From Empirical Measurements to AI Fusion—A Holistic Review of SOH Estimation Techniques for Lithium-Ion Batteries in Electric and Hybrid Vehicles

Abstract

Accurate assessment of lithium-ion battery state of health (SOH) represents a cross-disciplinary challenge that is critical for the reliability, safety, and total cost of ownership of electric vehicles (EVs) and hybrid electric vehicles (HEVs). This review systematically examines the evolutionary trajectory of SOH estimation methods, ranging from conventional experimental measurement approaches to cutting-edge data-driven techniques. We analyze how these techniques address critical challenges in battery aging and performance evaluation, while discussing their respective advantages across different application scenarios. The paper highlights emerging trends in artificial intelligence-integrated advanced technologies for SOH estimation, along with practical implementation considerations. Special emphasis is placed on key challenges of SOH estimation in EVs/HEVs applications with proposed alternative solutions. By synthesizing current research directions and identifying critical knowledge gaps, this work provides valuable insights for fundamental research and industrial applications in battery health management.

1. Introduction

In recent years, global warming and energy crisis issues have sparked intense worldwide discussions [1]. Electric Vehicles (EVs) and Hybrid Electric Vehicles (HEVs) can mitigate the greenhouse effect through reduced carbon emissions, and their improved performance and efficiency are recognized as effective means to achieve energy conservation and environmental protection goals [2,3,4]. Numerous countries are vigorously promoting the development of EVs and HEVs, with consumer market demand showing significant growth [5,6]. Lithium-ion batteries (LIBs), with advantages including high energy density, long cycle life, low self-discharge rate, and absence of memory effect, have become indispensable energy storage devices for EVs and HEVs [7].

However, with increasing vehicle mileage and continual cycling of LIBs, their performance inevitably undergoes gradual aging degradation [8]. As a core automotive component, LIBs are intrinsically linked to EV/HEV operation. Age-induced failures like short circuits, electrolyte leakage, or insulation damage can pose serious safety risks to drivers [9]. Therefore, real-time and accurate monitoring of LIB health status becomes particularly crucial. State of health (SOH) serves as a key battery health metric [10]. To achieve accurate and robust SOH estimation, researchers developed a battery management system (BMS). A BMS operates by collecting voltage, current, and temperature data to monitor critical indicators, including State of Charge (SOC), SOH, State of Power (SOP), and Remaining Useful Life (RUL) [11,12].

As one of the most expensive components in EVs/HEVs, accurate SOH assessment enables users to optimize charging/discharging strategies, mitigate battery aging, and extend service life. The scaling of electric vehicles faces resource constraints of critical minerals (lithium, cobalt, nickel)—with global lithium reserves supporting only ~2.5 billion EVs—and geopolitical concentration risks (>50% cobalt from Congo, 80% lithium processing reliant on China). Extending battery service life through accurate SOH assessment can partially mitigate over-dependence on and depletion of these resources. For automotive traction batteries, industry consensus recommends replacement at an 80% SOH threshold. Precise SOH prediction facilitates predictive maintenance scheduling, preventing sudden failures and costly repairs. SOH prediction enhances reliability and safety by providing early warnings of potential hazards like thermal runaway or capacity plunge, thereby reducing fire risks. Particularly for HEVs, battery condition critically affects powertrain coordination efficiency, where SOH prediction ensures drivetrain stability. Meanwhile, a BMS utilizes SOH data to refine management strategies, dynamically adjusting charge/discharge parameters (e.g., current/voltage thresholds) for enhanced efficiency while preventing overcharge/discharge [13,14,15].

The cascaded utilization of retired batteries necessitates SOH-based screening, where advanced SOH prediction enhances both sorting efficiency and economic viability in residual value assessment. Long-term perspective reveals that accurate SOH prediction mitigates premature battery retirement, alleviates recycling burdens, reduces environmental pollution, and contributes to carbon neutrality goals. Concurrently, prolonged battery lifespan decreases consumption of scarce resources (e.g., lithium, cobalt) and minimizes material wastage. Ultimately, advancements in SOH prediction facilitate standardization of battery diagnostics, provide data-driven support for next-generation battery design, and enable OEMs/battery manufacturers to optimize products through data accumulation [16,17].

The evolution of SOH estimation techniques for lithium-ion batteries has progressed through four distinct technological paradigms [18], each addressing the limitations of its predecessors [19,20].‌ ‌During the initial phase (approximately 1990s to 2005),‌ methods primarily relied on experimental measurement techniques. While capable of providing precise electrochemical information, these approaches were severely constrained by laboratory settings and proved difficult to implement in practical onboard scenarios. ‌Subsequently (approximately 2005 to 2015),‌ model-driven methods emerged, achieving the objective of real-time onboard estimation. However, they universally encountered inherent trade-offs between model complexity, real-time performance, and accuracy. ‌Entering the big data era (approximately 2015 to 2020),‌ data-driven techniques flourished. Leveraging vast amounts of battery data significantly enhanced estimation accuracy, yet these methods introduced new challenges, notably insufficient model interpretability. ‌Currently (2020 to present),‌ AI-fused advanced techniques represent the cutting-edge trend, aiming to overcome the constraints of traditional methods by integrating the strengths of physical mechanisms with data-driven approaches. ‌This clear evolutionary trajectory, particularly the progression from early experimental measurements to current AI-fused approaches, forms the core motivation for this systematic review. It aims to:‌ synthesize cross-paradigm knowledge and developmental patterns, analyze the applicability and intrinsic relationships among methods at various stages, and provide strategic guidance for identifying optimal fusion pathways for a next-generation BMS.

Despite significant progress in SOH prediction, substantial challenges remain for EV and HEV applications. Primarily, battery aging involves coupled effects from temperature, charge/discharge cycles, depth of discharge (DoD), and usage patterns, making in-vehicle SOH prediction an inherently complex problem. Furthermore, diverse battery types and brands impose stringent demands on model generalizability, while customized modeling requires extensive effort. Moreover, an onboard BMS necessitates lightweight SOH prediction algorithms to balance computational accuracy and speed, imposing critical efficiency requirements [21].

While previous reviews have outlined research progress in SOH prediction methods [22,23,24], systematic and comprehensive summaries of SOH estimation approaches for LIBs remain scarce, with insufficient in-depth discussion on critical challenges for EV/HEV traction batteries, particularly regarding the integration of artificial intelligence (AI) with emerging technologies. To address these gaps, this review comprehensively examines SOH estimation methods for LIBs in EV/HEV applications, highlighting AI-integrated cutting-edge techniques through detailed comparative analysis of methodological strengths/limitations, while identifying key challenges and influencing factors to guide future development of SOH estimation technologies for EV/HEV implementations.

The rest of this review is structured as follows: Section 2 introduces the specific definition of SOH. Section 3 systematically reviews the research history of SOH estimation techniques. Section 4 provides a detailed comparison of the advantages and disadvantages of various technologies. Section 5 identifies key challenges of SOH estimation methods in EVs and HEVs technologies and proposes actionable recommendations. Finally, Section 6 provides concluding remarks.

2. Definition of SOH

SOH quantifies the performance degradation of a battery relative to its initial condition, with SOH prediction serving as the core function in a BMS for real-time health monitoring. As a critical parameter for evaluating remaining useful life and reliability, SOH is typically expressed as a percentage (100% indicates pristine state). Industry standards define 80% SOH as the end-of-life threshold, at which point battery replacement in EVs/HEVs becomes necessary to maintain performance requirements [25]. Battery aging manifests as capacity fade and power deterioration, where capacity loss primarily correlates with active material depletion, while increased internal resistance during aging constitutes the dominant mechanism for power degradation [26]. Consequently, industrial SOH definitions are principally derived from two parameters: capacity and internal resistance. The SOH metrics based on lithium-ion battery capacity fade and resistance increase are respectively given by Equations (1) and (2):

SOH(t) = 100% × C/C₀,

(1)

where C₀ denotes the initial rated capacity (Ah) and C represents the maximum available capacity (Ah) at the current state.

SOH(t) = 100% × R₀/R,

(2)

where R₀ indicates the initial internal resistance (Ω) and R denotes the real-time internal resistance (Ω).

3. SOH Estimation Methods

3.1. Experimental Measurement Techniques

In the field of battery SOH evaluation, experimental measurement techniques refer to methods that conduct mandatory physical data acquisition processes to assess battery health status through directly observable physical parameters (e.g., voltage, current, temperature). These techniques typically require laboratory or specialized testing environments with complex equipment, resulting in high costs and limited real-time applicability. Their distinctive value lies in the intrinsic capability to interpret electrochemical degradation mechanisms through raw data analysis, whereas data-driven methods utilize features without mechanistic consideration. Unlike model-based approaches that depend on predefined physical equations or equivalent circuit constructs, experimental methods often rely on empirical rules while maintaining independence from theoretical modeling constraints. This characteristic, combined with their incorporation of physical/electrochemical mechanisms, ensures high accuracy, repeatability, and clear aging mechanism interpretation, making them foundational for battery theoretical research and engineering applications. This review categorizes experimental measurement techniques into: coulomb counting, electrochemical impedance spectroscopy, incremental capacity/differential voltage, internal resistance test, and other experimental techniques.

3.1.1. Coulomb Counting

The coulomb counting method involves fully charging the battery and discharging it to a preset cutoff voltage, then calculating SOH by dividing the total discharged capacity by the rated capacity [27]. Reference [28] proposed an improved Coulomb counting method for estimating SOC and SOH of LIBs. Compared with traditional Coulomb counting, this method considers charge/discharge efficiency and dynamically calibrates the maximum releasable capacity, achieving accurate SOH evaluation. In this paper, SOH is defined as the ratio of maximum releasable capacity to rated capacity. After complete battery discharge, the accumulated depth of discharge equals the current maximum releasable capacity, which can then be used to recalibrate SOH.

