From Laboratory to Real-World Application: A Comprehensive Study on Battery State of Health Assessment Methods

Abstract

Accurate state of health (SOH) assessment is the cornerstone for ensuring the safety, reliability, and lifecycle value prediction of electric vehicles. While extensive research has demonstrated the significant advantages of data-driven approaches in SOH evaluation, the vast majority of work still relies on standardized test data obtained under laboratory conditions. These ideal conditions, including complete charge–discharge cycles and constant temperatures, are often unattainable in real-world operation where EV batteries face highly irregular driving patterns, fragmented charging segments, and unpredictable environmental disturbances. This paper provides a comprehensive and systematic overview of data-driven SOH assessment based on real-vehicle data, aiming to address the current research gap in unified laboratory-to-vehicle transfer frameworks. This paper first reviews existing SOH evaluation methodologies and highlights the challenges encountered when transitioning to real-world vehicle data. It delves into core technical challenges and solutions across the entire real-world SOH assessment chain, closely examining the complex characteristics of real-world data. The paper thoroughly evaluates the role of cutting-edge paradigms including weakly supervised, self-supervised, and transfer learning in mitigating label scarcity. We summarize a unified evaluation framework tailored for real-world scenarios: Vehicles-Out, Time-Rolling, Domain-Stratified (VTDS). This framework aims to systematically assess models’ generalization limits and engineering deployability across vehicles, time, and operating conditions. This work provides systematic guidance for researchers and practitioners, advancing data-driven SOH evaluation methods from theoretical research to engineering applications.

1. Introduction

With the ongoing electrification of global transportation, lithium-ion batteries are expected to remain the dominant energy storage technology for electric vehicles (EVs) in the foreseeable future [1]. Owing to their high energy density, long cycle life, low self-discharge rate, and flexible modularity, lithium-ion batteries have become the primary power source for EVs [2,3,4]. However, the rapid expansion of EV deployment has been accompanied by recurring battery safety incidents, resulting in substantial property losses and posing severe risks to personal safety, which has intensified the demand for improved battery safety and durability [5,6]. During long-term cycling, lithium-ion batteries inevitably age, manifested by internal resistance growth, capacity fade, and power degradation [7,8]. Consequently, accurate state of health (SOH) monitoring is essential to ensure safe and reliable EV operation.

In practice, SOH is commonly defined as the ratio of the current available capacity to the initial capacity [9,10]. When SOH drops below a critical threshold, timely replacement becomes necessary to mitigate safety hazards. Accurate SOH estimation further enables battery management systems (BMSs) to perform predictive maintenance and health management, thereby extending service life and improving operational reliability [11,12,13,14]. In recent years, large-scale EV adoption has produced massive amounts of in-vehicle field data, offering new opportunities for practical SOH assessment while also introducing non-negligible challenges [15]. Real-world usage involves irregular parking, driving, and charging/discharging behaviors, as well as aging induced by uncertain storage voltages and ambient temperatures [16,17]. Compared with carefully controlled laboratory tests, BMS data collected in the field often suffer from severe quality issues, including high measurement noise, signal distortion, and discontinuities caused by data dropouts [18], as shown in Figure 1. Although numerous reviews have summarized battery health estimation methods from modeling or algorithmic perspectives, most emphasize laboratory-based studies and fully supervised learning. The fundamental mismatch between laboratory datasets and real-world EV data—including incomplete observations, distribution shifts, and sparse labels—remains insufficiently analyzed. More importantly, prior reviews often prioritize model architecture evolution while underestimating the decisive role of data governance and engineering constraints in real-vehicle SOH assessment. To bridge these gaps, this paper systematically reviews SOH assessment methods for real-world applications from a data-feature-model perspective.

The core contributions of this paper include:

(1)

Systematic Analysis: In-depth examination of core challenges and constraints at the data, feature, and environmental levels, spanning from laboratory to real-vehicle applications.

(2)

Cutting-Edge Synthesis: A first-of-its-kind focus on and comparative synthesis of frontier paradigms—weakly supervised, semi-supervised, self-supervised, and domain adaptation—designed to address label scarcity and domain drift issues. Discussion of their application evidence and potential on real fleet data.

(3)

Proposed Evaluation Framework: Advocates a deployment-oriented, real-world evaluation framework—the VTDS benchmark—providing a rigorous, unified framework for assessing future algorithms’ generalization, timeliness, and robustness.

2. Electrochemical Degradation Mechanisms and Health Indicators for Battery SOH Estimation

Battery aging is caused by the combined effects of multiple degradation mechanisms involving different battery components, including electrodes, electrolytes, and current collectors. These mainly include lithium content loss (LLI), loss of active material in electrodes (LAM), and increased impedance associated with interfacial side reactions. These degradation modes lead to capacity decay but exhibit different electrochemical characteristics and aging kinetics. Therefore, using only the scalar SOH index to represent battery health may not fully capture the underlying physical degradation processes [19].

Several studies have highlighted the importance of degradation mode analysis for battery health assessment. For example, Birkl et al. [20] showed that aging mechanism analysis can effectively mitigate battery capacity degradation. By analyzing LLI, active cathode material loss, and active anode material loss, a diagnostic method based on open-circuit voltage analysis was proposed to quantify these effects. Electrochemical diagnostic techniques such as incremental capacity analysis (ICA) and differential voltage analysis (DVA) have been widely used to differentiate aging modes by analyzing characteristic peaks in voltage–capacity curves. These techniques enable researchers to identify electrode-specific degradation mechanisms and understand how LLI and active cathode material loss lead to capacity loss [2]. Furthermore, certain degradation mechanisms, such as lithium deposition, can significantly impact battery safety even with relatively small capacity losses. Lithium deposition typically occurs under low-temperature or high-rate charging conditions and can accelerate degradation or trigger internal short circuits [21]. Therefore, incorporating health indicators oriented towards aging modes into battery diagnostics can significantly improve the interpretability of battery SOH assessments and provide a deeper understanding of battery aging behavior. Recent studies have combined electrochemical understanding with data-driven modeling approaches to improve battery health monitoring [22].

