Utafiti huu unapendekeza mfumo wa deep learning unaozingatia muundo wa kimwili kwa lengo la kutambua mapema nyufa zinazoweza kutokea katika vipengele vya rotor vya kasi kubwa vya injini za ndege, kama diski na shaft, kupitia vibration signals zinazopimwa kwenye casing ya injini. Badala ya kuichukulia injini kama black box moja ya namba, model imeundwa na parallel subnetworks zinazolingana na sehemu za kimwili kama inlet, fan, high-pressure compressor, combustion chamber, high-pressure turbine, low-pressure turbine na nozzle.

Katika approach iliyopendekezwa, kila target flight hutabiriwa kwa kutumia flights tano zilizotangulia. Model hufundishwa upya kwa flights zinazochukuliwa kuwa healthy katika recent past pekee na hujaribu kufuatilia normal vibration baseline inayobadilika kadiri injini inavyozeeka. Tofauti kati ya predicted vibration na actual vibration hutumika kama crack indicator.

Watafiti walipunguza training data kwa kuchagua operating ranges ambazo moving Pearson correlation coefficient kati ya rotor speed na casing vibration ilikuwa juu ya 0,8. Imeripotiwa kwamba sMAPE value ya asilimia 27,34 katika model iliyojengwa kwa full-speed-range data ilishuka hadi asilimia 7,36 kwa selected ranges. Kati ya points 645 za worst validation scenario, katika asilimia 97,5 ya hizo prediction error ilibaki chini ya asilimia 20.

Study pia inatumia compressor-exit total-pressure signal P₃^* kutenganisha transient vibration spikes zinazotokea wakati wa compressor instability au “surge” na actual rotor damage. Pressure inaposhuka rapidly huku vibration ikiongezeka, adaptive gate hufunguka na aerodynamic-source transient deviation hukandamizwa. Persistent na growing vibration differences bila pressure drop huhifadhiwa kama indicators za structural deterioration.

Katika real-flight validation, 101 consecutive flights zilichunguzwa. Kulingana na study, katika last part ya flight 99 actual vibration ilipanda juu ya predicted healthy baseline, katika flight 100 tofauti hii iliendelea throughout the flight, na katika borescope inspection baada ya flight 101 visible crack ilipatikana katika disk-shaft region. Watafiti wanaeleza behavior hii kwa unidirectionality, accumulation na irreversibility.

Hata hivyo, results zilizoonyeshwa hazithibitishi kwamba method inaweza kutambua crack kwa uhakika mapema katika all aircraft engines. Study ina only one validated crack event; engine type, sensor location, crack size, maintenance history, raw data na source code hazijashirikiwa. Ingawa slope-based warning rule inahitaji five consecutive threshold exceedances, kuonyeshwa kwa exceedance katika flights 98, 99 na 100 pekee kunafanya reported early-warning timing kuwa mathematically ambiguous.

Ni safety problem gani inayoshughulikiwa na study?

Disks na shafts za aircraft engines hukabiliwa na very high rotational speeds, centrifugal forces, temperature changes na aerodynamic loads wakati wa operation. Small fatigue crack inayoanza katika components hizi inaweza kubadilisha local stiffness na damping behavior ya rotor. Crack ikifikia critical size, rotor inaweza kuvunjika na kusababisha uncontrolled failure inayotoka nje ya engine casing.

Cracks katika rotor disks na shafts zinaweza kuonekana less frequently kuliko bearing au blade failures; lakini consequences zake zinaweza kuwa severe zaidi. Kwa hiyo, tatizo si tu kupata crack wakati tayari ipo, bali kutambua weak effect yake kwenye vibration kabla haijawa visible wakati wa maintenance.

Kwa sasa engine health inaweza kufuatiliwa kwa three main methods:

• Online vibration monitoring: Sensors kwenye engine casing hufuatilia kama certain vibration levels zimezidiwa.
• Nondestructive inspection: Engine husimamishwa na, inapohitajika, kuvunjwa kwa sehemu ili cracks zitafutwe kwa ultrasonic, eddy-current au other methods.
• Borescope inspection: Visual inspection hufanywa kupitia narrow access channels bila kuvunja engine completely.

Traditional vibration thresholds zinaweza kukosa weak na nonlinear effects za early cracks. Nondestructive inspection ni reliable lakini inahitaji engine kusimamishwa na costly maintenance processes. Borescope inaweza kuonyesha accessible surfaces pekee. Study inalenga kujaza gap hii kwa model inayotabiri continuously expected healthy vibration ya engine na kutoa warning wakati measured vibration inaondoka persistently kutoka baseline hii.

Swali kuu la utafiti ni nini?