This method has clear physical significance and a low implementation cost. However, it also has limitations, including difficult-to-eliminate cumulative errors, requiring complete charge/discharge cycles for accurate capacity calculation, and inability to diagnose root causes of degradation. In the EVs and HEVs field, the coulomb counting method is suitable for real-time monitoring in an onboard BMS but requires integration with dynamic operating condition compensation algorithms and temperature/aging models to improve accuracy. Future improvements should focus on multi-parameter fusion and online error correction mechanisms to meet SOH estimation requirements under complex operating conditions.

3.1.2. Electrochemical Impedance Spectroscopy (EIS)

EIS is a non-destructive testing technique that evaluates battery state by measuring impedance responses under AC excitation at different frequencies. For SOH assessment, EIS reflects battery aging by analyzing changes in internal electrochemical processes. Its core principle involves applying small-amplitude (typically 10–50 mV) sinusoidal AC voltage/current signals to the battery, then measuring the system’s impedance responses (magnitude and phase) across frequencies ranging from 10 mHz to 100 kHz. This yields Nyquist and Bode plots [29]. By fitting EIS data with equivalent circuit models, key parameters like ohmic resistance, charge transfer resistance, and Warburg coefficient can be extracted, which are then used to calculate SOH based on predefined models [30].

The literature [31] primarily investigates a lithium-ion battery SOH estimation method based on charge transfer resistance under varying temperatures and SOC conditions. Through EIS measurements and equivalent impedance model fitting, the authors analyzed the effects of temperature and SOC on charge transfer resistance and proposed an analytical calculation model that converts charge transfer resistance measured at different temperatures and SOC levels to standardized conditions for SOH estimation. The proposed SOH estimation method eliminates the requirement for EIS measurements at specific temperatures and SOC conditions, thereby enhancing practical application flexibility.

Yang et al. [32] proposed a lithium-ion battery SOH estimation method based on a fractional-order impedance model (FIM) and interval capacity. By measuring and analyzing battery EIS, they established a fractional-order impedance model for batteries, which can more accurately simulate battery dynamic performance compared with traditional integer-order models. Using a least squares genetic algorithm for model parameter identification combined with a voltage tracking system, they achieved high-precision parameter estimation with a voltage tracking error rate below 0.2%. Furthermore, through accelerated aging tests, they validated battery aging performance. Based on the identified model parameters and interval capacity, combined with a backpropagation neural network, they realized accurate battery SOH estimation without requiring special state observers or filtering methods, with errors controlled within 1.5%. Figure 1 illustrates the detailed EIS analysis process for lithium-ion batteries.

Reference [33] pioneered the combination of EIS measurements with battery short-term relaxation effects for SOH estimation. By measuring EIS at different SOH states and employing a particle swarm optimization algorithm to fit the EIS data, the researchers obtained parameters for the fractional-order impedance model. The method simultaneously accounts for short-term relaxation effects by analyzing impedance variations (particularly the change rate of charge transfer impedance) during battery rest periods. Based on the correlation between SOH and the relaxation-induced charge transfer impedance variation rate, an empirical model was established for SOH estimation. This model demonstrates simplicity, ease of implementation, and high accuracy, achieving an average SOH estimation error below 1%.

However, although numerous studies have demonstrated that analyzing EIS test results can non-destructively identify battery aging mechanisms and provide multi-timescale information, Reference [34] has proven that battery impedance varies significantly across different battery types. Furthermore, it should be noted that high onboard costs, implementation complexity, lengthy EIS testing durations, and high sensitivity to testing conditions further limit its application in EVs and HEVs. Currently, the EIS method is more suitable for laboratory-level offline diagnostics. To improve the feasibility of online monitoring for this technology, it is necessary to research simpler, more universal online EIS parameter acquisition methods that are compatible with onboard environments.

3.1.3. Incremental Capacity (IC) and Differential Voltage (DV)

The IC/DV method analyzes battery SOH based on charge/discharge curve characteristics. IC refers to analyzing the differential of capacity versus voltage (dQ/dV) during constant-current charging, extracting changes in the position, height, or area of characteristic peaks that reflect internal aging mechanisms. Aging causes these characteristic peaks to shift, diminish, or disappear, and comparing IC curve differences between new and aged batteries enables SOH quantification. DV analyzes the differential of voltage versus capacity (dV/dQ) during discharge, identifying characteristic inflection points corresponding to voltage plateaus to assess aging degree. The plateaus in DV curves undergo translation or deformation with aging, commonly used for analyzing anode degradation.

Researchers [35] proposed a fully embedded SOH estimator based on IC analysis for practical EV applications. The study conducted aging tests on C/LiFePO4 batteries through pure EV cycling patterns while monitoring via onboard BMS. By applying IC analysis to large-format batteries, a dynamic integration boundary detection algorithm was developed to determine the integration intervals in IC curves for capacity estimation, incorporating cost reduction and industrial constraints to meet real-world vehicle requirements. This achieved robustness against variations in temperature, current, and DoD. Under ideal conditions, the SOH estimation error remained below 2% while maintaining 3–4% error bounds across three robustness test scenarios: depth of discharge, charging current, and temperature variations.

Bian et al. [36] proposed a novel voltage reconstruction method that effectively eliminates noise in raw IC curves, achieving smoother profiles for more accurate feature extraction. By combining a Thevenin equivalent circuit model with a polynomial model of open-circuit voltage, they reconstructed the battery’s Q-V curve. Informative features, including peak height, peak voltage, and peak area, were extracted from the reconstructed IC curves. These features were then correlated with reference capacity through univariate linear regression to establish feature-SOH models. The proposed method demonstrates strong robustness against cell-to-cell variations, temperature uncertainties, and noise contamination, while maintaining applicability across different battery chemistries. For nickel cobalt aluminum (NCA) batteries, the root mean square error (RMSE) of SOH estimation ranged between 0.25 and 0.69% using discharge curves and 0.78–1.45% using charge curves. For Kokam pouch cells, RMSE values were 0.12–0.33% (discharge) and 0.20–0.41% (charge). For LiFePO4 batteries, estimated SOH showed excellent agreement with measured values, further validating the method’s robustness under high-temperature fluctuations.

Reference [37]‌ directly extracted features from raw IC curves by quantifying the correlation between average IC values and battery capacity across different voltage ranges, selecting features most consistent with capacity degradation to enhance SOH estimation accuracy. On the Battery A1 dataset, the Support Vector Regression (SVR) model achieved 0.889% MAE (Mean Absolute Error) and 1.145% RMSE, while the Backpropagation Neural Network (BPNN) model yielded 1.162% MAE and 1.331% RMSE. For the Battery A2 dataset, the SVR model demonstrated 0.622% MAE and 0.692% RMSE, versus the BPNN’s 0.957% MAE and 1.086% RMSE. Figure 2 presents IC curves across different voltage ranges for both batteries.

Liu et al. [38] conducted cyclic aging tests on retired batteries, collecting voltage, current, and capacity data during charge/discharge processes. They extracted SOH-related health features from the data using IC, DV, and probability density function (PDF) methods, including peak height, area, and width of IC/DV curves, as well as peak height and area of PDF curves. After dimensionality reduction via principal component regression, they established an SOH estimation model. The MAE for SOH evaluation reached 0.86% (IC-based) and 1.03% (DV-based). While extensive literature confirms IC/DV curves’ utility for SOH assessment, the mathematical definition of IC curves reveals that when dV approaches zero, IC values exhibit infinite spikes—noise that compromises SOH evaluation. To smooth the IC curves and mitigate this noise, researchers have employed various function models to simulate IC curve morphological characteristics [39,40,41].

The IC and DV methods evaluate SOH by analyzing differential features of battery charge/discharge curves. Their primary advantage lies in directly reflecting phase transition processes in electrode materials, offering high mechanistic interpretability and laboratory-grade accuracy. However, these methods demand stringent data acquisition precision and are significantly affected by charge/discharge rates, limiting their applicability under dynamic vehicular operating conditions. For automotive applications, developing anti-interference algorithms and lightweight computing modules is essential to enhance real-time performance.

3.1.4. Internal Resistance Test

The internal resistance of a battery refers to the opposition encountered by electric current flowing through the battery during charge/discharge processes. This resistance primarily consists of contributions from the electrolyte, electrode materials, separator, and interfacial contact resistances between components. As battery internal resistance progressively increases with usage time, internal resistance testing serves as an indirect method for SOH evaluation by measuring the battery’s opposition to current flow [42]. Typically measured by recording voltage variations upon current excitation [43], the internal resistance can be approximated using Ohm’s Law as shown in Equation (3):

R = ∆U/∆I,

(3)

where ∆U denotes the step variation quantity of battery voltage (V), ∆I denotes the step variation quantity of battery current (A).

Researchers [44] demonstrated that Hybrid Pulse Power Characterization (HPPC) testing can be employed for internal resistance measurement. By applying short-duration discharge/regeneration pulses at various DoD conditions and recording voltage variations before/after pulses, the internal resistance during charge/discharge processes can be directly calculated. The study revealed that battery systems with parallel-connected electric double-layer capacitors (EDLCs) exhibited significantly lower average internal resistance compared to standalone batteries, proposing an effective method for internal resistance reduction. This finding holds substantial importance for optimizing battery performance in high-power applications such as hybrid electric vehicles.

The direct current internal resistance method calculates resistance by measuring short-term voltage variations, offering simplicity but lower accuracy due to polarization effects [45]. Literature [46] has proposed an onboard resistance estimation method for vehicular applications without requiring laboratory equipment. During vehicle operation (particularly during HEV engine cranking), battery terminal voltage/current data are acquired. Using a derived application-specific model with a recursive least squares (RLS) algorithm, the method enables online internal resistance estimation. By fitting fresh batteries’ resistance-temperature characteristic curves, temperature effects on resistance estimation are compensated. Ultimately, a degradation index reflecting battery SOH is computed through the ratio between measured resistance and theoretical resistance (derived from fresh battery curves at current temperature).

Chen et al. [47] leveraged the linear relationship between ohmic resistance and battery capacity for SOH estimation. Employing a Thevenin equivalent circuit model with RLS algorithm, they simulated battery dynamic characteristics and identified model parameters (particularly ohmic resistance). Using discharge cycle data from two distinct aging phases, the established linear correlation between ohmic resistance increase and capacity degradation enabled estimation of ohmic resistance at both beginning-of-life (BOL) and end-of-life (EOL) states, significantly reducing computational complexity for online EV applications. The defined SOH formula based on estimated ohmic resistance achieved a maximum absolute error below 4%.

