Seli za mafuta za utando wa kubadilishana protoni zinazotumiwa kwenye magari hufanya kazi kila mara chini ya hali za mzigo unaobadilika kutokana na kuongeza kasi, kupunguza kasi na mahitaji ya nguvu yanayobadilika. Hitilafu kama kukauka kwa utando, kuongezeka kwa maji kupita kiasi ndani ya seli au kutofika kwa oksijeni ya kutosha kwenye cathode mwanzoni zinaweza kusababisha mabadiliko madogo tu katika voltage ya kutoka ya seli ya mafuta. Kwa sababu ishara hizi dhaifu zinaweza kupotea ndani ya kelele ya kipimo na mabadiliko ya kawaida ya uendeshaji, hitilafu mara nyingi hutambuliwa baada tu ya kushuka kwa voltage kuanza kuwa dhahiri.

Katika utafiti huu, watafiti walitengeneza mtandao wa neva wenye taarifa za fizikia unaoitwa LS-PINN ambao hujifunza “hali fiche” ya vipimo vichache inayowakilisha hali ya ndani ya kielektrokemia ya seli ya mafuta ambayo haiwezi kuonekana moja kwa moja kutoka kwenye vipimo vya voltage vya mfululizo wa muda. Modeli inaunganisha residual temporal convolutional network, multi-head attention mechanism na bidirectional GRU layer zinazochakata mfululizo wa voltage, pamoja na polarization equations za seli ya mafuta na equivalent-circuit impedance model.

Majaribio yalifanywa kwenye PEM fuel-cell stack ambayo kila cell ilikuwa na eneo hai la 347 cm² na kwa jumla ilikuwa na cells 10. Hali za uendeshaji wa kawaida, kukauka kwa utando, kujaa maji na upungufu wa oksijeni ziliundwa katika current densities za chini, za kati na za juu. Voltage ilirekodiwa kwa sampling rate ya 4 Hz na resolution ya 0,001 V, huku impedance spectra zikichukuliwa kama sparse steady-state measurements kati ya 0,1 Hz na 10 kHz.

Jumla ya time windows 10.785 ziliundwa na experimental runs ziligawanywa, kabla ya window generation, takribani kuwa %70 training na %30 test. Fault labels za %20 pekee ya training samples zilitumika; unlabeled samples zilichangia kwenye physical reconstruction na consistency losses.

Full LS-PINN model ilipata accuracy ya %98,67 kwenye original test set na %98,62 kwenye class-balanced test set. Ikilinganishwa na best comparison model V-ECM-Fusion, accuracy increase ilikuwa 4,96 na 7,17 percentage points, mtawalia. Kutoka time-series model isiyo na physical constraints hadi full model, accuracy iliongezeka kutoka %85,91 hadi %98,67.

Total polarization resistance iliyotolewa na modeli kutoka voltage ilitoa anomaly indicator mapema kwa 13,5–17 seconds katika membrane drying, 29,5–38,5 seconds katika water accumulation na 1–1,5 seconds katika oxygen deficiency, kulingana na values zilizoripotiwa katika utafiti. Hata hivyo, resistance hii si impedance value iliyopimwa continuously; ni physical proxy indicator inayotolewa kutoka voltage na neural network iliyokalibrishwa kwa sparse EIS measurements.

Ingawa utafiti una matumaini, results zimewekewa mipaka na laboratory stack moja, load levels tatu na fault types tatu zilizoundwa kwa udhibiti. Time-window length, detailed dimensions za network layers, threshold na smoothing parameters pamoja na idadi ya sparse EIS matches hazijaelezwa kikamilifu. Long-term vehicle tests, combined faults, stack ageing, sensor degradation na environmental conditions tofauti bado hazijatathminiwa.

Kwa nini early fault detection ni ngumu katika PEM fuel cells?

Proton-exchange-membrane fuel cell ni mfumo changamano unaozalisha umeme kupitia electrochemical processes na unaopaswa kuitikia rapid load changes katika automotive applications. Cell voltage inaweza kupimwa kwa urahisi; lakini processes kama water content ya membrane, reaction activity kwenye electrode surface, oxygen transport na internal resistance haziwezi kupimwa moja kwa moja na continuously.

Faults katika hatua za mwanzo zinaweza kutoa ishara ambazo ni vigumu kuonekana kutoka nje kwa sababu zifuatazo:

• Amplitude ya voltage change inaweza kuwa ndogo.
• Fault signal inaweza kufanana na normal fluctuations zinazosababishwa na load changes.
• Measurement noise inaweza kuficha small voltage trends.
• Faults tofauti zinaweza kusababisha voltage drops zinazofanana kwenye output voltage.
• Kunaweza kuwa na time lag kati ya internal electrochemical change na external voltage response.

