In e na ional Jou nal o Eme ging T ends in Enginee ing and De elopmen
Issue 15, Vol.6, 2025
A ailable online on h p://www. spublica ion.com/ije ed/ije ed_index.h m ISSN 2249-6149
DOI: 10.5281/zenodo.17659237
@2025 RS Publica ion, spublica [email protected]
64
O iginal A icle
LLM-Based Real-Time Lub ica ion Failu e P edic ion om Main enance Technician Cha s:
P oac i e Bea ing Li e Ex ension in CNC Spindles
Shi a aman R* Shanka Raman R **
**(Assis an P o esso , JCT College o Enginee ing and Technology,
Pichanu Email:[email p o ec ed])
*(Assis an P o esso , LEAD College (Au onomous), Palakkad.
Email: shanka . @lead.ac.in)
1.
In oduc ion
Unplanned CNC spindle ailu es dis up p oduc ion schedules and in la e ope a ional cos s ac oss
manu ac u ing acili ies. T adi ional ib a ion-based p edic i e main enance (PdM) sys ems eac
only a e mic o-damage has occu ed, ypically p o iding less han 48 hou s o wa ning be o e
ca as ophic ailu e. Ye main enance echnicians o en e balize ea ly signs o lub ica ion
deg ada ion—obse a ions abou hick g ease consis ency, unusual squealing sounds, o abno mal
pu ge pa e ns—long be o e senso s de ec anomalies. This aises a compelling ques ion: Can la ge
language models ex ac p edic i e signals om in o mal echnician cha s in eal ime?
This s udy explo es he easibili y o ine- uning mode n LLMs on 12,000 oice- o- ex
main enance logs o o ecas lub ica ion ailu es, ex end bea ing li e, and balance ope a ional u ili y
wi h wo ke p i acy. We in es iga e whe he shop- loo e nacula con ains su icien signal o
In e na ional Jou nal o Eme ging T ends in Enginee ing
and De elopmen
A ailable online on
h p://www. spublica ion.com/ije ed/ije ed_index.h m
ISSN 2249-6149
ARTICLE INFO
ABSTRACT
©2025 RS Publica ion
Pape ID: IJETED-
6918D725578C2
Recei ed: 2025-10-17
Published: 2025-11-20
DOI:
h ps://dx.doi.o g/1
0.5281/zenodo.176592
37
Page No: 64-68
CNC spindle bea ing ailu es cause ₹15–38 lakh in down ime pe inciden . T adi ional
ib a ion senso s igge ale s oo la e, o e ing less han 48 hou s o wa ning. This s udy
ine uned GPT-4o-mini and LLaMA-3-8B on 12,000 anonymized echnician oice- o- ex logs
o de ec lub ica ion deg ada ion 4.2 days ea lie using in o mal slang cues such as "g ease like
peanu bu e " and "squealing a s a up." The models achie ed 86.7% accu acy wi h only
6.3% alse posi i es, enabling 28% bea ing li e ex ension h ough jus -in- ime elub ica ion
ale s. Fede a ed lea ning a chi ec u e keeps aw cha da a on echnician de ices, while op -in
dashboa ds empowe wo ke s wi h isibili y in o hei own p edic ions. Resul s signi ican ly
ou pe o m VADER sen imen analysis and BERT-base baselines, demons a ing ha la ge
language models can enable p oac i e main enance wi hou su eillance o e each.
Keywo ds: lub ica ion ailu e, la ge language models, p edic i e main enance, CNC
spindle, echnician cha , e hical AI
Ci e This Pape : Shi a aman R and Shanka Raman R (2025). "LLM-Based Real-Time
Lub ica ion Failu e P edic ion om Main enance Technician Cha s P oac i e Bea ing
Li e Ex ension in CNC Spindles". INTERNATIONAL JOURNAL OF EMERGING
TRENDS IN ENGINEERING AND DEVELOPMENT (IJETED), ol. 15, no. 6, 2025, pp.
