From First Life to Second Life: Advances and Research Gaps in Prognosis Techniques for Lithium-Ion Batteries

Abstract

The growing use of lithium-ion batteries (LIBs) in electric vehicles has accelerated the need for efficient strategies to extend their lifespan through second-life applications, where retired batteries are repurposed for stationary storage and other less demanding roles. This paper reviews the most pertinent degradation mechanisms underlying battery aging and the most frequently occurring faults during battery operation. After establishing the correlation between degradation and fault occurrence, reliable state-of-health (SOH) and remaining useful life (RUL) predictions are identified as central to ensuring safety, reliability, and cost-effectiveness in repurposed systems. Next, we present a systematic review of the recently published studies on battery prognosis, with methods categorized into three groups: (i) physics-informed and hybrid models; (ii) purely data-driven approaches; and (iii) transfer learning and features extraction methods. A comparative analysis highlights the strengths and limitations of each group and identifies the most promising approaches for battery repurposing. Modeling heterogeneous second-life packs remains particularly challenging, as cells often enter repurposing with different usage histories and only partial BMS records. In this context, transfer learning and domain adaptation emerge as the most promising directions. In parallel, Generative Adversarial Networks (GANs) can help in addressing the challenge of data scarcity, particularly when integrated into hybrid frameworks for second-life applications. At the same time, systematic exploration of health indicators—including the possibility of stage-specific ones—remains essential. Finally, reinforcement learning offers a complementary yet still underexplored path, enabling real-time adaptation in dynamic scenarios, as batteries enter nonlinear regimes beyond the knee point.

1. Introduction

In response to climate change, the rising cost of traditional energy sources, and global policies aimed at reducing dependency on imported fossil fuels, both the integration of renewable energy and the adoption of electric vehicles have seen a significant increase [1,2]. As a result of this shift, lithium-ion batteries (LIBs) have become highly important. These batteries, renowned for their high power density, low self-discharge, long cycle life, and extended calendar life have emerged as a key enabler of both electrical mobility and renewable energy integration. The high power density is essential for electric vehicles, as it enables rapid acceleration and efficient energy use. The long cycle life and extended calendar life make them suitable for both mobile and stationary storage, ensuring reliability over long operation periods. Moreover, the low self-discharge rate allows LIBs to store energy with minimal losses, which is crucial for backup and intermittent supply scenarios. These characteristics collectively make lithium-ion batteries an important solution for optimizing grid performance and addressing the intermittency challenges associated with renewable energy sources such as solar and wind. Despite these many advantages, lithium-ion batteries (LIBs) have two main drawbacks. First, their performance must be carefully supervised to prevent over-charging or deep discharging, both of which can lead to safety problems, including thermal runaway or even explosions. Second, their gradual degradation over time causes a decay in their performance until it no longer meets the specifications required for the application. When this occurs, LIBs are typically retired. After retirement, they are either recycled or repurposed for second-life applications that are less demanding than their primary ones. Unfortunately, the recycling infrastructure for LIBs in Europe remains underdeveloped, making second-life applications an increasingly attractive alternative.

To address these challenges, various diagnosis [3,4] and prognosis techniques [5,6,7,8,9,10] have been proposed. These techniques are essential for maintaining safe and reliable performance during the first life of LIBs and for supporting their repurposing in second-life applications. On the one hand, diagnosis models aim at detecting and identifying faults and anomalies such as short circuits or abnormal temperature rises occurring in the meantime. On the other hand, prognosis models go a step further by predicting the future evolution of the battery state of health (SOH) and estimating its remaining useful life (RUL). This is particularly important for predictive maintenance and optimizing the control of the battery.

Hence, prognosis plays a dual role. First, it allows safe operation, prevents catastrophic failures, and optimizes operational strategies such as fast charging or load management during the first life of the battery. Second, for second-life applications, it provides a quantitative assessment of whether retired batteries can be safely and economically reused, and for how long, thereby supporting business models and sustainability goals. Nevertheless, it is important to note that this does not eliminate the need to identify micro-health parameters that reflect the actual chemical changes occurring inside the battery prior to second-life use. In this context, emerging techniques that approximate the pseudo-two-dimensional (P2D) model using Padé approximation have been proposed in [11]. This method aims to balance between the computational time required for parameter identification and the accuracy of the results. More broadly, three main prognosis categories have been extensively studied in the literature: model-based techniques that rely on electrochemical models, data-driven models that use historical datasets, and hybrid techniques that combine both [12,13]. However, as these techniques have matured, research has gradually shifted away from purely model-based methods and has increasingly focused on data-driven and learning-based approaches, given their scalability and suitability for handling complex and heterogeneous battery systems, a trend that we will examine in detail throughout this review.

In this paper, the objective is to review recently published prognostic techniques for lithium-ion batteries (LIBs), with an emphasis on recent developments and emerging applications. In other words, we aim to outline the future trend of prognosis techniques for LIBs based on the new emerging challenges and applications in the domain. This paper focuses particularly on the growing use of second-life battery applications and their impact on the development of prognostic methods. The remainder of this article will be structured as follows: in Section 2, the most pertinent degradation mechanisms of lithium-ion batteries are presented, as well as the correlation between these mechanisms and the most frequently occurring faults. In Section 3, second-life batteries, their challenges, and the importance of prognostic techniques to ensure their safety and economic viability are briefly introduced. Section 4 is dedicated to the presentation and analysis of different types of prognosis techniques reported in the literature, highlighting current trends in the field. This analysis will allow us to understand the state of the art and offer recommendations for future work in light of second-life applications. Section 5 compares the most promising prognosis methods from all categories and introduces reinforcement learning (RL) approaches that use degradation predictions to improve battery management and safety. Finally, the conclusion and future recommendations are presented in Section 6.

2. Degradation Mechanisms of Lithium-Ion Batteries

The degradation of a LIB is a gradual decline in its capacity and performance over time due to chemical and physical changes. This degradation is primarily due to three factors. The first is the loss of lithium inventory in the battery caused by the formation of the solid electrolyte interface (SEI), dendrite, lithium plating, and electrolyte decomposition. These phenomena are the results of unintended internal side reactions. The second is the loss of active materials (LAMs) in the electrodes, and the third is conductivity loss (CL), which is represented by the increase in the internal resistance of the battery caused by the consumption and composition of electrolytes [14,15]. These degradation factors shown in Figure 1 are key contributors to the increase in the chance of faults occurring in the battery operation. Recent experimental studies [16] have identified the most common faults, such as internal and external short circuits, voltage deviations, and thermal runaway. These faults will be discussed in the following section, along with their relationships to the degradation mechanisms that either trigger them directly or increase their probability.

2.1. Internal Short Circuit

An internal short circuit is a high current that occurs when a negative electrode (the anode) and positive electrode (the cathode) come into direct or partial contact inside the battery without going through the external circuit [17]. This phenomenon can result from various causes that are directly linked to the degradation of the battery. As illustrated in Figure 2, several degradation mechanisms at both the anode and the cathode can contribute to creating such failure conditions [18]. Recent experiments have confirmed these degradation-induced origins of internal short circuits. In particular, using in situ cross-sectional imaging, the penetration of lithium dendrites through the separator and the subsequent formation of metallic bridges leading to internal short circuits were observed experimentally [19]. Moreover, it is experimentally demonstrated in [20] that electrodes and electrolyte degradation, including SEI breakdown, particle cracking, and separator collapse, collectively accelerate the onset and severity of internal short circuits. These findings prove that microscale degradation is not only a precursor to capacity fade but also a direct initiator of catastrophic electrical failures.

