Utafiti huu unapendekeza mfumo wa EMS–FDI–FTC unaounganisha usimamizi wa nishati, utambuzi na utengaji wa hitilafu, pamoja na udhibiti unaostahimili hitilafu ndani ya usanifu mmoja wa mzunguko uliofungwa kwa magari ya umeme mseto ya hidrojeni yanayotumia kwa pamoja fuel cell (FC), betri (Bat) na supercapacitor (SC). Tofauti kuu ya mfumo ni kwamba hauishii tu kwenye kugundua hitilafu. Baada ya hitilafu kuainishwa, matokeo hutumwa kwenye mfumo wa usimamizi wa nishati na mzigo wa chanzo chenye hitilafu hupunguzwa huku nguvu ya traction inayohitajika ikigawiwa upya kwa vyanzo vilivyo salama.

Badala ya standardized synthetic driving cycles, utafiti umetumia njia tatu tofauti za kuendesha zilizonaswa kwa GPS kwenye barabara halisi nchini Tunisia. Taarifa za njia kama kasi ya gari na mteremko wa barabara zilitumika kukokotoa traction-power demand; tabia za umeme za fuel cell, betri na supercapacitor ziliwakilishwa kwa simplified models zinazofaa kwa real-time control.

Tabaka la utambuzi wa hitilafu linaunganisha mechanisms tatu zinazokamilishana: EKF-based residual/innovation monitoring, CUSUM change detection na CNN–LSTM-based AI classification. CNN layers hutoa local short-term fault patterns, wakati LSTM layer hutathmini jinsi signals hizi zinavyobadilika kwa muda. Ili kupima uwezo wa model kugeneralize katika route conditions tofauti, leave-one-route-out (LORO) validation ilitumika; yaani model ilitrainiwa kwa routes mbili na kutestiwa kwenye route ya tatu ambayo haikuiona wakati wa training.

Window-level LORO classification accuracy iliripotiwa kuwa %87,04 kwa Route 1, %74,78 kwa Route 2 na %68,52 kwa Route 3. Route 2 yenye dynamics zaidi ilitoa higher power peaks na hitaji kubwa la energy redistribution, huku Route 3 iliyokuwa smoother ikionyesha lower energy deficit na hydrogen consumption. Result hii pia inaonyesha engineering implication muhimu: dynamic driving profiles zinaweza kufanya faults zionekane zaidi kwa diagnosis, lakini huweka energy-management system katika demanding conditions zaidi.

Kwa embedded implementation, EMS–FDI–FTC controller iliendeshwa kama processor-in-the-loop (PIL) kwenye STM32F407VG, ARM Cortex-M4 168 MHz processor. Control loop iliendeshwa kwa 10 Hz, yaani sampling interval ya 100 ms; mean computation time ilikuwa 1,6 ms na worst-case computation time ilipimwa kuwa chini ya 6,2 ms. Result hii inaonyesha kwamba control architecture iliyotumiwa inaweza kukamilisha 100 ms supervisory-control loop kwa wakati kwenye processor hiyo.

Kwa mtazamo wa Uturuki: Study haichunguzi driving routes za Uturuki wala specific hydrogen-vehicle platform iliyotengenezwa nchini Uturuki. Hata hivyo, method inatoa applicable architecture example kwa research na engineering teams zinazofanya kazi kwenye hydrogen fuel-cell vehicles, electric powertrains, battery–supercapacitor hybrid storage, onboard fault diagnosis na embedded automotive electronics. Kwa transfer kwenda vehicle ya Uturuki, local driving data, real FC–Bat–SC component characteristics, temperature effects, aging behavior, converter losses na real fault data lazima zibainishwe upya. Energy values katika source hazipaswi kutumiwa kama experimental consumption values za specific commercial vehicle.

Kwa nini route-aware fault diagnosis inahitajika?

