Abstraktný
The State of Health (SOH) estimation technology for lithium-ion batteries is crucial for the safety and reliability of electric vehicles. With the development of artificial intelligence (AI) and machine learning (ML) technologies, the field of battery management is beginning to adopt these methods to improve efficiency and stability. Especially, neural networks have shown advantages in high efficiency, low energy consumption, high robustness, and scalability in SOH simulation and prediction. The hybrid model, combined with equivalent circuit models (ECMs) and deep learning, has been proven to have potential in improving the accuracy and real-time performance of SOH estimation. Future research directions include utilizing more on-site data for health feature screening and model construction, as well as intelligent screening and combination of battery parameters to more accurately characterize actual SOH. The development of these technologies will further enhance the scientific, reliable, stable, and robust management of electric vehicle batteries.
1. Stručne
1.1 dôležitosť lítium-iónových batérií pre elektrické vozidlá a kritický význam SOH odhad
Lítium iónové batérie sú rozhodujúce pre prevádzka elektrické vozidlá, a ich výkon je ovplyvnený rôznymi degradáciou procesmi. Presné odhadnutie stav zdravotné (SOH) batérie je rozhodujúce pre zaistenie bezpečné, spoľahlivé, a ekonomické prevádzkové elektrické vozidlá. Ako dopyt pre elektrické vozidlá rastie, SOH monitorovanie stáva sa čoraz dôležitejšie, ako lítium-iónové batérie typicky pokles na 80% ich pôvodná kapacita pred koncom ich životnosťou. Pridanie, Stav Nabitie (SOC) je tiež a kľúč parameter, a je zmeny môže odrážať starnutie a degradácia batérie kapacita. Presné SOC predpoveď je užitočné pre SOH odhad, ktoré za postupne určuje zostávajúci životnosť z z batéria.
1.2 Vývoj SOH odhad metódy
Prehľad a pokrok existujúcich metód: Multiple SOH estimation methods have been developed, among which SOC based methods integrate real-time data such as current, voltage, and temperature to achieve more accurate SOH prediction in multiple charge and discharge cycles, optimize battery performance, prevent faults, and extend battery life. The latest advances in machine learning methods have further enhanced SOH estimation, and neural networks such as feedforward and convolutional neural networks perform well in battery modeling, outperforming traditional regression methods in complexity and accuracy, with an average error deviation of about 0.16% and a root mean square error of 5.57mV at the battery cell level.
1.3 Klasifikácia a Charakteristiky Batéria Modelovanie Metódy
Model prístup založený na modeli
Biela krabica model: Založené na podrobné elektrochemické princípy, it simuluje batéria správanie prostredníctvo základné parametre s vysoká presnosť. Avšak, jeho vysoká výpočtová požiadavka a zjednodušené predpoklady pre reálny svet dynamika reduce jeho presnosť pod dynamické podmienky, robí to nevhodné pre reálny čas aplikácie.
Sivá krabica modely (taká ako ECM): Kombinovanie fyzické prehľady a empirické úpravy, použitie obvod analógia k približné batéria správanie, môže odhad SOC s vysoká presnosť (zvyčajne v rámci 3% chyba), a sú užitočné pre reálny čas SOH odhad a zostávanie užitočné život (RUL) predpoveď, ale čela výzvy v údajoch kvalita a výpočtové požiadavky. A jednoduchý ekvivalentný obvod model pre lítium-iónové batérie (vrátane séria rezistory a a a dva RC prvky) môže môže byť použitý pre spoľahlivý simulácia, zatiaľ čo viac komplexný ECM (vrátane viac RC vetiev alebo konštantné fázové prvky CPE) môže simulovať vysoko dynamické procesy (ako elektrický vozidlo prevádzka), ale zvyšujúci výpočtový výpočtový dopyt riadi vývoj viac pokročilý SOH odhad metódy.