2.2. External Short Circuit

An external short circuit occurs when the current flowing from the battery bypasses the load and goes directly from the anode to the cathode during discharge. This fault generally occurs due to mechanical accidents or damage. The immediate consequence in such cases is a significant generation of heat inside the battery due to the high current. This heat raises the temperature of the battery, which can lead to an explosion of the battery or at least disrupt its operation [21]. Based on the above, it can be deduced that this type of fault is more expected in electric vehicle applications rather than in stationary ones [22].

2.3. Voltage Deviations

Voltage deviations refer to the cases when the terminal voltage of the battery is either higher or lower than its nominal voltage. This is a common result of over-charging and over-discharging of the battery. Another factor that can increase the likelihood of voltage deviations is cell-to-cell or module-to-module imbalances in the battery [23]. A direct consequence of the divergence of the battery voltage from its nominal values is an increased risk of explosions or fire eruptions as well as power output reduction. In addition, voltage deviations create conditions that make the occurrence of an electric shock for persons near the battery more probable [24]. Finally, it is important to note that over-charging and over-discharging have lasting consequences, with over-charging degrading the internal capacity of the battery and over-discharging causing electrode degradation and irreversible capacity loss [25].

2.4. Thermal Runaway

Thermal runaway is a dangerous fault that releases an excessive amount of heat that can cause fire or explosions [18]. All the aforementioned faults in this section can cause thermal runaway. However, there are also others such as manufacturing defects, high ambient temperature, and chemical degradation within the battery [26]. Hence, preventing this problem is a significant challenge in the context of batteries, and it is particularly critical for lithium-ion batteries due to their high power density and because the decomposition of their electrolytes, which is necessary for operation, occurs at high temperatures [27]. Given the severe consequences of thermal runaway, proper health management for LIBs is more critical compared to other battery types, such as lead-acid and nickel–cadmium batteries. This includes continuous monitoring of both the state of charge (SOC) and temperature for the early detection and prevention of critical faults in lithium-ion batteries. In this regard, an ultrasonic reflection wave for joint SOC–temperature estimation is proposed in [28]. This non-destructive approach is a remarkable low-cost, high-precision solution that has been recently proposed to enhance battery safety.

3. Second-Life Batteries: Opportunities and Challenges

Second-life batteries are a new form of energy storage systems that are expected to witness high integration into smart grids. In 2023, the purchase of electric vehicles has witnessed a 35% year-on-year increase, and by 2035, all new cars sold in the EU should be zero-emission vehicles. These cars use lithium-ion batteries to store energy. These batteries last between 8 and 10 years before becoming inadequate to the EV application. However, after their retirement, they still preserve up to 80% of their initial capacity, which remains acceptable for the smart grid application [29].

Some statistics show that approximately 250,000 tons of lithium-ion batteries will reach their end of life by 2025 [30]. To reduce environmental impacts, these batteries should either be reused or recycled. Consequently, the concept of reusing these batteries, particularly in second-life applications such as battery energy storage systems (BESSs), is gaining traction, especially after the adoption of the Glasgow Climate Pact [31]. In fact, second-life battery modules could drastically lower the cost of energy storage and reduce the environmental footprint by decreasing the demand for new battery production and raw material extraction. However, the economic and technical viability of these applications is based on the ability to effectively manage these batteries, especially considering the fact that the price of new LIBs is decreasing [32]. Unfortunately, accurate modeling of second-life battery packs remains relatively underdeveloped because of the nonlinear behavior of lithium-ion batteries. As shown in Figure 3, battery degradation is approximately linear before the knee point, after which it becomes nonlinear, complicating its prediction. For example, a recent comparison of two models, each evaluated with and without considering the knee point, was presented in [33]. Accounting for the knee point improves the prediction accuracy of the remaining lifecycle, with RMSE improvements of about 3.8% for the first model and 18.3% for the second one.

In second-life battery applications, where batteries operate closer to the knee point, the probability of the occurrence of the faults mentioned in Section 2 becomes higher. This is due to the fact that three of these most frequent faults are correlated to the degradation of the battery. In addition, as the battery enters the nonlinear degradation region, its predictability decreases, particularly when interactions between heterogeneous cells and modules are not adequately considered. Recent studies have also shown that some capacity degradation might be reversible [34], which also furthers the complexity of remaining useful life (RUL) prediction. A comparison of RUL prediction errors before and after the knee point, as reported in [35], shows that prediction accuracy significantly decreases once the battery enters the accelerated degradation region. Hence, the development of accurate prognosis models becomes increasingly critical, while also facing significant design challenges. The next section presents and compares the latest contributions in battery prognosis, highlighting strengths, limitations, and research gaps.

4. Battery Prognosis and Second-Life Applications

A prognosis model predicts the remaining useful life (RUL) or the state of health (SOH) of the battery based on its degradation patterns. An overview of a prognosis model is shown in Figure 4. These models can benefit second-life applications in two different ways. The first is an indirect benefit, as a prognosis model can be operated in the first-life application of the battery and collect historical SOH and RUL data. For instance, collecting these data will help the model in understanding the first-life degradation roots, which in turn helps predict the battery’s aging in its second life. Moreover, prognosis indirectly extends the overall lifecycle of batteries by enabling optimized first-life usage by helping to implement adaptive charging strategies that slow degradation. This ensures that more packs reach the end-of-first-life with a sufficient remaining capacity (often above 70%) to be repurposed, thereby reducing refurbishment costs, enhancing the economic viability of second-life applications, and supporting circular economy objectives by lowering raw material demand and delaying recycling [36]. The second benefit is in the second-life application by avoiding faults occurring and the failure of the battery. In fact, the failure of one module in the second-life pack can accelerate the degradation in other modules within the pack [37]. Hence, by incorporating degradation knowledge, the prognosis model can assist the battery pack’s control system in preventing imbalances between modules, thereby optimizing overall system performance and avoiding such scenarios.

In general, prognosis techniques can be divided into three main families: model-based techniques, data-driven techniques, and hybrid techniques. While model-based techniques were the first to be explored and multiple filtering techniques already exist in the literature [38,39,40,41,42,43], they have been phased out in recent years because their accuracy relies heavily on the availability of an explicit model representing the battery. In [5], the authors present a comprehensive literature review demonstrating that integrating data-driven models into battery prognosis offers a highly promising perspective, whether applied alone or within hybrid models. Hence, in this section, we focus on the most recent published papers in the domain, and the section is organized accordingly. It should be noted that some techniques can belong to more than one category, depending on the specific implementation. For example, a transfer learning approach might incorporate a hybrid model, but it is categorized under the transfer learning section since its main contribution lies in the transfer learning strategy itself. Finally, it is important to note that some studies classify empirical methods as model-based [8]; however, in this paper, we consider them data-driven. Building on this classification, we present three groups of prognosis models to which most recent papers belong: physics-informed and hybrid models, data-driven neural networks and regression models, and transfer learning and feature extraction models.

4.1. Physics-Informed and Hybrid Models

To address the challenge of relying on explicit LIB models that accurately capture battery behavior across all operating conditions, many studies in the literature have focused on physics-informed data-driven approaches. These studies can be grouped into three different categories: physics-informed feature (PIF) methods, physics-guided algorithm structure (PGAS), and physics-informed training constraint techniques (PITC) [44].