To bridge the gap between purely data-driven “black-box” models and physically interpretable approaches, the electrochemical mechanisms underlying commonly extracted health indicators must be clearly understood. Features derived from ICA and DVA are widely used to characterize battery aging mechanisms. Peaks in the incremental capacity curve correspond to thermodynamic phase transitions occurring within electrode materials during lithiation and delithiation. Variations in peak position, amplitude, and area therefore provide insight into underlying degradation modes. Peak shifts are typically associated with LLI caused by parasitic side reactions, whereas peak attenuation often indicates LAM in the electrodes. Time domain features, including voltage slopes and voltage drops observed at characteristic points during charging, also contain important electrochemical information [23]. These variations reflect changes in thermodynamic equilibrium and internal impedance rather than simple statistical fluctuations. In particular, changes in voltage slope may indicate stoichiometric shifts in electrode materials caused by lithium inventory loss, as well as increased polarization resulting from internal resistance growth. Understanding the electrochemical significance of these health indicators improves the interpretability and robustness of data-driven SOH estimation and supports the integration of machine learning methods with physics-informed battery diagnostics.

3. Laboratory-Based SOH Assessment Methodology and Limitations

3.1. Definition of SOH

The aging of lithium-ion batteries is a long-term, gradual, and inherently nonlinear process. SOH describes the battery’s health condition and is typically expressed as a percentage. In practice, SOH is commonly quantified either by capacity retention or by internal-resistance growth, i.e., the ratio of the current capacity (or internal resistance) to its initial value [24,25], as shown in Equation (1):

$$\left\{\begin{array}{l}SOH={\displaystyle \frac{C_{aged}}{C_{fresh}}}\times 100\%\\ SOH={\displaystyle \frac{Q_{aged-max}}{Q_{new-max}}}\times 100\%\\ SOH={\displaystyle \frac{R_{EOL}-R_{cur}}{R_{EOL}-R_{new}}}\times 100\%\\ SOH={\displaystyle \frac{C_{{\mathrm{n}\mathrm{t}}_{remain}}}{C_{{\mathrm{n}\mathrm{t}}_{total}}}}\times 100\%\end{array}\right.$$

(1)

where $C_{aged}$ and $C_{fresh}$ denote the battery’s current available capacity and initial rated capacity, respectively [26]. However, the initial rated capacity calibrated for batteries may deviate from the actual rated capacity. Therefore, a capacity definition based on charge quantity has been proposed [27]. This is primarily defined through the battery’s discharge capacity, where $Q_{aged-max}$ denotes the maximum discharge capacity of the current battery, and $Q_{new-max}$ represents the maximum discharge capacity of a new battery. Some studies also define it based on internal resistance, where $R_{EOL}$ is the internal resistance at the end of battery life, $R_{new}$ is the internal resistance at the battery’s initial state, and $R_{\mathrm{C}\mathrm{u}\mathrm{r}}$ is the aged internal resistance measured at the current time [28]. Additionally, a SOH definition method based on remaining cycle count has been proposed, where $C_{{nt}_{remain}}$ denotes the battery’s remaining cycle count and $C_{{nt}_{total}}$ represents its total cycle count [29]. However, the above definitions largely rely on idealized assumptions, such as complete charge–discharge cycles and stable operating conditions. In real-world use, battery aging arises from the coupled effects of multiple degradation mechanisms, including active material loss, interfacial film growth, electrode structural degradation, and increased polarization. Their contributions to capacity fade and resistance rise vary substantially across operating regimes, making it difficult for any single indicator to fully capture the true health state. Moreover, in automotive applications, fragmented charging/discharging events and irregular driving profiles prevent batteries from following standardized test protocols, rendering conventional capacity- or resistance-based SOH metrics difficult to obtain directly in the field. Therefore, SOH should be treated as a key latent variable that must be inferred from multi-source operational data. This practical requirement has driven the growing adoption of data-driven SOH assessment and, in turn, raises stricter demands on model design, feature construction, and evaluation frameworks.

Existing SOH estimation methods primarily fall into three categories: direct measurement techniques, model-based SOH estimation methods, and data-driven SOH estimation methods; Figure 2 illustrates the advantages and disadvantages of the different approaches.

3.2. Direct Measurement

Direct measurement methods constitute a typical class of offline SOH assessment techniques. The underlying principle is to obtain physical indicators that directly reflect battery aging—such as capacity, ohmic internal resistance, or impedance characteristics—through standardized test procedures, and then compute SOH according to predefined formulas [30]. These methods are widely regarded as the “gold standard” for SOH evaluation and provide essential references for model validation and algorithm calibration. Among them, capacity testing, open-circuit voltage (OCV) curve analysis, and electrochemical impedance spectroscopy (EIS) are the most commonly used techniques. Tang et al. [31] and Su et al. [32] agreed that the actual capacity measured via constant-current charge–discharge cycles offers the most reliable benchmark for quantifying battery degradation. Nevertheless, this approach has very limited online applicability. In real-world operation, users typically follow shallow charge–discharge patterns, and batteries rarely experience a complete discharge from 100% state of charge (SOC) to 0% SOC. As a result, direct capacity-based SOH evaluation becomes nearly infeasible outside laboratory settings. Furthermore, Liu et al. [33] analyzed impedance responses and showed that EIS is highly sensitive to interfacial resistance, mass transfer impedance, and polarization losses across different frequency bands, highlighting its unique advantage in characterizing degradation mechanisms. However, the required instrumentation and testing cost remain high, which limits large-scale practical adoption.