Čierna skrinka model (založená na údajoch prístup): Založené na vstupe a výstupe dáta, model je skonštruovaný bez spoliehania na interný pracovný princíp vedomosti. Stroj učenie techniky môže predpovedať batéria stav od a veľké množstvo merania dáta. Stroj učenie exceluje v identifikácii vzorov v komplexných množinách, takých ako viackanálových neurálnych sieť ktoré majú majú vysokú presnosť v kapacite odhadu, ale spoliehanie na vysokokvalitné a rôznorodé školenie údaje. Avšak, v praktické vozidlo aplikácie, mnohé interné premenné nemôžu byť priamo merané, a dáta riedkosť a nedostatok interpretovateľnosti urobiť model ťažké pochopiť a udržiavať.
1.4 Evolúcia Model Metódy a Vývoj Hybrid Modely
Evolúcia modelových metód: V minulom desaťročí, modelové metódy majú nepretržite vyvinuté, vrátane Kalman filtrovanie (KF) a jeho rozšírenia (také ako rozšírené Kalman Filter EKF, Neparfumované Kalman Filter UKF). Tieto metódy majú vysoká presnosť v batérii stav odhad, ale vyžadujú presné dynamické modely a sú komplexné na implementovať.
vzostup z hybridných modelov: V poradí a riešiť obmedzenia reálneho sveta dáta a zlepšiť výpočtová účinnosť, hybridné modely majú vznikajú, kombinovanie založené na modeloch a dátami metódy na trénovať stroj učenie modely prostredníctvom podrobné simulácie. V rovnakom čase, strojové učenie techniky dosiahli významný pokrok v minulých piatich rokoch, vrátane pravdepodobnostné metódy, meta učenie, nepriateľské učenie, semi supervízia učenie, atď. Hlboké učenie (a podmnožina stroj učenie) vykonalo dobre v spracovaní štruktúrované a neštruktúrované dáta. Fyzické Informácie Neurónové Siete (PINNs) Kombinovať empirické degradačné modely s neurónové siete na zlepšiť SOH odhad, zväčšenie adaptabilita metódy pod rôzne batérie typy a podmienky. S rozvojom automobilového priemyslu 2c tieto technologickým pokrokom sú rozhodujúce pre optimalizácia batéria výkon, prevencia poruchy, a podpora vývoj elektrických vozidiel.
1.5 Prehľad nasledujúce kapitoly v tomto článku
Section 2 provides a detailed introduction to the methods for screening and selecting review literature, ensuring the systematic and comprehensive nature of the research methodology. Section 3 provides an in-depth analysis of state of charge estimation techniques, exploring the impact of battery degradation mechanisms on modeling methods for electric vehicle batteries, including Kalman filtering and its improved methods, as well as integration with aging models. Section 4 focuses on SOH estimation techniques, compares traditional methods with new methods, and emphasizes methods applicable to electric vehicles. Section 5 demonstrates the role of deep learning in SOH estimation, such as long short-term memory (LSTM) networks and hybrid models, as well as how convolutional neural networks (CNN) consider practical factors to improve health assessment accuracy. Finally, Section 6 summarizes and looks forward to future research directions for battery health management systems to support the development of the electric vehicle market and other energy storage applications.
2. Materiály a Metódy
2.1 Definícia Výskum Otázka
Táto štúdia navrhuje päť kľúčových otázok na sprievodcu aplikáciu strojové učenie technológiu v SOH odhad lítium-iónových batérií v elektrických vozidlách.
Clarify the main machine learning techniques currently used for estimating the state of health (SOH) of lithium-ion batteries in electric vehicles, and explore the specific algorithms and models developed and used by researchers.
Preskúmajte vplyv rôzne údaje zdroje (laboratórium, vozidlo, a pole údaje) na presnosť a robustnosť SOH odhad stroj učenie modely, analýza ako údaje zdroje ovplyvnenie model výkon, a určiť ktoré údaje je najprínosnejšie pre presné SOH predpoveda.
Identifikujte kľúč výzvy aplikácie stroj učenie techniky v SOH odhad lítium-iónových batériách, ako ako variácie týchto výziev v rôznych environmentálnych podmienkach a a aplikácia scenáre, také ako teplota výkyvy, starnuti, a vplyv rôzne použitia režimy na presnosť SOH odhad.
