In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 1, pp.01-09, July 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i1.1 1
AI-based P edic i e Main enance in
Mechanical Sys ems
1Sayyed Nagulmee a, 2G Rajesh, 3Bandi Rajasekha ,
4Shaik Hedaya h Basha
1Assis an P o esso , Depa men o CSE, DVR & D . HS MIC College o Technology, Vijayawada, India.
syednagulmee [email protected]
2Assis an P o esso , Depa men o CSE, NBKRIST, Vidyanaga , India. ajeshg@nbk is .o g
3HOD, Depa men o CSE, S ee Venka eswa a College o Enginee ing, Nello e, India. sekha [email protected]
4Associa e P o esso , Depa men o ECE, R.M.K. College o Enginee ing and Technology, Chennai, India.
shaikhedaya hbasha@ mkce .ac.in
Abs ac : P edic i e main enance has eme ged as a c i ical solu ion o minimizing unplanned down ime, ex ending equipmen li espan, and
enhancing ope a ional e iciency in mechanical sys ems. Recen ad ancemen s in a i icial in elligence (AI) ha e enabled he de elopmen o
in elligen main enance s a egies ha le e age machine lea ning (ML), deep lea ning (DL), and hyb id algo i hms o an icipa e equipmen
ailu es wi h high accu acy. While nume ous AI-d i en p edic i e main enance solu ions ha e been p oposed, mos a e applica ion-speci ic,
esul ing in agmen ed me hodologies wi h limi ed ans e abili y ac oss domains. This pape p oposes a uni ied concep ual amewo k o AI-
based p edic i e main enance ailo ed o mechanical sys ems. D awing insigh s om di e se sec o s—including HVAC, gas u bines,
pho o ol aic sys ems, manu ac u ing, and unneling in as uc u e— he amewo k in eg a es essen ial laye s: da a acquisi ion, p ep ocessing,
modeling, decision-making, and ac ion. The p oposed model emphasizes modula i y, scalabili y, and adap abili y, and suppo s in eg a ion wi h
eal- ime da a sou ces, including edge compu ing pla o ms. This pape aims o consolida e cu en ad ancemen s, add ess c oss-domain
limi a ions, and o e a eusable amewo k ha can guide he implemen a ion o p edic i e main enance ac oss a a ie y o mechanical
en i onmen s.
Keywo ds: AI-based p edic ion, HVAC, Gas Tu bines, Mechanical Sys ems, Tunneling.
1 INTRODUCTION
Mechanical sys ems a e a he co e o c i ical in as uc u e in indus ies anging om ene gy and anspo a ion o
manu ac u ing and cons uc ion. Ensu ing hei op imal pe o mance and eliabili y equi es p oac i e s a egies o de ec aul s
and schedule main enance be o e ailu es occu . T adi ional main enance app oaches— eac i e and p e en i e—a e o en
insu icien , leading o cos ly down ime and educed equipmen li espan. P edic i e main enance (PdM), powe ed by a i icial
in elligence (AI), o e s a mo e in elligen al e na i e by le e aging ope a ional da a o an icipa e ailu es and ecommend imely
in e en ions.
Wi h he ise o Indus y 4.0, AI-based PdM has gained momen um h ough he in eg a ion o senso s, he In e ne o Things
(IoT), and ad anced analy ics. Machine lea ning (ML) models such as Suppo Vec o Machines (SVM), A i icial Neu al
Ne wo ks (ANN), and Ex eme G adien Boos ing (XGBoos ) ha e demons a ed hei u ili y in p edic ing ene gy usage and
es ima ing main enance imelines in HVAC sys ems [1]. Simila ly, deep lea ning a chi ec u es like Long Sho -Te m Memo y
(LSTM) ne wo ks ha e been employed in u bine heal h moni o ing and pho o ol aic sys em anomaly de ec ion [2], [3].
Despi e he p o en bene i s o AI-d i en PdM, cu en esea ch emains siloed wi hin domain-speci ic solu ions. Fo ins ance, gas
u bines [2], manu ac u ing sys ems [4], and unnel bo ing machines [5] ha e all seen unique implemen a ions, o en wi h dis inc
algo i hms, inpu ea u es, and deploymen s a egies. This agmen a ion limi s he scalabili y and in e ope abili y o PdM
solu ions ac oss indus ies.
The absence o a uni ying amewo k hampe s knowledge ans e inc eases deploymen cos s, and c ea es ba ie s o
o ganiza ions a emp ing o adop AI-based main enance a scale. As such, he e is a compelling need o a gene alizable, modula ,
and scalable concep ual amewo k ha can guide he de elopmen o AI-based PdM sys ems o di e se mechanical sys ems.
Such a amewo k should accommoda e a ying senso da a s eams, suppo bo h cen alized and edge compu ing, and in eg a e
seamlessly wi h exis ing main enance wo k lows. This pape add esses his gap by p oposing a concep ual amewo k o AI-
based p edic i e main enance in mechanical sys ems. The amewo k is syn hesized om an ex ensi e li e a u e su ey co e ing
ele en s a e-o - he-a s udies ac oss mul iple indus ial domains. I is designed o s anda dize he key componen s o PdM—da a
acquisi ion, p ep ocessing, model aining, decision-making, and ac ion planning—while emaining adap able o indus y-speci ic
cons ain s.