The internal resistance measurement method enables rapid SOH estimation by correlating battery impedance with aging degree, featuring millisecond-level response and low-cost advantages, making it particularly suitable for online monitoring in electric vehicles. However, its measurement accuracy is susceptible to temperature fluctuations, and different aging patterns may lead to similar internal resistance variations. Real-time measurements under dynamic operating conditions also present challenges. Potential improvements include: developing temperature-SOC coupled compensation models, employing machine learning to decouple aging path characteristics, and optimizing dynamic measurement strategies based on vehicle operational data. Multidimensional technology integration can enhance the method’s reliability and accuracy in complex vehicular environments.

3.1.5. Other Experimental Techniques

The cycle counting method indirectly reflects battery SOH by recording the number of complete charge-discharge cycles experienced by the battery [48]. This approach is simple and practical, making it suitable for real-time BMS recording. Cycle counting can serve as an auxiliary method for SOH estimation in EVs and HEVs, particularly when rapid approximate assessment of battery degradation is required. However, due to its inherent limitations in ignoring other critical factors and the varying impacts of shallow cycling versus deep cycling in automotive applications, this method cannot provide sufficiently accurate SOH evaluation results.

Ultrasonic testing is a novel battery SOH assessment technique. Its non-destructive nature allows for battery reuse, making it highly attractive. This method is sensitive to internal structural changes and can detect physical alterations in electrode materials [49]. Figure 3 demonstrates the acoustic signal responses observed by researchers. However, as this method heavily relies on acoustic signals, and the battery encapsulation process in EVs may interfere with signal transmission, the correlation between acoustic characteristics and SOH cannot be well established. Currently, its accuracy is inferior to traditional electrochemical SOH evaluation methods, and commercial applications remain immature.

Open Circuit Voltage (OCV) refers to the terminal voltage of a battery at rest without charge/discharge. Different SOC levels correspond to distinct OCV values, and the SOC-OCV curve of aged batteries may exhibit shifts. Analyzing these shifts enables indirect SOH estimation [50]. For batteries with steep voltage plateaus like lithium nickel manganese cobalt oxide battery (NMC), the distinct OCV-SOC relationship makes this method particularly suitable. However, for batteries with flat voltage curves such as lithium iron phosphate (LFP), the method proves less effective and is significantly influenced by temperature and rest time. Consequently, it primarily serves as a secondary health indicator in a vehicle BMS rather than a standalone SOH assessment method.

3.1.6. Brief Summary

The selection of optimal SOH estimation methods depends critically on application-specific requirements. Coulomb counting demonstrates broad applicability across all battery types due to its fundamental charge measurement principle, though it necessitates periodic full calibration to maintain accuracy. EIS offers precise electrochemical characterization, making it particularly suitable for laboratory environments and stationary storage systems where specialized equipment is available. IC/DV analysis proves most effective for electric vehicles, as it leverages standard charge/discharge protocols without requiring additional hardware. Internal resistance testing emerges as the preferred approach for hybrid systems and emergency monitoring applications, where its ability to reflect gradual degradation trends provides critical safety insights. Each method’s distinct advantages align with specific operational constraints and accuracy requirements across these diverse implementation scenarios.

3.2. Model-Based Techniques

In EV and HEV applications, model-driven SOH estimation techniques primarily employ mathematical formulas or physical models to describe battery behavior, combining the battery’s physical/chemical characteristics or statistical data patterns for health state prediction. This review categorizes model-based SOH assessment methods into the following types: equivalent circuit model (ECM)-based approaches, electrochemical model-based methods, and mathematical model-based techniques.

3.2.1. ECM-Based Methods

ECM-based methods represent a widely adopted technical approach in a BMS for EVs and HEVs. The core concept involves simulating battery dynamic characteristics through electrical circuit components, coupled with parameter identification and state estimation algorithms, to indirectly quantify battery health status. Common ECM-based methods include the Rint model, Thevenin model, PNGV model, and n-order RC models (n ≥ 2) [51,52]. Figure 4 illustrates schematic diagrams of three typical ECMs.

Researchers [53] developed a voltage-controlled ECM-based SOH estimation method. The study employed two ECMs for SOH estimation: a basic model containing only a voltage source and resistor, and an extended model that added resistor-capacitor circuits and solid-phase diffusion components to more accurately represent battery dynamics. This method shows robustness to battery aging, enables SOH estimation from arbitrary load profiles, and has a simpler numerical implementation than existing model-based methods. For fresh batteries under full-cycle and shallow-cycle conditions, the basic model achieved MAEs of 0.18% and 0.61% respectively. Under dynamic Worldwide Harmonized Light Vehicle Test Procedure (WLTP) cycling conditions, the basic model showed 6.7% MAE while the extended model reduced it to 1.23%.

Li et al. [54] improved the fitting performance of the existing ECM by adding an additional capacitor, particularly enhancing the fitting accuracy in low-frequency regions. They combined EIS data with ECM and a Gaussian Process Regression (GPR) algorithm for lithium-ion battery SOH estimation. This method fully utilizes the internal battery information contained in EIS data, improving SOH estimation accuracy. Validation using different temperature conditions and lithium-ion battery types showed the method achieved an average RMSE of only 1.77%. Even for the worst-performing battery in estimation, the RMSE was merely 2.95%. Reference [55] integrated ECM with a Kalman filtering algorithm to accurately estimate SOH during battery degradation. Using both the Thevenin model and the second-order Thevenin model for SOH evaluation, the MAE remained below 5%. While traditional ECMs typically only consider the effects of SOC and temperature on parameters, reference [56] systematically investigated the impact of SOH on Thevenin model parameters for the first time. Based on experimental data, the authors proposed an empirical model to quantitatively describe the influence of SOH, SOC, and temperature on ECM parameters. Through various current profiles and temperature conditions, the model demonstrated accuracy superior to conventional ECMs, while its relatively low complexity ensured feasibility for integration into a BMS.

The ECM-based approach simplifies battery systems through resistor-capacitor (RC) networks, achieving both engineering practicality and mechanistic interpretability. Its advantages include: RC networks can quantitatively characterize polarization effects across different time scales; the mathematical relationship between parameters and capacity degradation is well-defined; and the real-time identification algorithms require minimal computational resources, making them suitable for onboard BMS applications. However, limitations exist: difficulty in modeling structural evolution of electrode materials, coupling between temperature and aging effects, and lack of universal topology due to variations in battery chemistries. For EV/HEV applications, key development directions include: variable-topology ECMs, temperature-aging decoupling algorithms, cross-model parameter transfer techniques, and impedance spectroscopy integration to improve high-frequency accuracy, ultimately enabling further reduction of SOH estimation errors.

3.2.2. Electrochemical Model-Based Methods

Electrochemical models simulate battery behavior by describing the internal electrochemical reaction processes. These models typically involve key processes such as solid-phase diffusion, liquid-phase transport, charge transfer, and charge conservation, enabling more accurate reflection of dynamic changes during battery charge/discharge cycles. In electrochemical models, SOH is intrinsically linked to variations in internal battery parameters, such as active material loss, increased internal resistance, and electrolyte decomposition. Monitoring these parametric variations enables assessment of battery SOH. The most prevalent electrochemical models are the pseudo-two-dimensional (P2D) model and the single particle (SP) model [57,58,59], with Figure 5 illustrating the schematic of a classic lithium-ion battery P2D model.

Reference [60] integrates electrochemical, thermal, and aging effects within a unified model, comprehensively accounting for factors influencing battery performance. By dynamically updating model parameters based on aging indicators, including internal temperature, solid electrolyte interface (SEI) film thickness, and cathode crack depth, simulation accuracy is significantly enhanced. The incorporation of these aging parameters enables high-precision SOH estimation, achieving an RMSE below 0.5% in SOH prediction. Wu et al. [61] employed an adaptive particle swarm optimization algorithm to identify SP model parameters and utilized GPR for battery SOH prediction.

Zhou et al. [62] extracted seven electrochemical features from battery data using a modified SP model and developed a hybrid model integrating temporal convolutional networks with bidirectional long short-term memory (LSTM) for SOH prediction based on these features. This approach achieved an MAE of 1.71% for SOH prediction on the Oxford University dataset. Xu et al. [63] realized simultaneous estimation of SOC and SOH by combining a minimalistic electrochemical model with an equivalent circuit model, with the SOH prediction MAE reaching approximately 2%. Chen et al. [64] acquired mechanism features related to battery aging using IC and DV analysis techniques and theoretically eliminated polarization effects in typical discharge curves via electromotive force extraction, enabling electrochemical voltage spectroscopy analysis in high C-rate charging scenarios. By integrating electrochemical modeling with data-driven techniques, they elucidated battery degradation mechanisms and achieved online SOH estimation with a maximum error below 3%.

The greatest advantage of battery SOH estimation methods based on electrochemical models is their ability to directly associate microscopic degradation mechanisms with macroscopic capacity fade and provide early warnings for safety hazards such as lithium plating. However, this method has high computational complexity, expensive parameter calibration costs, and is sensitive to manufacturing process differences. For vehicle applications, possible improvements include developing neural network reduced-order models to reduce computational load, constructing cloud-vehicle collaborative architectures for dynamic parameter updates, and using machine vision technology to quickly detect electrode microstructure parameters. These innovations are expected to meet real-time requirements for vehicles while maintaining mechanistic accuracy.

3.2.3. Mathematical Model-Based Methods

Mathematical models, as an important tool, can simulate and predict the battery aging process without relying on electrochemical mechanisms, making them suitable for battery SOH evaluation in engineering practice. Their core principle is to extract degradation features from historical data using empirical degradation models or probabilistic models, establishing a mapping relationship between SOH and time or cycle count [65].