Swali kuu la study linaweza kufupishwa hivi:

Ikiwa healthy vibration inayotarajiwa kutokana na engine operating conditions inaweza kutabiriwa kwa usahihi wa kutosha kwa artificial intelligence, je, persistent na growing departure ya actual vibration kutoka baseline hii inaweza kutumika kama early indicator ya rotor crack?

Approach hii haikokotoi directly crack length au geometry. Model hutabiri healthy engine behavior na kutathmini deviation kutoka behavior hii kama anomaly indicator.

Flight-to-flight prediction kwa sliding window

Katika strategy iliyoonyeshwa katika Kielelezo 1, window width imechaguliwa kuwa flights tano. Flights tano za kwanza zilizo healthy hutumiwa kutabiri vibration ya flight ya sita. Kisha window inasogezwa forward kwa flight moja:

• Flights 1–5 hutumika kutabiri flight 6.
• Flights 2–6 hutumika kutabiri flight 7.
• Flights kati ya i na i+4 hutumika kutabiri flight i+5.

Design hii ina objectives mbili. Kwanza, model ijifunze recent healthy behavior pekee na kuzuia training baseline kuchafuliwa na data yenye crack indication. Pili, kufuatilia normal vibration inayobadilika kutokana na long-term engine ageing, wear na performance degradation.

Fixed model iliyofundishwa katika early life ya engine inaweza kutoa systematic error wakati inatathmini aged engine. Sliding window inalenga kufuatilia normal ageing na kufanya sudden na irreversible structural changes kuwa visible zaidi.

Input na output variables za model ni zipi?

Input vector katika kila time step ina seven engine na environmental parameters:

\[ X_i(t)=[T_0,\ P_0,\ N_1,\ N_2,\ P_3^*,\ T_5^*,\ W_f] \]

Output ya model ni vibration response V inayotarajiwa kupimwa na sensor katika engine casing au support region. Lengo ni kwamba tofauti kati ya actual measurement na prediction itokane na change katika structural condition, si engine operating conditions.

Kwa nini Bi-LSTM ilitumika?

Engine data zimeunganishwa over time. Rotor speed, pressure, temperature na vibration hazibadiliki suddenly na independently kutoka time step moja kwenda nyingine. Long Short-Term Memory network, au LSTM, inaweza kuhifadhi important past information katika cell state na kujifunza long-term relations katika time series.

Study imetumia bidirectional LSTM au Bi-LSTM. Network moja huchakata sequence forward na nyingine backward, kisha hidden states mbili huunganishwa:

\[ H_t=[\overrightarrow{h_t},\overleftarrow{h_t}] \]

Method hii inaweza kutumia both past na future context katika offline sequence analysis. Hata hivyo, kutokuwepo kwa future samples kwa prediction ya current second katika real-time system ni practical issue muhimu. Study haielezi jinsi bidirectional structure inavyotumika wakati wa one-second real-time inference bila kuhitaji future information.

Physical-topology-driven network architecture ilijengwaje?

Katika Kielelezo 3, network imegawanywa katika three main layers:

1. Component network layer,
2. Physical-law matching network,
3. Mapping network.

Component network layer

Seven physical sections za engine zimewakilishwa kwa separate subnetworks:

• Inlet,
• Fan,
• High-pressure compressor,
• Combustion chamber,
• High-pressure turbine,
• Low-pressure turbine,
• Nozzle.

Inlet temperature na pressure zimeelekezwa kwenye inlet subnetwork, N₁ kwenye fan na low-pressure-turbine subnetworks, na N₂ kwenye high-pressure-compressor na turbine subnetworks. Fuel flow imeunganishwa na combustion chamber, huku T₅^* ikiunganishwa na nozzle section.

Physical-law matching network

Local features kisha zimepitishwa kupitia Bi-LSTM modules zenye majina yafuatayo:

• Boundary condition,
• Work interaction,
• Aerodynamic load,
• Energy conversion,
• Torque balance,
• Low-pressure-turbine transmission,
• Thrust feedback.

Watafiti wanasema arrangement hii inaiga mechanical na aerodynamic information flow ya engine. Hata hivyo, hakuna explicit physical constraints zilizotolewa kwa modules kulingana na mass, momentum, energy, torque au rotor-dynamics equations. Modules kimsingi ni neural-network blocks zilizopangwa kwa majina yenye physical meaning.

Katika future-work section ya study, inakubaliwa kwamba current model haiingizi explicitly rotor-dynamics laws na inajifunza component interactions implicitly kupitia fully connected layers. Kwa hiyo, method inapaswa kuelezwa zaidi kama data model iliyopangwa kwa domain knowledge au kugawanywa kwa physical architecture kuliko physics-informed neural network ambamo physical equations zimeingizwa katika loss function.