In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 1, pp.01-09, July 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i1.1 2
The emainde o he pape is o ganized as ollows: Sec ion 2 p esen s a ca ego ized li e a u e e iew; Sec ion 3 iden i ies key
gaps and limi a ions in cu en sys ems; Sec ion 4 de ails he p oposed concep ual amewo k; Sec ion 5 maps eal-wo ld case
s udies o he amewo k; Sec ion 6 discusses i s bene i s and challenges; Sec ion 7 ou lines u u e esea ch di ec ions; and Sec ion
8 concludes he pape .
2 LITERATURE REVIEW
This sec ion su eys ecen ad ances in AI-based p edic i e main enance, ca ego ized by applica ion domains, AI echniques
employed, and no able implemen a ion s a egies. The e iew spans key indus ial sec o s including HVAC, gas u bines,
pho o ol aic sys ems, manu ac u ing, unnel in as uc u e, and au onomous ehicles. These case s udies p o ide c i ical insigh s
in o he capabili ies and limi a ions o exis ing models, o ming he ounda ion o he p oposed uni ied amewo k.
2.1 Applica ion Domains
2.1.1 HVAC and Ene gy Sys ems
Amin e al. [1] employed a da a-d i en app oach o p edic ene gy consump ion and plan main enance ac i i ies o Ac i e
Chilled Beam (ACB) sys ems in o ice buildings. Using a da ase o 2,500 samples, machine lea ning models such as Gaussian
P ocess Reg ession (GPR) and XGBoos we e used o accu a ely p edic bo h elec ici y consump ion and main enance ime. Thei
wo k demons a es he dual bene i o AI models in op imizing ene gy usage and p e en i e main enance scheduling.
2.1.2 Gas Tu bines
In [2], B ahimi e al. de eloped an in elligen moni o ing sys em o MS5002C gas u bines using Adap i e Neu o-Fuzzy
In e ence Sys ems (ANFIS) and Long Sho -Te m Memo y (LSTM) ne wo ks. T ained on decades o his o ical da a (1985–2021),
hei model e ec i ely p edic ed componen deg ada ion and enabled eliabili y-cen e ed main enance planning.
2.1.3 Pho o ol aic Sys ems
Syamsuddin e al. [3] p oposed a p edic i e main enance app oach o sola powe plan s based on anomaly de ec ion using
SCADA da a. A Long Sho -Te m Memo y Au oencode (LSTM-AE) model was ained o econs uc no mal sys em beha io ,
wi h de ia ions lagged as po en ial aul s. The model achie ed high accu acy in iden i ying anomalies and demons a ed he
e ec i eness o unsupe ised lea ning in eal-wo ld ene gy sys ems.
2.1.4 Manu ac u ing Sys ems
Ma i-Puig e al. [4] explo ed p edic i e main enance in a wooden piece manu ac u ing se ing. They p edic ed mo o
empe a u e in he ac o y's ex ac ion sys em using Ex eme Lea ning Machines, le e aging IoT senso da a and no el da a-
cleaning echniques. Thei app oach emphasized as aining and implemen a ion eadiness, sui able o dynamic indus ial
en i onmen s. A iushenko e al. [6] ad anced his u he by implemen ing a esou ce-e icien Edge AI sys em o ool condi ion
moni o ing (TCM) du ing milling p ocesses. The sys em execu ed mul iple ML and DL models on low-powe edge de ices,
demons a ing he easibili y o eal- ime PdM deploymen in cons ained en i onmen s.
2.1.5 Tunnel and In as uc u e Main enance
Zou e al. [5] add essed p edic i e main enance o Tunnel Bo ing Machines (TBMs) using an a en ion-based G aph
Con olu ional Ne wo k (a -GCN) o p edic machine wea and pe o mance. Thei model achie ed supe io p edic i e accu acy
by in eg a ing geo echnical and ope a ional da a, wi h an online lea ning a ian imp o ing pe o mance du ing eal- ime
ope a ions. Lu e al. [7] e alua ed he mechanical pe o mance o shield unnel s uc u es pos - i e using a Backp opaga ion (BP)
neu al ne wo k. The model co ela ed a ious s uc u al damage indices o p edic pos - i e unnel capaci y, aiding in eme gency
main enance decision-making.
2.1.6 Au onomous Sys ems
Aeddula e al. [8] ocused on au onomous haulage ehicles and in eg a ed AI-based p edic i e main enance wi hin P oduc -
Se ice Sys em (PSS) de elopmen . Using a ia ional au oencode s, he sys em o ecas ed bo h expec ed and unexpec ed ailu es,
accoun ing o complex in e dependencies in ehicle componen s and enabling p oac i e planning du ing ea ly PSS phases.
In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 1, pp.01-09, July 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i1.1 3
2.2 AI Techniques in P edic i e Main enance
A di e se ange o AI echniques has been adop ed ac oss domains:
• Supe ised Lea ning: SVM, GPR, and XGBoos models we e used o p edic ing main enance ime and ene gy
consump ion [1], [3].
• Deep Lea ning: LSTM [2], LSTM-AE [3], BP Neu al Ne wo ks [7], and a en ion-based GCNs [5] enabled dynamic
sequence modeling and mul i-ou pu p edic ions.
• Hyb id App oaches: ANFIS-LSTM [2] and a -GCN [5] illus a e he g owing end o combining mul iple
a chi ec u es o imp o ed accu acy and obus ness.
• Unsupe ised Lea ning: Au oencode s [3], clus e ing echniques [7], and online lea ning [5] allowed anomaly
de ec ion in sys ems lacking labeled ailu e da a.
• Edge AI: As shown in [6], unning models on embedded de ices nea he da a sou ce o e s educed la ency,
enhanced p i acy, and be e scalabili y.