The SOH evaluation method based on empirical models assumes that battery performance degradation follows specific mathematical rules. It determines equation parameters from battery historical data and uses these mathematical equations to fit the decay patterns of battery capacity/internal resistance with cycle count or time, establishing a degradation trajectory equation to extrapolate future performance changes. Generally, the equation includes aging factors such as temperature and rate [66], and increasing these factors typically improves the accuracy of SOH evaluation. Reference [67] uses empirical models to construct calendar aging and cycle aging models for lithium-ion batteries, with the calendar aging model shown in Equation (4) and the cycle aging model shown in Equation (5).

$$C_{loss,cal}=A\times exp\left({\displaystyle \frac{{-E}_A}{R·T}}\right)\times t^B\times {SOC}^C$$

(4)

where A, B, C are different coefficients, the calendar aging model’s capacity loss is related to temperature T, time t, and SOC, determined by the Arrhenius law method.

$$C_{loss,cycl}=f\left(SOC,T,C\right)\times {EN}^D,$$

(5)

where the cyclic aging model is a weighted energy throughput model, with the weighting function determined by SOC, temperature T, and charge-discharge current C.

The SOH evaluation method based on probabilistic models quantifies the uncertainty of the battery aging process through PDF, with the core idea of treating SOH as a random variable rather than a deterministic value. This approach is particularly suitable for addressing uncertainties arising from measurement noise and individual differences. Reference [68] uses the PDF method to extract features from charging voltage data, avoiding the need for complex filtering of IC curves required by traditional IC analysis methods. By combining the PDF method with the GPR algorithm, the accuracy and robustness of SOH evaluation are improved. The proposed method was validated on the NASA dataset, with all batteries achieving an MAE within 1.1% and an RMSE within 1.2%. For the CALCE dataset, the MAE and RMSE for all batteries were below 3%.

The method based on empirical models has the outstanding advantage of extremely low implementation cost and almost zero computational requirements for the BMS, making it particularly suitable for scenarios like low-cost electric micro-vehicles where precision is not critical. The probabilistic model-based method not only provides point estimates but also offers critical risk assessment indicators such as confidence intervals, which have unique diagnostic value for the complex aging patterns caused by frequent start-stop in HEVs. However, both methods have inherent limitations: empirical models have poor extrapolation capabilities, and probabilistic models are sensitive to prior distribution assumptions. In practical onboard applications, these should be integrated with other technologies to ensure the accuracy of SOH evaluation.

3.2.4. Brief Summary

The selection of battery SOH estimation models depends fundamentally on application-specific requirements and operational constraints. ECMs demonstrate optimal suitability for EVs/HEVs, where their computational efficiency enables seamless real-time integration with the BMS. Electrochemical models provide superior accuracy for laboratory and stationary storage applications, as their detailed physicochemical representations facilitate comprehensive offline analysis, albeit at higher computational costs. Mathematical models emerge as the most practical solution for low-cost systems, where empirical aging models based on simplified mathematical formulations offer an effective balance between implementation complexity and performance. Each modeling approach exhibits distinct advantages that align with the specific demands of these operational environments, ranging from real-time processing requirements to analytical depth and cost considerations.

3.3. Data-Driven Techniques

Undoubtedly, data-driven battery SOH prediction technology is one of the current research hotspots. It predicts future SOH by analyzing historical operational data of batteries. This approach does not require a deep understanding of the internal physicochemical mechanisms of batteries, making it more flexible and convenient in practical applications for EVs and HEVs [69,70,71]. This review categorizes data-driven SOH evaluation methods into the following types: filtering-based methods, fuzzy logic, artificial neural networks, and support vector machines.

3.3.1. Filtering-Based Methods

Filtering-based methods typically use the Kalman filter (KF) [72], particle filter (PF) [73], or their variants to handle noise and uncertainty in battery operation data, thereby estimating the battery’s SOH. These methods achieve accurate SOH prediction by recursively updating battery state estimates, combining prediction and correction steps. The KF is a recursive optimal estimation algorithm used to estimate the state of a dynamic system from noisy observation data. It is based on the assumption of a linear Gaussian system, meaning both the system model and observation model are linear, and the noise follows a Gaussian distribution. The PF is a recursive Bayesian filtering algorithm based on the Monte Carlo method, suitable for state estimation in nonlinear and non-Gaussian systems. It approximates the posterior probability distribution of the system through a set of weighted random particles. Each particle represents a possible state of the system, and its weight reflects the probability of that state.

The Extended Kalman Filter (EKF) is a nonlinear extension of the KF that linearizes the nonlinear dynamic model of the battery system and recursively updates the battery state estimates by integrating real-time measurement data. Reference [74] applies EKF in the battery management system of HEVs to achieve accurate estimation of key parameters such as SOC, SOH, power fade, and capacity fade of the battery pack.

The Dual Extended Kalman Filter (DEKF) can be considered as consisting of two coupled EKFs, one for state estimation and the other for parameter estimation. Reference [75] achieves online estimation of lithium-ion battery SOH through DEKF. Researchers [76] combined an improved sine-cosine algorithm with DEKF, optimizing the noise covariance matrix to enhance the accuracy and robustness of SOH estimation. Li et al. [77] used a Dual Adaptive Extended Kalman Filter (DAEKF) for real-time estimation of SOH, dynamically adjusting the window size based on the residual innovation sequence to better match the actual noise characteristics of the current system, achieving an RMSE of 0.93%.

The Unscented Kalman Filter (UKF) is an extension of the KF for nonlinear systems, retaining the KF framework (predict-update) but replacing the linearization step with sigma points, offering higher accuracy and suitability for complex systems. Fang et al. [78] estimated SOH using Adaptive Weighted Unscented Kalman Filter (AWUKF), which adaptively adjusts weights based on state and measurement residual vectors, reducing sensitivity to noise. Researchers [79] proposed a UKF-EKF joint estimation method based on a multi-innovation principle, achieving simultaneous estimation of SOC and SOH, with SOH estimation error remaining within 2% under complex conditions.

Reference [80] addresses the particle degradation issue in PF by using an improved firefly algorithm to enhance particle diversity, thereby optimizing PF performance. Compared to EKF, UKF, and PF, the improved PF algorithm has a smaller SOH estimation error, with an average error of 1.34%, outperforming other algorithms. This method is significant for extending battery life and improving the safety of electric vehicle operation. Zhu et al. [81] combined UKF with an Improved Unscented Particle Filter (IUPF) algorithm, using UKF to estimate SOC and IUPF to estimate ohmic resistance, enhancing the accuracy and speed of SOH estimation.

Filter-based battery SOH estimation methods show unique advantages under dynamic conditions. KF and its variants achieve efficient parameter tracking through state space modeling, with low computational complexity, suitable for estimating low-frequency parameters like the ohmic resistance of EVs. PF methods use non-parametric sampling to accurately capture nonlinear degradation processes, making them particularly suitable for predicting capacity drops under HEV pulse conditions. However, KF is sensitive to model mismatch, and PF suffers from particle degradation. Improvements include developing an LSTM-KF hybrid framework to predict process noise, constructing a vehicular network collaborative PF architecture for particle load balancing, and embedding electrochemical models into PF proposal distributions. These innovations can enhance estimation accuracy and reduce PF computational time, offering new solutions for a highly reliable BMS.

3.3.2. Fuzzy Logic

The battery SOH evaluation method based on fuzzy logic is a technique that uses fuzzy mathematics theory to handle the uncertainty and nonlinear characteristics of battery systems. Its core is to transform the complex battery degradation process into a quantifiable evaluation model through fuzzy rules and membership functions [82,83].

The literature [84] extracts key features from impedance data, such as the imaginary part of the impedance at specific frequencies, and uses Mamdani and Sugeno fuzzy inference systems to establish a relationship model between features and SOH. Specifically, for lithium-sulfur batteries, a Sugeno-based three-input one-output fuzzy logic system is used. Researchers [85] collected charging data from real-world EVs, performed data filling and error correction, extracted constant current-constant voltage charging segments, ensured the SOC range was between 30 and 90%, and filtered data with suitable charging currents. They extracted IC peaks on the high voltage platform as battery health features and constructed a fuzzy logic-radial basis function neural network (RBFNN) model, using fuzzy logic to handle uncertainties in influencing factors like temperature, combined with the nonlinear mapping capabilities of RBFNN, to achieve battery aging assessment.

The core advantage of the battery SOH estimation method based on fuzzy logic lies in its ability to integrate expert experience and experimental data, performing well in scenarios lacking precise mathematical models. However, this method has inherent drawbacks such as strong dependency on rule sets, insufficient adaptability, and high computational complexity. For applications in EVs and HEVs, it is crucial to focus on breakthroughs in adaptive fuzzy systems, hierarchical inference architectures, and hybrid modeling frameworks to enhance dynamic adaptability and computational efficiency. Through system optimization, fuzzy logic methods are expected to become an effective supplementary solution for power battery health management.

3.3.3. Artificial Neural Network (ANN)

ANN is a mathematical model that mimics the structure and function of biological neural networks, consisting of multiple processing units (neurons) connected to each other through weighted connections. Neural networks learn complex relationships between input and output data by adjusting connection weights. Battery SOH evaluation technology based on ANN has broad application prospects in fields like electric vehicles, energy storage systems, and smart grids. By monitoring and evaluating the health state of batteries in real-time, the safety and reliability of battery systems can be improved, battery lifespan extended, and maintenance costs reduced [86,87,88]. Figure 6 shows the ANN model structure diagram for lithium-ion battery SOH/RUL estimation constructed in the literature.

In reference [89], features such as cycle number, SOH, and characteristic time are automatically extracted from charge-discharge voltage curves to form a 68-dimensional feature vector. LSTM and gated recurrent units (GRU) are used to train on a fast-charging dataset, and a new loss function is developed to emphasize SOH regression while penalizing its degradation. The LSTM model has a mean error of 5.49% in predicting SOH. Luo et al. [90] proposed a method to estimate SOH online using alternating current impedance data when the battery is fully charged. The amplitude of the alternating current impedance is selected as the input feature for the ANN model, which can reflect changes in internal resistance with aging. A BPNN is used to train the collected alternating current impedance data to establish a mapping between impedance and SOH. The trained BPNN model predicts SOH based on the current alternating current impedance data, achieving an MAE of 0.35%, compared to 0.77% with the linear interpolation method.