Mapping network

Local na matched features zote zinaunganishwa katika fully connected layers mbili ili kutoa final vibration prediction. Katika layer hii, multiple engine parameters hubadilishwa kuwa single vibration output.

Kwa nini only certain operating ranges zilichaguliwa?

Engine vibration haihusiani na rotor speed kwa kiwango sawa katika entire operating envelope. Combustion instability katika idle, low variability katika steady cruise, resonance near critical speeds na different thermal states za flight zinaweza kuvuruga vibration-speed relationship.

Study inakokotoa Pearson correlation kati ya N₁ rotor speed na casing vibration ndani ya moving window:

\[ \rho_{N_1,V}=\frac{\operatorname{cov}(N_1,V)} {\sigma_{N_1}\sigma_V} \]

Correlation coefficient ikiwa juu ya 0,8, rotor speed na vibration zinachukuliwa kuwa strongly coupled na range hiyo inachaguliwa kwa training.

Pia, ikiwa local change ya rotor speed ni smaller kuliko asilimia 2 ya entire operating range, window inachukuliwa kuwa invalid:

\[ \Delta N_{1,\mathrm{local}}>0.02\times \operatorname{Range}(N_1) \]

“Dead-band” rule hii inalenga kuzuia artificial correlations zinazotokana na measurement noise katika cruise segments ambapo rotor speed ni almost constant.

Inconsistency kuhusu correlation window

Katika data-preprocessing section, kwa 1 Hz sampling moving correlation window ya sekunde 150 imetajwa. Katika results section, same data selection imeelezwa kwa window ya sekunde 30.

Windows za sekunde 150 na 30 huenda zisitoe same results:

• Window ya sekunde 150 hupima correlation iliyo stable zaidi lakini slower-changing.
• Window ya sekunde 30 ni more sensitive kwa transient events lakini more exposed kwa noise.

Kwa kuwa haijaelezwa ni window ipi ilitumika katika final experiments, data selection haiwezi kureproduciwa independently.

Data amount na data groups

Study ina three different data definitions:

• Katika data-preprocessing section, consecutive flights 100 na zaidi ya samples milioni 6 zinatajwa.
• Katika model-accuracy section, flights 24 zimegawanywa katika groups nne B1–B4, kila group ikitumia flights tano kwa training na flight moja kwa validation.
• Katika real-crack validation, 101 consecutive flights zimechunguzwa.

Haijulikani kama hizi ni different subsets za engine ileile, different engines au independent datasets. Distinction hii ni muhimu kwa kutathmini kama model inageneralize kwa different engines.

Min–max normalization

Selected time series zili-scale kwenda common numerical range:

\[ X'=\frac{X-X_{\min}}{X_{\max}-X_{\min}}\,(a-b)+b \]

Equation inalingana na standard min–max scaling. Hata hivyo, text inaeleza X′ kama original data na X kama normalized value. Explanation hii ni reverse ya matumizi katika equation.

Katika real-time application, haijaelezwa kama X_min na X_max values zilizopatikana kutoka training data zilitumika, au kila new flight iliscalewa kwa own minimum na maximum. Kutumia minimum na maximum zilizopatikana kutoka entire new flight kunaweza kusababisha use ya future information katika real-time prediction.

Prediction residual inakokotolewaje?

Main health indicator ya model ni difference kati ya measured vibration na predicted healthy vibration. Relative residual imefafanuliwa katika study hivi:

\[ e_{\mathrm{rel}}= \frac{1}{n}\sum_{i=1}^{n} \frac{|V_{\mathrm{actual},i}-V_{\mathrm{pred},i}|} {V_{\mathrm{pred},i}}\times100\% \]

Absolute residual ni mean absolute difference kati ya measurement na prediction:

\[ e_{\mathrm{abs}}= \frac{1}{n}\sum_{i=1}^{n} |V_{\mathrm{actual},i}-V_{\mathrm{pred},i}| \]

Kutumia predicted vibration katika denominator ya relative residual kunaweza kutoa very large percentage values wakati prediction inakaribia zero. Sehemu ya maximum errors zinazozidi asilimia 1.000 katika full-speed-range model inaweza kuhusishwa na mathematical sensitivity hii.

Dynamic warning threshold

Robust statistical threshold inajengwa kutoka relative residuals za healthy flights:

\[ T_{\mathrm{static}}= \operatorname{Median}(e_{\mathrm{rel}}) +m\,\operatorname{NMAD}(e_{\mathrm{rel}}) \]

Katika study, m=3 imechaguliwa. Watafiti wanaihusisha na three-sigma rule. Hata hivyo, threshold inayotegemea median absolute deviation haitoi directly confidence level ya asilimia 99,73 wakati distribution si normal. Pia haijaelezwa NMAD ilinormalize kwa scaling coefficient gani.