For instance, to predict the remaining useful life (RUL) of lithium-ion batteries during fast charging with limited early-cycle data, a physics-informed machine learning (PIML) framework was proposed in [45]. The model couples a physics-informed branch, embedding SEI growth dynamics, with a task-specific data-driven branch based on dilated CNNs. A three-stage training strategy is performed, ensuring that both physical consistency and predictive accuracy are achieved. Then, the trained model is tested on the Stanford–MIT–Toyota dataset; the method achieved superior accuracy using only four initial cycles, outperforming state-of-the-art approaches. Another similar approach is presented in [46], where a hybrid Grey forecasting model that considers the regeneration phenomenon of the capacity of the battery is proposed. This model assumes that the overall capacity degradation is exponential until it reaches 70% of its nominal capacity. Hence, the used model for features is quasi-exponential. And as the regeneration of the capacity is random and the degradation becomes faster after its occurrence, the proposed Grey forecasting model is discrete and non-homogeneous. Once a prediction of the RUL is obtained using the Grey model, an ensemble Kalman filter is used to correct this prediction based on the system’s measurement. This model is validated using four different batteries from the NASA datasets, then compared with several other state-of-the-art prognosis methods.

Nevertheless, these two methods do not consider the fact that data-driven models are highly sensitive to noise in training data, especially in the case of limited data, such as in [46], and in highly parameterized models, such as in [45]. To address this challenge, another physics-informed neural network called PI-TNet is proposed in [47]. This method uses a Local Information Processor (LIP) that leverages an ensemble of convolution neural networks that allows multiscaling feature extraction, enhancing the robustness of the PI-TNet. The model predicts the SOH of the battery based on the measured voltage, current, and temperature during charging and discharging. Nevertheless, these data being time series require a model that has long-term sequence capabilities to analyze them. This is achieved by integrating a transformer-based ViT module into the system. For further enhancement of the generalization of the PI-TNet, a mathematical model [48] is used for loss computation and constraint imposition that mimics actual bounded and nonlinear degradation processes and ensures that the learned patterns respect physics laws. However, it is important to note that these constraints are not directly based on electro-thermal equations. Finally, the optimized loss function is a composite function including three errors as shown in Equation (1), $L_u$, $L_t$, and $L_f$, where $L_u$ and $L_t$ ensure that the PI-TNet predictions match the measurements and respect the governing dynamical equations, while $L_f$ enforces consistency with the system dynamics. The PI-TNet model has been validated fo four different batteries of the NASA dataset [49]. For three of these batteries, the PI-TNet shows an average improvement in the mean absolute error (MAE) of 94.69%, 68.73%, and 84.64%, respectively.

$$L={\lambda }_u{L}_u+{\lambda }_t{L}_t+{\lambda }_f{L}_f.$$

(1)

Another work that addresses the challenge of the limited amount of available data is presented in [50]. However, in this paper, another main characteristic of the behavior of the battery is considered: the inconsistencies among the cells in the battery pack. In this article, the author proposed using a combination of two data-driven models to estimate the end of life (EOL). The first is used to extract two features that are representative of the battery from the discharge curve; then, these features are mapped to the capacity using linear regression. The second is a gradient module that is responsible for accurately capturing the aging trends of the capacity and uses this trend to predict the capacity of the battery. The two predictions are combined to get the final prediction of the ’Tell-me’ model. The weights of this model are tuned based on the historical error between the predicted values and the ground truth. The training is performed offline explicitly using the historical data of the target battery, which consists of voltage and capacity curves. Finally, this model is validated using three different public datasets, where two correspond to lithium batteries and the third corresponds to a different chemistry. The results show that Tell-me achieves higher accuracy compared with existing approaches on these datasets. In addition to its accuracy, the model’s other major advantage is that it enables physical interpretability by directly linking trend mean and seasonality area to known degradation mechanisms—internal resistance growth and electrode phase transitions. Furthermore, because the framework needs only early-cycle discharge curves and no large labeled datasets, it is ideally suited for second-life prognosis. However, since the model is trained on high-quality cycling profiles, its performance may deteriorate when exposed to noisy or imperfect measurement data.

While earlier studies in this section do not consider that cells in a battery pack can follow different degradation trajectories due to internal interactions [51], a method is proposed in [52] to estimate the degradation of multiple cells within a pack. First, the paper presented a method dedicated for single cells. The degradation trends of cells are considered as vectors and the euclidean distance between each cell and a reference cell is calculated for normalization. The reference cell is chosen as the one that has a degradation trajectory with minimal sum distances compared with the others. Second, as the author aimed to transfer the method to pack level, he opted to use a two-stage adaptive differential model decomposition (ADMD) to address the limitation of classical (DMD), which struggles when dealing with amplification or shrinking transfer coefficients. The first divides the degradation trajectories into stabilized degradation trajectories and extracted noise dynamics. After that, a differential operation is applied to further minimize the fluctuations in the degradation dynamics of the cells. Then, the uniform uncertainty is quantified and the degradation model is stabilized using the second decomposition. Next, this model is validated on these individual cells but fails to accurately predict their degradation when operating in groups. In consequence, an online adaptive updating coefficient is proposed, which tries to minimize the minimal euclidean distance between each cell operating in a pack and the predicted degradation for all cells when operating individually. This optimization problem is solved online using the Lagrange method. Afterwards, to address the uncertainty problem, the author uses a classical unscented transform. Finally, the obtained model is validated on the entire lifespan of the packs and the absolute percentage error is calculated (APE) between the average predicted failure time and the actual one. The online precision index (OPI) and the online accuracy index (OAI) are two metrics considered for comparison with other prognosis methods. The obtained results proved better performance on the pack level compared with empirical methods [53], the LSTM network [54], and the basic DMD approach [55]. Another method that lies within the physics-informed framework and that successfully reduced the time required by the prognosis model to predict the capacity and the SOH of LIB is presented in [56]. This work was inspired by previous studies that have proven that simplified equivalent circuit models (ECMs), which are already considered the simplest models used for representing the behavior of the battery, can in fact be used to generate the data required for the training of neural networks [57]. For instance, the technique proposed in [57] used parameters that are generated by a simplified ECM and that are highly correlated with the growth of the SEI and achieve an accurate prediction of the capacity of LIBs with an average relative error below 4% for three different discharge rates (1C, 2C, and 3C). This work was followed by [58], who proposed the usage of an improved ECM with an added capacitor with its parameters identified using the EIS technique. Then, these parameters were used to train a Gaussian process regression (GPR) model. This model was able to predict the SOH with an average root mean square error (RMSE) of 1.77%, with the highest one being 2.95%. Then, ref. [59] used further simplification of the ECM, as this article proposed a model consisting only of an internal resistance $R_{in}$, with an approximation of the increase in this resistance assuming its linearity. However, this method was validated on nickel-cobalt batteries (NC) and not LIBs, and it was capable of estimating—not predicting—the SOH of the battery with an error ranging between 2 and 5%. Now, as a continuity of these approaches, ref. [56] proposed the use of EIS to identify the parameters of a first-order ECM shown in Figure 5 only in the high- and mid-frequency range, avoiding the time consumption of the low-frequency EIS. Then these parameters were used to train a back-propagation neural network (BPNN), which is a feedforward neural network trained by the back-propagation algorithm. This algorithm was chosen because it requires shorter learning and running periods compared with other used algorithms (random forest and gradient boosting). Finally, this model was validated using four different cells and it proved that it is capable of predicting the SOH and the capacity of the battery, with an average error of 1.4%, while also reducing the training time by 6% compared with full-range EIS.