Katika hybrid hydrogen electric vehicle, power haitoki kwenye source moja. Katika architecture iliyochunguzwa, proton-exchange-membrane fuel cell ndiyo main energy source, betri hushughulikia medium-term power changes, na supercapacitor hushughulikia fast transient power demands na regenerative-braking pulses.

Structure hii ina advantage muhimu; lakini pia inaongeza complexity ya control problem kati ya energy sources, DC/DC converters na sensors. Fuel-cell voltage ikishuka, battery internal resistance ikiongezeka au supercapacitor capacitance ikipungua, kuendelea na nominal energy-distribution strategy bila mabadiliko kunaweza kuvuruga power balance, availability na operating conditions za components.

Kwa hiyo, main approach ya watafiti ni kujenga chain hii:

Real route → traction-power demand → sensor measurements → residual signals → EKF/CUSUM → CNN–LSTM classification → fault decision → EMS reconfiguration → power transfer kwenda healthy sources.

Jambo muhimu katika structure hii ni kwamba diagnosis na control hazijaachwa kama modules mbili tofauti. Fault decision hubadilisha moja kwa moja operating mode ya energy-management system.

Four-layer system architecture

Proposed structure imefafanuliwa katika main layers nne:

1. Sensing na data acquisition: voltages, currents, battery SOC, temperatures, hydrogen flow na relevant measurement channels hufuatiliwa.
2. Route-aware demand prediction: GPS-derived speed na road-grade information hutumika kukokotoa traction-power demand.
3. Fault detection na isolation: model-based residuals, CUSUM, EKF innovation indicators na CNN–LSTM classification hutumiwa pamoja.
4. Energy management na fault-tolerant control: power sharing kati ya FC, battery na supercapacitor hupangwa upya kulingana na diagnosis result.

Traction power inakokotolewaje kutoka route data?

Vehicle traction-power demand huhesabiwa kwa kutumia speed \(v(t)\), acceleration na road grade zinazotokana na GPS. Longitudinal vehicle model iliyotumiwa katika source ni:

\[ P_{Load}(t)= \left[ m\dot{v} +mg\sin\theta +\frac{1}{2}\rho C_d A v^2 +mgC_r\cos\theta \right]v(t) \]

.

Equation hii inaunganisha force inayohitajika kwa acceleration, gravity component kutokana na slope, aerodynamic drag na rolling resistance katika single traction-power demand. Hivyo energy-management system ileile inaweza kupewa different load excitations kwenye different road profiles.

Fuel-cell model

Fuel-cell terminal voltage ilimodeliwa kwa lumped equivalent circuit kwa control purposes:

\[ V_{FC}=E_N-R_{FC}I_{FC} \]

na output power:

\[ P_{FC}=V_{FC}I_{FC} \]

.

Hydrogen consumption huhesabiwa kupitia relation kati ya electrical output ya fuel cell, lower heating value ya hydrogen na assumed fuel-cell efficiency:

\[ \Delta m_{H_2}= \frac{\max(P_{FC},0)\Delta t} {\eta_{FC}\cdot LHV_{H_2}} \]

Kwa hiyo, hydrogen consumption katika article si experimental hydrogen consumption iliyochukuliwa moja kwa moja kutoka fuel meter ya real commercial vehicle; ni model-based indicator.

Battery model

Battery terminal voltage:

\[ V_{Bat}=E_0-R_{Bat}I_{Bat} \]

na battery power:

\[ P_{Bat}=V_{Bat}I_{Bat} \]

huhesabiwa.

State of charge huupdate kwa Coulomb counting:

\[ SOC_{Bat}(t+\Delta t) = SOC_{Bat}(t) - \frac{I_{Bat}\Delta t}{C_{Bat}} \]

Supercapacitor model

Supercapacitor voltage imewakilishwa na:

\[ V_{SC}(t)=V_{SC0}+\frac{1}{C_{SC}}\int I_{SC}(t)\,dt \]

. Main role ya supercapacitor ni kubuffer fast power changes ambazo si desirable kuachwa kwa fuel cell na battery.