Researchers [91] used a hybrid network of convolutional neural networks (CNN) and GRU to extract features from raw data. CNN is used to extract spatial features, while GRU extracts temporal features. Each pair of source domain data is classified to calculate domain classification loss, guiding the model to extract domain-invariant features. A linear layer regresses the extracted features to predict battery SOH. The final loss function combines domain classification loss, SOH prediction loss, and hybrid domain loss, optimizing model parameters through backpropagation. Compared to methods without domain generalization, this approach reduced RMSE from 0.01782 to 0.00649 and MAE from 0.01431 to 0.00519 on the NASA Group 1 dataset.

Liu et al. [92] reconstructed the original V-Q curve through importance sampling to enhance detail capture. DV analysis was performed on the reconstructed V-Q curve to extract features highly correlated with SOH. A fusion estimation model based on improved SVR and CNN was established, with weights of sub-models determined by the least squares method for accurate SOH estimation. Under optimal conditions, RMSE can be reduced to 0.001789 and MAE to 0.001270, demonstrating extremely high prediction accuracy.

Wu et al. [93] added a polynomial to the traditional RBFNN to describe the overall decline trend of SOH, addressing the issue of traditional RBFNN deviating from true values in long-term predictions. Features such as partial constant current charging time, area near the peak, peak, and its corresponding position were extracted from the charging curve. The optimized, improved RBFNN model was used to estimate battery SOH. On the NASA dataset, the proposed method achieved an average MAE of 0.54% and an average RMSE of 0.69%; on the CALCE dataset, the average MAE was 0.90% and the average RMSE was 1.15%.

Battery SOH estimation methods based on ANN do not require precise mechanistic models and are robust to data noise, making them particularly suitable for the variable environments and load conditions in EVs and HEVs. However, ANN methods depend on large amounts of high-quality training data, and the model’s interpretability is poor, making it difficult to analyze specific aging mechanisms. Under dynamic conditions, generalization ability may be insufficient, leading to decreased estimation accuracy. Future improvements should focus on model lightweighting and adaptability to enhance real-time performance and robustness.

3.3.4. Support Vector Machine (SVM)

SVM is a supervised learning model based on statistical learning theory, widely used in classification and regression tasks. It maps the input space to a high-dimensional feature space through kernel functions, effectively handling the nonlinear relationship between battery characteristics and SOH [94,95]. The literature [96] employs the weighted least squares support vector machine (WLS-SVM) algorithm for SOH estimation, weighting error variables through a weighting function to improve model robustness, with the RMSE of the WLS-SVM method being less than 1.85%.

Researchers [97] fused battery data under different working conditions as a training set to improve the model’s generalization ability, using the least squares support vector machine (LSSVM) algorithm for SOH estimation. Compared to traditional SVM, LSSVM offers faster computational speed and higher estimation accuracy, with the RMSE of the LSSVM method generally below 0.02 for multiple test batteries. SVR is a variant of SVM for regression analysis. Reference [98] selected charging time, discharging time, and cycle count as the feature set by analyzing characteristics within different voltage ranges. An improved sparrow search algorithm was used to optimize SVR parameters, resulting in an accurate SOH prediction model with prediction errors limited to within 1.5%.

The SVM-based battery SOH estimation method performs excellently under small sample conditions, effectively handling nonlinear features, and is particularly suitable for precise modeling in laboratory environments. Its kernel function techniques can capture subtle feature changes in the early stages of battery aging, and the structural risk minimization principle ensures model generalization ability. However, this method lacks adaptability to dynamic conditions, with computational complexity increasing significantly with feature dimensions, and fixed kernel functions struggle to adapt to long-term aging evolution. To make SVM an effective tool for balancing accuracy and efficiency in automotive BMSs, it is necessary to optimize the dynamic adjustment mechanism of kernel functions and develop lightweight architectures, while also enhancing real-time performance through feature selection.

3.3.5. Brief Summary

The choice of computational methods for battery SOH estimation depends on specific operational requirements and data characteristics. Filtering techniques, particularly Kalman filters, demonstrate optimal performance in EVs due to their inherent capability to handle dynamic operating conditions through recursive state estimation. Fuzzy logic systems prove most effective for hybrid systems, where their tolerance for imprecise inputs and ability to incorporate expert knowledge address the inherent uncertainties in such applications. ANNs exhibit superior performance in stationary storage systems, leveraging their big data processing capacity and nonlinear mapping capabilities for accurate long-term predictions. SVMs emerge as the preferred choice for a small/medium BMS, where their structural risk minimization principle provides robust performance advantages with limited training samples. Each method’s distinctive algorithmic properties align with the specific demands of these diverse operational scenarios.

3.4. AI-Fused Advanced Techniques

The rapid development of AI has led to many cutting-edge battery SOH estimation technologies. While traditional data-driven methods (e.g., filtering, fuzzy logic, ANN, SVM) emerged around 2015, focusing on statistical pattern recognition, the more recent AI-fused advanced techniques (since ~2020) represent a paradigm shift toward system-level intelligent solutions. These newer approaches—including digital twin, physics-informed neural networks (PINN), federated learning, explainable AI (XAI), edge-cloud collaborative computing, and large language models—demonstrate three key advantages: (1) deeper system integration beyond standalone algorithms, (2) novel computing paradigms, and (3) enhanced accuracy and robustness, particularly suited for complex EV/HEV scenarios. Notably, some AI-fused techniques (e.g., LLM applications) remain in exploratory stages with limited publications (only 10–20 papers in Web of Science). Given the current lack of comprehensive reviews on these emerging AI-fused techniques for SOH assessment, this systematic review aims to bridge this critical gap in the literature.

3.4.1. Digital Twin

Digital twins use digital modeling, real-time data synchronization, and simulation technology to create dynamic mirrors of physical entities in virtual space, enabling virtual-real interaction, state monitoring, and predictive optimization. Digital twin dynamic evaluation systems are an important development direction for intelligent battery management, contributing to performance optimization, efficiency improvement, and intelligent advancement in EVs and HEVs [99].

Researchers [100] proposed a real-time SOH estimation method based on digital twins, utilizing innovative cyclic data synchronization, temporal attention mechanisms, and online data reconstruction technology to achieve accurate SOH estimation under partial discharge cycle data. For most sampling times, the proposed method achieves SOH estimation errors of less than 1% during continuous cycles. Figure 7 [101] illustrates a digital twin-based electric vehicle battery management framework, which combines cloud SOH prediction and vehicle-side SOC estimation, introducing incremental learning technology to enhance model generalization and prediction accuracy. This achieves efficient resource utilization and precise battery state monitoring, with a mean squared error (MSE) of 0.022 in the NASA battery dataset.

The literature [102] uses wavelet analysis to extract frequency domain features from battery voltage and current signals, thereby estimating the internal impedance of the battery, providing critical data for the construction of battery digital twins. By training neural networks to simulate changes in the OCV and internal resistance with SOC and SOH, accurate modeling of battery behavior is achieved. Saba et al. [103] combined digital twin technology with the twin delayed deep deterministic policy gradient algorithm for energy management in extended-range electric vehicles, achieving real-time state (SOC, SOH, SOE, RUL) estimation accuracy of up to 99.8%. Meanwhile, the twin delayed deep deterministic policy gradient algorithm dynamically adjusts energy allocation strategies based on real-time state information provided by the digital twin, optimizing energy use efficiency and vehicle performance in extended-range electric vehicles.

The battery SOH estimation method based on digital twins achieves high-precision health state monitoring through real-time interaction between virtual models and physical batteries. Its advantage lies in integrating multi-source data to accurately reflect the battery aging process, making it particularly suitable for scenarios with high safety and reliability requirements, such as EV battery pack management. However, this method has high computational complexity, strict hardware requirements, and requires a large amount of data to support model construction. In the field of EVs and HEVs, digital twin technology has potential applications, but issues of real-time performance and cost must be addressed for large-scale deployment.

3.4.2. PINN-Based Methods

PINN incorporates physical laws (such as partial differential equations and conservation laws) directly into the loss function of neural networks, allowing the network to not only fit the given data but also adhere to physical principles during training. This approach makes PINN excel in solving complex physical problems, especially in situations where data are scarce or noisy.

Wang et al. [104] cleaned, transformed, and segmented the collected EV battery data, categorizing vehicle states into three types: parking, driving, and charging. Then, based on multi-stage constant current charging data, they calculated the current battery capacity using IC curve analysis to obtain SOH. Next, they used a latent Dirichlet allocation model to cluster driving behaviors and analyze their impact on battery SOH. Key features affecting battery SOH, including driving behavior and other relevant state features, were extracted from the preprocessed data. Finally, they constructed an uncertainty-weighted PINN model, using the extracted features as input to train a predictive model for battery SOH. The mean absolute percentage error (MAPE) of the SOH prediction was 2.6862%, demonstrating an effective method for handling real-world EV data.

Researchers [105] extracted statistical features from short-term data before battery charging, including mean, standard deviation, kurtosis, and skewness, as inputs for the model. They constructed a PINN model comprising a solution function f(⋅) for mapping features to SOH and a nonlinear function g(⋅) for modeling battery degradation dynamics. By minimizing data loss, monotonicity loss, and PDE constraint loss, the PINN model was trained to accurately estimate battery SOH, achieving MAPE values of 0.85%, 1.21%, 0.65%, and 0.78% across four datasets. Sun et al. [106] transformed the monotonic relationship between IC curve peaks and SOH into physical constraints embedded in the neural network’s loss function to guide model training. On the NASA dataset, compared to feedforward neural networks (FNN), LSTM, and CNN, the proposed PINN’s MAE and RMSE were reduced to 0.2164% and 0.2685%, respectively. Figure 8 shows the flowchart of the SOH estimation method.