2.3 Obse a ions and T ends
F om he e iewed s udies, se e al key obse a ions eme ge:
• Domain-Speci ic Implemen a ions: Mos s udies ocus on highly speci ic use cases, leading o non- ans e able
models.
• Da a-D i en Emphasis: SCADA, IoT, and senso da a a e cen al o almos all implemen a ions, wi h p ep ocessing
playing a c i ical ole.
• E alua ion Me ics: Common me ics include RMSE, MAE, R², and MAPE, enabling compa a i e pe o mance
e alua ion.
• Pla o m Di e si y: Implemen a ion ools ange om MATLAB [1] o open-sou ce edge de ices [6], indica ing
lexibili y bu also a iabili y in s anda diza ion.
These ends highligh he inno a i e bu agmen ed na u e o cu en esea ch, jus i ying he need o a uni ied concep ual
amewo k.
3 GAPS IN EXISTING SYSTEMS
Despi e no able p og ess in AI-based p edic i e main enance ac oss a ious mechanical sys ems, cu en implemen a ions o en
emain agmen ed, domain-speci ic, and limi ed in scalabili y. The ollowing key limi a ions a e e iden om he li e a u e:
3.1 Lack o Gene alizabili y Ac oss Domains
Mos exis ing p edic i e main enance solu ions a e de eloped o speci ic indus ial con ex s, wi h limi ed conside a ion o
c oss-domain applicabili y. Fo ins ance, models ailo ed o gas u bines [2] o HVAC sys ems [1] a e no easily ans e able o
manu ac u ing lines [4], pho o ol aic sys ems [3], o unnel in as uc u e [5]. Each implemen a ion ypically employs cus omized
da ase s, ea u e se s, and lea ning s a egies, hinde ing model euse o adap a ion. This siloed app oach inc eases de elopmen
cos s and delays b oade adop ion. A gene alizable amewo k would acili a e eusabili y o co e componen s (e.g., p ep ocessing,
modeling laye s), educing he ime and expe ise equi ed o deploying AI-based PdM sys ems in new domains.
3.2 Challenges in Da a Quali y and A ailabili y
AI models hea ily depend on he a ailabili y and quali y o ope a ional da a. Howe e , many indus ial en i onmen s su e
om issues such as:
• Senso blockages and noise (e.g., add essed h ough p ep ocessing in [4])
• Lack o labeled ailu e da a, especially in a e-e en sys ems like pho o ol aic a ms [3]
• He e ogeneous da a o ma s om SCADA sys ems, IoT de ices, and legacy equipmen
In [3], unsupe ised lea ning ia LSTM-AE was used as a wo ka ound o unlabeled da a, bu he absence o s anda d da a
cleaning and labeling p ocedu es emains a bo leneck. Mo eo e , high dependency on domain-speci ic da a u he educes model
in e ope abili y.
In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 1, pp.01-09, July 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i1.1 4
3.3 Limi ed Adop ion o Real-Time and Edge Deploymen
Only a ew s udies, such as [6], explo ed Edge AI— he deploymen o models on ligh weigh , low-powe ha dwa e nea he
equipmen . Mos p edic i e models s ill ely on cen alized compu a ion, which may in oduce la ency, inc ease communica ion
o e head, and aise da a p i acy conce ns. Real- ime decision-making is essen ial in mechanical sys ems like mo o s o
au onomous ehicles [8], ye many cu en models a e ained o line and lack online lea ning o in e ence capabili ies. The
adop ion o edge-compu ing in as uc u e and eal- ime adap i e lea ning models emains unde de eloped.
3.4 F agmen ed Model In eg a ion wi h Main enance Wo k low
Many AI models ocus on p edic ing ailu es bu all sho in ac ionable in eg a ion wi h exis ing main enance managemen
sys ems (e.g., CMMS). Fo ins ance, while [1] and [3] p edic ed main enance ime and anomalies espec i ely, he e is li le
discussion on how hese p edic ions eed in o ac ual main enance planning, scheduling, o esou ce alloca ion.
Addi ionally, use in e ace design, in e p e abili y o model ou pu s, and collabo a ion wi h human decision-make s a e la gely
o e looked. Wi hou in eg a ion in o b oade decision suppo sys ems, AI ou pu s may be unde u ilized.
3.5 Limi ed Model Explainabili y and T us
Indus ial adop ion o AI models equi es anspa ency and explainabili y, especially in sa e y-c i ical sys ems. Howe e , deep
lea ning models such as LSTM [2], [3] o a -GCN [5] o en unc ion as “black boxes.” Few s udies a emp o explain p edic ions,
quan i y unce ain y, o in ol e domain expe s in he modeling loop.
A lack o explainabili y unde mines us , making s akeholde s hesi an o ely on au oma ed decisions. Fu u e amewo ks mus
inco po a e in e p e able AI me hods and in eg a e isual analy ics o explana ion modules o b idge his gap.
3.6 Inconsis en E alua ion and Benchma king
The e is no consis en benchma king amewo k ac oss p edic i e main enance s udies. Models a e e alua ed using di e en
me ics—RMSE, MAE, MAPE, R²—on a ying da ase s. Fo example, [5] used MAPE o unnel sys ems, [1] used RMSE and
R² o HVAC p edic ions, and [3] used MAE o sola a m anomalies. This lack o s anda diza ion hinde s ai compa isons and
makes i di icul o iden i y uni e sally e ec i e app oaches. A concep ual amewo k mus p omo e e alua ion consis ency and
de ine baseline me ics o di e se applica ion a eas.
