So wa eX 29 (2025) 102039
A ailable online 22 Janua y 2025
2352-7110/© 2025 The Au ho s. Published by Else ie B.V. This is an open access a icle unde he CC BY-NC license (h p://c ea i ecommons.o g/licenses/by-
nc/4.0/).
Con en s lis s a ailable a ScienceDi ec
So wa eX
jou nal homepage: www.else ie .com/loca e/so x
O iginal so wa e publica ion
PHMD: An easy da a access ool o p ognosis and heal h managemen
da ase s
Da id Solís-Ma ína,b,∗
, Juan Galán-Páeza,b, Joaquín Bo ego-Díaza
aDepa men o Compu e Science and A i icial In elligence, Uni e sidad de Se illa, Se ille, Spain
bDa ik In elligence S.A., Se ille, Spain
A R T I C L E I N F O
Keywo ds:
Da ase managemen
P ognos ics and heal h managemen
P edic i e main enance
Condi ion moni o ing
A B S T R A C T
This wo k in oduces a comp ehensi e open-sou ce Py hon lib a y designed o seamless access and handling
o P ognos ics and Heal h Managemen (PHM) da ase s. The lib a y cu en ly suppo s 59 da ase s om
di e se domains, and has been de eloped o simpli y, da ase s sea ch, e ie al, load, and p ep ocessing while
s anda dizing da a o ma s o easy in eg a ion in machine lea ning wo k lows. Wi h buil -in me ada a handling
and ask-speci ic expe imen se ings o diagnosis, p ognosis, and de ec ion, use s can e icien ly p epa e and
analyze da a wi hou needing o manage aw ile o ma s o di ec o ies. A ailable h ough Gi Hub and PyPI,
he lib a y p o ides a obus ounda ion o PHM esea ch and applica ion, o e ing use ul esou ces o boos
he p ojec s o p ac i ione s and esea che s alike.
Code me ada a
Cu en code e sion 2024.0.01
Pe manen link o code/ eposi o y used o his code e sion Fo example:
h ps://gi hub.com/Else ie So wa eX/SOFTX-D-24-00587
Pe manen link o Rep oducible Capsule
Legal Code License GNU Gene al Public License (GPL)
Code e sioning sys em used gi
So wa e code languages, ools, and se ices used py hon
Compila ion equi emen s, ope a ing en i onmen s & dependencies Py hon 3 on Linux, OSX o Windows. Dependencies lis ed in se up.py
and equi emen s. x in code eposi o y.
I a ailable Link o de elope documen a ion/manual –
Suppo email o ques ions [email p o ec ed]
1. Mo i a ion and signi icance
P ope main enance o equipmen and machine y in he indus y is essen ial o ensu e ope a ional e iciency and ex end hei li espan. Thus,
educing eplacemen and epai cos s, esul s in a long- e m posi i e economic impac . The implemen a ion o p edic i e main enance s a egies
in he indus y no only educes cos s associa ed wi h b eakdowns and unexpec ed shu downs bu also imp o es esou ce planning, enhances
p oduc i i y and p oduc quali y, and s eng hens ma ke compe i i eness, gene a ing a a o able economic impac . The gene al aim o p edic i e
main enance is o iden i y and add ess, h ough he use o da a analysis, po en ial p oblems in he equipmen be o e hey become c i ical.
In he ealm o machine lea ning, his o ical da a holds pa amoun impo ance as i se es as he ounda ion upon which machine lea ning
models acqui e knowledge. Howe e , i is commonly obse ed ha da a sca ci y and spa se labeling pose signi ican challenges ac oss mul iple
domains. This challenge is pa icula ly p onounced in p edic i e main enance, whe e de ec ing eal aul s in indus ial se ings, despi e hei cos ly
implica ions, is exceedingly a e, especially in he case o O iginal Equipmen Manu ac u e s (OEMs). Consequen ly, da ase s pe aining o such
scena ios hold immense alue. Howe e , hey a e o en inaccessible o he public due o a mul i ude o conce ns and cons ain s. Fo ins ance, eal
∗Co esponding au ho a : Depa men o Compu e Science and A i icial In elligence, Uni e sidad de Se illa, Se ille, Spain.
E-mail add esses: [email p o ec ed] (Da id Solís-Ma ín), [email p o ec ed] (Juan Galán-Páez), [email p o ec ed] (Joaquín Bo ego-Díaz).
h ps://doi.o g/10.1016/j.so x.2025.102039
Recei ed 3 No embe 2024; Recei ed in e ised o m 3 Janua y 2025; Accep ed 6 Janua y 2025
So wa eX 29 (2025) 102039
2
Da id Solís-Ma ín e al.
Table 1
TL: T ans e Lea ning, DL: Deep Lea ning, CV: C oss-Valida ion, Well-Known: Indica es whe he he ool includes well-known
da ase s, Mul i- ask: Speci ies i he ool de ines mul iple asks pe da ase , Me ada a: Indica es whe he he ool includes
me ada a associa ed wi h each da ase .
Da ase gi [8] gi [9] pyPHM [3] P ogPy [10] gi [7] phmd (ou s)
Sea ch No No No No No Yes
Download No No Yes Yes No Yes
Load Yes Yes Yes Yes Yes Yes
CV No No No No No Yes
#Da ase s 6 3 4 2 3 59
Well-Known Yes Yes Yes Yes No Yes
Me ada a No No No No No Yes
Mul i- ask No No No No No Yes
Focus TL DL Rep oducibili y P ognosis Pape Rep oducibili y
Discon inued No Yes Yes No Yes No
da ase s could con ain sensi i e in o ma ion such as ailu e equencies and he ypes o senso s ins alled [1]. The e o e, mos s udies demons a ing
success ul applica ions o machine lea ning in manu ac u ing wi hhold hei aining and es ing da ase s om public access. This p ac ice hinde s
e ec i e compa ison be ween di e en app oaches and ep oducibili y o such s udies.