The literature [107] integrated the partial differential equations of Fick’s diffusion law, which describe solid-phase lithium-ion diffusion behavior, into the neural network’s loss function. Under different battery and temperature conditions, the RMSE of SOH estimation ranged from 1.1% to 2.3%. Fu et al. [108] embedded the ECM into a recurrent neural network (RNN) to construct a PINN model. By training the PINN model, they identified ECM parameters, including ohmic resistance and RC network parameters. Using the linear relationship between the identified ohmic resistance and battery capacity, they estimated the battery’s SOH with an error of less than 2%.

The PINN-based battery SOH estimation method combines the advantages of data-driven and mechanistic modeling. Its benefits include achieving high-accuracy predictions with minimal data while maintaining physical interpretability, making it particularly suitable for studying battery aging mechanisms in laboratory settings. However, this method has high computational complexity and is sensitive to boundary conditions, with room for improvement in real-time performance under complex automotive conditions. Although applicable to the EV domain, it requires optimization for onboard computing resources and enhanced robustness of the model under varying temperature and load conditions.

3.4.3. Federated Learning

Federated learning is a distributed machine learning method that allows multiple devices or data centers to collaboratively train a shared model while keeping data localized [109]. The core of this method is that each participant uploads only model updates to a central server, rather than raw data, thereby protecting user privacy and data security. Battery data in EVs and HEVs include sensitive information such as user driving habits and routes. Federated learning enables battery SOH estimation without collecting this data, thus protecting user privacy. Automakers and battery manufacturers can share knowledge without sharing sensitive data, solving the data silo problem and protecting trade secrets [110,111]. Figure 9 specifically shows how researchers used federated learning to estimate SOH.

Chen et al. [112] improved the adaptability and generalization of the global model through Contribution-Aware Federated Strategy (CAFS), which performs weighted aggregation based on the contribution of each client model. The battery health prediction model incorporates feature-enhanced autoencoders and a Time Period Coupled Attention (TPCA) mechanism to effectively extract battery aging features and learn the relationship between time and periodic fluctuation information. The TPCA mechanism helps the model better capture the periodic fluctuations in battery capacity curves, improving SOH estimation accuracy, with an MAE of 0.87% at room temperature.

Wang et al. [113] proposed a battery SOH estimation method based on adaptive personalized federated learning, which combines importance weight differences, adaptive personalization layers, and clustering techniques to improve the model’s adaptability to individual differences, achieving more accurate SOH estimation. Additionally, the study integrated a SOH-related differential privacy protection mechanism to enhance local battery data protection while ensuring robust model performance. Compared to traditional federated learning, the MAE decreased by 0.14%; compared to local training methods, the MAE decreased by 6.01%.

Reference [114] combined federated learning and ensemble learning for battery SOH estimation, integrating data from multiple stakeholders such as charging stations, roadside units, and battery manufacturers. It used each EV trip as an edge scenario to update the SOH estimation model in real-time, allowing it to adapt to different operating conditions, with an average error of 3.24%.

The federated learning-based battery SOH estimation method ensures data privacy protection, making it particularly suitable for scenarios requiring cross-enterprise data collaboration. However, it faces challenges such as high communication overhead and model bias due to uneven client data distribution. While this method facilitates collaborative analysis of battery health data among automotive companies, its practicality can be enhanced by optimizing communication efficiency and improving the model’s adaptability to heterogeneous data.

3.4.4. XAI-Based Methods

In traditional black-box models, although they can provide accurate predictions, their decision-making processes are opaque to humans. XAI technology provides the rationale and logic behind model decisions, allowing people to understand and trust the predictions [115]. Using XAI technology for battery SOH assessment is significant for EVs as it not only enhances decision transparency and credibility but also optimizes battery management and maintenance, and improves fault prediction and prevention [116]. As XAI technology continues to develop, its application prospects in the EV field will be even broader.

Researchers [117] applied XAI technology to the assessment of lithium-ion battery SOH, enhancing the transparency and interpretability of model predictions. As shown in Figure 10, Shapley Additive Explanations (SHAP) values were used to analyze in detail the importance and impact direction of features such as temperature, current, cycle count, and voltage in predicting the discharge capacity of LIBs, with an SOH prediction MAE of 0.103.

Huang et al. [118] also used the SHAP method to analyze the contribution of different features to the prediction of SOH by a temporal convolutional neural network model. Based on the SHAP values, the three features with the greatest contribution to SOH prediction were selected. These features are prioritized in subsequent model training, while others with lesser contributions are considered redundant and excluded. This not only improves the accuracy and scientificity of feature selection but also enhances the model’s interpretability. Reference [119] proposed a novel SOH estimation method based on inverse ampere-hour integration and the natural gradient boosting algorithm, validated using real operational data from Vehicle-to-Grid EVs, with an MAPE of 1.484%. Additionally, XAI analysis revealed that the number of charge and discharge cycles contributed the most to the model output, increasing model transparency and providing new insights for lithium-ion battery health assessment.

XAI has significant advantages in enhancing model transparency. Through feature importance analysis and attention mechanisms, it can intuitively demonstrate the impact of key factors like temperature and charge/discharge rates on battery aging, making it particularly suitable for high-reliability energy storage systems and second-use scenarios. However, XAI methods still face a balance between interpretability, accuracy, and computational efficiency, and their stability under dynamic conditions needs improvement. XAI methods help enhance the credibility of BMSs but require optimization for real-time performance to suit onboard environments.

3.4.5. Edge-Cloud Collaborative Computing

Edge-cloud collaborative computing is a technological architecture that combines edge computing (local processing near the data source) with cloud computing (centralized high-performance processing) [120,121]. The edge can quickly process real-time battery data, achieving millisecond-level preliminary SOH assessment, avoiding the impact of cloud communication delays on safety warnings. Meanwhile, energy-intensive deep learning models are deployed in the cloud, with the edge running only lightweight algorithms, reducing onboard hardware costs [122]. Figure 11 illustrates the SOH estimation framework based on edge-cloud collaborative computing conducted by Chen et al. [123]. Researchers used a variational mode decomposition optimized by a multi-population evolutionary whale optimization algorithm to decompose capacity data, obtaining global components representing capacity decay trends and fluctuating components representing capacity regeneration. The decomposed components are simultaneously input into a Transformer model, followed by model training in the cloud. The trained model is deployed to edge devices for real-time SOH estimation, and prediction results are transmitted to the onboard BMS for data exchange, showing a maximum MSE of no more than 0.009% on three public datasets.

The battery SOH estimation method based on edge-cloud collaborative computing achieves a balance between real-time performance and computational accuracy through a distributed architecture. Its advantage lies in ensuring real-time response at the edge and providing high-precision analysis in the cloud, making it particularly suitable for operational scenarios of EVs/HEVs that require continuous monitoring. However, this method is limited by the stability of vehicular network communication and the computational power bottleneck of edge devices. It is crucial to improve dynamic resource scheduling mechanisms and automotive-grade communication protocol design to adapt to the automotive field.

3.4.6. Large Language Model (LLM)

LLM refers to a class of natural language processing models with massive parameters and powerful data-driven structures. These models are trained using deep learning techniques on large volumes of text data, learning the statistical patterns of language and enabling the generation and understanding of natural language text [124,125]. By integrating multi-source data and optimizing feature engineering, LLMs can significantly enhance the accuracy of battery SOH evaluation, providing strong support for the safe operation of EVs and HEVs.

Zhang et al. [126] proposed a multi-cycle charging information fusion feature combination method to address the challenges of feature extraction caused by fragmented single-cycle charging data, enhancing the flexibility and complexity of feature engineering. Bidirectional Encoder Representations from Transformers (BERT) was chosen as the pre-trained LLM, and Figure 12 shows the model architecture. By modifying input types, output layer architecture, and dataset partitioning for model fine-tuning, cross-domain knowledge transfer was achieved, resulting in excellent estimation performance across multiple datasets, with RMSE no greater than 0.0203 and MAPE no greater than 1.99%.

Reference [127] transfers the strong generalization capability of the Generative Pre-trained Transformer (GPT-4) to the single-variable supervised regression problem of battery SOH estimation. Through knowledge distillation from teacher-student models and lifelong learning, GPT-4 can adaptively estimate SOH while minimizing the need for fine-tuning on new data. Test results on embedded systems show that this method maintains high accuracy (RMSE less than 1%) with low computational cost, demonstrating potential for large-scale data processing and complex task learning.

LLM-based battery SOH estimation methods have unique advantages, capable of handling unstructured data and achieving small sample learning through pre-training, making them suitable for scenarios like battery recycling evaluation. However, they require significant computational resources, can accumulate errors in temporal predictions, and lack electrochemical interpretability. While applicable in EVs and HEVs fields, challenges remain in onboard deployment and mechanism integration.

3.4.7. Brief Summary

The selection of advanced battery analytics methodologies exhibits distinct domain-specific advantages based on operational requirements. Digital twin technology proves optimal for EVs and high-precision BMS applications, providing real-time virtual replicas for accurate state estimation. PINNs demonstrate particular efficacy in hybrid systems by effectively integrating physical laws with data-driven approaches through their dual neural network architecture. Federated learning emerges as the preferred solution for fleet management scenarios, enabling privacy-preserving collaborative model training across multiple vehicles. XAI techniques are indispensable for stationary storage systems facing regulatory compliance demands, as they provide interpretable decision-making processes. Edge-cloud collaborative computing offers superior performance for large-scale EV fleets by distributing computational loads between local and centralized resources. LLMs show unique potential in research prototype development, where their text-data fusion capabilities facilitate comprehensive analysis of complex battery phenomena. Each approach addresses specific challenges across these diverse operational contexts through its specialized technical characteristics.

4. Comparison of Different SOH Estimation Methods

Detailed analysis of various SOH estimation methods mentioned in some of the reviewed literature is shown in

Table 1. Comparison of recent studies on different methods of SOH estimation in the literature.