Porovnať analýzu metódy SOH odhad, rozdiely medzi tradičnými metódami, a ich evolúcia proces, štúdia ako stroj učenie metódy môžu byť integrované s tieto tradičné metódy, identifikovať ich rešpektujúce výhody, nevýhody, a potenciál synergie.
Looking ahead to future research directions to improve the accuracy, adaptability, and computational efficiency of machine learning based SOH estimation models in lithium-ion batteries for electric vehicles, identifying research gaps, technical requirements, and innovative methods.
2.2 Literatúra vyhľadávanie a skríning
Databáza výber a vyhľadávanie stratégia: Conduct a comprehensive literature search using Scopus database, determine keywords based on research questions, and use Boolean search strings (TITLE-ABS-KEY (electric AND vehicle) AND KEY (battery AND state AND of AND health) AND TITLE-ABS-KEY (lithium AND ion) AND PUBYEAR>2003 A PUBYEAR<2025) to retrieve papers and patents published between 2003 and 2024. A total of 882 documents and 16286 patents were obtained, nearly half of which were published between 2020 and 2024, reflecting the industrial demand and progress in this field. The search results are distributed by year, major journals, national and patent offices, showing the time trend of research, journal distribution, regional diversity, and industry development priorities (such as battery management systems, modular architecture, vehicle control systems, and low resistance materials).
Literatúra skríning a zameranie: Získaná literatúra pokrýva viaceré disciplinárne odbory, s inžinierstvo oblasť ma najvyšší podiel (730 článkov), sledované energia, počítač veda, a matematika. Po zameranie na oblasť z počítača veda, 209 relevantné dokumenty boli identifikované, z ktoré 183 boli publikované medzi 2019 a 2024, indikujúce aktuálnosť z údaje. Tieto dokumenty zahŕňajú konferenciu príspevky, články, recenzie, a kniha kapitoly, s 72 článkov publikované medzi 2009 a 2024 ako hlavné prehľadné základ, priačo manuálne začlenenie relevantné články a kniha kapitoly z iných inžinierskyh odborov na zabezpečiť komplexné pokrytie výskumu oblastí a zachytiť inovatívne technologické pokroky v používaní umelej inteligencii na zlepšiť batériu manažment systémy.
3. Stav Nabíjanie (SOC) Odhad Technológia
3.1 batéria degradácia mechanizmus a jej vplyv na výkon
Lítium iónové batérie hlavne degradácia prezraz dva mechanizmy: lítium inventár strata (LLI) a aktívny materiál strata (LAM). LLI súvisí tvorba tvorba pevný elektrolyt rozhranie (SEI) vrstva na anóda, ktorá pochádza z z strana reakcia medzi lítium ióny a elektrolyt. LAM je spôsobený vnútorným mechanickým napätím v batérii, taký ako opakovaný expanzia a kontrakcia elektróda materiály počas nabíjanie a vybíjanie, čo vedie mikrotrhliny a oddelenie elektróda častice, redukcia aktívny povrch plocha dostupný pre elektrochemické reakcie, tým redukcia batéria kapacita, zvyšovanie vnútorný odpor, a nakoniec ovplyvňujúci batéria výkon. Tieto degradácia mechanizmy sú zrýchlené faktormi ako vysoká nabíjanie stav, vysoká teplota, a a agresívna cyklická podmienky. Podrobné informácie a modelovanie podrobnosti o rôznych starnutiu mechanizmoch (tepelné, elektrochemické, atď.) môže byť nájdený v relevantnej literatúre.
3.2. SOC odhad a modelovanie technológia pre elektrické vozidlo batérie
V dennom používaní elektrických vozidlá, batéria je často nabitá na 20% -40% SOC na údržba batéria zdravie, ale nelineárne a degradácia charakteristiky batéria kapacita môže výsledok v nepresné SOC hodnoty, ovplyvnenie odhad batéria plná kapacita. Výkon a údržba lítium-iónových batérií sú tiež ovplyvnené klímou, s teplotou a elektrolytom čerstvosťou (určené výrobou a plnením dátumy) ovplyvnené batéria účinnosť a životnosť. Vlastnosti nové elektrolyt batérie môžu líšiť sa pod rôzne pod odlišné klímy, a tepelné manažment stratégie môžu pomoc adresa teplota súvisiace výkon problémy a zlepšiť batéria odolnosť.