Static threshold inarekebishwa kwa engine-life au ageing coefficient:

\[ \lambda_{\mathrm{life}}= 1+\alpha\frac{F-F_{\mathrm{avg}}}{F_{\mathrm{overhaul}}} \]

Hapa F ni flight cycle, F_overhaul ni maintenance period na α=0,15 ni empirical coefficient. Jinsi variable F_avg ilivyoamuliwa haijaelezwa. Katika Kielelezo 14, maintenance period ya cycles 3.000 imetumika na dynamic threshold imeonyeshwa kuwa takribani 0,095560.

Kwa nini aerodynamic instability husababisha false alarm?

Wakati wa compressor instability, compressor-exit pressure inaweza kushuka rapidly na vibration kwenye engine casing inaweza kuongezeka temporarily. Healthy-baseline model ikiwa haiwezi kutabiri sudden event hii, measured vibration huwa kubwa kuliko prediction na positive residual inayofanana na rotor crack hutokea.

Study inaweka distinction ifuatayo kati ya aerodynamic instability na crack:

Adaptive vibration compensation inafanyaje kazi?

System inafuatilia jointly derivative ya P₃^* pressure kwa time na vibration residual. Pressure inaposhuka rapidly huku residual pia ikizidi threshold, high-suppression mode ina-activate:

\[ \alpha_t= \begin{cases} \alpha_{\mathrm{high}}, & \nabla P_3^*<\delta\ \text{ve}\ e_{\mathrm{abs}}>\gamma \\ \alpha_{\mathrm{low}}, & \text{diğer durumlar} \end{cases} \]

Moving mean na variance za residuals zinaupdate adaptively:

\[ \mu_t=\alpha_t R_t+(1-\alpha_t)\mu_{t-1} \]

\[ \sigma_t^2= \varepsilon(R_t-\mu_t)^2+ (1-\varepsilon)\sigma_{t-1}^2 \]

Compensated health indicator ni difference kati ya raw residual na adaptive mean:

\[ CHI_t=e_{\mathrm{abs},t}-\mu_t \]

Wakati wa vibration spike inayolingana na pressure drop, α_t huongezwa ili adaptive baseline ifuatilie event haraka na suppress residual signal. Wakati wa crack, kwa kuwa pressure haishuki, gate hubaki closed na persistent vibration increase inaendelea kuonekana katika health indicator.

Attention-based gate module

Study inabadilisha fixed conditional rule kuwa trainable attention module. P₃^* signal inapelekwa kwenye gate kupitia direct path inayoskip Bi-LSTM layers. Lengo ni kuzuia short-duration pressure drops zisifanyiwe excessive smoothing na long-term memory layers.

Attention weight inakokotolewa kwa general structure ifuatayo:

\[ \alpha_t= \operatorname{Sigmoid} \left( W_a[ \operatorname{ReLU}(-\nabla P_{3,t}^*) \oplus e_{\mathrm{abs},t-1} ]+b_a \right) \]

Final vibration prediction ni combination ya hidden features za physical network na compensation term inayozalishwa kutoka pressure:

\[ \widehat{V}_{\mathrm{pred},t}= f_{\mathrm{map}}(H_{\mathrm{phys},t}) +\alpha_t g(P_{3,t}^*) \]

Haijaelezwa gate ilifundishwa kwa surge labels zipi, training ilikuwa na aerodynamic-instability examples ngapi, au false-gate-opening rate ilikuwa kiasi gani.

Model training settings

Kwa sababu study inatumia 1 Hz sampling, system ina takribani sekunde moja kuchakata kila data point. Kwa hiyo, inference time ya chini ya 1 ms inaonekana sufficient. Hata hivyo, haijaelezwa kama measurement inajumuisha GPU synchronization, warm-up runs, repetition count, preprocessing time na data transfer.

Comparison ya full-speed range na selected operating range

Katika validation groups nne B1–B4, flights tano zilitumika kwa training na flight moja kwa validation. Katika kila group, full-speed data ilikuwa na takribani samples 50.000, huku baada ya correlation filter takribani samples 8.000 zikibaki.

Study inaeleza mean error reduction kuwa takribani asilimia 70. Ikiwa values 27,34 na 7,36 zinachukuliwa kuwa same type ya statistic, mathematical reduction ni takribani asilimia 73:

\[ \frac{27.34-7.36}{27.34}\times100\approx73.1\% \]

Hata hivyo, katika results section asilimia 7,36 imeelezwa kama maximum sMAPE value ya selected model, huku asilimia 27,34 ikiwa mean sMAPE value ya full-speed model. Katika abstract na conclusion section, values zote mbili zimewasilishwa kana kwamba ni same type ya average. Kwa hiyo, haiko clear direct percentage reduction inategemea statistics zipi.