Kwa hiyo, approach ya “toa alarm tu voltage ikishuka chini ya value fulani” inaweza kukosa degradation iliyoanza mapema zaidi ndani ya mfumo. Lengo kuu la utafiti ni kufichua kupitia modeli mabadiliko yanayohusiana na internal resistance na mass transfer kabla voltage haijaonyesha wazi fault signal.

Ni faults zipi zilichunguzwa katika utafiti?

Mbali na normal operation, watafiti waliunda hali tatu kuu za fault:

Kukauka kwa utando

Water amount katika membrane na catalyst layer ikipungua, proton conductivity inaweza kudhoofika. Katika utafiti, hali hii iliundwa kwa kushusha relative humidity hadi %0 na kuongeza stack temperature ukilinganisha na normal condition. Membrane-drying fault inatarajiwa kusababisha slower cumulative change katika internal resistance na reaction polarization.

Kujikusanya kwa maji

Kujikusanya kwa liquid water ndani ya fuel cell kunaweza kuzuia oxygen kusafirishwa kupitia porous layers hadi catalyst surface. Katika utafiti, water accumulation iliundwa kwa lower stack temperature na %70 relative-humidity condition. Imeripotiwa kwamba fault hii huongeza oxygen mass-transfer resistance kwa haraka, hasa katika high load.

Upungufu wa oksijeni

Kupunguza kiwango cha hewa kinachopelekwa kwenye cathode hupunguza oxygen concentration na limiting-current capacity. Katika utafiti, oxygen deficiency iliundwa kwa kupunguza air stoichiometric ratio chini ya normal value. Kwa sababu fault hii huendelea kwa ghafla zaidi, time difference kati ya internal-mechanism indicator na external voltage response ilipatikana kuwa fupi zaidi kuliko faults nyingine.

Approach kuu ya utafiti ni ipi?

Online diagnostic input ya modeli ni voltage time series ya fuel-cell stack. Watafiti walitengeneza channels mbili kutoka voltage:

1. Raw voltage sequence,
2. First-order difference ya voltage.

Raw voltage sequence ya sample yenye namba s inaonyeshwa hivi:

\[ V^{(s)}=[v_1^{(s)},v_2^{(s)},\ldots,v_T^{(s)}] \]

Voltage-difference sequence ni hii:

\[ \Delta v_t^{(s)}= \begin{cases} 0, & t=1 \\ v_t^{(s)}-v_{t-1}^{(s)}, & t=2,\ldots,T \end{cases} \]

Channels mbili zikiunganishwa, model input huundwa:

\[ X^{(s)}=[V^{(s)},\Delta V^{(s)}]\in\mathbb{R}^{T\times2} \]

• \(T\): Ni idadi ya samples katika time window.
• \(v_t\): Ni stack voltage katika wakati t.
• \(\Delta v_t\): Ni mabadiliko kati ya voltage measurements mbili zinazofuatana.

First-difference channel inalenga kukamata local slope changes na abrupt transitions ambazo zinaweza kuonekana kabla ya mabadiliko wazi ya absolute voltage value.

Kauli ya “modeli inayotumia voltage pekee” inapaswa kuelewekaje?

Diagnostic na classification branch ya modeli itakayofanya kazi kwenye gari inaanzia voltage. Hata hivyo, hatua ya model development si voltage pekee kabisa:

• Current density na required operating quantities zilitumika kama boundary conditions za polarization equations.
• Sparse EIS measurements zilitumika kufundisha impedance-reconstruction branch.
• Fault labels zilitoa classification supervision katika %20 ya training samples.

Kwa hiyo, kauli “voltage ndiyo input pekee” inaeleza online diagnostic observation ya modeli. Haielezi sources zote za information zilizotumika katika learning process ya modeli. Ingawa EIS requirement huondoka wakati wa inference, physical impedance mapping hujifunzwa kwa EIS data zilizopimwa awali.

Feature-extraction network inafanyaje kazi?

Kwenye ukurasa wa 8 wa utafiti, Kielelezo 1 kinaonyesha LS-PINN workflow katika sehemu tatu kuu: feature extraction, physics-informed constraints na fault diagnosis.