In summary, direct measurement methods provide irreplaceable benefits in accuracy and physical interpretability. However, their dependence on stringent test conditions, long testing durations, and specialized equipment makes them unsuitable for online monitoring and large-scale deployment in EVs. Consequently, these methods are primarily used as laboratory benchmarks rather than being directly applied to onboard SOH estimation. These limitations motivate the development of SOH assessment approaches based on operational data, forming the foundation for subsequent data-driven modeling and real-vehicle applications.

3.3. Model-Based Method

Within the current SOH research framework, model-based approaches continue to hold significant importance. These methods construct mathematical models that reflect the internal physical or electrochemical behavior of batteries, thereby characterizing the evolution of capacity decay and internal resistance at a mechanistic level. Compared to purely data-driven approaches, model-based methods inherently offer superior interpretability, making them long regarded as essential tools for understanding battery aging mechanisms. Conceptually, existing research can be categorized into three main types: electrochemical models (EMs), equivalent circuit models (ECMs), and empirical models [34,35,36].

3.3.1. Electrochemical Model

EMs directly describe the migration, diffusion, and interfacial reaction processes of lithium-ions in electrodes and electrolytes. The Doyle–Fuller–Newman (P2D) model serves as the theoretical cornerstone in this field [37]. As shown in Figure 3, Li et al. [38] characterized the SEI film growth process based on a single-particle model, enabling quantitative prediction of capacity decay trends. Building upon this, Zhang et al. [39] incorporated side-reaction mechanisms, enhancing SOH estimation accuracy by modeling interfacial impedance and deposition layer thickness. Moritz Streb et al. [40] conducted a systematic sensitivity analysis of the Doyle–Fuller–Newman model parameters, providing theoretical justification for the observability and identifiability of these parameters. Although these studies have progressively advanced electrochemical modeling toward engineering applications, their high dependence on parameter accuracy and computational resources still limits direct deployment in vehicle-based systems.

3.3.2. Equivalent Circuit Model

In contrast, ECMs approximate battery dynamic characteristics using electrical components such as resistors and capacitors. Most literature selects the equivalent current model (ECM) as either the traditional Thevenin model or the second-order RC model. The structures of both models are shown in Figure 4; they strike a reasonable balance between modeling complexity and computational efficiency, and are the mainstream choices for current BMSs. Their core principles involve employing system identification algorithms to track changes in ohmic internal resistance or polarization capacitance online, thereby deriving SOH. Studies by Amir et al. [41] and Zhang et al. [42] consistently reported that the second-order RC model achieves a better accuracy–computational cost balance than lower-order alternatives. Nevertheless, the applicability boundaries of different model orders under highly dynamic operating conditions remain inconclusive. In particular, fixed RC structures may fail to represent all operating regimes, which can cause parameter drift, observer divergence, or delayed state tracking.

3.3.3. Empirical Model

Empirical models bypass complex physical derivations, primarily relying on extensive experimental data to establish mathematical relationships between SOH and aging factors such as cycle count, temperature, and charge/discharge rates through polynomial or exponential functions [43]. Due to their relatively simple modeling process and low computational cost, such methods were widely adopted in early research. Sivalertporn et al. [44] and Dai et al. [45] achieved high fitting accuracy under specific experimental conditions by introducing decay factors and the Eyring equation. These methods depend on offline data; under dynamic operating conditions, the model’s generalization error increases continuously. Krupp et al. [46] investigated calendar aging by analyzing capacity recovery and resistance reduction induced by periodic characterization tests; however, such measurement-induced interference contaminates baseline empirical data, leading to systematic bias when models calibrated on laboratory datasets are applied to real-world conditions where such testing is absent. Even when data biases are mitigated, analytically capturing multifactor aging mechanisms remains challenging. Redondo-Iglesias et al. [47] introduced an Eyring-based calendar aging model to couple temperature and dynamic state-of-charge (SOC) effects, but incorporating multiple stress factors significantly increases model complexity and becomes difficult to generalize under highly variable operating conditions. Consequently, these limitations substantially degrade model performance in practical applications. Jin et al. [48] showed that empirical degradation models can exhibit prediction errors of up to 71% under uncalibrated out-of-distribution scenarios, requiring frequent recalibration during real-world deployment. These challenges highlight the fundamental gap between laboratory-based empirical models and practical applications, motivating the development of more adaptable data-driven evaluation frameworks and dynamic feature extraction strategies. Overall, different model types exhibit clear trade-offs between accuracy, complexity, and engineering applicability. Electrochemical models possess clear physical meaning but incur high computational costs, while empirical models offer simplicity but suffer from limited generalization capabilities. ECM provide a compromise between the two, though their robustness under complex dynamic conditions remains to be enhanced. This has spurred increasing research efforts to integrate modeling approaches with data-driven techniques, offering new avenues for the engineering application of SOH estimation [49,50].