Table 1. Comparative summary of physics-informed and hybrid prognosis models for lithium-ion batteries.

Approach	Features/Metrics Used	Predicted Quantity + Error	Datasets/Validation	Adequacy for Second-Life Batteries
Physics-informed ML (CNN + SEI dynamics) [45]	SEI growth dynamics + voltage/current cycles	RUL; superior accuracy using only 4 initial cycles (MAE = 8.3–8.4%)	Stanford–MIT–Toyota dataset	Limited: requires explicit SEI modeling, which is complicated in repurposed packs
Hybrid Grey + ensemble Kalman filter [46]	Quasi-exponential degradation + random regeneration; correction via EnKF	RUL; improved prediction vs. benchmarks (MAPE = 0.65–0.93%)	NASA dataset; 4 cells	Limited: sensitive to noise and depends heavily on early-cycle data
PI-TNet [47]	Voltage, current, temperature time series	SOH; MAE improved up to 94.69% vs. baselines	NASA dataset; validated on 4 cells	Limited: constraints require physical knowledge not tracked by BMS
Tell-me model (dual data-driven hybrid) [50]	Discharge-curve features + gradient module (voltage and capacity)	SOH + EOL; higher accuracy than benchmarks (MAPE = 3.27–7.94%)	3 public datasets (LIB + other chemistries)	Promising: relies on discharge features that can be extracted from BMS data, handles inconsistencies
Transfer-driven ADMD for packs [52]	Degradation trajectory vectors + adaptive differential model decomposition	RUL; MPE = 5%	NASA dataset; pack-level validation	Promising: explicitly considers cell heterogeneity in packs, relevant for second-life scenarios
ECM + BPNN hybrid [56]	Mid/high-frequency EIS parameters from 1st-order ECM	SOH + capacity; average error 1.4%	4 cells; reduced training time vs. normal EIS	Limited: requires EIS measurements, rarely feasible in second-life

All reported performance metrics (MAE, MAPE, MPE) are as stated in the referenced studies.

presents a summary of all the physics-informed papers discussed in this section. As shown in this table, most physics-informed papers cannot be used in the second-life application for many reasons. First, some require a physical model for the SEI growth which is hard to obtain in the case of heterogeneous packs [45], while others are very sensitive to data noise [46]. The PI-TNet model proposed in [47] also requires physical constraints that are hard to generalize across packs. However, this is not the case for the Tell-me model [50] that has been proven to be effective in handling inconsistencies among cells, making it promising for second-life applications. A similar conclusion can be drawn for the method proposed in [52], since it was explicitly designed to address groups of heterogeneous cells, closely resembling the conditions of second-life packs. By contrast, the approach presented in [56] requires an ECM model with parameters identified through EIS for each individual cell, which is impractical when large variations exist between cells. One potential improvement to overcome this limitation could be to use the aforementioned (P2D) model. This will allow the prognosis model to capture the internal degradation of the battery without requiring EIS measurements for each individual cell, while being much faster to implement, potentially making it more suitable for heterogeneous second-life packs.

4.2. Data-Driven Neural Networks and Regression Models

For data-driven techniques, we consider two groups of methods: neural networks and regression models. The first group relies on the weighted structure of neural networks that tries to learn patterns based on historical degradation of the battery. The latter group uses predefined functions [60] and tries to fit the data to this function by calculating the coefficients of the function in a way that minimizes the error between the label of the data and the prediction obtained after applying the function. Some already existing studies on the prediction of the capacity degradation use the current and voltage data which can be obtained using procedures that accelerate the degradation of the battery. Hence, other studies opted to use Electrochemical Impedance Spectroscopy (EIS), but this method requires a long time especially at lower frequencies. To bridge this gap, a prognosis method that focuses only on frequencies that are tightly linked to the degradation of the battery is proposed in [61]. The goal is to reduce the time required by the EIS, while preserving the accuracy of the prediction. The salient frequency extraction (SFE) algorithm is used to extract the most pertinent frequencies from the datasets. Then, a back-propagation neural network (BPNN) is used to predict the capacity, as they are good in handling nonlinearity. The weights of the neural network are chosen using the GridSearchCV to ensure optimal parameters. This method is validated with four different cells and is shown to predict the SOH and the capacity of the cells with an error as low as 0.94%. The time of the acquisition of the EIS data was reduced by 93%.

A creative method is proposed in [62], especially for electric vehicles, as it tries to find the correlation between the driving behavior and the degradation of the battery. This method combines two models: a fully connected (FC) layer, and an exponential smoothing transformer (ETSformer) decomposition layer. The ETS layer is responsible for capturing the trend and the seasonality, and the FC layer is a safety layer that helps the model adapt when transferring knowledge from one dataset to another. The combined model is compared with the standalone ETSformer and demonstrates superior performance, even when relying only on minimal discharging profiles. It achieves prediction errors of 2% for aggressive driving and 4% for moderate driving. Furthermore, incorporating driving behavior reduces the MAE by 63.48% and 27.83%, the MSE by 79.51% and 43.58%, and the RMSE by 54.74% and 24.88%.

Another method that is also dedicated to the RUL prediction in electric vehicle applications is proposed in [63]. In this article, a combination of a quantile regression model that is more adequate for uncertain systems compared with classical linear regression and a self-encoder neural network with a temporal convolutional network (TCN) for extracting battery data is proposed. This model is used for the prediction of both the SOH and the RUL of the battery. The main strength of the temporal convolutional network lies in its suitability for time-series data, and when coupled with the regression model, it provides high robustness against noise and variations in charging and discharging protocols. This method was validated on the TRIBD and NASA datasets. The proposed model achieved up to 88% SOH prediction accuracy, improved SOH assessment by 5%, extended battery life by 45 cycles, maintained high-capacity fidelity (91.08%) under varying temperatures with minimal prediction error, and outperformed comparable models on both the TRIBD and NASA-PBD datasets.

In contrast to the two aforementioned techniques, in [64], the author chose to compare different techniques to estimate the RUL of the LIB specifically in the electrical grid, with the aim of maintaining stability and ensuring battery reliability and environmental sustainability. The compared algorithms are as follows: Levenberg–Marquardt (LM), Scaled Conjugate Gradient (SCG), Bayesian regularization (BR), random forest (RF), and Gaussian process (GP). In this comparison, the algorithms are used to train a feedforward neural network (FFNN) with ten hidden layers, whereas an RBF kernel is used for the GP to help build the predictive model, and a 100-tree ensemble is used for the RF technique. The obtained results using the NASA prognosis datasets show that both the random forest (RF) and Scaled Conjugate Gradient (SCG) models achieved superior performance for battery RUL estimation, with RF obtaining the lowest error (MSE = 0.0016 in training, 0.0020 in testing, $R^2$ = 0.9854/0.9807) and SCG yielding the lowest MAE (0.0295) and MAPE (1.098%). These results outperform or match commonly used models like LSTM, RNN, and GPR, demonstrating the strong potential of machine learning for accurate and efficient battery health monitoring. However, it is difficult to draw a general conclusion based on a single comparison between an FFNN with ten hidden layers and an RF model with 100 trees.