Residual temporal convolutional network

ResTCN module hutoa short- na medium-scale local patterns katika voltage sequence. Residual-connection structure inaonyeshwa kwa relation hii:

\[ H_l^{(s)}= \sigma\left( \mathcal{F}_{TCN}^{(l)}(H_{l-1}^{(s)}) +\mathcal{P}^{(l)}(H_{l-1}^{(s)}) \right) \]

• \(H_{l-1}\): Ni features za previous layer.
• \(\mathcal{F}_{TCN}\): Ni transformation inayojifunzwa na temporal-convolution block.
• \(\mathcal{P}\): Ni dimension-matching au identity connection.
• \(\sigma\): Ni activation function.

Utafiti pia unasema kwamba Squeeze-and-Excitation channel weighting imeongezwa kwenye convolutional feature-extraction process. Hata hivyo, layer dimensions na reduction ratio za SE module hazijatolewa.

Multi-head self-attention

Multi-head self-attention mechanism hujaribu kujifunza mahusiano kati ya sehemu zilizo mbali katika time sequence. Kwa njia hii, small voltage deviation inayoonekana katika wakati fulani inaweza kutathminiwa pamoja na operating behavior ya awali.

Bidirectional GRU

BiGRU huchakata current observation window katika directions mbili na kuunganisha temporal context:

\[ \overrightarrow{h_t}= GRU_f(a_t,\overrightarrow{h_{t-1}}) \]

\[ \overleftarrow{h_t}= GRU_b(a_t,\overleftarrow{h_{t+1}}) \]

\[ h_t=[\overrightarrow{h_t};\overleftarrow{h_t}] \]

Backward-direction processing hapa haimaanishi lazima kutumia actual future vehicle data. Modeli pia huchakata kwa reverse direction time window ambayo tayari inapatikana kikamilifu katika wakati huo. Hata hivyo, utafiti haujatoa window length na exact position ya target time ndani ya window.

Latent state inawakilisha nini?

High-dimensional output ya feature-extraction network inabanwa kuwa vector ndogo:

\[ z_{raw}^{(s)}= f_{lat}(r^{(s)};\theta_z) \in\mathbb{R}^{d_z} \]

• \(r^{(s)}\): Ni feature representation iliyotolewa kutoka time sequence.
• \(z_{raw}^{(s)}\): Ni initial latent state.
• \(d_z\): Ni dimension ya latent space.

Latent state hii si physical variable iliyopimwa moja kwa moja. Ni low-dimensional representation inayoundwa na neural network kutoka common patterns katika voltage. Watafiti walilazimisha representation hii sio tu kutenganisha classes, bali pia kutoa parameters zinazolingana na physical equations za fuel cell.

Physical constraint katika time domain iliundwaje?

Parameters tatu za time domain hutolewa kutoka latent state:

• Exchange current density \(i_0\),
• Ohmic resistance ya polarization branch \(R_{\Omega,pol}\),
• Effective mass-transfer parameter \(k_{mt}\).

Fuel-cell voltage inareconstructiwa kupitia losses tatu kuu za polarization:

\[ \widehat{V} =E_{rev} -\eta_{act} -\eta_{ohm} -\eta_{conc} \]

• \(E_{rev}\): Ni reversible theoretical voltage.
• \(\eta_{act}\): Ni activation loss ya electrode reactions.
• \(\eta_{ohm}\): Ni ohmic loss inayotokana na membrane, contact na conductivity.
• \(\eta_{conc}\): Ni concentration loss inayotokana na limitations za reactant transport.

Activation overpotential ilimodeliwa kwa symmetric form ya Butler–Volmer approach:

\[ \eta_{act} =\frac{R_g\Theta}{\alpha F} \sinh^{-1}\left(\frac{i}{2i_0}\right) \]

• \(R_g\): Ni gas constant.
• \(\Theta\): Ni absolute temperature.
• \(\alpha\): Ni charge-transfer coefficient.
• \(F\): Ni Faraday constant.
• \(i\): Ni operating current density.

Ohmic loss inaonyeshwa hivi:

\[ \eta_{ohm}=iR_{\Omega,pol} \]

Relations zilizotumika kwa limiting current density na concentration loss ni hizi:

\[ i_L=n_eFc_{O_2}k_{mt} \]

\[ \eta_{conc} =-\frac{R_g\Theta}{n_eF} \ln\left(1-\frac{i}{i_L}\right) \]

• \(n_e\): Ni electron-transfer number.
• \(c_{O_2}\): Ni characteristic oxygen concentration.
• \(i_L\): Ni limiting current density.

Reconstructed voltage inalinganishwa na mean raw voltage ya window. Pia separate loss term hutumika kuzuia ratio ya \(i/i_L\) kukaribia boundary isiyofaa kifizikia.