3.4. Data Driven Method

In recent years, artificial intelligence technologies such as deep learning, deep reinforcement learning, and Bayesian optimization have demonstrated significant potential in battery state of health estimation and fast-charging optimization [51,52,53,54,55]. Compared with model-driven approaches that depend on detailed electrochemical mechanism modeling, data-driven methods learn degradation patterns directly from large-scale operational data and can achieve effective SOH and aging prediction without requiring highly accurate physical models. This has become a key development direction in battery management system research [56,57,58,59]. The core of data-driven SOH estimation lies in extracting meaningful features from time series data, typically recorded during charge–discharge cycles [60]. With the release of public datasets such as NASA and CALCE, data-driven approaches based on experimental data have been extensively studied, achieving high prediction accuracy under ideal conditions [61,62,63,64,65]. Although these datasets do not fully capture battery behavior under real EV operating conditions, they nevertheless provide a solid and widely used foundation for developing and benchmarking data-driven methods [66]. Common data-driven algorithms include Support Vector Regression (SVR) [67] (Figure 5a), Long Short-Term Memory (LSTM) [68] (Figure 5b), Extreme Learning Machine (ELM) [69] (Figure 5c), Transformer [70] (Figure 5d), and Convolutional Neural Network (CNN) [71] (Figure 5e).

3.4.1. Machine Learning

Within traditional machine learning frameworks, battery SOH is commonly estimated by combining hand-crafted features with regression models. Gaussian process regression (GPR) [72] and SVM are particularly popular due to their effectiveness with small datasets and their ability to capture nonlinear relationships. Accordingly, existing studies mainly concentrate on two aspects: feature engineering and model optimization. For feature engineering, Wang et al. [73] applied wavelet transforms to incremental capacity curves and extracted peak amplitudes and peak locations as aging-sensitive features. By further integrating a conjugate-gradient strategy with a multi-island genetic algorithm, they achieved notable improvements in SOH estimation accuracy. Richardson et al. [74,75] constructed feature vectors using variables such as capacity, time, and temperature, enabling effective modeling of capacity evolution. Hu et al. [76] extracted multi-stage charging features—including charge capacity, voltage slope, and voltage drops at inflection points—and enhanced aging characterization by optimizing GPR hyperparameters. Building on these efforts, subsequent work has emphasized computational efficiency and generalization. Deng et al. [77] extracted statistical descriptors from incremental capacity sequences within segmented voltage windows under partial-charge conditions and developed a sparse GPR model, which reduced computational complexity while alleviating overfitting commonly observed in standard GPR. Chen et al. [78] employed particle swarm optimization to tune LSSVM model parameters, achieving high-precision capacity estimation. Similarly, Tong et al. [79] optimized SVR model parameters through Variational Modal Decomposition (VMD) and the Dung Beetle Optimization (DBO) algorithm, achieving RMSE of 0.84% and MAE of 0.71% for SOH estimation on experimental data. He et al. [80] proposed an improved GPR framework integrating feature selection and intelligent optimization. By reducing feature redundancy and adaptively optimizing kernel hyperparameters, they significantly enhanced SOH estimation accuracy and model generalization under small-sample conditions. Both the mentioned capacity increment features and the frequency-domain features extracted via VMD heavily rely on complete and standardized charging segments. When real-vehicle data suffers from missing values, noise, or inconsistent operating conditions, these high-quality features become difficult to extract, leading to reduced model accuracy.

3.4.2. Deep Learning

With the rapid development of deep learning, researchers have increasingly introduced CNNs, recurrent neural networks (RNNs), and their variants into SOH estimation to address the limitations of traditional machine learning methods in feature representation and temporal dependency modeling. Unlike conventional approaches that rely on hand-crafted features, deep models can learn hierarchical representations directly from raw measurements, offering clear advantages in capturing complex nonlinear degradation patterns.

Early research primarily focused on constructing SOH estimation models based on CNNs. Chemali et al. [81] proposed a direct SOH estimation framework using a convolutional neural network, utilizing voltage, current, and temperature sequences during the charging process as inputs to achieve rapid SOH estimation. This approach avoids the training complexity associated with recurrent networks, making it more suitable for in-vehicle deployment due to its structural simplicity and inference efficiency. However, relying solely on convolutional structures remains insufficient for fully capturing the long-term temporal dependencies across cycles during battery aging, often neglecting the cumulative effects of historical states when processing long sequence data. To enhance the model’s ability to model temporal correlations, researchers began integrating convolutional neural networks with recurrent neural networks, forming hybrid deep learning architectures. Tian et al. [82] proposed an SOH estimation method integrating CNNs, bidirectional long short-term memory (BiLSTM) networks, and attention mechanisms. The CNN extracts local features, the BiLSTM captures long-term temporal dependencies, and the attention mechanism assigns higher weights to critical time segments, significantly enhancing the model’s ability to characterize complex degradation processes. Xu et al. [83] further incorporated feature importance analysis and added skip connections within the CNN–LSTM structure to mitigate gradient vanishing during deep network training, effectively enhancing model stability and generalization performance.

Building upon this foundation, subsequent studies integrated deep learning with physical information or auxiliary modeling processes. Chen et al. [84] proposed an SOH estimation framework integrating temperature prediction with GRU networks. It first employs a support vector machine to predict complete temperature curves, then derives stable health features through differential processing and Kalman filtering, finally feeding these into a GRU network for SOH estimation. This approach mitigates incomplete sensor data issues to some extent, enhancing the model’s applicability under real-world operating conditions. Ansari et al. [85] introduced the jellyfish optimization algorithm for parameter tuning of recurrent neural networks, validating its effectiveness on multi-cell samples using the MIT–Stanford battery dataset.