4 PROPOSED CONCEPTUAL FRAMEWORK
The inc easing complexi y o mechanical sys ems, along wi h he as amoun s o da a gene a ed by senso s and con ol
sys ems, necessi a es a sys ema ic and adap able amewo k o AI-based p edic i e main enance (PdM). Based on insigh s om
he li e a u e e iewed, we p opose a mul i-laye ed concep ual amewo k ha in eg a es da a collec ion, p ep ocessing, modeling,
decision-making, and deploymen in o a cohesi e s uc u e. This modula amewo k is designed o be adap able ac oss domains,
ensu ing ha AI-d i en main enance solu ions can be scaled and op imized o mee he speci ic equi emen s o a ious mechanical
sys ems.
The i s laye o he amewo k is da a acquisi ion, which se es as he ounda ional inpu laye . I encompasses bo h senso -
based eal- ime moni o ing sys ems and his o ical da a logs. This laye suppo s a wide a ie y o da a ypes such as ib a ion,
empe a u e, p essu e, humidi y, acous ic signals, sys em logs, main enance his o y, and ope a ing condi ions. A obus da a
acquisi ion sys em ensu es consis en , high- esolu ion inpu s ha a e c i ical o de ec ing sub le anomalies and long- e m
deg ada ion pa e ns. In eg a ion wi h Supe iso y Con ol and Da a Acquisi ion (SCADA) sys ems and In e ne o Things (IoT)
pla o ms enables con inuous s eaming and a chi ing o mechanical pe o mance me ics.
Once da a is collec ed, i is p ocessed h ough he da a p ep ocessing and ea u e enginee ing laye . This componen includes
noise il e ing, missing da a impu a ion, no maliza ion, and synch oniza ion ac oss mul iple senso s o sys ems. Equally impo an
is he ex ac ion o meaning ul ea u es, such as s a is ical indica o s (mean, RMS, ku osis), spec al cha ac e is ics (FFT, wa ele
ans o ms), and domain-speci ic pa ame e s. In some applica ions, ad anced me hods like P incipal Componen Analysis (PCA)
o au oencode s a e used o educe dimensionali y while p ese ing c i ical in o ma ion. Clean, well-s uc u ed da a enhances
model pe o mance and in e p e abili y, especially when used in eal- ime in e ence.
The modeling laye is a he co e o he amewo k. I suppo s bo h supe ised and unsupe ised lea ning echniques, chosen
based on he a ailabili y o labeled ailu e da a. Supe ised models such as Suppo Vec o Machines (SVM), Random Fo es s
(RF), G adien Boos ing Machines (GBM), and Deep Neu al Ne wo ks (DNN) a e sui able o applica ions wi h well-anno a ed
da ase s, while unsupe ised me hods such as clus e ing algo i hms o au oencode s a e p e e ed when ailu e labels a e
una ailable. Time-se ies models like LSTM and a en ion-based mechanisms can cap u e empo al dependencies in deg ada ion.
In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 1, pp.01-09, July 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i1.1 5
In some cases, hyb id models—such as ANFIS o ensemble me hods—a e employed o combine p edic i e accu acy wi h
in e p e abili y. Model selec ion is guided by e alua ion me ics such as Mean Squa ed E o (MSE), Roo Mean Squa ed E o
(RMSE), Mean Absolu e E o (MAE), R-squa ed, o classi ica ion accu acy. Following he modeling s age, he decision and
ac ion laye in e p e s model ou pu s o gene a e ac ionable main enance insigh s. This may include p edic ing he Remaining
Use ul Li e (RUL), lagging po en ial aul s, o classi ying equipmen condi ion in o heal h s a es. Th esholds can be p ede ined
o dynamically adjus ed using anomaly de ec ion echniques o isk-based p io i iza ion. Decision ules a e o en in eg a ed in o
main enance managemen sys ems (CMMS), ale dashboa ds, o isual analy ics in e aces o ensu e imely and da a-d i en
in e en ions. This laye also suppo s anking o main enance p io i ies, spa e pa s planning, and scheduling o in e en ions o
minimize down ime.
A he op o he a chi ec u e is he deploymen and in eg a ion laye , which add esses he sys em-le el implemen a ion o he
p edic i e main enance pipeline. Models can be deployed in cen alized cloud en i onmen s, on local se e s, o on edge de ices
depending on la ency, p i acy, and scalabili y needs. This laye includes API-based model se ing, in eg a ion wi h ERP/SCADA
pla o ms, eal- ime dashboa ds, and human-machine in e aces (HMI). In edge compu ing scena ios, ligh weigh models a e
op imized o pe o mance-cons ained de ices o suppo in-si u decision-making. The deploymen a chi ec u e mus also suppo
model e sion con ol, eedback loops, and e aining pipelines o ensu e long- e m adap abili y.
A c oss-cu ing heme o he amewo k is explainabili y and human-in- he-loop in e ac ion. Fo AI-based decisions o be
ac ionable in indus ial en i onmen s, use s mus unde s and he a ionale behind p edic ions. Visualiza ion ools, SHAP alues,
ea u e impo ance ankings, and in e p e able model s uc u es (e.g., decision ees o uzzy sys ems) enable human ope a o s o
us and alida e he sys em. Mo eo e , inco po a ing domain expe ise in model aining and eedback loops imp o es accu acy
and adop ion.
The p oposed amewo k is designed o be lexible, scalable, and explainable, o e ing a uni ied e e ence a chi ec u e ha
can guide he de elopmen and implemen a ion o AI-based p edic i e main enance sys ems in a a ie y o mechanical se ings.