Power balance kwenye DC link

Energy sources zote tatu zimeunganishwa kwenye traction system kupitia common DC-link. Basic power continuity ya system imefafanuliwa kama:

\[ P_{dem}=P_{FC}+P_{Bat}+P_{SC}-P_{loss} \]

.

Upper power limits zilizotumiwa ni 35 kW kwa fuel cell, 30 kW kwa battery na 15 kW kwa supercapacitor. Nominal voltages zimetolewa kama 60 V kwa FC na 48 V kwa battery na SC. Battery energy capacity ni 500 Wh, huku equivalent supercapacitor capacitance ikiwa 120 F.

Fault inatambuliwaje?

Katika hatua ya kwanza ya model-based diagnosis, difference kati ya measured signal na signal iliyopredicted na model huhesabiwa:

\[ r_i(t)=y_i(t)-\hat{y}_i(t) \]

Katika normal operation, residual signals zinatarajiwa kubaki karibu zero. Sudden load changes zinaweza kutoa short-term deviations; katika fault scenarios za study, persistent residual changes hutokea.

Kwa nini CUSUM ilitumika?

CUSUM hufuatilia accumulation ya small lakini persistent mean changes kwa muda. Positive na negative CUSUM statistics katika source zimetumika kama:

\[ g^+(k)=\max\{0,g^+(k-1)+r(k)-k_d\} \]

\[ g^-(k)=\max\{0,g^-(k-1)-r(k)-k_d\} \]

. CUSUM drift parameter imewekwa 0,3 na decision threshold 3,5.

Role ya EKF ni nini?

Extended Kalman Filter (EKF) hutumika kama lightweight observer anayefuatilia nominal electrical behavior online. EKF state vector ni:

\[ x=[V_{FC},V_{Bat},SOC_{Bat},V_{SC}]^T \]

na measurement vector:

\[ y=[V_{FC},V_{Bat},V_{SC},V_{dc}]^T \]

.

Equivalent resistance ya fuel cell, battery internal resistance au supercapacitor capacitance hazikadiriwi directly katika kila step. Reflection ya changes zilizowekwa kwenye parameters hizi katika residual na EKF innovation signals hutumika kwa diagnosis.

CNN–LSTM AI layer inafanyaje kazi?

Six input channels za AI classifier katika kila time step ni:

\[ x(t)= [r_{FC},r_{Bat},r_{SC},I_{FC},I_{Bat},P_{NS}]^T \]

.

Signals zilisampleiwa kwa 10 Hz. Ingawa nominal window duration ilifafanuliwa kuwa 2 seconds na window advance step 0,5 seconds, minimum samples 32 zilitumika kutokana na network input requirement; hivyo effective window length ilikuwa karibu 3,2 seconds.

Kila window ilinormalized channel-wise kwa z-score:

\[ X_w= \frac{X_w-\mu_w}{\sigma_w+\varepsilon} \]

CNN–LSTM network ina one-dimensional convolution blocks mbili, zikifuatiwa na LSTM na SoftMax classifier. Kernel size katika convolution blocks ni 3 na filter counts ni 16 na 32 mtawalia. Hizi hufuatiwa na batch normalization, ReLU na stride 2 max pooling. LSTM layer ina 64 hidden units, ikifuatiwa na 0,3 dropout.

Training ilitumia Adam optimization, initial learning rate ya \(10^{-3}\), 32 mini-batch na maximum 30 epoch. Loss function ni categorical cross entropy.

Kwa nini LORO ilitumika kuzuia data leakage?

Katika road data, consecutive time windows zinaweza kufanana sana. Neighboring windows kutoka route ileile zikiwa katika training na test groups kwa pamoja zinaweza kufanya true generalization performance ya model ionekane juu kuliko ilivyo.

Ili kuzuia hili, study ilitumia leave-one-route-out approach. Model ilitrainiwa kwa data za routes mbili, huku route iliyobaki ikiwekwa kama completely unseen test route. Process ilirudiwa kwa three different test routes.