Con e sely, o e he pas decade, esea che s and indus y s akeholde s ha e begun c a ing da ase s speci ically designed o asks like aul
p edic ion and diagnosis, emaining use ul li e es ima ion, o assessing wea in componen s, wi hin he ealm o p edic i e main enance. These
da ase s ci cula e wi hin he esea ch communi y o p opel ad ancemen s in he ield. Howe e , hey equen ly p esen di e se o ma s (such as
CSV, HDF5, TXT, MAT, e c.), may be spli ac oss a ious iles ha equi e consolida ion, and, in ce ain cases, lack in o ma ion such as, how he
da a is o ganized wi hin he ile ( his is he case o he MAT o ma ), he eby limi ing i s euse in o he wo ks.
I is a ac ha many scien i ic s udies a e di icul o impossible o ep oduce [2,3]. Despi e he exis ence o mul iple public da ase s o
p edic i e main enance, he p ocess o loca ing and managing hem emains ime-consuming due o hei di e se o ma s. This is likely one
o he easons why mos published wo ks only alida e hei esul s wi h one o wo da ase s, despi e he public a ailabili y o many mo e.
P o iding a ool ha simpli ies he ga he ing and p ocessing o public da ase s could signi ican ly bene i he esea ch communi y, in ad ancing
echnology. Collec ing mul iple da ase s o ca y ou a esea ch wo k can be ime-consuming and include asks such as loca ing he da a sou ce,
ga he ing desc ip i e in o ma ion (such as me ada a, ea u e desc ip ions, and da a o ganiza ion), downloading, and loading in o a well-known
and s anda dized in-memo y da a s uc u e o ma , ob aining ela ed bibliog aphy, e c. Addi ionally, such a ool would acili a e benchma king
wi h mo e han jus wo da ase s, enabling esea che s o e alua e hei wo k agains a b oade ange o scena ios. Mo eo e , alida ing esea ch
hypo hesis on he same da ase s (and same e sions) would acili a e esul compa ison be ween s udies.
Exis ing wo ks, such as [4,5], and [6], p o ide su eys o publicly a ailable da ase s in PHM, bu none o e a p ac ical ool o seamless da ase
access. While [7] add esses his gap by p o iding a ool o download, load, and spli da ase s, i s unc ionali y is limi ed o h ee da ase s c ea ed
by he au ho s o a published pape . The ool mos simila o ou s is pyPHM [3], which ocuses on he ep oducibili y o esea ch. Howe e , his
ool only p o ides access o ou da ase s, and he eposi o y sugges s ha main enance could be discon inued.
The Table 1p o ides a compa ison o hese ools and o he s wi h ou s. The mos ema kable ea u e o ou ool is he la ge numbe o da ase s—
59 a he ime o w i ing— ha i p o ides access o. The sea ch capabili y is ano he di e en ia ing ea u e, hough i is no as necessa y when he
numbe o da ase s is e y low, as in he case o he o he wo ks. All he me ada a in o ma ion associa ed wi h each da ase is ano he impo an
ea u e, ensu ing ha use s can ully unde s and he con ex and cha ac e is ics o he da a. Addi ionally, he c oss- alida ion capabili ies, oge he
wi h o he elemen s ha will be e iewed la e , make his ool pa icula ly aluable o ep oducible esea ch. By acili a ing easy compa ison
o models, hese ea u es con ibu e signi ican ly o ad ancing he s a e o he a in PHM, enabling esea che s o quickly es and alida e new
app oaches wi h mul iple da ase s.
The es o he pape is s uc u ed as ollows. Sec ion 2p o ides a de ailed desc ip ion o he ool, including an enume a ion o i s unc ionali ies
and he a chi ec u e o he published so wa e package. Sec ion 3p esen s a ious sou ce code examples o showcase he so wa e’s unc ionali ies,
such as da ase sea ch, da ase in o ma ion and me ada a e ie al, and da a loading. Addi ionally, his sec ion includes wo examples ha
demons a e he aining o neu al ne wo ks o diagnosis and p ognosis asks. Finally, Sec ion 4ou lines conclusions and sugges s di ec ions
o u u e esea ch.
2. So wa e desc ip ion
This wo k is accompanied by a ee, open-sou ce so wa e lib a y designed o acili a e access o da ase s in he ield o PHM. The p ima y goal
is o s eamline au oma ed p ocessing, allowing esea che s and p ac i ione s o easily ob ain, manipula e, and analyze da a ele an o hei wo k.
This lib a y p o ides ools o da a sea ch, downloading, and loading wi hou conce ns abou da ase loca ions o sou ce o ma s. In summa y, he
lib a y o e s a uni ied in e ace o accessing PHM da ase s (59 a he ime o w i ing) om di e se sou ces and he e ogeneous o ma s, signi ican ly
educing he ime and e o needed o p epa e da ase s o machine lea ning p ojec s and o he analy ical asks.
Rega ding he o igin o he 59 da ase s ha ha e been ga he ed, 10 a e p o ided by NASA [11–20], while 8 a e o e ed by he PHM Socie y
h ough hei Ame ican, Eu opean, and Asian challenges [21–28]. Uni e si ies wo ldwide ha e con ibu ed 19 da ase s [29–47], and 7 mo e we e
p oduced by o he esea ch ins i u ions [48–54]. The one da ase belongs o he Socie y o Machine y Failu e P e en ion Technology (MFPT) [27],
and he emaining we e published by a ious companies.
Rega ding he a ailable asks, 31 da ase s suppo diagnosis asks, 35 can be used o p ognosis asks, and 6 a e con igu ed o de ec ion asks.
So wa eX 29 (2025) 102039
3
Da id Solís-Ma ín e al.