Category	Method	Battery Chemistry/Type	Ref./Year	Operation Mode	Evaluation Index	Estimation Error
Experimental measurement techniques	Coulomb counting	CGR 18650 LIB	[28]/2009	Offline	N.A.	N.A.
	EIS	NMC	[33]/2021	Offline	RMSE	<1.74%
	IC/DV	LiFePO4	[35]/2019	Online	Estimation error	<4%
	Internal resistance test	NMC	[47]/2018	Online	Maximum error	<4%
Model-based techniques	ECM	LiCoO2	[54]/2024	Offline	RMSE	<2.95%
	Electrochemical model	LiCoO2	[63]/2022	Online	Mean error	~2%
	Mathematical model	LiCoO2 NCA LiFePO4	[68]/2023	Offline	MAE	1.55% 0.82% 0.81%
Data-driven techniques	Filtering	LiCoO2	[79]/2025	Online	Estimation error	<2%
	Fuzzy logic	LiFePO4	[85]/2025	Online/Offline	N.A.	N.A.
	ANN	LiCoO2	[90]/2022	Online	MAE	0.35%
	SVM	NMC LiFePO4	[96]/2021	Online	RMSE	<1.85%
AI-fused advanced techniques	Digital twin	LiFePO4	[100]/2023	Online	Estimation error	<1%
	PINN	NCA NMC	[106]/2025	Online	MAE RMSE	<0.4%
	Federated learning	NMC	[111]/2025	Offline	Estimation error	2.80%
	XAI	LiCoO2	[115]/2025	Offline	MAE	0.104
	Edge-cloud collaborative computing	LiCoO2 LiFePO4	[123]/2024	Offline	MAE	<0.6%
	LLM	NCA LiFePO4 NMC	[127]/2025	Online/ Offline	Mean RMSE	0.64%

, while

Table 2. Evaluation of SOH estimation methods.

Category	Method	Advantages	Drawbacks
Experimental measurement techniques	Coulomb counting	Simple and intuitive Low cost High real-time performance	Accumulated errors exist Full cycle dependency Aging mechanism blind spots
	EIS	Non-destructive testing characteristics Strong mechanism correlation Sensitive to early degradation	High testing complexity Significant environmental interference Strong model dependency
	IC/DV	Strong mechanism correlation Quantify aging patterns No additional hardware required	C-rate limitations Long testing time Not suitable for platform voltage batteries
	Internal resistance test	Fast and simple measurement Low cost Suitable for online monitoring	Weak correlation with capacity Sensitive to measurement conditions Weak diagnostic capability
Model-based techniques	ECM	High real-time performance Intuitive physical meaning Simple implementation	Limited accuracy Poor adaptation to dynamic conditions Parameter coupling issues
	Electrochemical model	Complete mechanism Reveal aging mechanisms High evaluation accuracy	Extremely high computational cost Numerous and hard-to-obtain parameters High modeling difficulty
	Empirical model	Computationally efficient Rapid deployment	Poor generalization Data dependency
	Probabilistic model	Quantify uncertainty Adapt to complex aging data	Computational complexity Poor mechanistic interpretability
Data-driven techniques	Filtering	High real-time performance Strong robustness	Dependence on precise initial model High computational cost(PF)
	Fuzzy logic	Suitable for uncertain systems Can incorporate expert experience	Limited generalization ability Rule base requires manual design
	ANN	Automatically learn complex relationships Suitable for high-dimensional data	Requires large amounts of training data Black box characteristics
	SVM	Good generalization with small samples Kernel functions handle nonlinear problems	Computationally complex with large data volumes Sensitive to kernel functions and hyperparameters
AI-fused advanced techniques	Digital twin	Real-time simulation of aging process Support virtual debugging and optimization	Dependence on high-precision sensors Huge computational power consumption
	PINN	Integrate electrochemical mechanisms Enhance long-term prediction ability	Complex differential equation solving Difficult training convergence
	Federated learning	Protect data privacy Cross-device collaborative training	Significant communication delays Difficult to align heterogeneous data
	XAI	Transparent decision-making process Meet automotive reliability standards	Reduces model accuracy Increases computational overhead
	Edge-cloud collaborative computing	Low-latency local inference Dynamic model updates	Complex load balancing Sensitive to network stability
	LLM	Natural language interaction Knowledge reasoning assistance	Lack of specialized data Risk of hallucinations

provides a high-level summary of the advantages and disadvantages of different SOH estimation methods. A single method is challenging to balance accuracy, cost, and real-time performance. In practical applications, for experimental measurement techniques, the Coulomb counting method is the mainstay for online SOH estimation in current BMSs, but it requires regular calibration. Internal resistance testing alone is not highly reliable as an SOH indicator, but can be integrated into a BMS as a supplement to Coulomb counting, aiding judgment or triggering calibration. EIS and IC/DV have high mechanistic depth and are the preferred options for aging diagnosis and aging model research, but are limited by cost and real-time performance, mainly used in laboratory or offline scenarios.

Three model-based SOH estimation techniques each have unique features in terms of accuracy, real-time performance, and mechanistic interpretability. Real-time performance and robustness are key requirements for a BMS, making methods that combine ECM and probabilistic models the most practical and mainstream in current onboard BMSs. Electrochemical models are foundational for understanding mechanisms; although they are currently difficult to apply in real-time, they are essential for understanding aging, generating simulation data, and guiding the design of other models. Empirical models are effective under specific conditions; when the application scenario is fixed and data are abundant, simple and efficient empirical models are a practical choice.

Data-driven SOH evaluation techniques are a hot topic in current research. Filtering-based methods achieve dynamic updating capabilities through state space models, making them particularly suitable for the real-time requirements of an onboard BMS. Fuzzy logic methods, with their semantic rule systems, show unique value in simple systems with clear rules, especially suitable for integrating expert knowledge. ANN, as a typical black-box model, has significant advantages in scenarios with large-scale multi-source integration, effectively handling multi-sensor fusion data. SVM demonstrates good generalization performance in small sample scenarios through kernel functions.

AI-fused advanced techniques are leading the trend in future battery SOH assessment development, providing new viable solutions for battery SOH evaluation. Digital twins are suitable for scenarios requiring high-precision, full lifecycle virtual monitoring, with advantages in balancing modeling capabilities and computational resources. PINN is adapted for mechanism-driven extreme condition prediction by embedding physical laws to compensate for data deficiencies. Federated learning addresses collaborative modeling needs under distributed data silos, balancing privacy protection and knowledge sharing. XAI meets transparency decision-making scenarios with strong regulatory requirements, needing to balance interpretability demands and accuracy loss. Edge-cloud collaborative computing is suitable for latency-sensitive real-time monitoring scenarios, with its core in dynamically optimizing computational load and network stability. LLM reduces technical barriers through natural language interaction and can be integrated into battery maintenance knowledge base systems. For real-time monitoring with 5G conditions, edge-cloud collaborative computing is preferred; for multi-agent data, federated learning is adopted; for stringent physical explanations, XAI is chosen; other scenarios can be evaluated comprehensively to choose between digital twins or PINN. Figure 13 provides a visual comparison of four SOH estimation approaches across five critical evaluation dimensions: real-time capability, complexity, data dependency, adaptability, and accuracy. The normalized scoring system enables clear cross-technique trade-off analysis. We integrate all technique comparisons into a summary table (

Table 3. Comparative summary of different SOH estimation methods.

Method	Accuracy (SOH Error)	Computational Cost	Training Epochs	Application Suitability
Coulomb counting	Low (±8–15%)	Very Low (Ah integration)	N/A	All types (requires periodic full calibration)
EIS	High (±2–5%)	Very High (impedance spectroscopy)	100–200+	Lab/stationary storage (requires specialized equipment)
IC/DV	Medium-High (±3–7%)	Medium (high-precision differentiation)	30–80	Electric vehicles (relies on standard charge/discharge protocols)
Internal resistance test	Low-Medium (±5–10%)	Low (Ohm’s law calculation)	N/A	Hybrid systems/emergency monitoring (reflects degradation trends)
ECM	Medium (±5–10%)	Moderate (parameter tuning)	50–150	EVs/HEVs (real-time BMS integration)
Electrochemical model	High (±2–5%)	Very High (PDE solving)	N/A (physics-based)	Lab/stationary storage (offline analysis)
Mathematical model	Low-Medium (±8–15%)	Low (curve fitting)	20–50	Low-cost systems (empirical aging models)
Filtering	Medium (±5–10%)	Low (real-time updates)	N/A	EVs (suitable for dynamic operating conditions)
Fuzzy logic	Medium (±7–12%)	Low-Moderate (rule inference)	N/A	Hybrid systems (suitable for imprecise inputs)
ANN	High (±3–6%)	High (requires GPU acceleration)	100–500+	Stationary storage (big data + long-term prediction)
SVM	Medium-High (±4–8%)	Moderate (kernel computation)	50–200	Small/medium BMS (advantages with small samples)
Digital twin	High (±3–7%)	Very High (multi-physics simulation)	N/A (continuous calibration)	EVs and high-precision BMS (real-time digital replica)
PINN	Very High (±2–5%)	High (PDE-constrained training)	200–500+	Hybrid systems (combining physics + data)
Federated learning	Medium-High (±4–8%)	Moderate (distributed training)	100–300	Fleet management (privacy-preserving multi-vehicle)
XAI	Medium (±5–10%)	Moderate (interpretability overhead)	100–400	Stationary storage (regulatory compliance needs)
Edge-cloud collaborative computing	Medium (±6–12%)	Low-Edge / High-Cloud	50–200 (edge)	Large-scale EV fleets (real-time edge + cloud AI)
LLM	Low-Medium (±10–20%)	Extreme (pre-training + fine-tuning)	500+	Research prototypes (text-data fusion analysis)

) with standardized metrics including accuracy, computational cost, training epochs, and application suitability.

5. Challenges and Recommendations

This review systematically outlines the development of battery SOH assessment technologies and deeply analyzes the advantages and disadvantages of each key technology. However, these technologies still face many challenges and bottlenecks in specific EVs and HEVs application scenarios. Section 5.1 will discuss the challenges of each technology in detail, and Section 5.2 will propose alternative solutions. We hope the suggested framework can provide a reference for the development of an advanced onboard BMS.