Method hii ni evaluation ngumu zaidi kuliko random window splitting kwa sababu wakati wa test model hukutana na new driving dynamics na different residual distribution.

Fault decision haikuachwa kwa classification moja

CNN–LSTM output huzalisha probability vector. Katika source, fault class inakubaliwa tu ikiwa maximum probability ni angalau 0,80 na class ileile inadumu katika three consecutive overlapping windows.

Kwa window advance step ya 0,5 seconds, decision mechanism hii hutafuta continuity ya karibu 1,5 seconds. Lengo ni kuzuia short-term load changes au measurement fluctuations kuwekewa alama moja kwa moja kama persistent fault.

Energy management inafanya nini fault inapogunduliwa?

Baada ya fault kuthibitishwa, system hubadilika kutoka normal operating mode kwenda degraded operating mode. Power demand kwenye faulty energy source hupunguzwa na remaining demand huhamishiwa healthy sources:

\[ \begin{bmatrix} P^*_{FC}\\ P^*_{Bat}\\ P^*_{SC} \end{bmatrix} = \begin{bmatrix} \alpha_{FC}\\ \alpha_{Bat}\\ \alpha_{SC} \end{bmatrix} P_{dem} \]

Coefficients za \(\alpha\) hubadilishwa kulingana na diagnosed fault type, severity na source availability.

Katika multi-objective cost function ya energy management, hydrogen consumption, battery SOC change, supercapacitor power use na fault penalty huzingatiwa pamoja:

\[ J= \int_0^T \left( w_1\dot{m}_{H_2} +w_2(\Delta SOC_{Bat})^2 +w_3(P_{SC})^2 +w_4F_{fault} \right)dt \]

Optimization hufanywa chini ya battery SOC limit, DC-link voltage limit na FC, Bat na SC power limits.

Mbinu na Matokeo ya Utafiti

Real route data

Study ilitumia GPS-based speed data kutoka real drives kati ya Faculty of Sciences of Tunis (FST) na National School of Engineers of Tunis (ENSIT) nchini Tunisia.

GPS data ziliresampleiwa hadi 10 Hz, short data gaps zikajazwa kwa linear interpolation na 0,5-second moving-average filter ikatumika. Road grade ilitolewa kutoka filtered elevation information kwa finite differences na kupitishwa kwenye low-pass filter ili kupunguza fluctuations zinazotokana na GPS quantization.

Injected faults

Three main fault scenarios ziliingizwa kwenye system kama parameter changes. Kila fault ilitumika peke yake na ilibaki persistent hadi mwisho wa driving cycle baada ya kuanza.

Approach hii inatoa repeatable controlled fault test; lakini si real aging au field fault iliyotokea kwenye physical component.

CNN–LSTM na energy results kwa route

Route 1 ilitoa highest classification accuracy. Route 2, kutokana na higher speed changes na traction-power peaks, iliunda most demanding energy-management scenario; highest ENS, mean PNS na hydrogen consumption ziliripotiwa kwenye route hii. Route 3 ilitoa lowest ENS, mean PNS na hydrogen consumption kutokana na smoother driving profile; lakini Route 3 pia ndiyo route yenye lowest LORO classification accuracy.

Kwa nini smoother route inaweza kutoa lower AI accuracy?

Watafiti hawatafsiri lower accuracy ya Route 3 kama model failure pekee. Ingawa smoother speed na power demand hupunguza normal variation katika residual signals na kufanya threshold-based diagnosis iwe easier, pia hupunguza data diversity ya jinsi faults zinavyoonekana katika different operating points. Kwa maneno mengine, less dynamic excitation inaweza kupunguza diversity ya fault signatures ambazo CNN–LSTM inaweza kujifunza.

Route 2 huongeza fault visibility kupitia stronger accelerations na power peaks; lakini wakati huo huo hufanya iwe difficult zaidi kwa healthy energy sources kufidia faulty source. Mojawapo ya important engineering findings za study ni balance hii kati ya fault observability–energy-management load.