Fig. 1. The download package e ie es he da ase om he in e ne , i i has no been downloaded ye , and s o es i in a local di ec o y. The eade package con ains sou ce
code speci ic o eading each da ase o ma , while he me ada a di ec o y holds JSON iles wi h he da ase me ada a. The main unc ionali y is p o ided by he da ase s package
allowing o sea ch, display da ase in o ma ion, and load da ase s. The Da ase and Task classes encapsula e such unc ionali y, enhancing he usabili y o he lib a y.
2.1. So wa e unc ionali ies
The main ea u es o he de eloped lib a y can be summa ized as ollows:
•Lis ing and sea ch. The lib a y allows use s o sea ch o da ase s ailo ed o speci ic pu poses. Each da ase is associa ed wi h a se o
me a-a ibu es acili a ing he selec ion o he mos sui able da ase o a pa icula esea ch ask.
•Au oma ic download. The download p ocess is au oma ically igge ed when a da ase is equi ed o load only i i has no been p e iously
downloaded. The da ase in eg i y is con olled h ough a hash code o ensu e ha he da ase has no been co up ed du ing he download
p ocess.
•Loading. One o he mos impo an ea u es o his lib a y is i s da ase loading capabili y. The loading unc ion abs ac s away he o iginal
sou ce o ma o he da ase , making i o ma -agnos ic o he use and allowing hem o ocus di ec ly on da a analysis a e loading. The
loading unc ion accep s he da ase name and ask ype as pa ame e s. Based on he speci ied ask, he a ge a iable is au oma ically
compu ed, and ele an ea u es a e il e ed acco dingly. When he da ase is ead i is also pa i ioned, p o iding c oss- alida ion olds and
a ese ed es se . To acili a e expe imen ep oducibili y, a andom seed can be se o con ol he spli . This will make esul s o di e en
esea ch s udies easily ep oducible and compa able by jus anno a ing he seed used in each expe imen .
2.2. So wa e a chi ec u e
The s uc u e o he lib a y is ou lined in Fig. 1. Each da ase is de ined by wo iles: a JSON ile con aining all me ada a, and a sou ce code
ile esponsible o he loading logic o he da ase . JSON iles a e s o ed in he me ada a di ec o y, while he loading code esides in he eade s
package. The download package includes he code dedica ed o downloading unc ionali ies. The end-use unc ionali y is p o ided by he da ase s
package, which p o ides unc ions and classes o eading me ada a, loading da ase s, and u ili ies o desc ibing and lis ing a ailable da ase s.
All unc ionali ies can be accessed h ough bo h he Da ase and Task classes. The Da ase class includes he s a ic me hod sea ch,
which p o ides he sea ch unc ionali y. This unc ionali y allows use s o lis all a ailable da ase s o il e hem based on a ious a ibu es, such
as he da ase ’s name o code-name, ask name (e.g., diagnosis, p ognosis, wea ), a ge ask name (e.g., aul , ul, s age), applica ion and domain
o applica ion, ype o ea u es, publishe , and he na u e o he da a.
An ins ance o he Da ase class ep esen s a speci ic da ase and equi es only he code-name o he da ase o be speci ied in he cons uc o .
Addi ionally, he use can op ionally speci y a cache di ec o y, whe e he da ase iles will be s o ed.
The download me hod allows use s o manually download he da ase , p o iding ull con ol o e he downloading p ocess, which may be
necessa y in scena ios such as mul ip ocessing en i onmen s. Howe e , he loading p ocess au oma ically e i ies whe he he da ase has al eady
been downloaded and downloads i i needed. The desc ibe me hod gene a es a de ailed desc ip ion o he da ase . Addi ionally, use s can
di ec ly access he asks de ined wi hin he da ase using indexing.
The Task ins ances, ob ained by indexing he da ase , allow use s o con igu e a ious a ibu es o eading old se s:
• olds (In ): De aul s o 5 o o he maximum numbe o olds depending on he numbe o uni s p esen in he da ase .
•p ep ocess (Objec ): De aul s o None. This a ibu e can be se o a alid Sciki -lea n ans o me , o any o he objec ha implemen s
bo h he i and ans o m me hods.
•no malize_ou pu (Bool): De aul s o False. In asks wi h con inuous a ge s, such as RUL, no malizing he ou pu is common. When
se o T ue, he a ge column will be no malized.
• es _pc (Floa ): De aul s o 0.3, speci ying he pe cen age o da a used o he es se .
So wa eX 29 (2025) 102039
4
Da id Solís-Ma ín e al.
• e u n_ es (Bool): De aul s o T ue. I se o False, only he alida ion and aining se s a e e u ned.
• andom_s a e (In ): Se s he andom seed used o spli he da a. Changing his alue esul s in a di e en da a spli .
Once hese a ibu es ha e been se , each c oss- alida ion old can be accessed h ough indexing. This e u ns Pandas Da aF ames con aining he
aining, alida ion, and, i speci ied, es se s. The ep oducibili y o expe imen s is ensu ed hanks o he olds, es _pc , and andom_s a e
a ibu es, which con ol he c oss- alida ion spli .
No e ha , cu en ly, his ool does no p o ide au oma ic p ep ocessing pipelines, excep o a ge no maliza ion in RUL asks. Howe e , use s
can p o ide a cus om p ep ocessing pipeline by se ing he p ep ocess a ibu e o he Task class.
3. Illus a i e examples
To p o ide a be e idea o he con ibu ions o his wo k, his sec ion p o ides examples o illus a e he lib a y use and i s main ea u es.
3.0.1. Sea ch da ase s
The da ase Sea ch module allows lis ing da ase s ma ching ce ain c i e ia. Fo example sea ching o da ase s ha con ain ib a ion da a, use s
can iden i y a ious da ase s ac oss di e en domains and applica ions, such as mechanical o manu ac u ing con ex s. The Example 1illus a es
how o pe o m a da ase sea ch using he phmd lib a y, showcasing he ele an de ails, including da ase names, domains, applica ion a eas, ask
names, and he na u e o he da a and ea u es p o ided.