Instead of estimating the RUL, the work presented in [65] focused on estimating the SOH while simultaneously predicting the knee point shown in Figure 3. The author pioneered the use of a data-driven method based on time–frequency feature maps using continuous wavelet transformation. The motivation for this technique is twofold. First, most state-of-the-art prognosis and health management (PHM) techniques assume constant current (CC) and constant voltage (CV) charging protocols and use discharging curves that are hard to obtain; hence, this technique considers complex discharging profiles which are closer to real-world scenarios, especially with the growing adoption of fast-charging techniques. Second, feature extraction heavily depends on expert knowledge or the use of 1D CNNs, which leads to information loss in the frequency domain [66]. This limitation increases the need for more automated techniques that consider both time and frequency information. Based on this, the continuous wavelet transform (CWT) is used in this paper to leverage both temporal and frequency features for battery prognostics. The major advantage of (CWT) compared with other frequency transform like the Fourier Transform (FT) is the fact that the CWT captures the frequency features of a degradation and its occurring time while the FT cannot do so. Finally, these features are fed to a neural network that combines CNN, Bi-LSTM, and an attention mechanism, making this model more capable of capturing time dependencies for battery degradation. The data used for the training and validation of this method are produced experimentally using 124 commercial LFP batteries, then 48 cells that use 18 charging protocols are chosen for the training while another 6 with different protocols are chosen for testing for the SOC interval between 30 and 60%, and a current rate between 3.7 and 8C. These data are filtered using a Gaussian filter (GF) to reduce the impact of noises. After that, the SOH is calculated for these data as the ratio of actual to nominal capacity of the battery [67]. After validation, the proposed CNN-BiLSTM-Attention model achieved the best overall accuracy, with an RMSE of 0.74%, outperforming alternative architectures such as CNN-BiLSTM (1.01%), CNN (1.08%), CNN-BiGRU-AM (0.81%), CNN-LSTM-AM (1.12%), CNN-GRU-AM (1.13%), and CNN-BiGRU (1.06%). The integration of attention and bidirectional recurrent structures proved essential, enabling the model to capture nonlinear degradation patterns and maintain errors within 2% before the knee point, with $R^2$ exceeding 90% in most cases. These results are further confirmed by cross-dataset experiments on NASA and CALCE datasets.

Another technique that uses a similar approach for features extraction is presented in [68], where discrete wavelet transform (DWT) is applied to extract health indicators (HIs) from voltage curves in multistage constant current (MCC) charging profiles. The extracted features were proven to be robust against the changes in fast-charging profiles. Then a hybrid neural network that combines both Long Short-Term Memory (LSTM) and a Gaussian process regression model (GPR) is used for SOH and RUL prediction. The first helps in capturing sequential dependencies, and the second is essential to consider uncertainties in the prediction process. Here, it is important to note that while the training time of the GPR model usually increases exponentially with the increase in the number of training batteries, this is not the case for the GPR-LSTM. The proposed framework predicts SOH based on the HIs, and the historically estimated SOH to the RUL. To train this model, 4 cells are randomly selected from a large dataset with one-step and two-step MCC profiles, and then the testing of the model is performed using 88 cells from the same datasets. The obtained results show that the proposed GPR-LSTM outperforms standalone GPR and LSTM models. Three quantification metrics are used to evaluate the performance of the three models, and the GPR-LSTM achieved the lowest error rates in all scenarios for SOH estimation: RMSE = 0.91%, MAE = 0.79% for one-step MCC, and RMSE = 1.02%, MAE = 0.81% for two-step MCC, with a significantly narrower 95% confidence interval (0.523–0.667%) compared to GPR (0.727–0.778%). Then, another comparison is performed with six state-of-the-art techniques for SOH estimation: Bi-ViT-DSA [69], RNN-JFO [70], RBFNN [71], LSTM-TF [72], WGAN [73], and GLSTM [74]. The comparison confirms the superiority of the proposed technique on all other ones with an RMSE = 0.99% and MAE = 0.81% across 88 test cells. Now for RUL prediction, a third comparison is performed between the GPR-LSTM and the LSTM alone. Both models are trained offline on historical SOH and the testing is performed using online estimated SOH. For one-step predictions in the MCC scenario, the GPR-LSTM model has an overall lower RMSE compared to the plain LSTM model. However, for two-step predictions in the same setup, GPR-LSTM’s median RMSE is a bit higher than LSTM’s but still achieves a lower maximum RMSE across all cycles. Unlike LSTM, GPR-LSTM accounts for uncertainties. Hence, overall, the GPR-LSTM outperformed the LSTM model both in estimating the SOH and in predicting the RUL.

In [75], an improved technique is proposed based on a previous study [76] that presents a non-parametric degradation model that solves the problem of sparse or fragmented data for accurate RUL prediction. The model relies on degradation data and employs functional principal component analysis (FPCA) to extract degradation patterns. The improvement proposed in this article is the use of a Pseudo-Huber loss function in the optimization, and an optimization algorithm coupled with a Bayesian update of the distribution of the FPC component. In other terms, this model will estimate the RUL as a distribution composed of an average and a covariance. The proposed FPCA-based method has been validated and the method achieved the lowest median and variance in RUL prediction errors compared with three other models. Compared with Method 1 (two-step axis transform), it provided more accurate and stable predictions by accounting for the randomness of FPC scores and being robust to outliers. Compared with Method 2 (conventional FPCA-based state-dependent model), it avoided deviations caused by reliance on the original time scale, yielding higher accuracy under incomplete data. Finally, relative to Method 3 (parametric Wiener process model), it significantly outperformed the method, as Method 3 suffered from inaccurate functional assumptions and the absence of degradation time, resulting in the worst prediction accuracy. Overall, across different missing data rates, the FPCA-based method consistently delivered the most accurate and stable RUL predictions.

In the previous studies, we have seen that the limited amount of data that can be used to train data-driven methods is one of the main challenges they face. A method based on a recurrent conditional generative adversarial network (RCGAN) is presented in [77] to bypass this obstacle. This method adds synthetic data to the training data to increase the reliability of capacity prediction. This process is performed using an LSTM-based generator and discriminator. Coupled together, these two will try to learn patterns in the temporal and spatial distribution of the data. The first will generate data that abide by the dynamic of the capacity degradation data that are already available. The dynamic of the data is learned using two LSTM layers. The second will differentiate between the generated (voltage, current, and temperature) data and the real ones. The validation of the model is performed using MIT and NASA datasets, and both datasets are augmented using the generated new data. Upon acquiring the augmented datasets, a comparison is performed as follows: a gated recurrent unit (GRU) is trained once on the original real data and once on the augmented datasets, and the same process is performed on an LSTM model. Now, for validation, three batteries were selected from the NASA datasets and another three from the MIT datasets. The results demonstrate that across all examined scenarios, training both the GRU and LSTM models on augmented datasets yields higher accuracy in predicting the capacity of batteries compared to training on the original datasets alone, which validates the validity of the RCGAN to address the data limitation issue. Finally, across both datasets, GRUs always outperform LSTMs.

Another similar approach is presented in [78], but instead of using an LSTM-based generator and discriminator, this time, a CNN-LSTM generator and an Echo State Network (ECN) discriminator enhanced with triple attention (spatial, temporal, and contextual) are used. In addition, instead of generating voltage, current, and temperature data as before, the generator in this case generates the SOH trajectories. The GAN is trained using the Fennec Fox Optimization (FFA) algorithm. A comparative study that is similar to the one presented in the previous paper is performed. The proposed approach is compared with a neural network (NN), an ESN standalone, Bi-LSTM, and classical GAN models. The comparison is performed on four different batteries from the NASA datasets. The proposed GAN model outperforms the other techniques, achieving the lowest MAE (0.0075), MSE (5.96 ×${10}^{-5}$), RMSE (0.0113), and MAPE (1.1).

