equal
o ±2 GHz [7,15]. Fo he eg ess scena io, we ob ained a e e ence o
he OSNR/ASE noise es ima ion, e alua ing he ampli ie NF, as ollows.
We conside ed he cascade o a numbe o ibe spans wi h EDFAs a he
end o each one o hem. We assumed he EDFAs o ha e a NF equal o 5
dB, wi h luc ua ions o ±0.5 dB [30,31]. This assump ion yielded a noise
e e ence e o wi h a 1 dB ange.
Then, elying on he il e and signal- ela ed pa ame e s es ima ion
e o s epo ed in Sec ion 3 and in Sec ion 4, we calcula ed he wo s
case e o anges and he ela i e e o s wi h espec o he abo e
e e ences. The esul s o hese calcula ions a e epo ed in Table 5.
Conce ning he expe imen al esul s, we ound ha he ing ess
moni o ing s a egy imp o ed he il e cen al equency es ima ion by
a ac o g ea e han he 84% o oll-o alue o 0.1 and by a ac o
g ea e han he 87% o 0.2 oll-o . These imp o emen s a e wi h
espec o he scena io whe e no moni o ing s a egy is implemen ed
and he nominal pa ame e s p o ided by endo s/da ashee s a e used.
Likewise, he il e 6-dB bandwid h es ima ion was imp o ed by he
75% o 0.1 oll-o and by a ac o g ea e han he 76% o 0.2 oll-o .
On he o he hand, he eg ess moni o ing s a egy imp o ed he ASE
noise es ima ion accu acy o he 22%, wi h espec o he case whe e no
op ical moni o s a e employed. Based on he abo e, we clea ly see ha
he ing ess scena io e u ns highe bene i s in e ms o educed
unce ain ies, wi h espec o he eg ess one.
I is wo h no ing ha he ela i e e o is a aluable me ic, bu no
pe ec ly sui able o e en ually e alua ing he impac o he ne wo k
elemen s on he QoT (e.g. he OSNR/SNR) o he connec ions. The e o e,
o e alua e he QoT es ima ion ela ed bene i s, we ansla ed he
pa ame e s es ima ion e o s in o SNR es ima ion e o s. In he eg ess
scena io, since we e alua ed he imp o emen s in OSNR/noise
es ima ion, he e was no need o such a ansla ion. Ou p oposed
eg ess moni o ing and p ocessing me hod esul ed in 0.22 dB
imp o emen in OSNR es ima ion pe link. To ob ain a simila me ic o
he ing ess scena io, we simula ed in VPIpho onics [27] he
ansmission o a 64 GBd QPSK modula ed signal, wi h oll-o ac o
equal o 0.1 and 0.2, c ossing a single il e . We measu ed he SNR
penal y in oduced by he il e , which we implemen ed as a 3.5 h o de
Gaussian TF wi h 75 GHz bandwid h, as unc ion o i s cen al equency
shi and o i s bandwid h a ia ion. The esul s o hese simula ions a e
plo ed in Fig. 11. Then, using he ob ained cu es, we calcula ed he
imp o emen on he SNR penal y es ima ion. The e e ence il e -
ela ed pa ame e s e o anges (±2 GHz) ansla ed in o SNR penal y
es ima ion equal o 0.12 dB o he il e cen al equency shi and o
0.16 dB o he il e 6-dB bandwid h. Since he penal y a ia ions o he
wo di e en oll-o ac o s we e negligible, we conside ed hem as a
unique case. Applying he ing ess moni o ing s a egy and ou
p oposed p ocessing me hod, we we e able o educe he es ima ed
il e -in oduced SNR penal ies down o 0.01 dB and 0.03 dB, o he
cen al equency shi and o he 6-dB bandwid h pa ame e s,
espec i ely. The e o e, ou solu ion yielded a o al educ ion o he
es ima ed SNR penal y equal o 0.24 dB, o he wo conside ed
pa ame e s.
6.CONCLUSION
We s udied di e en scena ios o moni o s placemen wi hin
DWDM and lex-g id op ical ne wo ks. In pa icula , we de ined an
ing ess and an eg ess scena io, in which he moni o s a e placed be o e
and a e he nodes o he ne wo k, espec i ely. In ac , ou goal is o
minimize he numbe o employed OCMs op imizing hei placemen
and o enhance he moni o ing ea u es wi h app op ia e spec al
p ocessing echniques. To his end, we p esen ed wo spec al
p ocessing echniques which le e aged a cu e i ing p inciple and a
ML eg ession algo i hm o e ie e he missing pa ame e s o each
scena io: he il e bandwid h and he il e cen al equency shi in he
ing ess, and he ASE noise o he signal in he eg ess scena io.
We alida ed he p oposed solu ions on spec al da a gene a ed
h ough simula ions and expe imen s. The ob ained esul s con i med
he alidi y o he p oposed echniques. In pa icula , in he ing ess
scena io we obse ed a MAE lowe han 0.98 GHz o he 6-dB
bandwid h es ima ion, and lowe han 0.5 GHz o he il e cen al
Table5.E o  angesand ela i ee o scompa isono  hedi e en scena ios
Se up Scena io Pa ame e Roll-o
ac o
Pa ame e e o
ange
Re e ence e o
ange
Rela i e
e o
Simula ion Ing ess Fil e cen al eq. shi 0.1 0.17 GHz 4 GHz 4.3 %
Fil e 6-dB bandwid h 0.1 0.44 GHz 4 GHz 11 %
Eg ess ASE noise 0.1 0.91 dB 1 dB 91 %
Expe imen al Ing ess
Fil e cen al eq. shi 0.1 0.63 GHz 4 GHz 15.8 %
0.2 0.5 GHz 4 GHz 12.5 %
Fil e 6-dB bandwid h 0.1 1 GHz 4 GHz 25 %
0.2 0.93 GHz 4 GHz 23.3 %
Eg ess ASE noise 0.1 0.78 dB 1 dB 78 %
Fig. 11. SNR penal y in oduced by a 3.5 h o de Gaussian il e , as
unc ion o i s cen al equency shi and i s bandwid h a ia ion, o
an inpu 64 GBd QPSK signal, wi h oll-o ac o equal o 0.1 and 0.2.
-4 -3 -2 -1 0 1 2 3 4
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
71 73 75 77 79
Fil e cen al equency shi [GHz]
SNR penal y [dB]
Fil e bandwid h [GHz]
Fil e BW, o=0.1
Fil e BW, o=0.2
Fc shi , o=0.1
Fc shi , o=0.2