Example 1 Da ase sea ch example
1>>> impo phmd
2>>> da ase s . Da ase . sea ch ( ea u es=’ ib a ’ )
3
4name domain na u e app ask name [ a ge ] da a na u e ea u es
5−−−−−−− −−−−−−−−−−−−− −−−−−−−−−−− −−−−−−− −−−−−−−−−−−−−−−−−− −−−−−−−−−−−− −−−−−−−−
6CWRU Mechanical ime−se ies Bea ing Diagnosis [ aul ] ime−se ies ib a ion
7DFD15 Manu ac u ing ime−se ies D ill Diagnosis [ aul ] ime−se ies ib a ion
8DFD15 Manu ac u ing ime−se ies D ill S age [ s age ] ime−se ies ib a ion
9. . .
10 . . .
11 . . .
12 UPM23 Mechanical ime−se ie s Bea ing Diagnosis [ aul ] ime−se ies ib a ion
13 XJTU−SY Mechanical ime−se i e s Bea ing P ognosis [ ul ] ime−se ies ib a ion
14 XJTU−SY Mechanical ime−se i e s Bea ing Diagnosis [ aul ] ime−se ies ib a ion
While diagnosis and p ognosis a e he mos common asks, some da ase s allows con igu ing seconda y asks. Fo example, he DF15 da ase
includes a ‘‘S age’’ ask, in which he goal is o classi y he s age o he d illing p ocess based on ib a ion signals. Al hough he s age o he d illing
p ocess is known and, in p ac ice, i may no be needed o be de ec ed by an a i icial in elligence model, his kind o ask could be pa icula ly
in e es ing in con ex s such as mul i- ask lea ning.
Mo e de ails abou he sea ch capabili ies can be ound in he documen a ion wi hin he code eposi o y.
3.0.2. Desc ibe da ase
Unde s anding he speci ics o a da ase is i al o e ec i e analysis and model de elopmen . The phmd lib a y simpli ies his p ocess by
p o iding a s aigh o wa d way o e ie e de ailed in o ma ion o each o da ase wi hin he collec ion. The Da ase module allows use s ob aining
comp ehensi e desc ip ions, sys em in o ma ion, ea u es, asks, esou ces, and e e ences associa ed wi h a pa icula da ase .
In Example 2, we demons a e how o access de ailed in o ma ion o he well-known CWRU (Case Wes e n Rese e Uni e si y) da ase , which
ocuses on bea ing aul diagnosis. The ou pu includes essen ial de ails such as he desc ip ion o he da ase , he ypes o senso s used, he na u e
o he da a, and s o age equi emen s o he da ase . This allows use s o e alua e he sui abili y o he da ase o hei speci ic applica ions and
esea ch needs. Mo eo e , i p o ides in o ma ion use ul when w i ing epo s o pape s such as ci a ion e e ence, license and ela ed wo ks using
he da ase .
Example 2 Da ase desc ip ion example
1>>> om phmd impo da ase s
2>>> ds = da ase s . Da ase ("CWRU" )
3>>> p in ( ds . desc ibe ())
4
5Desc ip ion
6===========
7In his enowned da ase , expe imen s we e conduc ed u ilizing a 2 HP Reliance Elec ic
8mo o , whe e accele a ion da a was measu ed a loca ions bo h nea o and emo e om he
9mo o . . .
10
11 Designa ion : Bea ing Faul Diagnos ic
12 Publishe : Case Wes e n Rese e Uni e si y
13 Domain: Mechanical componen
14 Applica ion : Bea ing
15 License : CC BY−SA 4.0
16
17 Sys em in o
18 ===========
19 1. ype : Ro a o y machine : bea ing
20 2. senso s : Vol me e , amme e and he mocouple senso sui e
21 3. bea ing : 6205−2RSL JEM SKF deep−g oo e ball bea ing (and NTN equi alen )
22
So wa eX 29 (2025) 102039
5
Da id Solís-Ma ín e al.
23 Fea u es
24 ========
25 BA : desc ip ion : base accele ome e da a (no a ailable in al l expe imen s )
26 ype : ib a ion
27 DE : desc ip ion : d i e end accele ome e da a
28 ype : ib a ion
29 FE : desc ip ion : an end accele ome e da a (no a ailable in a l l expe imen s )
30 ype : ib a ion
31
32 Tasks
33 =====
34 Diagnosis :
35 ea u es : DE
36 iden i ie : uni
37 min_ s_len : 63788
38 na u e : ime−se ies
39 num_uni s : 161
40 a ge : aul
41 a ge _dis ibu ion : 0.24 ,0.51 ,0.23 ,0.021
42 a ge _labels : IR ,OR,BA,NO
43 ype : classi ica ion : mul iclass
44
45 Resou ces
46 =========
47 1. s o age :
48 a) zipped : 246MB
49 b) unzipped : 689MB
50 c ) RAM :
51 Da a se ( u l l ) : 5.2GB
52 2. load ime (SSD disk ) :
53 a) unzipped :
54 Da a se ( u l l ) : 3s
55 a) zipped :
56 Da a se ( u l l ) : 7s
57
58 Re e ences
59 ==========
60 ci a ion : K.A. Lopa o , Bea ings ib a ion da a se . The Case Wes e n Rese e
61 Uni e si y Bea ing Da a Cen e . h ps :// enginee ing . case . edu
62 manual download : h ps :// enginee ing . case . edu/bea ingda acen e
3.0.3. Load da ase
In Example 3, we demons a e how o load he CWRU da ase . By ini ializing he Da ase class wi h he da ase name, use s can e ie e speci ic
asks associa ed wi h he da ase . The asks a ailable o each da ase a e speci ied in i s desc ip ion. In his case, he da ase is loaded o wo k in
he aul ask. The subsequen command loads he subse s associa ed wi h his ask, whe e he index 0 e e s o he i s old o he da ase spli s
o c oss- alida ion.