5.1. Technical Challenges in EVs and HEVs Applications

First, the battery SOH assessment technology based on experimental measurements faces inherent limitations in dynamic onboard environments, primarily due to accumulated errors in the coulomb counting method. These errors stem from the precision limits of current sensors and insufficient sampling frequency, leading to quantization errors. As the number of battery cycles increases, capacity estimation deviations grow larger without recalibration. The unique partial state of charge (PSOC) operating mode of HEVs [128] results in a lack of complete charge-discharge cycles, causing a loss of capacity calibration benchmarks. The core challenge faced by EIS technology is its “onboard real-time dilemma”. The method requires applying small amplitude alternating current excitation signals across a wide frequency domain and collecting responses, resulting in a single complete measurement taking 5–8 min, which fundamentally contradicts the BMS requirement for second-level updates under dynamic vehicle conditions [129]. This primarily arises from the inherent sequential nature of frequency domain scanning and the phase synchronization precision requirements of the hardware system.

The core drawback of IC/DV analysis is its severe dynamic condition adaptability issue in actual onboard applications, manifested as characteristic peak shifts; for example, higher charging rates can cause shifts in the phase transition plateau potential of the battery [130]. The physical essence of this problem lies in the disruption of the thermodynamic equilibrium assumption. IC/DV analysis relies on quasi-static charge-discharge data (requiring currents below 0.05 C), while actual onboard pulse conditions (peak currents of 2–5 C) induce combined concentration and electrochemical polarization, causing phase transition features to be obscured by noise. The main engineering challenge of internal resistance testing is the interference from multi-physics coupling. Battery internal resistance shows a strong nonlinear relationship with SOC and is simultaneously affected by temperature, historical charge-discharge rates, and other cross-influencing parameters, leading to excessively high SOH estimation errors with traditional threshold methods. The physical mechanism amplifying this error is that LIB aging is essentially a multi-mechanism coupling process, and a single internal resistance parameter cannot differentiate the contributions of different degradation paths.

Although the ECM method is widely deployed in BMSs due to its high computational efficiency, its application in complex onboard environments faces sensitivity issues with time-varying parameters, highlighting the fundamental contradiction between static parameter assumptions and dynamic aging processes. Traditional ECMs (such as second-order RC models) rely on fixed parameters to describe battery behavior, whereas in actual battery aging processes, the electrode interface reaction kinetics and electrolyte diffusion characteristics exhibit strongly coupled time-varying features. The deep-seated cause of this model mismatch is that the lumped parameter nature of ECM cannot reflect the spatially non-uniform aging of distributed parameter systems, leading to a blurring of the physical significance of model parameters.

The challenges faced by electrochemical models are mainly reflected in computational complexity and parameter identifiability. Although the P2D model can describe internal battery processes from first principles, its massive computational load results in excessively long simulation times [131], making it completely unsuitable for real-time onboard monitoring needs. Multiscale coupled modeling is not feasible with the current limited onboard computational resources. SOH estimation methods based on mathematical models still face key bottlenecks in generalization capability and a lack of physical consistency in the complex conditions of EVs. These methods perform extrapolation predictions solely through input-output mapping relationships, neglecting the time-varying characteristics of electrochemical aging mechanisms. Additionally, the parameter optimization process overly relies on specific data distributions, while real onboard environments present nonlinear disturbance factors such as sudden condition changes and historical path dependencies, causing model failure in out-of-domain scenarios.

Traditional filtering algorithms still face fundamental issues of model error accumulation and noise sensitivity in real onboard environments. For instance, EKF relies on local linearization to approximate nonlinear state equations, but the actual battery aging process involves multi-physics coupling, causing the prediction covariance matrix to rapidly distort over long iterations. A significant limitation of fuzzy logic methods is the static nature of rule bases and the bottleneck in knowledge acquisition. Specifically, traditional fuzzy inference systems depend on fixed rule bases built from expert experience, making it difficult to adapt to degradation path bifurcations under complex conditions. Membership function parameter optimization is usually based on offline data, resulting in insufficient dynamic adaptability in real onboard environments.

The biggest challenge faced by ANN is the lack of mechanistic interpretability and excessive data dependency. Neural networks learn only the statistical features of data, completely ignoring the constitutive equation constraints of electrochemical systems, resulting in their extrapolation prediction capabilities being fundamentally limited by the completeness of training data. The technical difficulty of SVM is the fixed feature space and insufficient dynamic adaptability. SVM is essentially a batch learning algorithm, with kernel function parameters and support vectors fixed after training, whereas actual battery aging involves multi-time scale coupling effects, requiring continuously updated feature extraction strategies.

5.2. Recommended Solutions for Current Challenges in EVs/HEVs Applications

In response to the deficiencies of the various SOH evaluation methods mentioned in Section 5.1, we have outlined specific solutions in

Table 4. Detailed solutions for technical challenges of SOH estimation methods.

Method	Challenge	Solution
Coulomb counting	Accumulated error	Develop dynamic Coulomb efficiency compensation algorithm
EIS	On-board real-time dilemma	Use machine learning-optimized sparse EIS measurement strategy
IC/DV	Feature peak shift	Develop dynamic baseline correction system
Internal resistance test	Multi-physical field coupling interference	Construct dynamic baseline internal resistance map
ECM	Parameter time-varying sensitivity issue	Build spatiotemporal-resolved hybrid modeling framework
Electrochemical model	Dual dilemma of computational complexity and parameter identifiability	Develop differentiable computational electrochemical model
Mathematical model	Lack of generalization ability and physical consistency	Develop physics-constrained meta-learning
Filtering	Model error accumulation and noise sensitivity	Construct variational Bayesian-deep reinforcement learning hybrid architecture
Fuzzy logic	Static nature of rule base and knowledge acquisition bottleneck	Couple neuro-fuzzy systems with online incremental learning
ANN	Lack of mechanistic interpretability and excessive data dependency	Develop physics-informed neural networks
SVM	Fixed feature space and insufficient dynamic adaptability	Develop dynamic kernel learning-online SVM hybrid architecture

. Additionally, we have fully integrated the latest advancements in the AI field, proposing multiple technological approaches from different dimensions to develop a systematic solution.

Considering the practical engineering constraints in the industry, especially the challenges in battery SOH estimation for EVs and HEVs, we propose a comprehensive solution with promising technical feasibility and industrial prospects by integrating cutting-edge AI technologies. Methodologically, we suggest overcoming the limitations of traditional PINN frameworks by developing a new paradigm of mechanism-guided self-learning. This paradigm does not simply use physical equations as constraints but constructs a dynamically adjustable mechanism-data weighting mechanism, allowing the algorithm to autonomously discover potential patterns in data-rich areas, while enhancing mechanism guidance in data-sparse regions. Additionally, we propose developing a “layered decoupling” modeling strategy, decomposing the battery aging process into multiple independently modellable sub-processes, and integrating the outputs of these sub-models through an adaptive fusion mechanism. This approach avoids reliance on a complete mechanistic model while maintaining physical interpretability.

In terms of data utilization, it is recommended to build a cross-platform, multi-model collaborative learning mechanism, achieving data value sharing through privacy computing technologies like federated learning. Additionally, establish standardized battery lifecycle data collection protocols to address issues such as incompatible data formats and inconsistent sampling frequencies among different manufacturers. For practical engineering applications, a layered and progressive SOH evaluation system should be designed: deploy lightweight real-time monitoring modules on the vehicle end to perform basic feature extraction and simple state judgment; run moderately complex local models on edge computing nodes to handle state corrections under specific operating conditions; deploy high-precision digital twin systems in the cloud for continuous model optimization and lifespan prediction.

Additionally, it is recommended that the industry jointly promote the standardized application of multimodal sensing technology, combining traditional voltage and current monitoring with novel acoustic and thermal imaging methods to construct a multidimensional battery health status evaluation index system. Finally, during the implementation of technology, special attention should be paid to the model’s transferability and adaptability. Develop a universal evaluation framework for different battery chemistries and vehicle platforms, and achieve continuous updating of model parameters through online learning mechanisms to ensure accurate SOH estimation performance throughout the battery’s lifecycle.

6. Conclusions

This paper comprehensively reviews the current status and challenges of SOH estimation technology for LIBs in EVs and HEVs, aiming to provide a deep understanding and reference for future development directions. From experimental measurements to AI integration, various technologies exhibit different advantages and limitations. Firstly, experimental measurement techniques hold a significant position in laboratory research due to their high precision and clear physical significance, but are limited by cost and real-time capability, making large-scale application in real vehicles challenging. Secondly, model-based technologies, especially ECM, are mainstream in onboard BMSs due to their simplicity and strong real-time performance, but their accuracy and generalization need improvement. Thirdly, data-driven technologies, particularly ANN and SVM, show great potential in big data processing and multi-source integration, but their black-box nature and dependency on training data limit their application. Finally, cutting-edge AI integration technologies, such as digital twins and PINN, offer new ideas and methods for SOH estimation, exhibiting high evaluation accuracy and robustness, and are important directions for future development.

However, various technologies still face numerous challenges in the practical application of EVs and HEVs, such as multifactor coupling effects, model generalization, and computational efficiency. To address these issues, this paper proposes several alternative solutions from multiple dimensions, including developing a new paradigm of mechanism-guided self-learning, building cross-platform collaborative learning mechanisms, and designing a hierarchical and progressive SOH evaluation system. For cost-constrained BMS deployments, the integration of ECMs with Kalman filtering is recommended, achieving ±5% estimation accuracy with processors operating at clock speeds ≤50 MHz. In edge-cloud computing architectures, federated learning demonstrates significant advantages by enabling localized processing of 98% raw data, though its implementation depends on 5G network infrastructure. For high-precision applications requiring prediction errors below 2%, PINNs constitute the optimal approach, necessitating GPU acceleration for computationally feasible execution times. These suggestions aim to enhance the accuracy, real-time capability, and generalization of SOH estimation technology, promoting its widespread application in the field of EVs and HEVs. In the future, as AI and IoT technologies continue to develop, SOH estimation technology will become more intelligent and efficient, providing strong support for the sustainable development of EVs and HEVs.