The traditional equivalent circuit model (ECM) is commonly used to estimate SOC, but requires frequent calibration. The article provides a detailed introduction to the SOC calculation equations based on ECM (including continuous and discrete forms), involving state space equations, open circuit voltage and SOC relationship equations, discrete-time domain SOC update equations, and voltage update equations. Relevant parameters (such as resistance, capacitance, open circuit voltage, etc.) are closely related to SOC. Standard laboratory testing (such as mixed pulse power characteristic testing at different temperatures) is commonly used for battery model parameter identification, but model inaccuracy and measurement noise can lead to small errors in SOC estimation. To improve the accuracy of SOC estimation, various techniques such as Kalman filtering and its extensions, PI based observer, sliding mode observer, etc. have been used to compensate for these effects, and integral correction methods have also been developed to handle initial model uncertainty and measurement noise. In addition, although electrochemical impedance spectroscopy (EIS) can evaluate battery characteristics (including SOC and SOH), it is time-consuming and impractical for large-scale applications (such as electric vehicle fleets), making it difficult to capture the dynamic and changing operating conditions of electric vehicle batteries. Therefore, a more adaptive and efficient method is needed.
3.3. Zlepšenie Technológia
Kalman filter a jeho zlepšenie metódy: Kalman filter (KF) and its extensions (such as Extended Kalman Filter EKF, Unscented Kalman Filter UKF, Volume Kalman Filter CKF) are widely used for SOC estimation. KF provides the optimal SOC estimation by minimizing the mean square error, solving the problems of cumulative error and initial SOC uncertainty. However, it is suitable for linear time-varying systems where the nonlinear dynamics of batteries require linearization approximation. Although EKF extends the KF framework to handle nonlinear models, linearization may affect accuracy and lead to estimator divergence. New methods such as UKF and CKF use sigma point estimation to estimate nonlinear transformation statistics, while CKF uses the spherical radial volume rule to calculate multivariate moment integrals to improve the accuracy of nonlinear Bayesian filtering. However, these filters typically assume that the noise characteristics are known and constant, and in practical applications, the noise is variable (such as non Gaussian noise generated by external interference). Therefore, robust adaptive filtering strategies have been developed, such as using Gaussian mixture models (GMM) to model non Gaussian noise to improve state estimation accuracy. Relevant studies have shown the applications and advantages of these methods in different fields. In addition, distributed and distributed filters (such as distributed Kalman filter DKF, distributed Kalman filter and covariance cross DKF-CI) are used to optimize state estimation of large-scale interconnected systems. Robust and nonlinear filters (such as robust Kalman filter) have superior performance in dealing with complex nonlinearities in battery systems (such as electrochemical processes). Adaptive techniques (such as adaptive EKF and adaptive UKF algorithms) dynamically adjust filter parameters to adapt to noise changes and improve SOC estimation accuracy. Relevant studies and examples have verified the effectiveness of these methods.
Iné zlepšenie metódy: such as the Adaptive Integral Correction State of Charge Estimation (AIC-SE) method proposed in 2022, which is based on the ECM model and improves the accuracy of SOC estimation through real-time correction mechanisms (including resistance and battery capacity correction) (maximum error ± 0.8%, RMS error less than 0.3%). The computational efficiency is higher than UKF (AIC-SE about 5n operations, UKF about n ^ 2 operations), effectively addressing the challenges of battery aging and degradation. The Variational Bayesian Maximum Correlation Entropy Volume Kalman Filter (VBMCCKF) in 2023 combines advanced filtering and statistical techniques to improve measurement error covariance estimation using the Variational Bayesian method. The Maximum Correlation Entropy criterion is used to handle non Gaussian noise measurement outliers, significantly improving SOC estimation accuracy (compared with EKF, CKF, and Variational Bayesian Volume Kalman Filter, the average absolute error is reduced by 77%, 68%, and 49%, respectively), and enhancing the robustness of the battery management system.