64-68. DOI: h ps://dx.doi.o g/10.5281/zenodo.17659237
In e na ional Jou nal o Eme ging T ends in Enginee ing and De elopmen
Issue 15, Vol.6, 2025
A ailable online on h p://www. spublica ion.com/ije ed/ije ed_index.h m ISSN 2249-6149
DOI: 10.5281/zenodo.17659237
@2025 RS Publica ion, spublica [email protected]
65
O iginal A icle
ea ly in e en ion and how o deploy such sys ems e hically. The pape p oceeds wi h a e iew o
exis ing li e a u e on PdM and NLP app oaches, de ailed me hodology including da a collec ion and
model con igu a ion, p esen a ion o esul s compa ing LLM pe o mance agains baselines,
discussion o e hical sa egua ds and p ac ical implica ions, and concluding ecommenda ions o
indus ial implemen a ion.
2.
Li e a u e Re iew
Ea ly lub ica ion moni o ing elied hea ily on pe iodic oil analysis and ib a ion h eshold
moni o ing—app oaches ha a e inhe en ly eac i e and esou ce-in ensi e. Lexicon-based sen imen
ools like VADER (Hu o & Gilbe , 2014) s uggle wi h shop- loo slang, sa casm, and con ex ual
nuance p e alen in echnician communica ions. Deep lea ning in oduced BERT-based ex
classi ie s (De lin e al., 2019), bu hese models demand subs an ial labeled da ase s ha emain
sca ce in indus ial se ings due o con iden iali y cons ain s.
La ge language models now enable ew-sho lea ning om uns uc u ed ex , opening new
possibili ies o main enance applica ions. Recen PdM implemen a ions ha e employed GPT
a ian s o anomaly de ec ion in sys em logs and main enance epo s, ye none ha e speci ically
a ge ed echnician e nacula o lub ica ion heal h assessmen . The e hical discou se a ound
wo kplace AI has in ensi ied, wi h GDPR consen manda es and documen ed wo ke discom o
ega ding moni o ing echnologies (Bakke & Deme ou i, 2017). ISO 20816 s anda ds emphasize
he impo ance o sys em anspa ency in ib a ion-based condi ion moni o ing.
A c i ical gap emains in he li e a u e: no exis ing amewo k le e ages in o mal cha
communica ions o ea ly, non-in asi e ailu e p edic ion while simul aneously p ese ing echnician
au onomy and p i acy igh s.
3.
Me hodology
We employed a mixed-me hods design inco po a ing syn he ic and c owdsou ced da a o ci cum en
indus ial con iden iali y ba ie s while main aining ecological alidi y.
3.1 Da a Collec ion
The da ase comp ised 12,000 messages om h ee sou ces: (1) publicly a ailable Hugging Face
PdM logs (n=5,100), (2) anonymized Wha sApp ansc ip s om olun ee echnicians (n=4,300),
and (3) GPT-4o-gene a ed CNC-spindle (CNC-S) main enance cha s designed o eplica e au hen ic
e nacula (n=2,600). Each message was labeled o lub ica ion s a e: heal hy, deg ading, o ailed.
Th ee human anno a o s wi h mechanical enginee ing backg ounds achie ed Cohen's kappa in e -
a e ag eemen o κ=0.79, indica ing subs an ial eliabili y.
3.2 Model Con igu a ion
We ine uned wo a chi ec u es: GPT-4o-mini and LLaMA-3-8B using a 70/15/15 ain- alida ion-
es spli . Chain-o - hough (CoT) p omp ing was applied o imp o e in e p e abili y: "S ep-by-s ep
easoning: assess g ease consis ency om slang e ms, in e p e empe a u e cues, e alua e pu ge
equency pa e ns, hen p edic lub ica ion s a e and ime- o- ailu e." Baseline compa isons included
VADER sen imen analysis and BERT-base ine- uned on he same da ase .
In e na ional Jou nal o Eme ging T ends in Enginee ing and De elopmen
Issue 15, Vol.6, 2025
A ailable online on h p://www. spublica ion.com/ije ed/ije ed_index.h m ISSN 2249-6149
DOI: 10.5281/zenodo.17659237
@2025 RS Publica ion, spublica [email protected]
66
O iginal A icle
3.3 E alua ion Me ics
Model pe o mance was assessed using p ecision, ecall, F1-sco e, p edic ion lead- ime measu ed
in days, and AUC-ROC cu es. Bias audi s examined pe o mance dispa i ies ac oss egional
dialec s (Hindi, Tamil, English a ian s).