Pia, kuondoa difficult operating regions kutoka dataset kunaweza kumaanisha kwamba model imetathminiwa tu katika ranges ambazo ni easier to predict. Result hii haionyeshi kwamba prediction performance imeboreshwa katika entire flight envelope, bali kwamba cleaner baseline imeundwa katika limited range yenye strong rotor-vibration relation.

Error distribution katika worst validation scenario

Katika B4 scenario, critical samples 645 zilichunguzwa:

Distribution inaonyesha kwamba model ilitabiri points nyingi katika selected operating range kwa low error. Hata hivyo, kwa kuwa percentage errors zinaweza kuwa kubwa kwa small vibration values, absolute error na physical unit ya vibration zilipaswa kuripotiwa pamoja. Study inatoa MAE equation lakini haitoi numerical MAE table kwa validation groups.

Result ya compressor-instability compensation

Katika Kielelezo 11, measured na predicted P₃^* pressures katika flight moja iliyoathiriwa na compressor instability zinalinganishwa. Sudden jump, rapid drop na large-amplitude oscillations zinaonekana katika measured pressure. Base model inafuatilia rapid pressure change kwa delay.

Katika Kielelezo 12(a), uncompensated model inaunderpredict vibration wakati wa surge na kuzalisha large positive residual. Katika Kielelezo 12(b), baada ya pressure-gated compensation kuongezwa, prediction curve imeonyeshwa kukaribia measured vibration.

Katika section hii, MAE, sMAPE, residual magnitude au false-alarm count kabla na baada ya compensation hazijatolewa. Result inategemea kwa kiasi kikubwa graphical comparison ya selected event moja. Idadi ya surge events zilizotumika kwa training na validation haijaelezwa.

Flight sequence katika real crack case

Study imechunguza 101 consecutive flights za engine ambayo crack iliripotiwa kuthibitishwa baadaye kwa borescope inspection.

Katika flights 25, 48, 66 na 81, predicted na measured vibration curves zinaonekana karibu. Short-term deviations zinarudi haraka kwenye healthy baseline.

Katika flight 99, baada ya takribani time step 18.000, measured vibration ilianza kubaki juu ya prediction. Deviation ilikua katika last part ya flight na haikurudi healthy baseline. Flight hii iliondolewa kutoka next training window.

Model iliyofundishwa kwa flights 94–98 ilitumika directly kwa flight 100 na actual vibration iliripotiwa kubaki juu ya predicted baseline throughout the flight. Katika ground inspection baada ya flight 101, visible crack ilipatikana katika disk-shaft region.

Behaviors tatu zilizofafanuliwa kwa crack

• Unidirectionality: Measured vibration inabaki continuously juu ya healthy prediction.
• Accumulation: Magnitude ya deviation inaongezeka katika consecutive flights.
• Irreversibility: Deviation hairudi baseline kama temporary operating change.

Behaviors hizi zinaweza kuendana na crack; lakini si specific kwa crack pekee. Persistent sensor drift, rotor imbalance, loose connection, rubbing, bearing degradation au other structural damage zinaweza pia kuzalisha similar persistent vibration increase. Study haijafanya differential validation dhidi ya other failure types.

Pointwise residual threshold

Kulingana na Kielelezo 14, relative residual ilibaki chini ya dynamic threshold katika flights 6–98. Katika flight 99 residual ilianza kuzidi threshold, katika flight 100 iliongezeka clearly, na value ya flight 100 ikafikia mara 6,4 ya pre-anomaly level.

Study inatafsiri result hii kwamba pointwise residual threshold ilitoa crack warning katika flight 99.

Slope-based progressive-degradation criterion

Ili kukamata early na small changes, first difference ya five-flight moving average hutumika. Equation inaweza kutafsiriwa hivi:

\[ \delta_k= \frac{1}{5} \sum_{i=k-4}^{k} \left(e_{\mathrm{rel},i}-e_{\mathrm{rel},i-1}\right) \]

Threshold ilijengwa kutoka slopes za healthy flights 11–80:

\[ \theta_{\mathrm{slope}}= \operatorname{Median}(\delta_{\mathrm{healthy}}) +3\,\operatorname{NMAD}(\delta_{\mathrm{healthy}}) =0.0046 \]

Slopes zilizotolewa katika study:

Watafiti wanasema first exceedance ilitokea katika flight 98 na slope criterion ilitoa warning one flight earlier kuliko pointwise threshold.