I s s uc u e suppo s bo h academic esea ch and indus ial deploymen , b idging he gap be ween algo i hmic inno a ion and
eal-wo ld main enance challenges.
Fig. 1. Wo k low o he p oposed me hod
In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 1, pp.01-09, July 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i1.1 6
The p oposed amewo k, as shown in Fig. 1, begins wi h da a acquisi ion, he ounda ion o any p edic i e main enance
sys em. This includes eal- ime senso da a om equipmen , his o ical logs om SCADA sys ems, and CMMS da a such as
main enance his o y and ope a o epo s. This laye ensu es ha bo h eal- ime and con ex ual da a a e a ailable o in elligen
decision-making. Nex , da a p ep ocessing and ea u e enginee ing is ca ied ou o clean he da a, handle missing alues, and
ex ac ele an ea u es. This s age is c ucial o imp o ing he signal- o-noise a io in da ase s and enhancing model pe o mance.
Fea u e ex ac ion me hods such as Fou ie ans o ms (FFT), wa ele analysis, and PCA a e pa icula ly use ul in mechanical
sys ems o de ec ing deg ada ion ends.
The AI modeling laye suppo s a ious machine lea ning and deep lea ning algo i hms. Supe ised models like Random
Fo es s and Deep Neu al Ne wo ks a e used when labeled da a is a ailable, while unsupe ised models such as au oencode s o
clus e ing echniques a e deployed o anomaly de ec ion in unlabeled da ase s. LSTM and g aph-based a en ion ne wo ks like
a -GCN a e le e aged o ime-se ies modeling in complex sys ems wi h empo al dependencies. Hyb id app oaches (e.g., ANFIS)
combine he s eng hs o mul iple algo i hms o obus p edic ions.
The decision and ac ion laye ans o ms model ou pu s in o ac ionable insigh s. This could in ol e p edic ing he emaining
use ul li e (RUL), de ec ing anomalies, o ecommending speci ic main enance ac ions. These decisions a e logged, isualized, o
sen o main enance eams h ough in eg a ed CMMS pla o ms. In he deploymen and in eg a ion laye , he ained models a e
deployed o he cloud, local se e s, o edge de ices depending on sys em equi emen s. This laye also in eg a es model ou pu s
in o en e p ise pla o ms such as ERP, SCADA, and HMI dashboa ds, enabling seamless low o insigh s in o ope a ional
wo k lows. Explainabili y and human-in- he-loop in e ac ion ensu es ha p edic ions a e in e p e able and usable by ield
enginee s and decision-make s. Visual ools (e.g., SHAP, LIME), ea u e impo ance sco es, and in ui i e dashboa ds help build
us and acili a e ac ion. Ope a o eedback is inco po a ed o imp o e model lea ning and sys em pe o mance o e ime.
5 MAPPING THE FRAMEWORK TO REAL-WORLD CASES
To alida e he comp ehensi eness and lexibili y o he p oposed concep ual amewo k, i is essen ial o examine how i s
componen s align wi h eal-wo ld implemen a ions o AI-based p edic i e main enance ac oss di e en mechanical sys ems. The
ollowing discussion maps he key laye s o he amewo k o selec ed s udies om a ious domains, he eby demons a ing i s
applicabili y and gene alizabili y.
In he con ex o HVAC sys ems, he s udy by Amin e al. [1] exempli ies he ull ealiza ion o he da a acquisi ion, modeling,
and decision laye s. Thei wo k u ilized a s uc u ed da ase comp ising o e 2,500 samples ep esen ing building and sys em
cha ac e is ics, aligning di ec ly wi h he amewo k's da a acquisi ion laye . The p ep ocessing s age in ol ed he ca e ul selec ion
o inpu ea u es, including en i onmen al and his o ical main enance a iables. Machine lea ning algo i hms such as Gaussian
P ocess Reg ession (GPR) and XGBoos we e applied o p edic bo h ene gy consump ion and main enance ime, he eby
popula ing he modeling laye o he amewo k. The ou pu s—accu a e p edic ions o consump ion and se ice in e als—enabled
da a-d i en main enance planning, eeding di ec ly in o he decision and ac ion laye . Thei use o MATLAB and Ene gyPlus
demons a es pla o m-le el lexibili y, which i s well wi hin he in eg a ion expec a ions o he p oposed a chi ec u e.
Simila ly, B ahimi e al. [2] o e a s ong example om he ene gy sec o , whe e ANFIS and LSTM models we e deployed o
moni o and an icipa e aul s in gas u bines. The modeling laye is en iched by he combina ion o neu o- uzzy logic wi h deep
lea ning, showcasing he hyb id app oach suppo ed by he amewo k. The models used a as his o ical da ase , ul illing he
amewo k's equi emen o obus inpu om he acquisi ion laye . Though less emphasis is placed on eal- ime eedback, he
sys em e ec i ely in o ms long- e m main enance schedules, suppo ing he decision-making componen . Thei eliabili y analysis
u he con ibu es o in e p e abili y, aligning pa ially wi h he amewo k’s ision o explainable and ac ionable AI.
In he manu ac u ing sec o , Ma i-Puig e al. [4] and A iushenko e al. [6] e lec how p edic i e main enance can be localized
h ough edge deploymen . Ma i-Puig e al. ocused on p edic ing he ope a ional condi ion o mo o s used in wood esidue
ex ac ion sys ems. Thei p ep ocessing laye included an inno a i e algo i hm o emo e senso -blockage a i ac s, di ec ly
suppo ing he amewo k’s emphasis on da a quali y. Meanwhile, A iushenko e al. buil an Edge AI p o o ype ha implemen ed
se e al machine lea ning and deep lea ning models on a low-powe de ice o enable ool condi ion moni o ing du ing milling.