Impedance constraint katika frequency domain iliundwaje?

Group ya pili ya physical parameters hutolewa kutoka latent state:

• Equivalent-circuit ohmic resistance \(R_{\Omega,ecm}\),
• Resistances za polarization branches mbili \(R_1\) na \(R_2\),
• Constant-phase-element magnitudes \(Q_1\) na \(Q_2\),
• Constant-phase exponents \(\kappa_1\) na \(\kappa_2\).

Second-order constant-phase-element equivalent-circuit model inaonyeshwa hivi:

\[ \widehat{Z}(\omega) =R_{\Omega,ecm} +\sum_{b=1}^{2} \left[ \frac{1}{R_b} +Q_b(j\omega)^{\kappa_b} \right]^{-1} \]

• \(\widehat{Z}(\omega)\): Ni reconstructed complex impedance.
• \(\omega\): Ni angular frequency.
• \(j\): Ni imaginary unit.
• \(R_b,Q_b,\kappa_b\): Ni parameters za polarization branch namba b.

Ili kuzuia branches mbili kubadilishana nafasi, characteristic relaxation times hupangwa:

\[ \tau_b=(R_bQ_b)^{1/\kappa_b}, \qquad \tau_1<\tau_2 \]

Kulingana na interpretation inayoungwa mkono na DRT katika utafiti:

• \(R_1\): Ni effective polarization resistance ya shorter time scale, inayohusishwa zaidi na electrode interface na charge transfer.
• \(R_2\): Ni effective polarization resistance ya longer time scale, inayohusishwa zaidi na oxygen mass transfer.

Watafiti wanasisitiza kwamba mapping hii si absolute. Resistances hizi mbili ni aggregated effective parameters zinazojumuisha physical processes nyingi kwa pamoja.

Total polarization resistance ilifafanuliwaje?

Early-anomaly indicator ya modeli ni jumla ya polarization branches mbili:

\[ R_{total}=R_1+R_2 \]

Ohmic resistance haikuongezwa katika jumla hii. Sababu ni kwamba ohmic resistance huamua high-frequency intercept ya impedance curve; haiwakilishi diameter ya polarization arcs.

\(R_{total}\) ni model-based indicator inayowakilisha combined change ya interfacial reaction resistance na oxygen-transport resistance. Si sensor value iliyopimwa real time.

Time na frequency domains ziliunganishwaje?

Time-domain na frequency-domain branches hukadiria ohmic resistance independently. Ili kuzuia estimates mbili kutoa results zinazopingana kabisa, normalized consistency loss ilitumika:

\[ \mathcal{L}_{\Omega} =\frac{1}{N} \sum_{s=1}^{N} \left[ \frac{ R_{\Omega,pol}^{(s)}-R_{\Omega,ecm}^{(s)} }{ 0.5\left( R_{\Omega,pol}^{(s)}+R_{\Omega,ecm}^{(s)} \right)+\varepsilon } \right]^2 \]

Constraint hii hailazimishi physical models mbili kutoa exactly same resistance value; lakini inapunguza physically contradictory outputs kutoka common latent state.

Physical information ilihamishwaje kwenye classification?

Physical parameters hazikutumika tu kama training loss, bali ziliongezwa moja kwa moja kwenye final form ya latent state:

\[ z^{(s)} =z_{raw}^{(s)} +\psi\left( [z_{raw}^{(s)};g_{phy}(P^{(s)})] \right) \]

• \(P^{(s)}\): Ni time- na frequency-domain parameters zinazotolewa na modeli.
• \(g_{phy}\): Hutoa mechanism representation kutoka physical parameters.
• \(\psi\): Huunganisha raw latent state na physical representation.
• \(z^{(s)}\): Ni final latent state iliyosahihishwa kifizikia.

Fault class inakokotolewa kutoka final representation hii:

\[ \widehat{y}^{(s)} =\operatorname{softmax}(h_{cls}(z^{(s)})) \]

Total training loss ina components zipi?

Modeli ilifundishwa kwa combined objective function ifuatayo:

\[ \mathcal{L} =\lambda_{cls}\mathcal{L}_{cls} +\lambda_V\mathcal{L}_V +\lambda_{lim}\mathcal{L}_{lim} +\lambda_Z\mathcal{L}_Z +\lambda_{param}\mathcal{L}_{param} +\lambda_{\Omega}\mathcal{L}_{\Omega} \]

Katika classification branch, focal cross-entropy loss inayotoa weight kubwa zaidi kwa difficult samples na label smoothing zilitumika. Loss weights zilichaguliwa kwa Bayesian optimization kulingana na validation-set performance.