3.4 Integrácia s Starnutie Modely
The battery aging model is closely related to SOC estimation, and recent research has innovated in both aspects. The battery aging model proposed in 2024 comprehensively considers the effects of SOC, battery temperature, time, and fully equivalent cycle times (NFECs) on battery aging. The model consists of two parts: the first part focuses on SOC and temperature related aging (calculating capacity loss through specific formulas), and the second part considers the impact of NFECs on aging. This model innovatively integrates battery aging as an electric vehicle subsystem with the battery model, covering all operating modes such as parking, driving, and charging. It achieves accurate interaction simulation between different subsystems through the formal method of energy macroscopic representation (EMR) (a graphical tool developed in 2000 for organizing subsystem connections, representing power flow, and causal relationships). Research has shown that reducing the charging frequency (such as changing from daily charging to every four days) can significantly prolong the time for the battery to reach 80% SOH. This integrated approach provides important progress in optimizing battery management and understanding the impact of charging practices on battery aging.
New methods such as AIC-SE and VBMCCKF have significant advantages in SOC estimation accuracy and computational efficiency. AIC-SE performs well in computational efficiency, while VBMCCKF performs better in handling dynamic estimation of measurement errors and noisy environments. If accuracy and noise processing are given priority, combining variational Bayesian and maximum correlation entropy criteria may be the current best choice; If we focus on computational efficiency and real-time applications, AIC-SE is a good choice, indicating that ECM modeling methods still have advantages in this area. In addition, the battery aging model studied in 2024 comprehensively considers the impact of multiple factors on battery degradation, which is of great significance for optimizing battery life (based on charging practice). Overall, these developments not only improve the accuracy of SOC estimation, but also contribute to extending battery life and enhancing battery operational reliability.
4. Stav Zdravie (SOH) hodnotenie techniky
4.1 Tradičné SOH odhad metódy
tradičná SOH odhad metóda je široko používaná v akademických a priemyselných oblastiach, hlavne založená na základných parametroch ako ako kapacita degradácia, vnútorná odolnosť, a cyklus životnosť na vyhodnotiť batéria SOH (pozri Tabuľka 4 pre relevantné vzorce a parameter významy). Tieto metódy poskytujú a základ pre batériu zdravie hodnotenie a pomoc pochopenie batéria výkon. Pochopením tieto tradičné metódy, my môžeme lepšie pochopiť vylepšenia nové odhady metódy v nasledujúcich kapitolách. Nové metódy často použitie viac komplexné dáta analýza a prediktívne modelovanie techniky na riešenie obmedzení tradičných metód. Porovnanie dve môže objasniť vývoj a evolúcia SOH odhad technológie a demonštrovať a a ukázať a ako moderné metódy môžu zlepšiť presnosť a prispôsobiteľnosť batérie manažment systémy.
4.2 Nové Vývoj v Nahradenie Tradičné Metódy
Nové zdravie ukazovatele kombinované s strojové vzdelávanie: Na zlepšenie presnosť SOH predpoveď, výskum zaviedol nové zdravotné ukazovatele ako ako Degradácia Rýchlosť Pomer (DSR). Vzorec Pre Výpočet DSR Z Sklon Nabíjanie Napätie Krivka Je:
Porovnaním svahov viacnásobných nabíjania cyklov, degradácia rýchlosť (in mV/s) batéria v a špecifické napätie rozsah (ako [3.8-3.9V]) je určené, ktoré je úzko súvisí s batéria kapacita a môže byť použitý ako a a kľúč indikátor na určiť koniec batérie životnosť. Kombinácia Gaussovský Proces Regresia (GPR) a Viac Vrstva Perceptrón Neurálna Sieť (MLPNN) modely môžu viac presne odhadu kapacita strata a degradácia. Porovnané s tradičnými modelmi, citlivosť a presnosť sú významne zlepšené, efektívne riešenie problém tradičných modelov bytie ťažké odhaliť degradácia skoro.