Example 3 Da ase load example
1>>> ds = da ase s . Da ase ("CWRU" )
2>>> ask = ds [ ’ aul ’ ]
3>>> se s = ask [0]
4
5Da ase CWRU al eady downloaded and ex ac ed
6Please emembe o ci e he o iginal da ase publishe :
7@misc{caseBea ingDa a ,
8au ho = {} ,
9 i l e = {{B}ea ing {D} a a {C} en e | {C}ase {S}chool o {E} nginee ing
10 {C}ase {W}es e n {R}ese e {U} ni e si y −−− enginee ing . case . edu} ,
11 howpublished = { u l { h ps :// enginee ing . case . edu/bea ingda acen e }} ,
12 yea = {} ,
13 no e = {[Accessed 08−04−2024]},
14 }
15 You can download he da ase manually om : h ps :// enginee ing . case . edu/bea ingda acen e
16
17 ∗∗ I you ind his ool use ul , please ci e ou So wa eX pape .
18
19 Reading da a : 100%||||||||||||||||| 161/161 [00:03<00:00, 47.43 i /s ]
20 INFO : oo : Read in 5.96511435508728 seconds
21 INFO : oo : I is possible s a i ied spli ? T ue
22 INFO : oo : Read 3 se s : ain , al , e s
23 INFO : oo : Columns : DE, aul , uni
24 INFO : oo : T ain shape : (28567988, 3)
25 INFO : oo : Val shape : (7804030, 3)
26 INFO : oo : Tes shape : (979629, 3)
No e ha he loading unc ionali y eads he da ase s uc u ed o be used in c oss- alida ion, and ese ing pa o he da a as es se .
3.1. Benchma king example
The Example 4demons a es he p ocess o benchma king a p edic i e model using he lib a y and mul iple da ase s. Fo each da ase , he RUL
ask is con igu ed using a 3- old c oss- alida ion se up. Key p ep ocessing s eps include no malizing he a ge alues and applying aMinMaxScale
So wa eX 29 (2025) 102039
6
Da id Solís-Ma ín e al.
o he inpu ea u es. A con olu ional neu al ne wo k (CNN) is ained on ime-se ies da a gene a ed by spli ing he signals in o non-o e lapping
windows. Ea ly s opping is employed o op imize aining ime. A e comple ing he c oss- alida ion, a inal model is ained on he combined
aining and alida ion da a, and e alua ed using he es se .
Example 4 Benchma king example
1# Lis o da ase s o p ocess
2DATASETS = [’ARAMIS20 ’ ,’PRONOSTIA ’ ,’PHME20 ’ ]
3
4# I e a e o e each da ase in he l i s
5 o da ase in DATASETS:
6# Load he da ase objec using he Da ase cl ass
7ds = da ase s . Da ase ( da ase )
8
9# Access he RUL ask wi hin he da ase
10 ask = ds [ ’ ul ’ ]
11
12 # Con igu e he ask : numbe o olds , no maliza ion , and p ep ocessing
13 ask . olds = 3# Use 3− old c oss− alida ion
14 ask . no malize_ou pu = T ue # No malize he a ge (RUL) alues
15 ask . p ep ocess = MinMaxScale () # Apply MinMaxScale o ea u e scaling
16
17 # De ine he ime se ies window size based on da ase me ada a
18 TS =min(1024 , ds [ ’ ul ’ ] . me a[ ’min_ s_len ’] // 20) # Ensu e TS is easonable
19
20 # Lis o s o e he numbe o epochs o ea ly s opping in each old
21 ea_epochs = []
22
23 # Ex ac he uni iden i i e o g ouping s ignal s
24 uni _id = ds [ ’ ul ’ ] . me a[ ’ i de n i i e ’ ]
25 EPOCHS = 100 # Maximum numbe o aining epochs
26
27 # Loop h ough each old in he c oss− alida ion se up
28 o iin ange ( ask . olds ) :
29 # Access he da a o he cu en old
30 da a = ask [ i ]
31
32 # Spli he da a in o aining , alida ion , and e s se s
33 X_ ain , X_ al , X_ es = da a [ ’ ain ’ ] , da a [ ’ al ’ ] , da a [ ’ es ’ ]
34
35 # De e mine he minimum signal size ac oss ain , alida ion , and e s se s
36 sig n a l _ s i zes = [ X_ ain . g oupby( uni _id ) . size ().min() ,
37 X_ al . g oupby( uni _id ) . size ().min() ,
38 X_ es . g oupby( uni _id ) . size ().min()]
39
40 # Se a maximum allowable s ign al leng h o s p l i i n g
41 signal_max_leng h =min(np . min( s ignal _ s i z e s ) , 20000000)
42
43 # P epa e he da a o he model by c ea ing o e lapping windows
44 X_ ain , Y_ ain = window_spli ( X_ ain , uni _id , ds[ ’ ul ’ ] . me a[ ’ ea u es ’] ,
45 ’ ul ’ , TS , signal_max_leng h )
46 X_ al , Y_ al = window_spli ( X_ al , uni _id , ds [ ’ ul ’ ] . me a[ ’ ea u es ’] ,
47 ’ ul ’ , TS , signal_max_leng h )
48
49 # C ea e a 1D con olu ional neu al ne wo k o ime−se ies p ocessing
50 model = c ea e_con _1d_ne wo k ( X_ ain . shape [1:])
51
52 # Compile he model
53 model .compile ( op imize = . ke as . op imize s .Adam( l =0.0001),
54 me ics=[’mae ’ ] , lo ss= ’mse ’ )
55
56 # Se up ea ly s opping o p e en o e i i ng
57 es = . ke as . callbacks . Ea lyS opping ( moni o = ’ al_loss ’ , pa ience =8)
58
59 # T ain he model on he cu en old
60 esul s = model . i ( X_ ain , Y_ ain ,
61 epochs=EPOCHS,
62 ba ch _size =128,
63 e bose=1,
64 a lida io n_d a a =(X_ al , Y_ al ) ,
65 callbacks=[es ])
66
67 # Reco d he numbe o epochs be o e ea ly s opping
68 ea_epochs . append( len ( esul s . his o y [’ loss ’ ] ))
69
70 # T ain he i n a l model using a l l aining da a ( ain + alida ion )
71 model = c ea e_con _1d_ne wo k ( X_ ain . shape [1:])
72 model .compile( op imize = . ke as . op imize s .Adam( l =0.0001),
73 me ics=[’mae ’ ] , lo ss= ’mse ’ )
74 esul s = model . i (np . conca ena e (( X_ ain , X_ al )) ,
75 np . conca ena e (( Y_ ain , Y_ al )) ,
76 epochs=in (np . mean( ea_epochs )) ,
77 ba ch _size =128,
78 e bose=1)
79
So wa eX 29 (2025) 102039
7
Da id Solís-Ma ín e al.