Semi-supervised training ilitekelezwaje?

Fault labels zilihifadhiwa katika %20 pekee ya samples kwenye training set. Samples zilizobaki hazikushiriki moja kwa moja katika classification loss; lakini ziliendelea kutumika katika physical tasks zifuatazo:

• Voltage reconstruction,
• Ohmic-resistance consistency,
• Parameter regularization,
• Limiting-current constraint.

Ni sparse voltage windows pekee zilizolinganishwa na real EIS ndizo zilichangia impedance-reconstruction loss. Kwa windows zisizo na EIS, frequency-domain loss haikukokotolewa.

%2 Gaussian perturbation ilitumika kwenye training EIS data, huku test EIS measurements zikiachwa bila kubadilishwa. Process hii haitengenezi new physical experiment; inatumia tu small noisy derivatives za spectra zilizopo katika training.

Experimental stack iliundwaje?

Kwenye ukurasa wa 23 wa utafiti, Kielelezo 3 kinaonyesha PEM fuel-cell stack, hydrogen na air inlet-outlet, cooling loop, humidification system, electronic load na monitoring computer.

Fault conditions ziliundwaje?

Kila operating condition ilirudiwa mara tatu. Mfumo kwanza uliendeshwa kwa sekunde 1.800 katika steady normal state, kisha fault ikafuliwa hatua kwa hatua kwa kubadilisha control variables.

Samples ziligawanywaje?

Experimental runs ziligawanywa takribani kwa ratio ya 7:3 kuwa training na test sets kabla ya kuunda sliding windows. Neighboring windows kutoka experimental run ileile ziliwekwa katika subset ileile. Approach hii huzuia nearly identical neighboring time windows kuingia katika training na test set kwa pamoja.

Time windows hizi 10.785 hazimaanishi physical experiments huru 10.785. Windows nyingi zimetengenezwa kutoka sequential na partially overlapping sections za experimental runs zilezile. Independent experimental unit ni experimental run zaidi kuliko time window.

Voltage reconstruction ilionyesha nini?

Kwenye kurasa za 26 na 27 za utafiti, Kielelezo 4 kinalinganisha raw voltage na voltage iliyoreconstructiwa kutoka physical equations katika load levels tatu na fault conditions tatu. Reconstructed curves zilifuata overall voltage drops kwa karibu na zililainisha noise zaidi.

Imeripotiwa kwamba reconstructed voltage ilitambua anomaly takribani sekunde 1 mapema kwa wastani kuliko raw voltage. Hata hivyo, katika oxygen-deficiency experiment ya low current density, reconstructed voltage ilivuka threshold sekunde 0,5 baada ya raw voltage. Watafiti wanaeleza hili kwa oxygen demand kuwa ndogo katika low load na reconstructed curve kuwa smoother.

EIS reconstruction ilionyesha nini?

Kwenye ukurasa wa 28 wa utafiti, Kielelezo 5 kinalinganisha measured na reconstructed EIS curves katika low, medium na high current densities. Modeli ilikamata general shape na magnitude ordering ya polarization arcs katika hali nne.

Katika medium load, agreement katika low na medium frequency regions ilionekana kuwa stronger, wakati katika high load baadhi ya local mismatches ziliongezeka. Hata hivyo, reconstructed impedance curves za modeli zilihifadhi same general direction of change kama measured curves.

Comparison hii ilifanywa tu katika sparse steady-state points zilizolinganishwa na real EIS. Hakuna impedance curve iliyopimwa kila sekunde ya fault transition.

Total polarization resistance iliitikiaje faults?

Kwenye ukurasa wa 32 wa utafiti, Kielelezo 7 kinaonyesha \(R_{total}\) curve iliyotolewa na modeli kutoka kila voltage window katika load levels tatu.

• Katika membrane drying: Resistance iliongezeka polepole zaidi na kwa kuendelea.
• Katika water accumulation: Kulionekana rise kali zaidi iliyohusishwa na oxygen-transport pathways kuzuiwa na liquid water.
• Katika oxygen deficiency: Kulikuwa na continuous increase iliyofuata initial deviation.

Load ilipoongezeka, resistance differences kati ya fault conditions zilionekana wazi zaidi. Hasa water accumulation na oxygen deficiency zilizalisha sharper impedance changes katika high current density.

Modeli ilitoa onyo mapema kiasi gani?