Early methods for improving traditional equivalent circuit models (ECM) continued to develop, such as estimating SOH by analyzing the body capacitance of the equivalent RC circuit model in 2015, using innovative algorithms to calculate the body capacitance attenuation factor, and combining it with discrete nonlinear observers to improve accuracy and reliability; In 2024, a second-order hybrid equivalent circuit model combined with adaptive update rate and nonlinear observer was adopted to consider the influence of temperature, achieving high accuracy in SOH estimation (average absolute error less than 0.5%, RMS error less than 0.2%); The cloud solution for 2023 utilizes long-term monitoring data and real-time data to estimate battery model parameters by adjusting the moving window least squares algorithm. Based on the ECM model, high-precision SOH evaluation is achieved, indicating that the improved ECM method still has significant importance in SOH estimation, consistent with the trend of continuous improvement of ECM methods in SOC estimation technology.
The new framework integrates Linear Statistical k-Nearest Neighbor (LSKNN), Maximum Information Entropy Search (MIES), and Collective Sparse Variational Gaussian Process Regression (CSVGPR) for processing data interpolation, noise filtering, feature selection, and uncertainty management. LSKNN estimates missing data points and filters noise, MIES selects features with high correlation to SOH, and CSVGPR processes data uncertainty to improve prediction accuracy. This framework was tested using the NASA battery dataset, and compared with methods such as ElasticNet, Support Vector Regression (SVR), Random Forest, and Gradient Boosting, the Root Mean Square Error (RMSE) was reduced by 77.8% (from 0.0510 in ElasticNet to 0.0113). Compared with Gaussian process models with different kernels, the RMSE was reduced by 55.5% (from 0.0254 to 0.0113), confirming the robustness and high accuracy of the framework and providing a more accurate method for SOH estimation.
The development trend of SOH estimation technology is shifting from traditional methods to more complex models suitable for electric vehicles. New methods include combining degradation models with classical machine learning, ECM based methods, and hybrid methods. DSR is an important innovation that reduces reliance on a complete charging cycle (reducing waiting time by approximately 84%) and, when combined with machine learning, improves the accuracy of capacity loss estimation, overcoming the difficulty of early degradation detection in traditional models. The improved ECM method has achieved good results in SOH estimation, which is consistent with the importance of ECM method in SOC estimation. Hybrid technologies (such as the new framework mentioned above) have high accuracy. Although real-time applications pose challenges, effectively solving key data processing problems is a significant improvement over traditional SOH estimation methods. Overall, these developments focus on real-time applications and data-driven methods, significantly improving the reliability of electric vehicle battery management systems. Deep learning methods such as LSTM, CNNs, and hybrid techniques have become the mainstream methods for SOH estimation. Subsequent chapters will present relevant research results and contributions.
5. Aplikácia hĺbka učenie v SOH odhad
5.1 LSTM a Hybrid Modely
Viaceré štúdie použili zlepšili starnutie modely kombinované s hĺbkové učebné techniky na vylepšenie presnosti SOH odhadu. Hlboké učenie je nevyhnutné v predpovedaní zostávajúcom užitočnom živote (RUL). Napríklad, integráciou SOH degradáciou modelom a zvážením rôzne prevádzkové podmienky také ako nabíjanie/vybíjanie prúd a teplota, a špecifický vzorec môže byť použitý do:
Medzi nimi, I2 {c} a I2 {d} sú normalizované nabitie a vybitie prúdy, T3 {c} a T4 {d} sú normalizované batéria a okolitá teploty, T3 {c} a T4 {d}} are nabíjanie a výboj časy, a (d1-d4) je a hmotnosť), ktorý viac presne simuluje batéria degradácia. RUL predpoveď model založený na LSTM sieti zlepšuje predpoveda presnosť, ale the výpočtová zložitosť zvyšuje, a reálny čas aplikácie tvár výzvy. Neurálne siete môže zvládnu časovo sa mení batéria procesy, nepretržite učiť sa prispôsobovať zmeny v batéria správanie, a údržba model spoľahlivosť.