80 # Make p edic ions on a single e s uni
81 X_ es , Y_ es = window_spli ( X_ es , uni _id , ds[ ’ ul ’ ] . me a[ ’ ea u es ’] ,
82 ’ ul ’ , TS , signal_max_leng h )
83 es _ esul s = model . e alua e ( X_ es , Y_ es )
84
85 p in ( da ase , es _ esul s )
The implemen a ion o he unc ions window_spli and c ea e_con _1d_ne wo k can be consul ed in he ull no ebook examples in
he da a sou ce eposi o y.
4. Conclusions and u u e wo k
This pape desc ibes he unc ionali ies o he Py hon package phmd, which aims o acili a e he sea ch, downloading and loading o di e en
(cu en ly 59) PHM da ase s. To ou knowledge, his is he only Py hon package ha manages such a la ge collec ion o publicly a ailable da ase s
wi hin his ield. The package is eely accessible o download om i s Gi Hub eposi o y. We p o ide a se o examples demons a ing how o
sea ch o da ase s, e ie e in o ma ion on a speci ic da ase , and download and load i seamlessly. Since he de elopmen e sion is hos ed on
Gi Hub, use s can epo any unc ionali y issues, eques new ea u es by opening an issue in he eposi o y o include new da ase s.
While we belie e his ool is a aluable con ibu ion o he PHM esea ch communi y, i has signi ican po en ial o enhancemen . The numbe
o da ase s can be expanded, including hose ela ed o main enance in e en ions, which a e no cu en ly included in he ool. Addi ionally, u u e
wo k could ocus on inco po a ing ad anced ea u es such as au oma ed p ep ocessing pipelines ailo ed o speci ic machine lea ning models, such
as ea u e ex ac ion and selec ion o non-neu al ne wo k machine lea ning models. Fu he mo e, ex ending he ool o include mo e ad anced
benchma king capabili ies could b oaden i s applicabili y and usabili y.
CRediT au ho ship con ibu ion s a emen
Da id Solís-Ma ín: W i ing – e iew & edi ing, W i ing – o iginal d a , Supe ision, So wa e, Me hodology, Fo mal analysis, Concep ual-
iza ion. Juan Galán-Páez: W i ing – e iew & edi ing, Resou ces, Me hodology, Funding acquisi ion, Concep ualiza ion. Joaquín Bo ego-Díaz:
W i ing – e iew & edi ing, Resou ces, P ojec adminis a ion, Me hodology, Funding acquisi ion.
Decla a ion o compe ing in e es
The au ho s decla e he ollowing inancial in e es s/pe sonal ela ionships which may be conside ed as po en ial compe ing in e es s:
Da id Solis-Ma in epo s inancial suppo was p o ided by Uni e si y o Se ille. Juan Galan-Paez epo s inancial suppo was p o ided
by Uni e si y o Se ille. Joaquin Bo ego-Diaz epo s inancial suppo was p o ided by Uni e si y o Se ille. I he e a e o he au ho s,
hey decla e ha hey ha e no known compe ing inancial in e es s o pe sonal ela ionships ha could ha e appea ed o in luence he wo k
epo ed in his pape .
Acknowledgmen s
This wo k has been suppo ed by G an PID2023-147198NB-I00 unded by MICIU/AEI/10.13039/501100011033 (Agencia Es a al de In es i-
gación) and by FEDER, UE, and by he Minis y o Science and Educa ion o Spain h ough he na ional p og am ‘‘Ayudas pa a con a os pa a
la o mación de in es igado es en emp esas (DIN2019-010887/AEI / 10.13039/50110001103)’’, o S a e P og amme o Science Resea ch and
Inno a ions 2017–2020.
Re e ences
[1] Kha azian Z, Lindg en T, Magnússon S, S eine O, Reyna OA. SCANIA componen x da ase : A eal-wo ld mul i a ia e ime se ies da ase o p edic i e main enance. 2024,
a Xi p ep in a Xi :2401.15199.
[2] Raube TW, da Sil a Loca AL, de Assis Bold F, Rod igues AL, Va ejão FM. An expe imen al me hodology o e alua e machine lea ning me hods o aul diagnosis based
on ib a ion signals. Expe Sys Appl 2021;167:114022.
[3] on Hahn T, Meche ske CK. Compu a ional ep oducibili y wi hin p ognos ics and heal h managemen . 2022, a Xi p ep in a Xi :2205.15489.
[4] Jo icic E, P imo ac D, Cupic M, Jo ic A. Publicly a ailable da ase s o p edic i e main enance in he ene gy sec o : A e iew. IEEE Access 2023.
[5] Hbiche O, Boudnaya J, Sallak M, Mkhida A, Inouss H. Da ase s analysis in p edic i e main enance: p ognos ics and heal h managemen . In: 2022 2nd in e na ional con e ence
on inno a i e esea ch in applied science, enginee ing and echnology. IEEE; 2022, p. 1–7.
[6] Mau he F, B aun C, Raible J, Zeile P, Hube MF. O e iew o publicly a ailable deg ada ion da a se s o asks wi hin p ognos ics and heal h managemen . 2024, a Xi
p ep in a Xi :2403.13694.