In

Table 2. Comparative summary of data-driven neural networks and regression models for lithium-ion battery prognosis.

Approach	Features/Metrics Used	Predicted Quantity + Error	Datasets	Adequacy for Second-Life Batteries
BPNN with Salient Frequency Extraction [61]	Impedance (EIS salient frequencies)	SOH $\overline{RMSE}$ = 4.36%, Capacity $\overline{RMSE}$ = 1.16%	4 cells (EIS)	Limited: requires EIS, less feasible for aged packs
ETSformer [62]	Driving behavior and discharge profiles	SOH, $\overline{RMSE}$ = 0.061%	Driving datasets (EV)	Limited: relies on driving behavior (EV)
Quantile Regression + TCN [63]	Time-series current/voltage cycles	SOH (88% accuracy) + RUL (extended by 45 cycles)	TRIBD + NASA datasets	Promising: robust under uncertainty and protocol variation
FFNN (Trained using SCG) [64]	Voltage/current features	RUL; MAE = 0.0295%	NASA dataset	Promising: robust with limited data
FFNN (Trained using RF) [64]	Voltage/current features	RUL; MSE = 0.0020%	NASA dataset	Promising: robust with limited data
CNN-BiLSTM-Attention [65]	CWT time–frequency maps	SOH; RMSE = 0.74%	124 LFP cells + NASA + CALCE	Promising: handles nonlinear degradation and fast charging
GPR-LSTM [68]	DWT features from MCC voltage charging profiles + health indicators	SOH (RMSE = 0.91–1.02%, MAE = 0.79–0.81%); RUL	140 cells (1- and 2-step MCC)	Promising: scalable, uncertainty-aware, fits heterogeneous data
FPCA + Bayesian Updating [75]	Functional principal components of degradation data (discharge capacity)	RUL distribution (median error lowest among 3 baselines)	Multiple sparse datasets	Promising: robust under incomplete or sparse data
RCGAN [77]	Voltage, current, temperature (synthetic cycles via GAN)	Capacity prediction; GRU+GAN RMSE = 0.0077%, MAE = 0.0061%	NASA + MIT datasets	Promising: accurate prediction with limited data
GAN (CNN-LSTM Generator + ESN) [78]	SOH trajectories (synthetic via GAN)	SOH; MAE = 0.0075%, RMSE = 0.0113%, MAPE = 1.1%	NASA dataset (4 cells)	Promising: accurate prediction with limited data

All reported performance metrics (RMSE, MAE, MAPE) are as stated in the referenced studies.

, a summary of all previous methods is presented. Methods relying on Electrochemical Impedance Spectroscopy (EIS) such as the one in [45] demonstrate high accuracy but are hard to be implemented for aged and heterogeneous second-life packs. In addition, the method that uses driving behavior [62] for the prognosis of the battery which is oriented for EV applications is also not adequate for second-life applications. In contrast, approaches that are based on measurable metrics (voltage, current, temperature, or driving profiles) were labeled as promising, since these inputs are routinely collected by battery management systems (BMSs) and thus transferable to second-life contexts [63,64]. Moreover, the methods that use wavelets transform (continuous and discrete) [65,68] can effectively extract features that are relevant to the nonlinear degradation, all while maintaining robustness against noise and heterogeneity of the data. Hence, this approach can inspire techniques for second-life batteries. For the technique in [75], the uncertainty part is highly considered; this can be beneficial in the case of packs where interactions between different heterogeneous cells and their parameters are uncertain. Finally, GAN-based augmentation techniques [77,78] stand out as highly promising, as they directly address the data scarcity challenge that is critical in second-life batteries, while achieving state-of-the-art accuracy. However, ensuring that the generated data is consistent with electrochemical laws governing the system remains a challenging problem. For example, the RCGAN proposed in [77] generates synthetic battery cycles that capture temporal dynamics and realistic degradation by conditioning on smoothed capacity, but it does not explicitly enforce electrochemical constraints, such as the Peukert law, which may limit the physical consistency of the synthetic data. Similarly, the CNN-LSTM generator used in [78] can capture dependencies from historical data to ensure the realism of the generated one, but it is still not directly constrained by battery electrochemical laws. A possible solution for this problem is the use of constrained GANs. Nevertheless, the integration of constraints—especially hard constraints—in GANs and neural networks in general remains a very active research topic and still faces major challenges in terms of generalization, computation load, and training stability [79]. Fortunately, a recently published paper [80] demonstrated that hard constraints can be successfully imposed in neural networks, showing versatility and effectiveness across applications, such as learning with piecewise constraints, training optimization solvers with guaranteed feasibility, and optimizing control policies in safety-critical systems. Another possible approach is to use a Wasserstein GAN that is more stable in the training phase and capable of handling the problem of missing and noisy data. This method has already been proposed for SOC estimation of batteries in [81], where it demonstrated accurate results.

4.3. Transfer Learning and Feature Extraction

In the past, transfer learning approaches relied heavily on parameter-based methods requiring extensive labeled data (5–25% of battery lifecycle) for tuning [82,83,84,85,86], but due to practical constraints in acquiring early capacity measurements, the new trend has shifted toward feature-based and hybrid strategies that operate effectively under severe data scarcity [87,88,89,90,91]. By extracting informative features from raw measurements, models can focus on patterns that are closely related to the degradation processes of LIB, addressing two problems that we have seen in the previous section: the need for extensive labeled datasets and the impact of noise on the PHM module. For transfer learning, it complements this by enabling the reuse of knowledge from related datasets or operating conditions, hence allowing generalization of a prognosis method across batteries with different chemistries or operated under different scenarios. In this section, we will focus on this group of prognosis techniques, especially the ones that have been recently published. While some transfer learning strategies incorporate hybrid modeling and physics-informed networks, this section focuses specifically on techniques where transfer learning or feature extraction is the main driver of prognostic performance.

In this regard, a physics-informed framework is proposed in [92] to predict and detect the knee point of the degradation of LIB. This method has not been listed in the Section 4.1 because its main contributions are the feature extraction method and its transfer ability across different charge/discharge scenarios just by using a small amount of labeled data of the considered scenario. The goal of feature extraction is the reduction in the dimension of the needed data. To do so, a histogram-based method that has been validated [93,94] is used assuming that the current and the voltage for each cell are measured. Simply put, this histogram illustrates the time spent by the battery in a specific range of voltage and current. On top of this, a hybrid physics-informed model is developed combining a Deep Hidden Physics Model (DeepHPM) for degradation mode estimation and an XGBoost classifier for degradation phase detection. Five different histogram-based feature sets are tested, and the 2D current–voltage histogram is found to have the best transfer ability across operating conditions. To adapt the model to new scenarios, a fine-tuning transfer learning strategy is introduced: the dynamic NN is frozen and only the surrogate NN is updated with limited labeled data. Then, it is demonstrated that using even a single cell with knee occurrence in the target dataset is sufficient to fine-tune the pre-trained model and significantly improve accuracy in both degradation mode estimation and phase classification. The results highlight the strong linear correlation between the detected knee onset point and the knee point itself, enabling not only accurate online diagnosis but also reliable knee prediction. Furthermore, the framework supports aging-aware battery classification and second-life repurposing by dividing the degradation trajectory into three phases. By embedding this transferable framework into a cloud-based digital twin, advanced BMS functionalities such as online prognosis, early knee detection, and scenario-adaptive RUL prediction can be achieved.