Although the aforementioned methods demonstrate outstanding performance on publicly available experimental datasets, it should be noted that most of these studies are built upon highly idealized data assumptions. Their training and validation processes typically rely on complete charge–discharge data collected under laboratory conditions, featuring stable sampling frequencies, continuous time series, and relatively consistent operating condition distributions. Figure 6 illustrates the SOH degradation curve and corresponding data processing techniques. However, significant discrepancies exist between real-world vehicle operating environments and laboratory conditions. These differences are particularly pronounced at the data level, feature level, and application scenario level. While deep learning methods demonstrate exceptional modeling capabilities in laboratory data environments, their applicability in real-vehicle scenarios faces significant challenges. This practical constraint has shifted research focus from model architecture design toward enhancing the adaptability of the data-driven paradigm itself—specifically, achieving high-precision SOH estimation under suboptimal, weakly labeled, or even unlabeled real-world vehicle data conditions.

4. Transition from Laboratory to Real-World Vehicles: Data, Characteristics, and Environment

4.1. Data Layer Differences and Data Governance

In practical deployments, real-world battery data are mainly collected through onboard BMS logs and, when available, external measurement devices. A typical collection workflow includes sensor acquisitiom, which is the real-time monitoring of operating states using voltage, current, temperature, and SOC sensors. For some vehicle models, basic signals can be exported through the OBD interface [86,87]. Real-world EV battery data collected by the BMS and external sensing systems often suffer from quality issues, including measurement noise, recording gaps, and relatively low sampling frequencies compared with laboratory tests, which poses significant challenges to accurate SOH estimation [88].

Charging characteristics extracted under laboratory conditions are typically derived from well-defined constant-current phases and therefore exhibit strong regularity. In contrast, real-world operating conditions vary substantially across drivers and are further constrained by charging station protocols, voltage limits, and temperature effects. As a result, the commonly used multi-stage constant-current charging with current switching may obscure or distort peak-related characteristics, and non-stationary aging patterns further hinder the direct transfer of laboratory-derived features to field SOH estimation [89]. In addition, in practical applications, cells within the same battery pack often experience different charge/discharge depths. Together with manufacturing tolerances and heterogeneous usage histories, this leads to pronounced cell-to-cell SOH inconsistency, which undermines the robustness of pack-level estimation algorithms. Standard laboratory tests typically assume cell uniformity, whereas this assumption rarely holds in real-vehicle environments [90]. How to adapt laboratory methods to practical scenarios remains an urgent issue, as illustrated in Figure 7.

In real-world vehicle scenarios, due to the absence of complete charge–discharge cycles, the ampere-hour integration method is commonly employed, as shown in Equation (2):

$$SOH={\displaystyle \frac{{\int }_{t_0} I_{bat}\left(t\right)dt}{Q_{new}\left[SOC\right(t_1)-SOC(t_0\left)\right]}}\times 100\mathrm{\%}$$

(2)

${SOC(t}_1)$ and ${SOC(t}_0)$ represent the SOC before and after battery charging, respectively. $Q_{new}$ denotes the initial capacity, while I indicates the current sequence. The key technology for capacity testing involves time integration of the current during charging or discharging to calculate the energy delivered by the battery, known as ampere-hour integration [91]. However, this method relies on high-precision SOC [92]. Traditional capacity estimation methods based on ampere-hour integration struggle to achieve stable application. Ampere-hour integration is highly sensitive to current measurement accuracy and initial state, with errors accumulating over time, making it unsuitable for long-term SOH estimation. In contrast, Equivalent Full Cycle (EFC) provides a normalized representation of the battery’s cumulative charge–discharge energy [93], as shown in Equation (3). EFC equates irregular, fragmented operational data to equivalent full cycles, effectively mitigating the impact of variations in depth of charge and operating conditions. Compared to direct capacity estimation, EFC better reflects the actual aging load experienced by the battery and has been widely adopted in real-vehicle lifetime modeling and degradation analysis.

$$EFC={\displaystyle \frac{\int \left|I\left(t\right)\right|dt}{2\times Q_{nominal}}}$$

(3)

where $I\left(t\right)$ represents the current sequence of the battery during the charge–discharge process as a function of time, and $Q_{\mathrm{nominal}}$ denotes the rated capacity of the battery.

The coefficient 2 in the denominator indicates that a complete charge–discharge cycle consists of one charging process and one discharging process. Therefore, by integrating the absolute value of the current over time and normalizing it with respect to the rated capacity, any irregular charge–discharge process can be equivalently converted into a standard full cycle count. Based on the ampere-hour integration method, the accumulated charge capacity increment within a time segment, $\Delta \mathrm{Q}$, can be calculated.

Combined with the change $\Delta SOC$ obtained from the OCV curve, which reflects the variation in the current estimated capacity of the battery $Q_{\mathrm{est}}$, as shown in Equation (4), the following formula is used to determine the SOH label required for model training, denoted as $SO{H}_{\mathrm{Label}}$, and shown in Equation (5):

$$Q_{est}={\displaystyle \frac{\Delta Q}{SO{C}_{end}-SO{C}_{start}}}$$

(4)

$$SO{H}_{\mathrm{L}\mathrm{a}\mathrm{b}\mathrm{e}\mathrm{l}}={\displaystyle \frac{Q_{\mathrm{e}\mathrm{s}\mathrm{t}}}{Q_{\mathrm{n}\mathrm{o}\mathrm{m}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{l}}}}\times 100\mathrm{\%}$$

(5)

Through this approach, fragmented real-vehicle data can be transformed into time series samples annotated with SOH. However, such a reconstruction-based labeling approach is inherently a form of weak supervision, and its accuracy is highly dependent on the precision of $\Delta SOC$ estimation as well as the effectiveness of temperature calibration for the OCV curve. Consequently, mitigating the influence of label noise on data-driven SOH models has emerged as a critical issue when utilizing real-world operational data. Based on EFC, a reproducible implementation approach is proposed: utilizing EFC and capacity data concatenation as supervisory signals, ensuring fragment comparability through threshold-based data governance, and establishing VTDS as a standardized evaluation protocol, as illustrated in Figure 8.