This s ongly aligns wi h he p oposed amewo k’s p o ision o eal- ime deploymen and edge in eg a ion, indica ing he
modula and adap i e na u e o he a chi ec u e.
Fo pho o ol aic sys ems, Syamsuddin e al. [3] employed LSTM au oencode s o de ec anomalies in sola ene gy gene a ion
da a. Thei wo k ep esen s a scena io whe e labeled ailu e da a is una ailable, equi ing he amewo k o suppo unsupe ised
lea ning s a egies. The model econs uc s no mal pa e ns and lags de ia ions, which is consis en wi h he modeling and
decision laye s o he amewo k. Al hough he ac ion laye is no deeply in eg a ed in hei s udy, he de ec ion o ope a ional
anomalies is i sel a i al inpu o p e en i e s a egies, hus alida ing he sys em’s co e unc ionali y.
In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 1, pp.01-09, July 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i1.1 7
The case s udy p esen ed by Zou e al. [5] on unnel bo ing machines (TBMs) illus a es how complex mechanical sys ems
ope a ing in unce ain en i onmen s can bene i om ad anced AI a chi ec u es such as a en ion-based g aph con olu ional
ne wo ks (a -GCN). Thei model p ocesses bo h ope a ional and geo echnical da a o p edic equipmen deg ada ion and adap s
in eal ime h ough online lea ning. This implemen a ion con i ms he ele ance o he amewo k’s lea ning and decision laye s
in high-s akes, dynamic scena ios. The addi ion o an online a ian u he suppo s he amewo k’s capabili y o con inuous
lea ning and pe o mance imp o emen based on eal- ime da a.
In au onomous sys ems, Aeddula e al. [8] demons a ed he use o a ia ional au oencode s o p edic main enance needs o
au onomous haulage ehicles. The s udy’s s eng h lies in i s ocus on in eg a ing p edic i e main enance du ing he ea ly phases
o p oduc -se ice sys em (PSS) de elopmen . This aligns closely wi h he amewo k’s op-le el in eg a ion laye , whe e AI
insigh s eed in o b oade s a egic planning ools. Thei model’s abili y o iden i y hidden ailu e modes and enable ea ly
in e en ion suppo s he amewo k’s goal o embedding in elligence ac oss he li ecycle o mechanical asse s.
Lu e al. [7] de eloped a backp opaga ion neu al ne wo k o assess pos - i e mechanical pe o mance in shield unnels. The
s udy elied on clus e ing and nume ical simula ions o de i e damage indices, which we e hen mapped o pe o mance g ades.
This showcases he amewo k’s capabili y o suppo mul i-laye ed decision-making, whe e physical simula ions and AI
p edic ions wo k oge he o enable c i ical in as uc u e main enance.
Each o hese eal-wo ld cases e lec s elemen s o he p oposed amewo k, whe he in da a acquisi ion, ad anced modeling,
eal- ime in e ence, o in eg a ion in o ope a ional wo k lows. While mos s udies ocus on one o wo componen s in dep h, he
p oposed amewo k o e s a uni ied s uc u e ha encapsula es all hese elemen s, demons a ing i s po en ial o se e as a
e e ence a chi ec u e o scalable and in elligen p edic i e main enance in di e se mechanical sys ems.
6 DISCUSSION
The p oposed concep ual amewo k o e s a uni ied, modula solu ion o implemen ing AI-based p edic i e main enance
(PdM) ac oss di e se mechanical sys ems. By syn hesizing insigh s om domain-speci ic applica ions, he amewo k p omo es
c oss-sec o applicabili y and encou ages a s uc u ed app oach o deploying in elligen main enance s a egies. This sec ion
discusses he b oade signi icance o he amewo k, i s po en ial ad an ages, implemen a ion challenges, and oppo uni ies o
u he enhancemen .
One o he p ima y bene i s o he p oposed amewo k is i s modula i y, which allows di e en indus ial sec o s o adap he
a chi ec u e o hei speci ic needs wi hou eenginee ing he en i e sys em. Fo example, a manu ac u ing plan and a pho o ol aic
powe s a ion may di e as ly in e ms o ope a ional dynamics and da a o ma s, ye bo h can ollow he same a chi ec u al
bluep in , using in e changeable componen s such as da a p ep ocessing pipelines, modeling blocks, o isualiza ion in e aces.
This modula i y no only acili a es as e de elopmen and deploymen bu also suppo s he con inuous e olu ion o each
componen based on echnological ad ancemen s o domain equi emen s.
Ano he key s eng h o he amewo k is i s suppo o eal- ime analy ics and Edge AI deploymen . T adi ional p edic i e
main enance app oaches o en ely on o line models ha analyze his o ical da a, esul ing in delayed insigh s and eac i e
decisions. The in eg a ion o edge compu ing capabili ies, as seen in s udies like [6], enables local in e ence wi h educed la ency,
imp o ed da a p i acy, and dec eased ne wo k bandwid h consump ion. This is especially aluable in emo e o in as uc u e-
limi ed en i onmen s, whe e eal- ime aul de ec ion and on-si e esponse a e c i ical [9]-[11].