Another feature extraction method is presented [95]; however, this time, instead of using the histogram-based feature extraction method of voltage and current data as before, a statistical framework is proposed to extract features from the temperature of the surface of the battery in the first 10 cycles, avoiding the need for long-term data collection process. After collecting these data points, elastic net regularization is performed to optimize them and avoid redundant health indicators (HIs). In addition, this regularization reduces the coefficients of the obtained predictive model, which makes it less sensitive to noises. Then, the resulting linear model is validated on six different datasets of LIBs with different cathodes. The proposed statistical temperature HIs are designed to capture universal and chemistry-agnostic traits, ensuring the ability to generalize across various chemistry, cycling rates, and environmental conditions. The linear temperature model achieved strong performance with $R^2$ values often above 0.8, and the worst test error reported was 17.1% MAPE on the SNL-NMC dataset. For the challenging TRI fast-charging dataset, the model achieved a MAPE/RMSE of 15.7% with 257 cycles, while a hybrid version that combined temperature HIs with additional degradation features further improved the accuracy to 14.2%/203 cycles. In all cases, the proposed approach outperformed benchmark early-cycle models, highlighting the ability of universal temperature-based HIs to provide accurate early-cycle lifetime predictions.

In the previously discussed articles, the first used feature extraction primarily to reduce the dimension of the data, and the second used it to eliminate the need for long-term data collection. But what if we have incomplete fragmented data? In [96], a method is specifically proposed to solve this problem side by side to the classical ones such as generalization across different scenarios and the complexity of obtaining full charge/discharge cycles. A physics-informed method is proposed to reconstruct the discharge voltage curve, addressing the problem of fragmented data and enabling subsequent feature extraction from the reconstructed curves. First, the data are produced in a way that mimic actual operating conditions; for example, usually the battery is never fully charged or discharged in real storage systems; hence, the voltage range is chosen to be between 3 and 3.4 V instead of 2 and 3.7 V. The reconstruction model is inspired from the second-order Thevenin model (ECM) shown in Figure 6, and then the parameters of the model are optimized using the Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm to solve a nonlinear least squares problem to minimize the difference between the predicted terminal voltage (calculated using Kirchhoff’s law) by this model and a few measured points of the capacity voltage curve for each cycle. For instance, a discharge curve is obtained for each cycle, and feature extraction for each curve is performed using the Pearson correlation metrics. Once performed, these features are used to estimate the actual capacity of the battery using four regression models, including a linear model and an XGBoost model. The four models had acceptable performances validating the feature extraction approach. Now, using the completed datasets, a Long Short-Term Memory (LSTM) neural network combined with a fully connected layer (FC) and a model-agnostic meta-learning (MAML) approach is used to predict the SOH of the battery. The meta-learner was specifically designed for few-shot learning, enabling accurate short-term capacity predictions even when trained on limited, discontinuous data segments from random starting points in the battery’s lifespan. The framework demonstrated robust performance across three challenging real-world scenarios, achieving a low prediction error (MAE < 0.2), thus proving its ability to operate in data-deficient BMS applications.

Instead of feature extraction from voltage, current, and surface temperature, as we have seen before, in [98], a method that used incremental capacity analysis (ICA) is proposed to predict degradation metrics of the battery in fast-charging scenarios. The ICA is a quantitative degradation method that allows us to track the degradation of the battery based on the variation in $dQ/dV$ with respect to V, where Q is the capacity of the battery and V is the voltage. This technique is proven to be more adequate for electric vehicle applications compared with other methods such as EIS [99] due to its simplicity. The basic idea of the prognosis techniques is to track the packs on the ICA to localize phase transitions in the operation of the battery, and by phase here we mean a particular arrangement of lithium atoms inside the crystal structure during the intercalation process. Usually the highest peak on the ICA curve is the primary phase transition, and the voltage at which this peak occurs is the voltage that results in the biggest structural change in the battery. This voltage value changes as the battery degrades [100], and by tracking this change, the author tries to predict two degradation metrics of the battery. First, the charging voltages of the source and the target battery cells are measured, and a neural network extracts the features from each partial voltage signal using a two-layer bidirectional gate recurrent unit (BiGRU). Then, using a drop-out layer and two linear layers, the neural network estimates the IC curves for each. In this framework, the transfer learning process is achieved through a two-stage domain adaptation strategy rather than explicit parameter sharing between cells. The first stage focuses on feature-level alignment, where the BiGRU-based feature extractor learns domain-invariant temporal representations from the partial charging voltage profiles. During training, features from the labeled source domain and the unlabeled target domain are aligned by minimizing a Maximum Mean Discrepancy (MMD) loss shown in Equation (2), where $I{C}^S$ and ${\widehat{IC}}^S$ denote the true and estimated source curves, and $H^{SIC}$ and $H^{TIC}$ represent the source and target domain features, respectively. This loss allows us to train the model on the labeled source data while simultaneously reducing the discrepancy between source and target domain features.

$${\mathcal{L}}_{\mathrm{IC}}={\displaystyle \frac{1}{N}}\sum _{i=1}^N{(I{C}_i^S-{\widehat{IC}}_i^S)}^2+{\mathcal{L}}_{\mathrm{ICDA}}(H^{SIC},H^{TIC})$$

(2)

The second stage introduces uncertainty-aware adaptation within the degradation predictor. Here, the model predicts both the mean and variance of the degradation trajectory, where the variance represents the confidence in the prediction. An uncertainty-aware pseudolabeling (UADA) mechanism is then employed: provisional degradation labels are generated for the target domain and iteratively refined according to their associated uncertainty. This is performed using the improved loss function shown in Equation (3), where $\lambda$ represents the uncertainty of the model, $y_i^S$ represents the source labels, and $\mu \left({\mathbf{x}}^T\right)$ represents the target predictions. Hence, when $\lambda$ increases, it forces the model to decrease the other term, which means increasing the alignment between the source and target domain, which makes this techniques more robust.

$${\mathcal{L}}_{\mathrm{UADA}-\mathrm{PLDL}}=\lambda \parallel {\displaystyle \frac{1}{N}}\sum _{i=1}^N(1-y_i^S)\odot H_i^{SDF}-{\displaystyle \frac{1}{N}}\sum _{i=1}^N(1-\mu \left(x_i^T\right))\odot H_i^{TDF}{\parallel }_H^2.$$

(3)

This improved framework: UADA-PLDP (Uncertainty-Aware Domain Adaptation with Pseudolabel Degradation Prediction) enables the smooth transfer of degradation knowledge from the source to the target domain even when labeled data are scarce. In contrast to conventional fine-tuning approaches that rely on weight reuse, this framework transfers distributional knowledge and the uncertainty structure, thereby enhancing robustness against dataset heterogeneity. This enables the simultaneous prediction of two key degradation metrics—loss of lithium inventory (LLI) and loss of active material (LAM)—with quantified uncertainty, using only 8 min of partial charging voltage data. The method demonstrates superior accuracy and robustness compared to conventional health factor-based methods, particularly under varying fast-charging protocols and limited labeled data scenarios.