Pomocou extrahovania kľúča funkcií (napríklad 6 kľúčov funkcií) optimalizovať SOH odhad % 2c kombinované s strojovým učením algoritmami dosiahnuť vysokú presnosť a nízku výpočtovú záťaž % 2c napätie funkcie prehrávanie a významnú rolu v zlepšovanie presnosť batériu stav hodnotenie. Kombinovanie viaceré hlboké learningové modely (ako ako CNN, LSTM, GRU, a ich obojsmerné varianty) do a hybridný framework (taký ako CNN-LSTM-DNN, CNN-GRU-DNN) to predpovedať RUL, využitie a široký rozsah z funkcií na zlepšiť presnosť, znížený RMSE by 90.5% in NASA dataset testing, ale the computational strength and complexity limit real-time applications. Multi model metódy (ako LSTM model knižnice) a pokročilé optimalizácia stratégie (ako ako integrácia LSTM do rámca AI-BMS a implementácia to na FPGA) môže zlepšiť predpoveda presnosť a systém účinnosť, ale aplikácia FPGA v komerčných elektrických vozidlách čelí náklady a praktickosť výzvy.
Kombinácia GRU a mäkké snímanie metódy má potenciál dlhodobo RUL predpoveď v laboratóriu prostredia, ale praktické aplikácie vyžadujú adaptácia na rôzne nabíjanie podmienky. Používanie metód založené na údajoch také ako LSTM, DNN, a GRU na proces NASA datasets, GRU má silný výkon (RMSE z 0}.003, MAE z 0.003, R-na druhú chyba z 0.004), a kombinovanie GRU a LSTM siete výsledky v lepším výkone. Metóda založená na LSTM extrahuje funkcie (ako 5 manuálne funkcie) analýzou batéria nabíjanie vybíjanie krivka, a použitie optimalizácia algoritmy (ako ako Adam) na zlepšenie školenie účinnosť a predpoveď presnosť. Pod tréning jednoduchá batéria čiastočné údaje, SOH odhad chyba pre iné batérie je nízka, ktorá je lepšia ako tradičná modely.
The MDA-LSTM network combines multiple features and temporal information, and improves the accuracy of RUL prediction through multiple feature fusion modules and dual attention modules. It performs well in multi dataset validation, with robustness and generalization. The stacked BiLSTM network is used to predict SOH using constant current charging data, and the bidirectional structure improves prediction reliability, making it suitable for real-time SOH estimation during fast charging. The TCN-LSTM model utilizes synthetic data and Bayesian optimization to accurately reconstruct open circuit voltage (OCV) and estimate State of Health (SOH) (MAE below 22mV, MAPE below 2.2%). It can be extended to different battery chemical systems through transfer learning, but there are extrapolation limitations when data is insufficient. The deep fusion method (such as utilizing historical data and multiple health indicators) achieves high accuracy (MAPE below 2.97%) through full charge discharge testing, and both the global framework based on GPR and the DFTN model for individual electric vehicles have achieved good results.
5.2. CNN a CNN-LSTM Integrovaný Model
The CNN-WNN-WLSTM method integrates CNN, WNN, and WLSTM networks. CNN extracts features, WNN and WLSTM process features and estimate SOH. The RMSprop optimizer is used to improve performance and outperforms traditional methods in NASA dataset testing, providing a promising approach for battery health management. The CNN-LSTM-CRF model is inspired by natural language processing, with the CRF layer capturing output variable dependencies to improve the accuracy and intuitiveness of battery capacity prediction. However, the computational requirements are high and exceed the capabilities of onboard processors. In the future, research is needed to improve its practicality (such as through transfer learning). The LSTNet model improves battery capacity prediction performance by segmenting data, integrating ConvLSTM and AR components, and optimizing the structure (for example, in NASA dataset testing, RMSE was 0.65%, MAE was 0.58%, and MAPE was 0.435% when trained on 40% data).