4.2. Feature Extraction from Actual Vehicles

In existing SOH studies, many feature construction methods rely on complete and continuous charge–discharge cycles, particularly ICA and DVA [94]. These approaches have become the primary methods for characterizing aging mechanism features in laboratory-based datasets [95,96,97]. However, the effectiveness of these methods critically depends on continuous and sufficiently complete constant-current charging segments. In real-vehicle scenarios, user behavior and fragmented charging strategies cause the start and end SOC of each charging event to vary substantially and dynamically [98]. As a result, the required voltage window is often unavailable, making ICA/DVA feature extraction unstable or even infeasible. To mitigate noise interference and local fluctuations in real-world data, She et al. [99] and Tao et al. [100] proposed IC curve enhancement techniques based on Gaussian window smoothing and the S-transform, respectively. Although these methods can effectively suppress noise amplification in differential calculations, notable limitations remain. Both implicitly assume that key IC peaks are present within the observed interval. In response to the reality of real-world EV operations, recent studies have increasingly shifted their focus toward SOH estimation utilizing dynamic profiles and short, fragmented charging segments. As Zhou et al. [101] pointed out, under highly stochastic and fragmented charging behaviors, critical IC peaks may become fundamentally unobservable due to data truncation, which renders peak-based localization ineffective. Consequently, the research focus has gradually shifted from peak detection to feature utilization under partial and fragmented observations. Zhang et al. [102] refined this approach further by abandoning reliance on complete IC curves. Instead, they employed a multi-interval feature fusion strategy to extract voltage-to-capacitance change rates from local voltage intervals, significantly enhancing algorithm robustness under partial charging conditions. However, both IC peak restoration and multi-interval capacity change rate extraction inherently involve differential operations on voltage or capacity, maintaining high computational demands on data quality. To circumvent the instability of differential operations entirely, Chen et al. [103] proposed a simpler health indicator based on the voltage rise time within constant-current charging segments. The method identifies the voltage interval most correlated with SOH and directly uses the corresponding voltage rise time as the health factor. By leveraging the inherent integration effect of the time domain, this feature suppresses measurement noise, removes the need for complete cycles and differential calculations, and shows strong potential for online computation and engineering deployment.

4.3. Real-Vehicle Data-Driven Approach

To facilitate the deployment of data-driven methods in real-vehicle settings, Lu et al. [104] released a large-scale dataset comprising 464 EVs and more than 1.2 million charging segments. To address the absence of SOH labels in field data, they proposed an unsupervised label-generation strategy. Unlike laboratory tests based on complete charge–discharge cycles, their method clusters fragmented charging segments using K-means and reconstructs complete charging curves by stitching segments with similar operating conditions. This enables automated capacity and SOH annotation for large-scale real-world data without manual intervention. Tao et al. [100] proposed an SOH estimation framework that combines incremental capacity ICA with the S-transform. By exploiting the high time–frequency resolution of the S-transform, the method enhances IC curve features during charging and mitigates the sensitivity of conventional ICA to noise and operating-condition variations. They further introduced cell-inconsistency indicators and built a GRU–LightGBM fusion model. Validated on long-term operational data from 37 EVs, the framework achieved strong generalization across materials and operating conditions, with SOH errors below 2%.

Traditional SOH estimation typically assumes the availability of relatively complete steady-state charging curves. However, in real-vehicle multi-stage fast-charging, pronounced polarization effects are common. Liu et al. [105] proposed a rapid SOH estimation method based on polarization features using battery pack data from 200 hybrid vehicles. Their study suggests that instantaneous current switching in multi-stage constant-current charging embeds critical polarization information associated with resistance-related aging. By extracting cross-stage polarization features at switching instants and combining them with a random forest regressor, the method achieves a MAPE of 1.83% without requiring complete charging curves, offering a practical pathway for fast health screening on large-scale fleet data.

To exploit abundant unlabeled BMS data while alleviating label scarcity in field applications, Hadzalic et al. [106] proposed a semi-supervised SOH estimation framework for real-world deployment. By incorporating semi-supervised learning to learn distributional representations from large volumes of unlabeled in-vehicle data, the approach reduces reliance on laboratory-only supervision and improves generalization under field operating conditions. Experimental results indicate that it can effectively mitigate the domain shift between laboratory and real-vehicle data, enhancing estimation accuracy without costly real-vehicle calibration. For new battery development cycles where full lifecycle data are unavailable, Lu et al. [107] proposed a deep transfer learning framework that avoids additional aging experiments. Using domain-adversarial neural networks to align features between source and target domains, the method enables accurate SOH estimation across different material systems without requiring target-domain aging labels, thereby reducing dependence on costly labeled data and supporting rapid BMS development.