Another study that uses the IC curve is presented [101], but this time it is used for the prediction of the nonlinear health degradation of the battery. The approach starts by splitting the battery lifetime into two parts: a source domain, covering the more gradual and predictable stage of aging (linear part), and a target domain, where nonlinear effects dominate. To decide where this split occurs, the authors compute temporal distribution characteristics (TDCs), which are statistical summaries of how voltage or incremental capacity (IC) curves evolve over time. The split is chosen where the difference between these temporal distributions is largest, which usually corresponds to the onset of nonlinear aging. Once this division is made, a deep learning model based on bidirectional gated recurrent units (BiGRU) is trained with three objectives: (1) predicting multi-step (SOH), (2) estimating IC curves as short-term electrochemical indicators, and (3) aligning features from the source and target domains so that knowledge from early life can be transferred to later life. This alignment is achieved by reducing the statistical distance between source- and target-extracted feature distributions, with adaptive weights that emphasize the most relevant signals near the knee point. In other words, we are forcing the model to obtain accurate predictions using features that exist in the source and target domains. An automatic weighting scheme also balances the three objectives so that the model does not rely on manual tuning. Tested on 151 cells of different chemistries and operating conditions, the framework consistently reduced prediction errors and provided early warnings of knee points, showing that temporal transfer learning can flexibly adapt to nonlinear aging without requiring massive amounts of labeled data from every scenario.

A summary of all aforementioned transfer learning and features extraction techniques is presented in

Table 3. Comparative summary of transfer learning and feature extraction methods for lithium-ion battery prognosis.

Approach	Features/Metrics Used	Predicted Quantity + Error	Datasets	Adequacy for Second-Life Batteries
DeepHPM + XGBoost [92]	Voltage–current histograms	SOH + knee point (Accuracy = 82–89%)	Multiple LIB datasets; few-shot transfer validation	Promising: requires few samples and is adaptable across different scenarios
Linear + Elastic Net Regression [95]	Early surface temperature (first 10 cycles)	RUL; MAPE as low as 14.2%	TRI dataset with multiple chemistries	Promising: chemistry-agnostic health indicators and robust early prediction
LSTM + MAML [96]	Reconstructed discharge voltage curves	SOH; MAE < 0.2% across fragmented datasets	Real-world fragmented datasets	Promising: effectively handles incomplete or fragmented data
BiGRU + ICA [98]	Incremental capacity (dQ/dV) curves	LLI (RMSE = 3.93%) + LAM (RMSE = 9.31%)	EV fast-charging datasets	Promising: fast, robust, handles uncertainty well
BiGRU + TDC [101]	Temporal distribution characteristics of IC and voltage curves	SOH (MAPE = 0.55–2.42%, RMSE = 0.72–2.52%); Knee point (MAPE = 0.54–1.44%, RMSE = 0.72–1.38%)	151 cells; multi-chemistry datasets	Promising: adapts to nonlinear regimes and generalizes across chemistries

All reported performance metrics (MAE, MAPE, RMSE) are as stated in the referenced studies.

. It is clear that this group of approaches is the most appropriate for second-life batteries. This can be traced to different reasons. First, all these techniques use a limited amount of data [92]. Second, the method proposed in [95] is capable of being generalized across different chemistries which is highly useful for second-life packs that can encompass modules with different chemistries. In addition, the data reconstruction approach proposed in [96] is also of great benefit for second-life applications. In fact, batteries used in second-life applications often have fragmented or inconsistent historical data. So the reconstruction method can help normalize and complete these datasets. Finally, the domain adaptation framework used in [98,101] based on feature alignment is very effective in transferring the prediction ability of any model from the linear to the nonlinear degradation phases.

5. Comparative Analysis and Health Management

Based on the sections above, it is important to provide a comparative overview of all approaches to highlight the strengths and limitations of each.

Table 4. Evaluation of all promising prognosis methods based on key strengths for second-life application.

Method	Robust to Uncertainty	Handles Heterogeneous or Sparse Data	High Data Efficiency
Tell-me model [50]		X
Transfer-driven ADMD [52]	X	X
Quantile+TCN [63]	X
FFNN(SCG) [64]			X
FFNN(RF) [64]			X
CNN-BiLSTM-Attention [65]	X
GPR-LSTM [68]	X	X
FPCA+Bayesian [75]		X
RCGAN [77]			X
GAN(CNN-LSTM+ESN) [78]			X
DeepHPM+XGBoost [92]		X	X
Linear+Elastic Net [95]		X	X
LSTM+MAML [96]		X
BiGRU+ICA [98]	X	X	X
BiGRU+TDC [101]	X	X

X indicates that the method possesses the corresponding strength.

summarizes these characteristics across the three aforementioned categories, focusing on strengths that are particularly relevant for second-life battery applications. For instance, robustness against uncertainty is important in this context because second-life batteries often exhibit unpredictable degradation behavior due to prior usage, varying cycling histories, and environmental conditions. In addition, the ability to handle heterogeneous and sparse data is a crucial characteristic, as cells within repurposed packs can differ significantly in their state of health, and complete datasets are rarely available. Finally, high data efficiency eliminates the need for extensive datasets and decreases the computational load, enabling practical deployment of prognosis methods in real-time battery management systems for second-life applications.

Building on the previous sections, it is important to emphasize that the ultimate goal of accurate degradation prediction is not only to prevent faults but also to prolong battery lifetime. In this context, reinforcement learning (RL) techniques are gaining attention for battery health management and control applications in electrical systems. Their main advantage lies in their ability to learn optimal decision-making online through interaction with the environment, allowing adaptation to complex conditions without requiring explicit system models. For instance, an RL-based approach that leverages an empirical model of capacity degradation to extend EV battery life while maintaining driver comfort is proposed in [102]. Another study experimentally applies Q-learning to optimize EV battery management, demonstrating improvements in both energy efficiency and battery lifespan [103]. However, safety guarantees remain a key challenge, preventing full integration of these models. Alternatively, other studies have used RL to enhance the control of converters rather than replace existing controllers [104]. Overall, reinforcement learning for battery management remains an active and rapidly evolving research domain.

6. Conclusions and Future Directions

Based on the comparison Table 1, Table 2, Table 3 and Table 4 presented in this paper, multiple conclusions can be drawn. First, prognosis methods that use explicit physical models are difficult to implement in the case of second-life applications. Hence, the physical knowledge should be integrated to the prognosis model using historical data. In contrast, purely data-driven techniques show better applicability since they rely on metrics that are usually collected by the BMS, but their performance can still be limited by data scarcity, noise, and the heterogeneity of second-life packs. In this regard, GAN-based approaches and wavelet transform feature extraction are particularly promising, since the first can enrich limited datasets and the second can capture features that are particularly related to the nonlinear degradation of the battery. Nevertheless, the most appropriate group of methods remains transfer learning and feature extraction, as they are naturally suited to operate with limited or fragmented data, can generalize across chemistry, and enable effective domain adaptation from linear to nonlinear phases of degradation.

Reinforcement learning techniques are still underexplored in the context of battery prognosis. While a few studies such as [105] have demonstrated their feasibility for capacity prediction, the number of existing techniques is still limited compared to other machine learning methods. The active learning capability of RL, which enables models to adapt dynamically as batteries enter nonlinear degradation regions beyond the knee point, could make it a promising but insufficiently studied approach for second-life applications.

In conclusion, future research should focus on hybrid frameworks that integrate the robustness of transfer learning with data augmentation strategies such as GANs, while also incorporating feature extraction techniques capable of capturing health indicators that can be generalized across different stages of battery life. Finally, reversible degradation phenomena such as voltage/capacity recovery, which influence long-term performance assessment of the battery and provide new insights for improving prognosis models remain another underexplored direction.