The amewo k also emphasizes he impo ance o explainabili y and in eg a ion— wo aspec s equen ly o e looked in
exis ing sys ems. While highly accu a e models such as deep neu al ne wo ks and au oencode s o e powe ul p edic ions, hei
opaque na u e can limi us and accep ance among ield enginee s and decision-make s. By including a dedica ed laye o human
in e ac ion and explainable AI, he amewo k acknowledges he socio- echnical eali y o indus ial en i onmen s, whe e AI mus
complemen human expe ise a he han eplace i . Fu he mo e, seamless in eg a ion wi h en e p ise sys ems such as
Compu e ized Main enance Managemen Sys ems (CMMS) o En e p ise Resou ce Planning (ERP) pla o ms ensu es ha model
ou pu s di ec ly in luence ope a ional decisions, he eby b idging he gap be ween insigh gene a ion and ac ionable main enance
[12].
Despi e i s s eng hs, he amewo k also p esen s se e al implemen a ion challenges. One o he mos p ominen is he
equi emen o high-quali y, con inuous da a s eams, which a e o en lacking in adi ional o legacy equipmen se ups. While
mode n indus ial acili ies inc easingly adop IoT-enabled senso s and SCADA sys ems, many mechanical sys ems s ill ope a e
wi hou su icien ins umen a ion. Re o i ing such sys ems wi h app op ia e da a acquisi ion capabili ies in ol es conside able
in es men and e o [13].
In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 1, pp.01-09, July 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i1.1 8
Ano he challenge is he model managemen and li ecycle main enance. AI models equi e con inuous moni o ing, e aining,
and alida ion o ensu e accu acy o e ime. Changes in sys em beha io , equipmen wea , o ope a ional condi ions may in oduce
concep d i , which can deg ade model pe o mance. O ganiza ions mus es ablish clea p o ocols o e sion con ol, pe iodic
e alua ion, and e aining s a egies—p e e ably au oma ed— o main ain he eliabili y o hei p edic i e models.
Mo eo e , while he amewo k ad oca es o lexibili y in algo i hm selec ion and deploymen pla o ms, his e y lexibili y
can become a sou ce o complexi y. O ganiza ions may ace decision a igue in choosing he igh combina ion o algo i hms,
pla o ms, and ools, especially when expe ise in AI is limi ed. To mi iga e his, domain-speci ic implemen a ion empla es o
model selec ion guidelines could be de eloped as supplemen a y esou ces o he amewo k.
Cybe secu i y and da a go e nance a e c i ical conce ns, pa icula ly in en i onmen s ha in ol e emo e moni o ing and
cloud-based in e ence. As da a becomes inc easingly cen al o main enance decision-making, ensu ing i s secu i y, in eg i y, and
compliance wi h egula o y s anda ds is essen ial. The amewo k, he e o e, mus be implemen ed alongside obus da a p o ec ion
policies and secu e communica ion p o ocols.
The p oposed amewo k o e s a p omising ounda ion o gene alizing AI-based p edic i e main enance ac oss mechanical
sys ems. I b ings s uc u e o a cu en ly agmen ed ield, p omo es scalable and explainable deploymen , and encou ages
in eg a ion wi h en e p ise-le el decision p ocesses. A he same ime, success ul implemen a ion demands ca e ul conside a ion
o da a quali y, model main enance, use us , and sys em secu i y. Add essing hese challenges p oac i ely will be key o ealizing
he ull po en ial o p edic i e main enance in he e a o Indus y 4.0 and beyond.
7 CONCLUSIONS
This pape p esen ed a concep ual amewo k o AI-based p edic i e main enance (PdM) in mechanical sys ems, d awing on
a comp ehensi e e iew o ecen li e a u e ac oss di e se indus ial domains. While nume ous s udies ha e demons a ed he
po en ial o AI o enhance main enance s a egies in speci ic con ex s—such as HVAC sys ems, gas u bines, manu ac u ing,
pho o ol aic a ms, and unnel bo ing machines—mos exis ing solu ions emain domain-speci ic, agmen ed, and di icul o
scale. Ou p oposed amewo k add esses hese limi a ions by o e ing a modula , gene al-pu pose a chi ec u e ha in eg a es da a
acquisi ion, p ep ocessing, in elligen modeling, decision-making, and sys em-le el in eg a ion.
The amewo k emphasizes lexibili y in algo i hm selec ion, suppo s eal- ime in e ence ia edge compu ing, and
inco po a es explainable AI componen s o build us among human ope a o s. I s alignmen wi h eal-wo ld applica ions alida es
i s s uc u e, while i s gene alizabili y allows o ganiza ions o adap i o a ied mechanical sys ems wi hou ex ensi e
eenginee ing. Mo eo e , he amewo k highligh s he impo ance o in eg a ing p edic i e ou pu s wi h en e p ise sys ems o
ensu e ac ionable main enance planning.
Despi e i s po en ial, he deploymen o such a amewo k poses challenges ela ed o da a quali y, model main enance, use
us , and cybe secu i y. O e coming hese challenges equi es a mul idisciplina y e o in ol ing AI expe s, domain enginee s,
IT p o essionals, and o ganiza ional s akeholde s. Fu u e wo k may ocus on implemen ing domain-speci ic ins an ia ions o he
amewo k, de eloping au oma ed model li ecycle managemen ools, and c ea ing s anda dized benchma ks o e alua ing
p edic i e main enance sys ems.
By b idging he gap be ween ad anced AI models and p ac ical indus ial needs, his concep ual amewo k p o ides a
ounda ion o de eloping in elligen , scalable, and in eg a ed p edic i e main enance sys ems. As indus ies con inue o emb ace
he p inciples o Indus y 4.0 and beyond, such amewo ks will play a c i ical ole in enhancing asse eliabili y, educing
down ime, and op imizing ope a ional e iciency.