Mathematical contradiction katika five-consecutive-flight condition

Katika methods section, imeandikwa kwamba kwa progressive-degradation warning slope lazima ibaki juu ya threshold kwa five consecutive flights. Katika results zilizoonyeshwa, however, flights 98, 99 na 100 pekee ziko juu ya threshold.

Kwa hiyo, kulingana na data iliyowasilishwa:

• Flight 98 ni first threshold exceedance pekee.
• Katika flight 99 kuna two consecutive exceedances.
• Katika flight 100 kuna three consecutive exceedances.
• Five-consecutive-exceedance condition haijakamilika katika available flight sequence.

Ikiwa first exceedance ilianza flight 98, fifth consecutive exceedance ingepaswa kutokea earliest katika flight 102. Study inaonyesha slope hadi flight 100. Kwa hiyo, conclusion kwamba warning ilitolewa katika flight 98 au 99 kwa “five consecutive flights” rule haijaungwa mkono mathematically.

Other contradictions kuhusu warning timing

Sections tofauti za study zina statements zifuatazo:

• Slope threshold ilizidiwa first time katika flight 98,
• Progressive-degradation warning ilitolewa katika flight 99,
• Pointwise residual threshold ilizidiwa katika flight 99,
• Crack ilionekana katika inspection baada ya flight 101,
• Method ilitoa warning only one flight before borescope validation.

Kulingana na flight numbers, kuna more than one flight kati ya flight 98 na inspection baada ya flight 101. Flight 99 pia si one flight tu kabla ya inspection after flight 101; kulingana na numbering ni takribani cycles mbili mapema. Kwa kuwa haijaelezwa warning timing ilihesabiwaje kwa msingi wa flight start, flight end na maintenance time, phrase “one flight earlier” haiko clear.

Inconsistency kuhusu healthy training data

Study inatumia statement inayofafanua first 100 flights kama “normal flights.” Kinyume chake, inaonyesha flights 99 na 100 kama crack onset na persistent deterioration. Definitions hizi mbili hazipatani.

Pia, ikiwa slope criterion inaonyesha anomaly katika flight 98, kutumia flight 98 kama healthy data katika training window inayotabiri flight 99 ni questionable. Study inaeleza explicitly kwamba flight 99 imeondolewa kutoka training; lakini inaendelea kutumia flight 98, ambayo inadaiwa kuwa na earliest warning, kama healthy training data.

Physical interpretability iko katika kiwango gani?

Kugawanya inputs na lower layers za network kulingana na engine components kunatoa architecture inayoeleweka zaidi kuliko ordinary fully connected network. Inawezekana kueleza physically parameter ipi inaingia component ipi.

Hata hivyo, katika physical-law matching layer hakuna:

• Torque-balance equation,
• Mass na energy conservation,
• Rotor stiffness matrix,
• Shaft-disk-bearing dynamics,
• Crack breathing au harmonic model,
• Physics-based loss function

Kwa hiyo, claim kwamba residuals zinaweza “kuhusishwa directly na crack” ni too strong. Network inaweza kuonyesha other unexplained au unmeasured physical degradations kama crack residual pia.

Strengths za study ni zipi?

• Inatoa modular approach inayoreflect physical sections za engine katika neural-network architecture.
• Inatumia sliding window inayoupdate pamoja na flight cycles badala ya single fixed model.
• Inalenga kuzuia crack data kuchanganyika na model training.
• Inachunguza correlation-based data selection kwa numerical error comparison.
• Inatumia independent physical signal kutenganisha transient aerodynamic instabilities na structural deterioration.
• Inawasilisha si mean error pekee, bali maximum error na pointwise error distribution pia.
• Inachunguza one crack case katika real flight sequence ambayo imeripotiwa kuthibitishwa baadaye kwa borescope.
• Inatathmini inter-flight degradation trend pamoja na pointwise deviation.
• Inafafanua physically understandable criteria ili model itenganishe transient na persistent vibration deviations.

Main limitations za study ni zipi?