To compensate for the limited information content of single features under fragmented observations, Chen et al. [108] proposed a multimodal fusion-based SOH estimation framework. This study innovatively combines local voltage curves with histogram data, utilizing dual-stream CNNs to extract dynamic response and steady-state distribution information from batteries, respectively. Experimental results demonstrate that this multimodal fusion strategy effectively compensates for the shortcomings of single features when confronted with data noise or missing samples. Combined with transfer learning, it further enhances the model’s generalization accuracy in complex real-vehicle environments. Liu et al. [88] studied 300 different EVs, covering three years of operational data. They systematically compared these real-world operational data with laboratory test data, revealing challenges in real-world scenarios such as data incompleteness, strong noise, and inconsistent labels. To address the insufficient robustness of single features under complex real-world conditions, this study constructed a deep fusion network that effectively integrates temporal information (voltage, current, temperature) with statistical features. More importantly, this work validated the generalization advantages of multimodal fusion strategies across vehicle models and operating conditions using large-scale open-source data. It provides crucial theoretical and practical references for breaking industry data barriers and establishing unified SOH algorithm evaluation benchmarks.

Table 1. Comparison of data-driven methods using actual vehicle data.

Data/Scale	Model	Estimation Error	Ref.
464 vehicles, over 1.2 million segments	GCNN FT-GCNN	RMSE = 1.528 ± 0.010 MAPE = 3.011 ± 0.006 RMSE = 11.503 ± 0.009 MAPE = 2.970 ± 0.006	[104]
Real-vehicle footage data	KNN MLP	RMSE = 1.18 MAPE = 1.12 RMSE = 1.40 MAPE = 1.25	[100]
200 hybrid vehicles	RFR GPR	RMSE = 1.83 MAPE = 1.39 RMSE = 1.95 MAPE = 1.48	[105]
McMaster and Stanford	CNN Multimodal fusion	RMSE = 2.29 MAPE = 1.71 RMSE = 1.36 MAPE = 1.04	[108]
20 EVs and 300 EVs	CBAG	RMSE = 0.278 MAPE = 0.279	[109]
Three-year data for 10 EVs	FL-ANN	RMSE = 0.62437 MAPE = 0.70737	[110]

presents a comparison of actual vehicle data-driven methods. Traditional supervised methods provide interpretable industrial baselines, while weakly/semi-supervised approaches demonstrate stronger generalization and scalability. Future research should combine both approaches, retaining the reliable anchors of traditional methods while leveraging the cross-domain and scaling advantages of weakly supervised frameworks. Although weakly and semi-supervised methods show significant advantages in academic research, they have yet to achieve large-scale industrial application. Three primary reasons exist: First, computational constraints. Most self-supervised pre-training and domain adaptations rely on complex deep neural networks that run on GPUs but struggle to meet real-time and power consumption requirements in vehicle-mounted BMS environments. Second, data sharing and privacy concerns. While federated learning enables cross-fleet collaboration, differences in data standards, upload frequencies, and privacy protection mechanisms across enterprises limit large-scale deployment. Third, the inconsistent quality of weak labels and pseudo labels. The accuracy of weak label stitching and pseudo label generation varies significantly across different vehicle models, climates, and charging strategies, increasing deployment risks. Consequently, these methods currently remain largely confined to prototype validation and small-scale demonstration phases. Achieving industrialization requires overcoming both engineering and institutional barriers.

5. VTDS Framework

In recent years, an evaluation approach oriented toward real-world vehicle applications has gradually emerged, systematically characterizing model generalization capabilities across vehicle, time, and operating condition dimensions. As illustrated in Figure 9, this paper summarizes it as the VTDS (Vehicle Time Domain Stratified) evaluation framework.

(1)

Vehicles-Out: Cross-Vehicle Generalization Verification

Most existing studies validate algorithms on a single vehicle model or an experimental platform, for example, using standardized datasets such as NASA or CALCE, or conducting modeling on a specific battery pack under controlled test cycles. Such evaluations, however, often overlook distributional discrepancies caused by differences in vehicle platforms, BMS implementations, and sensor accuracy. Vehicles-Out enforces validation across distinct vehicles, i.e., training and testing data are drawn from different vehicles. By holding out an entire vehicle during training and testing on unseen vehicles, this protocol directly assesses cross-vehicle transferability and reduces the risk of overfitting to a particular vehicle or a limited laboratory dataset.

(2)

Time-Rolling: Time-Rolling Evaluation

Battery aging is a dynamic process affected by ambient temperature, driving habits, and seasonal variations, which can induce drift in both capacity-fade rates and internal-resistance evolution. Therefore, static SOH comparisons at a few time points are insufficient to reflect long-term stability. Time-Rolling evaluates performance through rolling (or sliding) time windows, continuously tracking prediction errors over the lifecycle. This design exposes robustness to capacity-recovery phenomena, long-term trend shifts, and cross-season operation, thereby quantifying a model’s temporal generalization under non-stationary aging.

(3)

Domain-Stratified: Stratified Domain Validation

Substantial distribution differences across climate zones, operating conditions, and battery chemistries can lead to significant performance degradation when SOH algorithms are transferred across domains. Domain-Stratified addresses this issue by stratifying data into interpretable subdomains (e.g., climate/temperature domains, operating-condition domains, chemistry domains) and performing training and validation within each stratum. This protocol enables a more transparent assessment of robustness under heterogeneous conditions, such as variations in temperature, vehicle platforms, charging strategies, and environmental regimes.

In summary, the VTDS framework systematically characterizes the applicability of SOH models in real-world scenarios across vehicle, temporal, and operating condition dimensions, offering a novel approach to transcend traditional laboratory evaluation paradigms. Compared to assessments focused solely on prediction accuracy, this framework aligns more closely with engineering practice and provides critical support for advancing data-driven methods from “algorithm validation” to “engineering implementation.”