By integrating enhanced CNN and ECSSA optimization algorithms to predict the RUL of solid-state lithium-ion batteries, CNN improves feature extraction and prediction accuracy by optimizing hyperparameters and structures (such as using advanced convolutional layers, activation functions, and residual connections), while ECSSA optimizes model parameters through innovative mathematical methods (such as Circle Chaotic Mapping, Nonlinear Absorption Coefficient, and Cauchy Mutation) to improve RUL prediction accuracy and robustness. Combining PCA and CNN for feature optimization and dimensionality reduction improves the accuracy and efficiency of SOH estimation (compared to traditional CNN and fixed dimensional PCA-CNN models, MAE increases by more than 20% and RMSE increases by more than 30%). The real-time SOH estimation model integrates 1D-CNN and BiGRU, using BMS data to avoid complex feature extraction, and achieving high accuracy through Bayesian optimization of hyperparameters (such as in NASA dataset testing, MAE is 2.080%, RMSE is 2.516%, and EOL index error is zero).
5.3. Optimalizácia Stratégie Pre Hlboké Učenie Modely
Firstly, the random forest algorithm was used to identify key health factors, and then the genetic algorithm particle swarm optimization (GA-PSO) technique was used to optimize the support vector regression (SVR) model parameters for estimating State of Health (SOH). The effectiveness was verified on four batteries, improving accuracy and convergence speed (RMSE of 0.40%, MAPE of 0.56%), which is superior to other related methods. The GWO-BRNN hybrid method utilizes grey wolf optimization (GWO) to select hyperparameters for Bayesian regularized neural networks (BRNN). Based on the NASA dataset, the SOH estimation error is less than 1%, but the computational complexity is high and practical applications are limited. Directly using the raw data of electric vehicles to evaluate SOH and predict RUL, improving accuracy by introducing new evaluation features and interpolation correction methods (reducing the relative error of current integration to 0.94%), combined with D-NSGA-II optimization method to further optimize SOH estimation and reduce computation time. In response to the difficulty in estimating State of Health (SOH) caused by incomplete charging and discharging of lithium-ion batteries in electric vehicles, an indirect estimation method (ATAGA-BP) is proposed. The method utilizes the characteristics of constant voltage charging stage as a health indicator and is validated through simulation with NASA data. The method has a high correlation with battery capacity (over 85%), with an SOH estimation error of 3.7% and an iterative efficiency improvement of 17.8%.
Deep learning has made significant progress in SOH estimation, and comprehensive models considering multiple factors provide a deeper understanding of battery degradation. LSTM networks are important in capturing temporal dependencies and predicting RUL, but their computational complexity poses challenges for real-time applications. Feature extraction methods are important and can optimize SOH estimation. The combination of hybrid models and different neural network architectures for processing battery data complexity has promising prospects, but high computational requirements limit practical applications. Optimization strategies such as GA-PSO, GWO-BRNN, and D-NSGA-II have improved accuracy and efficiency, but implementing complex algorithms is difficult and requires a balance between accuracy and execution simplicity. Advanced AI technology is crucial for the application of secondary batteries (lacking detailed usage data). Subsequent chapters will provide an overview of the current state of secondary application research, particularly in the area of battery reuse.
6. Summary
This article advances the development of SOH and SOC estimation for lithium-ion batteries in electric vehicles through innovative methods and models, covering various technologies from traditional machine learning to advanced deep learning models such as LSTM and CNN. However, each method has differences in accuracy, complexity, and applicability, making direct comparison difficult. Research has found that data processing and sources have a significant impact on model performance, and further validation is needed for actual deployment. Although deep learning models have shown advantages in processing complex data, they still face challenges such as high computational resource requirements and adaptability to practical application scenarios. Future research should focus on improving feature selection, anomaly detection, adapting to diverse environmental conditions, optimizing algorithms to enhance computational efficiency, achieving real-time applications, integrating multiple data sources to improve SOH estimation model performance, while also addressing challenges in secondary battery applications, developing effective solutions, and promoting the development of battery management systems to meet the growing demands in the fields of electric vehicles and energy storage.