Kuna numerical inconsistency katika membrane-drying row ya low load. Raw voltage kuvuka threshold katika sekunde 1830,5 na \(R_{total}\) indicator kuvuka katika sekunde 1805,5 kunatoa tofauti ya sekunde 25; lakini table inaandika sekunde 15. Kwa hiyo, exact early-warning time ya row hiyo haiwezi kuthibitishwa.

Katika rows nyingine, reported differences zinalingana na arithmetic difference kati ya raw-voltage time na \(R_{total}\) time.

Early warning ya sekunde 38,5 inamaanisha nini?

Result hii inamaanisha kwamba katika controlled water-accumulation experiment ya high current density, \(R_{total}\) curve iliyotolewa na modeli ilivuka threshold yake sekunde 38,5 kabla ya anomaly threshold iliyowekwa kwa raw voltage.

Result haimaanishi yafuatayo:

• Haimaanishi kwamba independent sensor ilithibitisha actual physical fault ya fuel cell ilianza exactly sekunde 38,5 mapema.
• Haimaanishi kwamba continuous EIS measurement ilitoa alarm sekunde 38,5 mapema.
• Haimaanishi kwamba muda uleule utapatikana katika kila water-accumulation event.
• Haimaanishi kwamba driver au vehicle-control system imepewa direct guaranteed intervention time ya sekunde 38,5.

Early-warning time inategemea threshold iliyokokotolewa kutoka normal period, smoothing ya curve na condition ya kubaki juu ya threshold. Kwa kuwa utafiti haujatoa kikamilifu details kama smoothing window na idadi ya consecutive threshold crossings, result hii haiwezi kukokotolewa independently kutoka data ileile.

Latent state iliboresha classification kiasi gani?

Watafiti walilinganisha LS-PINN yenye low-dimensional latent state na RV-PINN inayotumia feature extractor na physical branches zilezile lakini huclassify high-dimensional features moja kwa moja.

Kwenye ukurasa wa 35 wa utafiti, Kielelezo 8 kinaonyesha kwamba bila latent state kulikuwa na confusion zaidi hasa kati ya water accumulation na oxygen deficiency.

Hata hivyo, class-specific rates katika section text zinalingana na values za balanced test matrix iliyotolewa baadaye badala ya original test matrix katika Kielelezo 8. Katika Kielelezo 8 LS-PINN class rates ni takribani %98,6, %99,7, %98,1 na %98,4, wakati text inaandika %98,7, %99,5, %97,9 na %98,4. Small difference hii inaashiria kwamba class rates za test sets mbili huenda zilichanganywa katika narrative.

Ablation experiment ilionyesha components zipi zinafaa?

Kila model ilirudiwa kwa random initializations kumi tofauti. Box plots kwenye ukurasa wa 37 wa utafiti zinaonyesha kwamba accuracy iliongezeka kwa utaratibu kadiri components zilivyoongezwa.

Kutoka pure time-series network M3 hadi full physics-informed M6 kulitoa increase ya 12,76 percentage points. Impedance constraint peke yake ilitoa accuracy ya 1,70 percentage points juu kuliko polarization constraint. Constraints mbili zilipotumika pamoja, result ilikuwa juu kuliko matumizi yao kila moja peke yake.

Comparison na models nyingine

LS-PINN ilitoa accuracy ya 4,96 percentage points juu katika original test na 7,17 percentage points juu katika balanced test kuliko strongest comparison model V-ECM-Fusion.

Difference ya LS-PINN kati ya original na balanced test ni percentage points 0,05 tu. Result hii inaonyesha kwamba modeli ilinufaika kidogo kuliko methods nyingine kutokana na normal class kuwa nyingi. Hata hivyo, balanced test set iliundwa kwa random subsampling; variability katika subsample selections tofauti haijatolewa kando.

Nguvu za utafiti ni zipi?

• Controlled experiment ilifanywa kwenye real PEM fuel-cell stack.
• Load levels tatu tofauti na typical faults tatu zilipimwa.
• Experimental runs ziligawanywa kuwa training na test kabla ya windowing.
• Modeli haikutoa fault class pekee; ilifanya voltage na EIS reconstructions pia.
• EIS measurements hazikutolewa kama direct input kwa test classification.
• Semi-supervised training ilitumika kwa %20 labels pekee.
• Ablation models zilirudiwa kwa random initializations kumi.
• Original na class-balanced test sets zilitathminiwa tofauti.
• Physical interpretation ya total polarization resistance iliungwa mkono na DRT analysis.
• Utafiti unakubali wazi kwamba continuous resistance indicator si measured EIS bali ni model-based proxy.

Mapungufu makuu ya utafiti ni yapi?