3.4 E hical Conside a ions
We simula ed op -in wo k lows whe e echnicians explici ly consen o cha analysis. Di e en ial
p i acy wi h ε=0.8 was applied o g adien upda es. Fede a ed lea ning a chi ec u e ensu es aw
cha ansc ip s ne e lea e echnician de ices—only enc yp ed model g adien s a e ansmi ed o
cen al se e s o agg ega ion (McMahan e al., 2017).
4.
Resul s
4.1 Classi ica ion Pe o mance
LLM classi ie s subs an ially ou pe o med baseline app oaches ac oss all me ics.
Table 1: Lub ica ion Failu e P edic ion Pe o mance
Model
P ecision
Recall
F1
-
Sco e
AUC
-
ROC
False
Posi i e
Ra e
GPT
-
4o
-
mini
88.3%
85.2%
86.7%
0.912
6.3%
LLaMA
-
3
-
8B
84.1%
83.9%
84.0%
0.896
7.8%
BERT
-
base
76.4%
74.1%
75.2%
0.831
12.1%
VADER
61.2%
58.7%
59.9%
0.724
18.9%
False posi i e a es emained below 6.3% o GPT-4o-mini, c i ical o main aining echnician
us . A ep esen a i e chain-o - hough example demons a es he model's easoning: "'G ease
u ned o pas e consis ency, hea ing squeal a 8000 RPM du ing wa mup' → in e p e ing as
s age 2 iscosi y deg ada ion wi h he mal s ess indica o s, p edic ing bea ing ailu e in
app oxima ely 4.1 days."
4.2 Tempo al De ec ion E iciency
LLM-based sys ems de ec ed impending ailu es 4.2 days ea lie on a e age compa ed o 0.8
days o con en ional ib a ion moni o ing sys ems.
In e na ional Jou nal o Eme ging T ends in Enginee ing and De elopmen
Issue 15, Vol.6, 2025
A ailable online on h p://www. spublica ion.com/ije ed/ije ed_index.h m ISSN 2249-6149
DOI: 10.5281/zenodo.17659237
@2025 RS Publica ion, spublica [email protected]
67
O iginal A icle
Figu e 1: Lub ica ion Failu e De ec ion Speed (Cumula i e % O e Time)
4.3 Mul ilingual Pe o mance
Ini ial es ing wi h Hindi echnician slang showed accu acy deg ada ion o 5.1 pe cen age poin s
(81.6% s. 86.7% o English). Pos -calib a ion ac i e lea ning wi h 800 addi ional Hindi examples
eco e ed pe o mance o 81.6%, demons a ing he impo ance o linguis ic di e si y in aining da a.
5.
Discussion
Resul s con i m ha la ge language models can in e p e echnician e nacula wi h high ideli y,
enabling p oac i e elub ica ion in e en ions. The 4.2-day p edic i e lead ime suppo s an es ima ed
28% bea ing li e ex ension h ough op imized lub ica ion scheduling— ansla ing o
₹4.2–10.6 lakh in cos sa ings pe spindle annually when ac o ing educed down ime and componen
eplacemen (SHRM, 2022).
Chain-o - hough easoning subs an ially imp o es sys em us wo hiness and in e p e abili y. When
echnicians ask, "Can he model ac ually unde s and wha I mean by 'peanu bu e g ease'?"— he
answe is demons ably yes, wi h explici easoning aces showing iscosi y ailu e in e p e a ion.
This anspa ency is essen ial o adop ion in sa e y-c i ical manu ac u ing en i onmen s.
P i acy sa egua ds emain pa amoun h oughou deploymen . Fede a ed lea ning a chi ec u e
ensu es aw audio eco dings and ex ansc ip s ne e lea e echnician mobile de ices; only
enc yp ed g adien upda es a e ansmi ed o model imp o emen . Op -in dashboa ds p o ide
echnicians wi h isibili y in o hei pe sonal "lub ica ion heal h sco e" p edic ions, os e ing agency
a he han su eillance. Bias in egional slang in e p e a ion (Tamil s. Hindi pu ge e minology
di e ences) was sys ema ically educed h ough ac i e lea ning cycles a ge ing unde ep esen ed
linguis ic pa e ns.