[7] Soualhi M, Soualhi A, Nguyen KT, Medjahe K, Cle c G, Razik H. Open he e ogeneous da a o condi ion moni o ing o mul i aul s in o a ing machines used in di e en
ope a ing condi ions. In J P ogn Heal Manag 2023;14(2).
[8] Zhang J. TL aul diagnosis lib a y. 2022, h ps://gi hub.com/Feaxu e- esh/TL-Faul -Diagnosis-Lib a y.
[9] Fangzheng X. Bea ing aul diagnosis da ase s. 2023, h ps://gi hub.com/ onde xu/Bea ing- aul -diagnosis-da ase s.
[10] Teube C, G i i h KJ, Co be a M, Kulka ni C, Bane jee P, Daigle M. P ogpy py hon p ognos ics packages. 2024, URL h ps://nasa.gi hub.io/p ogpy.
[11] B. De B, Kulka ni C, Daigle M. Randomized ba e y usage da ase . 70, 2014, NASA AMES p ognos ics da a eposi o y.
[12] Saha B, Goebel K. Ba e y da a se . 2007, NASA AMES p ognos ics da a eposi o y.
[13] Capaci o elec ical s ess-2 - ca alog — ca alog.da a.go . 2024, h ps://ca alog.da a.go /da ase /capaci o -elec ical-s ess-2. (Accessed 3 Janua y 2025).
[14] Celaya JR, Saxena A, Saha S, Goebel KF. P ognos ics o powe MOSFETs unde he mal s ess accele a ed aging using da a-d i en and model-based me hodologies. In: Annual
con e ence o he PHM socie y, ol. 3, 2011, p. 0–10.
[15] F icke K, Nascimen o R, Co be a M, Kulka ni C, Viana F. P ognosis o li-ion ba e ies unde la ge load a ia ions using hyb id physics-in o med neu al ne wo ks. In: Annual
con e ence o he PHM socie y, ol. 15, 2023, p. 0–12.
[16] Saxena A, Goebel K, Simon D, Eklund N. Damage p opaga ion modeling o ai c a engine un- o- ailu e simula ion. In: 2008 in e na ional con e ence on p ognos ics and
heal h managemen . IEEE; 2008, p. 1–9.
[17] P ognos ics cen e o excellence da a se eposi o y - NASA — nasa.go . 2024, h ps://www.nasa.go /in elligen -sys ems-di ision/disco e y-and-sys ems-heal h/pcoe/pcoe-
da a-se - eposi o y/. (Accessed 3 Janua y 2025).
So wa eX 29 (2025) 102039
8
Da id Solís-Ma ín e al.
[18] Kulka ni CS, Gua ne os Luna A. Desc ip ion o simula ed small sa elli e ope a ion da a se s. Tech. ep., NASA; 2018.
[19] Kulka ni C, Hogge E, Quach C, Goebel K. HIRF ba e y da a se . 2025, NASA Ames P ognos ics Da a Reposi o y, h ps://www.nasa.go /in elligen -sys ems-di ision/disco e y-
and-sys ems-heal h/pcoe/pcoe-da a-se - eposi o y/. (Accessed 3 Janua y 2025), h p:// i.a c.nasa.go /p ojec /p ognos ic-da a- eposi o y.
[20] A ias Chao M, Kulka ni C, Goebel K, Fink O. Ai c a engine un- o- ailu e da ase unde eal ligh condi ions o p ognos ics and diagnos ics. Da a 2021;6(1):5.
[21] Seijun Chung S. Da a challenge a PHM Asia Paci ic 2021. Kaggle. 2021, h ps://kaggle.com/compe i ions/phmap21-classi ica ion- ask.
[22] Danilo Gio dano MS. PHME da a challenge. 2024, h ps://gi hub.com/PHME-Da achallenge/Da a-Challenge-2024, Eu opean con e ence o he p ognos ics and heal h
managemen socie y.
[23] 2010 PHM socie y con e ence da a challenge - phm socie y — phmsocie y.o g. 2024, h p://www.phmsocie y.o g/compe i ion/phm/10. (Accessed 3 Janua y 2025).
[24] 2023 PAP. Da a challenge a PHM Asia Paci ic 2023. 2023, h ps://phmap.jp/p og am-da a/.
[25] Annual con e ence o he p ognos ics and heal h managemen socie y 2018 - PHM socie y — phmsocie y.o g. 2024, h p://www.phmsocie y.o g/compe i ion/phm/18.
(Accessed 3 Janua y 2025).
[26] İnce K, Si keci E, Genç Y. Remaining use ul li e p edic ion o expe imen al il a ion sys em: A da a challenge. In: Medjahe ABK, edi o . P oceedings o he Eu opean
con e ence o he PHM socie y 2020. PHM Socie y; 2020, p. 0–10, A ailable a h ps://phmpape s.o g/index.php/phme/a icle/ iew/1317.
[27] Faul da a se s - socie y o machine y ailu e p e en ion echnology — m p .o g. 2024, h ps://www.m p .o g/ aul -da a-se s/. (Accessed 3 Janua y 2025).
[28] He J, Guan X, Peng T, Liu Y, Saxena A, Celaya J, e al. A mul i- ea u e in eg a ion me hod o a igue c ack de ec ion and c ack leng h es ima ion in i e ed lap join s using
lamb wa es. Sma Ma e S uc 2013;22(10):105007.
[29] Bosello M, Falcome C, Rossi C, Pau G. To cha ge o o sell? EV pack use ul li e es ima ion ia LSTMs, CNNs, and au oencode s. Ene gies 2023;16(6). h p://dx.doi.o g/10.
3390/en16062837, URL h ps://www.mdpi.com/1996-1073/16/6/2837.
[30] Cao P, Zhang S, Tang J. P ep ocessing- ee gea aul diagnosis using small da ase s wi h deep con olu ional neu al ne wo k-based ans e lea ning. Ieee Access
2018;6:26241–53.