• Study ni preprint ambayo haijapitia peer review.
• Authors, institutions, corresponding author na publication date hazipo katika study text.
• Only one borescope-confirmed crack case imewasilishwa.
• Exact crack location, length, depth, orientation na progression state hazijatolewa.
• Borescope image au independent nondestructive-inspection result haijaonyeshwa.
• Engine model, fleet size na kama kuna data kutoka different engines hazijaelezwa.
• Raw flight data, sensor specifications na source code hazijashirikiwa.
• Exact location, measurement axis, unit, sampling accuracy na calibration ya vibration sensor hazijaelezwa.
• Vibration results zime nondimensionalize na physical amplitudes hazijaripotiwa.
• Kuna inconsistency kati ya 150-second na 30-second Pearson windows.
• Mean ya full-speed model huenda imelinganishwa na maximum sMAPE value ya selected model.
• Kwa kuwa correlation filter inaexclude difficult operating regions, generalization katika entire flight envelope haijaonyeshwa.
• Model haijalinganishwa under controlled conditions na standard LSTM, single network, graph network au physics-informed alternatives.
• Hakuna ablation analysis inayotenganisha contributions za physical topology, correlation filter na pressure gate.
• Surge compensation imeonyeshwa kwa selected graphs pekee, bila numerical error au false-alarm rate.
• Idadi ya surge events zilizotumika haijaelezwa.
• Jinsi bidirectional LSTM inavyotumika katika causal real-time prediction haijaelezwa.
• Haijaelezwa kama training bounds au target-flight bounds zilitumika katika real-time normalization.
• Prediction curves katika Kielelezo 13 zimesemwa kuwa “moderately smoothed,” lakini method haijaelezwa.
• Five-consecutive-slope-exceedance condition haijatimizwa na flights tatu zilizoonyeshwa.
• Ingawa flight 98 inatafsiriwa kama anomaly, imetumika kama healthy training data katika next model.
• Early-warning timings katika different sections za study hazipatani.
• Validity ya three-sigma confidence level kwa NMAD-based threshold haijathibitishwa experimentally.
• False-positive na false-negative rates hazijapimwa katika independent engine fleet.
• Hakuna differential test dhidi ya other degradations kama persistent imbalance, bearing fault, sensor drift na rubbing.

Study ina-support nini?

Study ina-support kwamba expected vibration inaweza kujifunzwa kutoka engine operating parameters na kwamba persistent departure ya measured vibration kutoka prediction hii inaweza kuwa useful health indicator ya structural deterioration.

Kuchagua ranges ambapo rotor speed na vibration zina strong relation kulitoa lower prediction error katika validation groups zilizochunguzwa. Pia imeonyeshwa kwamba compressor-exit pressure inaweza kuwa useful physical reference ya kutambua aerodynamic vibration spikes.

Single real case inaonyesha kwamba persistent na growing vibration residual ilikuwepo katika two consecutive flights kabla ya crack validation. Result hii inatoa early-warning approach inayostahili kuchunguzwa kwa condition-based maintenance.

Study haithibitishi nini?

• Haithibitishi kwamba model inaweza kutambua rotor crack one au more flights mapema katika all engine types.
• Haionyeshi kwa independent failure comparisons kwamba detected residual ilitokana only na crack.
• Haitabiri crack location, length au remaining life.
• Haionyeshi kwamba five-consecutive-slope-exceedance condition ilitimizwa kweli.
• Haithibitishi consistently timing ya “one-flight-early warning” ndani ya study.
• Haionyeshi kwamba model inageneralize kwa different engines, different sensors na different flight routes.
• Haithibitishi kwamba aerodynamic-instability compensation haitasuppress actual crack signal katika condition yoyote.
• Haionyeshi kwamba gate itafanya correct decision katika situation ambapo surge na crack vinatokea simultaneously.
• Haionyeshi kwamba model iko ready kutumika kama certified flight-safety system.
• Haithibitishi kwa controlled ablation kwamba physical-topology layers ni better kuliko standard black-box model.

Inawezaje kutathminiwa kwa mtazamo wa Uturuki?

Condition-based maintenance inazidi kuwa muhimu nchini Uturuki katika civil aviation, military engine systems, unmanned aerial vehicles, helicopters na domestic turbomachinery projects. Kulinganisha engine si kwa fixed vibration limits pekee bali na own previous healthy behavior kunaweza kuboresha maintenance planning.

Main approach ya study inayoweza kuhamishwa kwenda Uturuki ina steps zifuatazo:

• Kukusanya synchronized gas-path na vibration data kupitia engine flight cycles.
• Kuunda individual healthy vibration baseline kwa kila engine.
• Kufuatilia ageing na normal performance degradation kwa sliding training window.
• Kuweka labels kwa known transient events kama compressor instability kwa separate physical signals.
• Kulinganisha persistent residual na borescope, eddy-current na other nondestructive-inspection results.
• Kuvalidate warning threshold across different engines, climates na flight missions.

Kabla ya deployment, local database yenye many healthy na faulty samples kutoka different engine types inapaswa kuundwa. False-alarm rate, missed-crack rate na effect ya model kwenye maintenance cost zinapaswa kupimwa katika fleet level. Critical decisions zinapaswa kutegemea si AI output pekee bali multi-sensor validation na expert engineering assessment.