6. Challenges

Estimating the health status of batteries in actual vehicles still faces numerous challenges today, with the core difficulties lying in data collection and enhancing model adaptability. Although data-driven approaches have made significant progress, obstacles remain in gathering reliable, high-quality data due to various external factors. The computational burden of the models themselves, coupled with their limited flexibility in handling unexpected conditions, hinders the practical application of these technologies. As shown in Figure 10, overcoming these hurdles will lead to more refined BMS in the future. Compared to laboratory standard cycle data, real-vehicle data authentically reflects battery degradation patterns under complex operating conditions, random loads, and multi-environment coupling scenarios. Constant-current discharge cannot represent actual operational discharge states [111]. The non-stationary nature of real-vehicle data—such as random variations in mileage and ambient temperature—significantly reduces the applicability of laboratory models [112]. Future research is shifting focus toward leveraging real-time big data collected by onboard BMS, while employing digital twin or synthetic data technologies to fill data gaps [113]. Scaling SOH estimation from single cells to the pack level presents a critical challenge for real-world EV applications. While existing models are primarily validated on individual cells under controlled conditions, practical EV battery packs comprise thousands of interconnected cells. Manufacturing tolerances, thermal gradients, and diverse usage histories inevitably cause heterogeneous cell-to-cell aging. Consequently, overall pack performance is typically dictated by the most degraded cell—a phenomenon known as the “weakest-cell limitation.” Directly applying cell-level models to pack systems therefore introduces significant estimation errors. Future research must prioritize pack-oriented SOH frameworks that account for topology, cell balancing, and thermal–electrical coupling, potentially leveraging cloud-based monitoring and digital twin architectures for large-scale anomaly detection.

Furthermore, a substantial challenge arises from the strong coupling between SOC and SOH estimation. Capacity-based SOH methods rely heavily on precise SOC variation $\Delta SOC$ calculations. However, conventional BMSs typically estimate SOC using an OCV–SOC relationship calibrated from fresh cells. As batteries age, degradation mechanisms like LLI and LAM alter the electrodes’ thermodynamic equilibrium, causing noticeable shifts in the OCV–SOC curve. Applying fresh-cell OCV curves to aged batteries introduces systematic SOC errors, which directly corrupt the $\Delta SOC$ calculations and the resulting SOH estimates. In turn, inaccurate SOH estimates further degrade SOC accuracy, as SOC algorithms depend on health-variable parameters like available capacity and internal resistance. This creates a reinforcing error propagation loop that severely compromises estimation reliability. Achieving accurate SOC–SOH decoupling remains a formidable hurdle under the highly dynamic operating conditions of EV systems.

7. Conclusions

This paper provides a systematic review of recent research on lithium-ion battery state of health estimation based on real-world EVs operational data. It summarizes and analyzes existing research findings across multiple dimensions, including methodological frameworks, data characteristics, engineering challenges, and development trends. The paper systematically reviews the application characteristics of experimental measurement methods, model-driven approaches, and data-driven methods in SOH assessment. It compares the advantages and limitations of different methods in terms of interpretability, accuracy, generalization capability, and engineering applicability, while further identifying key challenges in SOH estimation under real-world conditions. Compared to laboratory-based studies, real-world EVs data exhibits significant differences in sampling continuity, noise levels, operational diversity, and label availability. This paper highlights that data quality has become one of the core factors constraining SOH model performance improvement. Particularly in real-vehicle scenarios, data preprocessing, label construction, and feature extraction are no longer simple preliminary steps but critical components directly determining model effectiveness. Extracting stable, physically meaningful features from fragmented, non-stationary data represents an urgent research priority.

Methodologically, this paper systematically reviews progress in applying traditional machine learning and deep learning to SOH estimation. It argues that model performance improvement does not solely depend on increasing network complexity but critically relies on data quality, feature rationality, and the scientific rigor of evaluation frameworks. Different models exhibit distinct advantages across various application scenarios. Model selection should be comprehensively evaluated based on specific operating conditions, data scale, and engineering constraints, rather than solely pursuing theoretical accuracy. Furthermore, this paper emphasizes the necessity of expanding SOH assessment from a “laboratory problem” to a “real-world engineering problem.” With the rapid growth of EVs ownership, SOH assessment is evolving from a single-state metric into a critical foundation for comprehensive battery lifecycle management, encompassing operational safety evaluation, remaining life prediction, and end-of-life utilization decisions. In this process, establishing unified data standards, refining evaluation frameworks, and enhancing model generalization across vehicles and operating conditions are key drivers for technological implementation. Regarding future research directions, this paper identifies several areas warranting priority attention:

(1)

Establishing open, standardized real-vehicle datasets to facilitate fair algorithm evaluation and cross-comparison;

(2)

Developing weakly supervised, semi-supervised, and self-supervised learning methods to mitigate label scarcity in real-world scenarios;

(3)

Promoting the integration of data-driven models with physically based models to enhance interpretability and stability.

Overall, with the continuous expansion of real-vehicle data and the ongoing evolution of data-driven technologies, SOH estimation is transitioning from an experiment-centric research paradigm toward a new phase oriented toward engineering applications. This paper systematically summarizes the research trajectory and development trends in this field from a real-data perspective, aiming to provide researchers with a clear technical framework and research reference, while offering valuable theoretical support for the engineering application and large-scale deployment of lithium-ion battery health management technologies.