Fault-tolerant power redistribution

Katika source figures, baada ya fault kuthibitishwa power contribution ya related energy source hupunguzwa na contributions za Bat/SC au other healthy sources huongezwa. Hasa wakati wa FC degradation, battery na supercapacitor hujaribu kukidhi traction demand kwa kutoa power zaidi.

Katika Route 2 redistribution ni pronounced zaidi na SOC changes ni kubwa kutokana na higher power peaks. Katika Route 3 power transfer ni smoother na SOC changes za storage sources ni limited zaidi.

SIL na PIL validation

Katika SIL environment, closed-loop system iliendeshwa katika MATLAB/Simulink 2022b. Katika PIL stage, supervisory controller inayojumuisha FDI, energy-management logic na FTC reconfiguration ilihamishwa kwenda STM32F407VG processor; physical plant model ilibaki katika simulation environment kwenye host computer.

Worst-case computation time kuwa chini ya 6,2 ms dhidi ya available 100 ms control period inaonyesha kwamba supervisory diagnosis na energy-management algorithm zina significant timing margin kwenye hardware hii. Hata hivyo, result hii haimaanishi kwamba real-time embedded control ya entire vehicle imevalidated; katika PIL experiment, physical FC–Bat–SC plant yenyewe si real hardware.

Matokeo yanayoungwa mkono na chanzo

• Matumizi ya real GPS driving profiles yaliwezesha energy demand na diagnostic residuals kujaribiwa chini ya different route dynamics.
• Structure inayotumia EKF/CUSUM pamoja na CNN–LSTM iliunganisha model-based na data-driven diagnosis katika control chain ileile.
• Katika LORO validation, window-level classification accuracy ilibadilika kati ya %68,52 na %87,04 kulingana na route.
• Baada ya fault diagnosis, FC, battery na supercapacitor power references ziliweza kugawiwa upya katika SIL environment.
• PIL implementation kwenye STM32F407VG ilionyesha sufficient computation-time margin kwa 10 Hz supervisory loop.
• Route dynamics zimeonyeshwa kuathiri both diagnosis accuracy na energy-redistribution load.

Matokeo ambayo utafiti haujathibitisha

• System haijafanyiwa full-vehicle test kwenye real hydrogen electric vehicle.
• Physical faults za fuel cell, battery na supercapacitor hazikuundwa na kupimwa kwenye real components; faults ziliingizwa kwenye model parameters.
• Electrochemical na thermal behavior hazikuwakilishwa na high-fidelity physical models.
• Long-term aging ya battery, fuel cell au supercapacitor haijamodelliwa directly.
• Reported hydrogen consumption haiwezi kutafsiriwa kama field consumption ya specific commercial vehicle.
• Study haitoi systematic closed-loop benchmark comparison dhidi ya alternative EMS/FTC architectures.
• Haijaonyeshwa kwamba results zilezile zitapatikana chini ya traffic, road grade, climate na vehicle parameters za Uturuki.

Numerical consistency issue ndani ya source

Katika energy definition ya study yenyewe:

\[ E_{dem}=E_{served}+E_{ENS} \]

na:

\[ \eta_{service}= \frac{E_{served}}{E_{dem}} \]

relations zimetolewa.

Hata hivyo, abstract inasema kwamba zaidi ya %94 ya demanded energy iliserved kwenye routes zote. ES na ENS values katika source Table 4 zikitumiwa directly katika formulas hizi, approximate service ratios huwa %74,81 kwa Route 1, %73,64 kwa Route 2 na %76,33 kwa Route 3. Kwa hiyo, abstract statement na explicit numerical values za Table 4 hazionekani kuwa consistent chini ya interpretation ileile.