• Results zinahusu laboratory stack moja yenye cells 10.
• Faults ziliundwa kwa controlled temperature, humidity na air-stoichiometry changes.
• Vibration, ambient temperature, road conditions na control-system interactions katika actual vehicle hazijapimwa.
• Combined au simultaneous faults hazijatathminiwa.
• Stack ageing na long-term material degradation hazijajumuishwa katika modeli.
• EIS haikupimwa continuously wakati wa fault transition; ilichukuliwa sparsely tu katika steady states.
• \(R_{total}\) curve si direct measurement, ni model estimate iliyotolewa kutoka voltage.
• Total number ya windows zilizolinganishwa na sparse EIS haijaripotiwa.
• Time-window length na overlap ratio hazijaelezwa.
• ResTCN layers, filter counts, attention heads, BiGRU units na latent-state dimension hazijaelezwa.
• Optimizer, learning rate, batch size na number of training epochs hazijaripotiwa wazi.
• Exact formula ya early-anomaly threshold, smoothing window na persistence condition hazijatolewa.
• Kuna arithmetic inconsistency katika low-load membrane-drying row ya Table 4.
• Class-specific LS-PINN rates zilizotolewa katika sections mbili tofauti zina small differences.
• Hakuna open-access link ya source code, model weights na raw experimental data.
• Inference time, memory use na computational cost katika automotive control unit hazijapimwa.
• False-positive na false-negative rate za alarm hazijaripotiwa katika separate early-warning experiment.

Utafiti unaunga mkono nini?

• Unaunga mkono kwamba physics-informed latent-state learning inayotumika kwenye voltage inaweza kutoa strong result katika controlled fault classification,
• Unaunga mkono kwamba sparse EIS supervision inaweza kutumika katika training ya model inayotoa physical impedance parameters kutoka voltage,
• Unaunga mkono kwamba polarization na impedance constraints zinaweza kuboresha class separation ikilinganishwa na time-series features pekee,
• Unaunga mkono kwamba kuna experimental-model evidence kuwa total polarization resistance inaweza kubadilika mapema kuliko external voltage katika baadhi ya gradual water-management faults,
• Unaunga mkono kwamba unlabeled voltage samples zinaweza kutumiwa katika training kupitia physical losses.

Utafiti hauthibitishi nini?

• Hauthibitishi kwamba accuracy ya %98,67 itapatikana katika automotive PEM fuel cells zote.
• Hauonyeshi kwamba kila water-accumulation fault inaweza kutambuliwa sekunde 38,5 mapema.
• Hauthibitishi kwamba \(R_{total}\) curve ni sawa kila wakati na physical resistance iliyopimwa real time.
• Hauonyeshi kwamba modeli inaweza kutenganisha compound faults.
• Hauhakikishi same result katika aged stacks au membrane na catalyst structures tofauti.
• Hauthibitishi kwamba false-alarm rate inakubalika kwa real vehicle use.
• Hauonyeshi kwamba modeli inaweza kufanya kazi real time kwenye automotive-grade embedded hardware.
• Hauonyeshi kwamba thresholds zilizoamuliwa katika utafiti zinaweza kuhamishwa moja kwa moja kwenda stack nyingine.

Inapaswa kutathminiwaje kwa Uturuki?

Utafiti hautoi social au regional data maalum kwa nchi fulani; unatengeneza general engineering method kwa fuel-cell vehicles na hydrogen-based energy systems. Kwa hiyo, results ni pendekezo la method linaloweza kupimwa kwa research centers, automotive manufacturers na energy-technology developers nchini Uturuki.

Ili mfumo sawa utathminiwe nchini Uturuki, hatua zifuatazo zinahitajika:

• Kukusanya data katika local au commercial PEM stacks zenye power na cell counts tofauti,
• Kupima athari ya hot na cold climate conditions,
• Kuongeza long-term ageing na dynamic vehicle cycles,
• Kuunda combined water, air na temperature faults,
• Kufunga thresholds katika independent test runs,
• Kupima latency na resource consumption ya modeli katika actual vehicle-control hardware,
• Kutathmini false-alarm cost pamoja na safety strategies,
• Kuthibitisha model-based \(R_{total}\) indicator kwa online au fast EIS.

Numerical accuracy na early-warning times katika utafiti haziwezi kutumika moja kwa moja kama design au safety limit bila independent validation kwenye new stacks.