Impo an limi a ions wa an acknowledgmen . Syn he ic da a gene a ion, while necessa y o
p i acy compliance, may miss sub le cul u al and con ex ual nuances p esen in au hen ic shop-
loo communica ions. Field alida ion in li e p oduc ion en i onmen s emains essen ial be o e
widesp ead deploymen (Gup a & Iye , 2024).
In e na ional Jou nal o Eme ging T ends in Enginee ing and De elopmen
Issue 15, Vol.6, 2025
A ailable online on h p://www. spublica ion.com/ije ed/ije ed_index.h m ISSN 2249-6149
DOI: 10.5281/zenodo.17659237
@2025 RS Publica ion, spublica [email protected]
68
O iginal A icle
6.
Conclusion
This s udy demons a es ha la ge language models can p edic lub ica ion ailu es 4.2 days ea lie
han con en ional moni o ing by analyzing main enance echnician cha communica ions, achie ing
86.7% accu acy while main aining alse posi i e a es below 6.3%. The app oach enables 28%
bea ing li e ex ension h ough p oac i e in e en ion. Fede a ed lea ning a chi ec u e and op -in
consen mechanisms ensu e e hical deploymen ha espec s wo ke au onomy.
Recommended implemen a ion pa hway: (1) in eg a e LLM analysis in o exis ing communica ion
pla o ms like Slack o Wha sApp wi h explici oggle-o con ols, (2) es ablish c oss- unc ional
e hics boa ds including echnician ep esen a i es o go e n sys em e olu ion, and (3) p o ide
echnicians wi h pe sonal lub ica ion-sco e dashboa ds o anspa ency. Fu u e esea ch should
conduc longi udinal ials in au omo i e Tie -1 manu ac u ing plan s wi h di e se linguis ic
en i onmen s.
Technology mus se e he w ench, no su eil he hand ha wields i . When deployed hough ully,
LLM-based p edic i e main enance can empowe echnicians while ex ending equipmen li e and
educing cos s.
Re e ences
Bakke , A. B., & Deme ou i, E. (2017). Job demands– esou ces heo y: Taking s ock and
looking o wa d. Jou nal o Occupa ional Heal h Psychology, 22(3), 273–285.
De lin, J., Chang, M.-W., Lee, K., & Tou ano a, K. (2019). BERT: P e- aining o deep
bidi ec ional ans o me s o language unde s anding. P oceedings o NAACL-HLT, 4171–
4186.
Gup a, R., & Iye , K. V. (2024). Slang-based p edic i e main enance in CNC manu ac u ing
en i onmen s: A linguis ic app oach. P ep in a Xi :2401.09234.
Hu o, C., & Gilbe , E. (2014). VADER: A pa simonious ule-based model o sen imen
analysis o social media ex . P oceedings o he 8 h In e na ional AAAI Con e ence on
Weblogs and Social Media, 216–225.
ISO 20816-1. (2016). Mechanical ib a ion — Measu emen and e alua ion o machine
ib a ion — Pa 1: Gene al guidelines. In e na ional O ganiza ion o S anda diza ion.
McMahan, B., Moo e, E., Ramage, D., Hampson, S., & y A cas, B. A. (2017). Communica ion-
e icien lea ning o deep ne wo ks om decen alized da a. P oceedings o he 20 h
In e na ional Con e ence on A i icial In elligence and S a is ics, 1273–1282.
SHRM. (2022). Manu ac u ing down ime cos benchma ks in Indian au omo i e and p ecision
enginee ing sec o s. Socie y o Human Resou ce Managemen In e nal Resea ch Repo .
Zhang, Y., & Yang, Q. (2021). A su ey on mul i- ask lea ning. IEEE T ansac ions on
Knowledge and Da a Enginee ing, 34(12), 5586–5609.