Vivyo hivyo, caption ya Table 5 inasema kwamba relative differences kati ya SIL na PIL energy-based indicators ni chini ya %5. Lakini katika table hiyo hiyo, Energy Not Served values zimetolewa kama SIL = 0,02410 na PIL = 0,00505 kwa Route 1; SIL = 0,07208 na PIL = 0,01878 kwa Route 2; SIL = 0,02283 na PIL = 0,00362 kwa Route 3. ENS pairs hizi zinatofautiana kwa zaidi sana ya %5.

Pia, katika Table 5, values 30,338%, 50,699% na 27,489% zenye percent signs zinaonekana chini ya “Average Speed”; katika earlier route table ya study, mean speed ilifafanuliwa kwa km/h. Kwa hiyo, kuna pia uwezekano wa inconsistency inayotokana na table labeling au formatting hapa.

Katika Verianla explanation hii, values hizi hazijarekebishwa kwa reinterpretation au assumption; inconsistency ndani ya source imehifadhiwa wazi.

Maelezo ya Chanzo na Mbinu

Jina kamili asilia: Route-Aware AI-Assisted Fault Diagnosis and Fault-Tolerant Energy Management for Hybrid Hydrogen Electric Vehicles: SIL and PIL Validation

Mpangilio wa waandishi: Sihem Nasri; Aymen Mnassri; Nouha Mansouri; Abderezak Lashab; Juan C. Vasquez; Adnane Cherif.

Corresponding author: Juan C. Vasquez.

Taasisi: Analyze and Process Electrical and Energy Signals (ATSSEE) Research Laboratory, University of Tunis Manar, Tunisia; Higher School of Engineering Technologies, ESPRIT, Tunisia; Center for Research on Microgrids (CROM), Department of Energy Technology, Aalborg University, Denmark.

Jarida: Actuators.

Mchapishaji: MDPI.

Publication: 2026, Volume 15, Article 126.

Submission date: 15 December 2025.

Revision date: 29 January 2026.

Acceptance date: 4 February 2026.

Publication date: 16 February 2026.

Aina ya chanzo: Peer-reviewed research article.

DOI: 10.3390/act15020126.

Official link:https://doi.org/10.3390/act15020126

Leseni: Creative Commons Attribution (CC BY).

Funding: Waandishi waliripoti kwamba research haikupokea external funding na article-processing charge pia haikulipwa na external organization.

Patent: Waandishi waliripoti kwamba hakuna patent iliyotokana na work hii.

Data availability: Driving profiles zilizotumiwa zilitolewa kutoka GPS-based speed na elevation information. Raw route data hazikuchapishwa publicly kwa sababu zina location-related information na ziko chini ya privacy na local data-sharing restrictions. Imeelezwa kwamba simulation models, processed datasets na trained AI models zinaweza kupatikana kutoka corresponding author kwa reasonable request.

Conflict of interest: Waandishi walitangaza kwamba hakuna conflict of interest.

Scientific evidence boundary: Real-world element katika study ni GPS-based driving profiles. FC–Bat–SC power system, fault injections na energy indicators zilitengenezwa kupitia mathematical models kwa control purposes. PIL stage inapima computation time na behavioral consistency na SIL kwa kuendesha controller kwenye STM32F407VG; si full physical vehicle validation.

Source-internal data-consistency note: Kuna numerical inconsistency kati ya statement ya abstract kwamba “zaidi ya %94 ya demanded energy iliserved” na results zinazopatikana kutoka ES/ENS values za Table 4 kwa kutumia service-efficiency equation ya study yenyewe. Statement ya Table 5 kwamba SIL–PIL differences ni chini ya %5 pia haionekani consistent na explicit values katika ENS column ya table hiyo. Inconsistencies hizi hazijasahihishwa kwa kuongeza external data.

Method summary: Research inaunganisha traction-demand generation kutoka real GPS route data, control-oriented FC–Bat–SC electrical modeling, EKF residual/innovation analysis, CUSUM change detection, CNN–LSTM time-series classification, LORO route-based generalization testing, synthetic parameter-based fault injection, diagnosis-driven energy redistribution na PIL validation kwenye STM32F407VG ndani ya single closed-loop workflow.