[31] Lu C, Wang Y, Ragulskis M, Cheng Y. Faul diagnosis o o a ing machine y: A me hod based on image p ocessing. PLoS One 2016;11(10):e0164111.
[32] Kollmeye P, Ca los V, Naguib M, Skells M. LG 18650HG2 Li-ion ba e y da a and example deep neu al ne wo k xEV SOC es ima o sc ip . 2020, URL h ps://da a.mendeley.
com/da ase s/cp3473x7x /3.
[33] Gea da ase s. 2024, h ps://www.kau-sdol.com/kaug. (Accessed 3 Janua y 2025).
[34] Wang B, Lei Y, Li N, Li N. A hyb id p ognos ics app oach o es ima ing emaining use ul li e o olling elemen bea ings. IEEE T ans Reliab 2018;69(1):401–12.
[35] Bea ing da a cen e | case school o enginee ing | case wes e n ese e uni e si y — enginee ing.case.edu. 2024, h ps://enginee ing.case.edu/bea ingda acen e . (Accessed
3 Janua y 2025).
[36] So o-Ocampo CR, Ma o o J, Cano-Mo eno JD, Me a JM. Bea ing da abase - e alua ion o isola ed cases — zenodo.o g. 2023, h ps://zenodo.o g/ eco ds/8241764. (Accessed
3 Janua y 2025).
[37] Nehasil O, Dobiášo á L, Mazanec V, Ši ok`
y J. Ve sa ile AHU aul de ec ion–design, ield alida ion and p ac ical applica ion. Ene gy Build 2021;237:110781.
[38] Lessmeie C, Kimo ho JK, Zimme D, Sex o W. Condi ion moni o ing o bea ing damage in elec omechanical d i e sys ems by using mo o cu en signals o elec ic mo o s:
A benchma k da a se o da a-d i en classi ica ion. In: PHM socie y Eu opean con e ence, ol. 3, 2016, p. 0–10.
[39] Co addu A, One o L, Ghio A, Sa io S, Angui a D, Figa i M. Machine lea ning app oaches o imp o ing condi ion-based main enance o na al p opulsion plan s. P oc Ins
Mech Eng, Pa M: J Eng Ma i En i on 2016;230(1):136–53.
[40] Cipollini F, One o L, Co addu A, Mu phy AJ, Angui a D. Condi ion-based main enance o na al p opulsion sys ems wi h supe ised da a analysis. Ocean Eng 2018;149:268–78.
[41] Yang X, Yan R, Gao RX. Induc ion mo o aul diagnosis using mul iple class ea u e selec ion. In: 2015 IEEE in e na ional ins umen a ion and measu emen echnology
con e ence (i2MTC) p oceedings. IEEE; 2015, p. 256–60.
[42] Bi kl C. Diagnosis and p ognosis o deg ada ion in li hium-ion ba e ies (Ph.D. hesis), Uni e si y o Ox o d; 2017.
[43] Li K, Ping X, Wang H, Chen P, Cao Y. Sequen ial uzzy diagnosis me hod o mo o olle bea ing in a iable ope a ing condi ions based on ib a ion analysis. Senso s
2013;13(6):8013–41.
[44] Ba bie i G, Sanchez-Londoño D, Ca aneo L, Fumagalli L, Rome o D. A case s udy o p oblem-based lea ning educa ion in aul diagnosis assessmen . IFAC-Pape sOnLine
2020;53(3):107–12.
[45] Se e son KA, A ia PM, Jin N, Pe kins N, Jiang B, Yang Z, e al. Da a-d i en p edic ion o ba e y cycle li e be o e capaci y deg ada ion. Na Ene gy 2019;4(5):383–91.
[46] So o-Ocampo CR, Me a JM, Cano-Mo eno JD, Ga cia-Be na do JL. Low-cos , high- equency, da a acquisi ion sys em o condi ion moni o ing o o a ing machine y h ough
ib a ion analysis-case s udy. Senso s 2020;20(12):3493.
[47] Ko alenko I, Saez M, Ba on K, Tilbu y D. SMART: A sys em-le el manu ac u ing and au oma ion esea ch es bed. Sma Sus ain Manu Sys 2017;1(1).
[48] Mao L, Jackson L. IEEE 2014 da a challenge da a. 2014, h p://dx.doi.o g/10.17028/ d.lbo o.3518141. 1, (Accessed 3 Janua y 2025).
[49] Von Bi gelen A, Bu a i D, Mage J, Niggemann O. Sel -o ganizing maps o anomaly localiza ion and p edic i e main enance in cybe -physical p oduc ion sys ems. P ocedia
Ci p 2018;72:480–5.
[50] UCI machine lea ning eposi o y — a chi e.ics.uci.edu. 2024, h ps://a chi e.ics.uci.edu/da ase /421/aps+ ailu e+a +scania+ ucks. (Accessed 3 Janua y 2025).
[51] Da ban ZZ, Webb GI, Pan S, Agga wal CC, Salehi M. Deep lea ning o ime se ies anomaly de ec ion: A su ey. 2022, a Xi p ep in a Xi :2211.05244.
[52] Helwig N, Pignanelli E, Schü ze A. Condi ion moni o ing o a complex hyd aulic sys em using mul i a ia e s a is ics. In: 2015 IEEE in e na ional ins umen a ion and
measu emen echnology con e ence (i2MTC) p oceedings. IEEE; 2015, p. 210–5.
[53] Ba e y da a | cen e o ad anced li e cycle enginee ing — calce.umd.edu. 2024, h ps://calce.umd.edu/ba e y-da a. (Accessed 3 Janua y 2025).
[54] Lee J, Qiu H, Yu G, Lin J, e al. IMS,‘‘Uni e si y o Cincinna i.‘‘bea ing da a se ’’, NASA ames p ognos ics da a eposi o y. 2007, NASA Ames Resea ch Cen e , Mo e Field,
CA.