Mbinu na Matokeo ya Utafiti

Research design

Performance metrics

Mean absolute error:

\[ MAE=\frac{1}{m}\sum_{i=1}^{m}|y_i-\widehat{y_i}| \]

Symmetric mean absolute percentage error:

\[ sMAPE= \frac{200\%}{m} \sum_{i=1}^{m} \frac{|y_i-\widehat{y_i}|} {|y_i|+|\widehat{y_i}|} \]

Maximum error imetolewa katika study hivi:

\[ MaxError= \max_i \left( \frac{|y_i-\widehat{y_i}|} {\widehat{y_i}} \right)\times100\% \]

Kugawa maximum-error formula kwa predicted value kunaweza kutoa extremely high ratios wakati prediction iko near zero. Kwa hiyo, maximum percentage error haipaswi kutathminiwa peke yake.

Main quantitative findings

Warning timeline

Internal-consistency check

Missing information kwa reproducibility

• Complete author na institution list.
• Engine type na operating history.
• Vibration-sensor model, location, direction na unit.
• Sample count ya kila flight na data-file structure.
• Flights zinazounda B1–B4 groups.
• Ni window ipi, sekunde 30 au 150, ilitumika.
• Bi-LSTM layer count, hidden-unit count na subnetwork sizes.
• Dropout, regularization na hyperparameter-selection method.
• Surge-labeling method na event count.
• Training target ya attention gate.
• NMAD normalization katika dynamic threshold na definition ya F_avg.
• Smoothing algorithm iliyotumika katika Kielelezo 13.
• Crack size na visual record iliyopatikana kwa borescope.
• Source code, model weights na anonymized raw data.

Reliable interpretation limit ya results

Result inayoungwa mkono kwa nguvu na study ni kwamba vibration baseline iliyojengwa kwa physically meaningful engine parameters na persistent residuals katika consecutive flights zinaweza kuwa useful kwa rotor-health monitoring.

Kinyume chake, claim kwamba model ilitoa crack alarm before a specific flight si conclusive kwa sababu ya consecutive-exceedance contradiction katika slope criterion. One confirmed case haitoshi kukokotoa sensitivity, specificity au fleet-level false-alarm rate.

Method inapaswa kutathminiwa si kama standalone maintenance-decision system, bali kama additional risk indicator ya kupanga borescope au other nondestructive inspections.

Maelezo ya Chanzo na Mbinu

Jina la asili la study: Application of Physical-Structure-Driven Deep Learning in Aircraft Engine Rotor Crack Prediction

Waandishi: Hakuna author name au complete author list katika study text.

Author field katika file metadata: gantian_nwpu@163.com.

Author order: Haijathibitishwa.

Equal contribution au equal first authorship: Haijaelezwa.

Corresponding author: Haijaelezwa.

Taasisi: Hakuna institution information katika study text.

DOI: DOI information haipo katika study text na haijaweza kufinalize kupitia verified official record.

Official source:SSRN abstract number 7201932

Platform: SSRN.

Publication au upload date: Haijaelezwa katika study text.

Page count: 35.

Source type: Preprint research yenye real flight-vibration data, physically structured Bi-LSTM, adaptive aerodynamic compensation na consecutive-flight anomaly analysis.

Peer-review status: Study haijapitia peer review. Kila page ya document ina preprint na peer-review warning.

Peer-reviewed journal: Haijaelezwa.

Original publisher: Hakuna accepted journal publication au finalized publisher.

Data source: Study imetumia 1 Hz flight data zinazosemekana kuchukuliwa kutoka civil turbofan engines. Engine model, operator, flight route na raw-data access hazijaelezwa.

Code na model access: Hakuna open link iliyotolewa kwa source code, trained model weights au experimental configuration files.

Funding na conflict of interest: Hakuna funding au conflict-of-interest statement katika study text.

Maelezo haya ya Kituruki yameandaliwa baada ya kuchunguza full text ya study, equations 21 kuu, sliding-window scheme, Bi-LSTM cell, physical-topology network, Pearson-correlation graphs, adaptive-compensation schemes, error distributions, surge comparisons, consecutive-flight predictions, dynamic residual threshold na slope graph. Scientific content inategemea uploaded study pekee.

Kwa kuwa author na institution information hazipo katika study, bibliographic identity haikuweza kukamilishwa fully. Upungufu huu unapaswa kuhifadhiwa wazi katika page inayoandaliwa kwa publication; unverified name au institution haipaswi kuongezwa.

Main methodological limits za study ni one validated crack case, missing raw data na code, 30–150-second contradiction katika correlation window, kutokuonyeshwa mathematically kwa physical-law integration na kutotimizwa kwa five-consecutive-threshold-exceedance condition katika flights zilizowasilishwa. Kwa hiyo, method inapaswa kutathminiwa kama promising research prototype na haipaswi kuwasilishwa kama certified flight-safety au maintenance-decision system.