FUNDING INFORMATION
This esea ch ecei ed no speci ic g an om any unding agency in he public, comme cial, o no - o -p o i sec o s.
ETHICS STATEMENT
This s udy did no in ol e human o animal subjec s and, he e o e, did no equi e e hical app o al.
STATEMENT OF CONFLICT OF INTERESTS
The au ho s decla e no con lic s o in e es ela ed o his s udy.
LICENSING
This wo k is licensed unde a C ea i e Commons A ibu ion 4.0 In e na ional License.
In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 1, pp.01-09, July 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i1.1 9
REFERENCES
[1] N. H. Amin, A. E emad, and A. Abdalisousan, “Da a-d i en pe o mance analysis o an ac i e chilled beam ai
condi ioning sys em: A machine lea ning app oach o ene gy e iciency and p edic i e main enance,” Resul s in
Enginee ing, ol. 23, p. 102747, Aug. 2024, doi: 10.1016/j. ineng.2024.102747.
[2] L. B ahimi, N. Had oug, A. I a ni, A. Ha ai a, and I. Colak, “Ad ancing p edic i e main enance o gas u bines: An
in elligen moni o ing app oach wi h ANFIS, LSTM, and eliabili y analysis,” Compu e s & Indus ial Enginee ing,
ol. 191, p. 110094, Ma . 2024, doi: 10.1016/j.cie.2024.110094.
[3] A. Syamsuddin, A. C. Adhi, A. Kusumawa dhani, T. P ahas o, and A. Widodo, “P edic i e main enance based on
anomaly de ec ion in pho o ol aic sys em using SCADA da a and machine lea ning,” Resul s in Enginee ing, p.
103589, Dec. 2024, doi: 10.1016/j. ineng.2024.103589.
[4] P. Ma i-Puig, I. A. Touhami, R. C. Pe a nau, and M. Se a-Se a, “Indus ial AI in condi ion-based main enance: A
case s udy in wooden piece manu ac u ing,” Compu e s & Indus ial Enginee ing, ol. 188, p. 109907, Jan. 2024,
doi: 10.1016/j.cie.2024.109907.
[5] X. Zou e al., “Ad ancing unnel equipmen main enance h ough da a-d i en p edic i e s a egies in unde g ound
in as uc u e,” Compu e s and Geo echnics, ol. 173, p. 106532, Jun. 2024, doi: 10.1016/j.compgeo.2024.106532.
[6] V. A iushenko, S. Lang, C. Le ez, T. Reggelin, and M. Hacke -Oschä zchen, “Resou ce-e icien Edge AI solu ion
o p edic i e main enance,” P ocedia Compu e Science, ol. 232, pp. 348–357, Jan. 2024, doi:
10.1016/j.p ocs.2024.01.034.
[7] Y. Lu, J. Jiang, B. Wang, W. Chen, and J. Ye, “AI-based e alua ion me hod o mechanical pe o mance o shield
unnel s uc u es a e i e,” Tunnelling and Unde g ound Space Technology, ol. 150, p. 105858, May 2024, doi:
10.1016/j. us .2024.105858.
[8] O. Aeddula, M. F ank, R. Ru ald, C. J. Askling, J. Wall, and T. La sson, “AI-D i en p edic i e main enance o
au onomous ehicles o P oduc -Se ice sys em de elopmen ,” P ocedia CIRP, ol. 128, pp. 84–89, Jan. 2024, doi:
10.1016/j.p oci .2024.06.008.
[9] C. Gnekpe, D. Tchuen e, S. Nyawa, and P. K. Dey, “Ene gy pe o mance o building e u bishmen s: P edic i e and
p esc ip i e AI-based machine lea ning app oaches,” Jou nal o Business Resea ch, ol. 183, p. 114821, Jul. 2024,
doi: 10.1016/j.jbus es.2024.114821.
[10] A. S oh , P. Ollig, R. Kelle , and A. Riege , “Gene a i e mechanisms o AI implemen a ion: A c i ical ealis
pe spec i e on p edic i e main enance,” In o ma ion and O ganiza ion, ol. 34, no. 2, p. 100503, Feb. 2024, doi:
10.1016/j.in oando g.2024.100503.
[11] P. D. Pa aschos and D. E. Koulou io is, “Lea ning-based p oduc ion, main enance, and quali y op imiza ion in sma
manu ac u ing sys ems: A li e a u e e iew and ends,” Compu e s & Indus ial Enginee ing, p. 110656, Oc . 2024,
doi: 10.1016/j.cie.2024.110656.
[12] Gusnul Mahesa, Ahmad Taj i, E lan Ananda Puji o, Muhammad Rizal, Dimas A ya Bagaska a, Mohammad Gi i
E gi ia, “In eg a ion o A i icial In elligence in Ne wo k Technology: A Li e a u e Re iew,” In e na ional Jou nal
o Eme ging Resea ch in Enginee ing, Science, and Managemen , ol. 4, no. 2, pp. 22-31, 2025. doi:
10.58482/ije esm. 4i2.4.
[13] Allu Venka a Sa yana ayana, Ye ubandi Venka a Ramana Mu hy, Chand a Mouli V.S.A, “E alua ion o Mechanical
P ope ies o Al6061 Composi e by S i Cas ing Me hod,” In e na ional Jou nal o Eme ging Resea ch in
Enginee ing, Science, and Managemen , ol. 3, no. 3, pp. 51-58, 2024. doi: 10.58482/ije esm. 3i3.9.