Mbinu na Matokeo ya Utafiti

Technical-method summary

Main quantitative findings

Additional studies zitakazoongeza reliability ya results

• Kuchapisha experimental data na source code,
• Kueleza time window na network hyperparameters zote,
• Kutoa idadi na distribution ya sparse EIS matches,
• Kuripoti threshold na smoothing algorithm kikamilifu,
• Kutathmini early-warning performance pamoja na false-alarm rate,
• Kufanya external validation kwenye fuel-cell stacks tofauti,
• Kupima \(R_{total}\) curve kwa real-time au more frequent EIS measurements,
• Kuchunguza combined faults na changing fault severities,
• Kuongeza ageing, sensor drift na environmental disturbances,
• Kupima inference time na energy consumption kwenye embedded hardware,
• Kukokotoa confidence intervals kutoka independent experimental runs,
• Kusahihisha early-warning arithmetic katika Table 4.

Maelezo ya Chanzo na Mbinu

• Jina asili la utafiti: Physics-Informed Latent State Learning with Polarization and Impedance Constraints for Early Anomaly Identification and Fault Diagnosis of Automotive PEM Fuel Cells
• Waandishi: Jiaming Zhang; Changqing Du; Fuwu Yan; Zhe Li; Mingyang Nie; Jing Zhu; Liangdong Li; Wenchao Zhu
• Mpangilio wa waandishi: Waandishi wametolewa hapo juu kwa mpangilio wa maandishi ya utafiti.
• Equal contribution au equal first authorship: Hakuna statement.
• Mwandishi wa mawasiliano: Fuwu Yan
• Barua pepe ya mwandishi wa mawasiliano: yanfw@whut.edu.cn
• Taasisi ya kwanza: Hubei Key Laboratory of Advanced Technology for Automotive Components, Wuhan University of Technology, Wuhan, China
• Taasisi ya pili: Foshan Xianhu Laboratory, Foshan, China
• Taasisi ya tatu: Nanyang Technological University, School of Mechanical and Aerospace Engineering, Singapore
• Mwaka wa utafiti: 2026
• Exact publication date: Hakuna tarehe kwenye title page.
• Aina ya chanzo: Preprint research text ambayo haijapitia peer review
• Jukwaa la uchapishaji: SSRN
• Preprint DOI: 10.2139/ssrn.7195879
• Rekodi rasmi:Ukurasa rasmi wa utafiti wa SSRN
• Permanent DOI link:10.2139/ssrn.7195879
• Jarida: Haijathibitishwa kwamba imechapishwa katika peer-reviewed journal.
• Hali ya mapitio ya kitaalamu: Kila ukurasa wa utafiti unaeleza kwamba haujapitia peer review.
• Ufadhili: National Key R&D Program of China, 2022YFB4003703; Foshan Xianhu Laboratory of the Advanced Energy Science and Technology Guangdong Laboratory, XHRD2024-11233100-01
• Mgongano wa maslahi: Hakuna separate conflict-of-interest statement katika text iliyochunguzwa.
• Data access: Hakuna open-access link iliyotolewa kwa raw experimental data.
• Code na model weights: Hakuna open-source link iliyotolewa.

Scientific content ya makala hii ya Verianla imeandaliwa kwa kutegemea tu text ya uploaded study, equations, tables, experimental conditions, architecture schematics, impedance graphs, early-warning analyses, ablation results na comparative experiments. External sources zilitumika tu kwa bibliographic verification ya author identity, institution, DOI, platform na publication status; hakuna scientific finding iliyoongezwa kutoka external sources.

Kauli ya utafiti ya “kutumia voltage kama diagnostic input pekee” ni sahihi kwa online inference stage. Wakati wa training, current density na operating quantities nyingine zilitumika kama boundary conditions katika physics equations, na sparse EIS measurements kama frequency-domain supervision.

Continuous \(R_{total}\) curve si impedance iliyopimwa real time. Ni model-based physical proxy inayotolewa kutoka latent state ya kila voltage window na equivalent-circuit parameter-mapping branch iliyofundishwa kwa sparse EIS measurements. Results section ya utafiti pia inakubali limitation hii.

Katika Table 4, times zilizotolewa kwa membrane drying katika low current density hazipatani kihisabati na “sekunde 15” early-warning value. Pia, class rates katika narrative text ya latent-state comparison zinafanana zaidi na balanced test matrix kuliko original test matrix ya section ileile. Pointi hizi hazijasahihishwa kimya kimya.

Utafiti haujapitia peer review. Kwa kuwa findings zimewekewa mipaka na laboratory stack moja na controlled fault scenarios, actual automotive application inahitaji additional validation chini ya independent stacks, long-term vehicle use, environmental disturbances, ageing na combined faults.