Deep Learning for State of Health Estimation of Lithium-Ion Batteries in Electric Vehicles: A Systematic Review

Abstract

Electric vehicles (EVs) play a crucial role in addressing the energy crisis and mitigating the greenhouse effect. Lithium-ion batteries are the primary energy storage medium for EVs due to their numerous advantages. State of health (SOH) is a critical parameter for managing the health of lithium-ion batteries, and accurate SOH estimation forms the foundation of battery management systems (BMS), ensuring the safe operation of EVs. Data-driven deep learning techniques are attracting significant attention because of their strong ability to model complex nonlinear relationships, which makes them highly suitable for SOH estimation in lithium-ion batteries. This paper provides a comprehensive introduction to the common deep learning techniques used for SOH estimation of lithium-ion batteries, with a focus on model architectures. It systematically reviews the application of various deep learning algorithms in SOH estimation in recent years. Building on this, the paper offers a detailed comparison of these deep learning methods and discusses the current challenges and future directions in this field, with the aim of providing an extensive review of the role of deep learning in SOH estimation.

1. Introduction

With the ongoing advancement of the global energy transition, electric vehicles (EVs) have emerged as a pivotal solution for reducing carbon emissions and fostering sustainable development [1]. Lithium-ion batteries have become the predominant energy source for EV power systems owing to their high energy density, rapid charge–discharge capabilities, and extended service life. As the primary power source of EVs, the performance degradation of lithium-ion batteries directly influences both the range and safety of these vehicles [2,3]. To accurately assess the health state of lithium-ion batteries, researchers have introduced the concept of state of health (SOH) as a critical metric for quantifying the extent of battery aging. Precise estimation of SOH is vital for optimizing battery management systems (BMS) and ensuring the safe and reliable operation of EV systems [4,5,6]. Accurate estimation of SOH is crucial for optimizing the operation of BMS, ensuring that the battery functions at its peak efficiency, and enhancing the overall performance of the system. Furthermore, accurate SOH estimation contributes to extending the lifespan of lithium-ion batteries and facilitates predictive maintenance. In conclusion, rapid and accurate SOH estimation plays a pivotal role in ensuring the safe and reliable operation of EVs [7].

The aging process of lithium-ion batteries is influenced by the interplay of numerous factors, including electrochemical mechanisms, operating load conditions, and ambient temperature, As a result, SOH estimation is both complex and nonlinear, with time-varying characteristics. Traditional SOH estimation methods typically rely on expert knowledge and experience, which limits their generalizability and effectiveness in addressing the complex real-world conditions encountered in EVs [8,9]. Thanks to rapid advancements in machine learning, an increasing number of studies have applied machine learning algorithms such as support vector machines (SVM) and artificial neural networks (ANN) to battery SOH estimation, yielding promising results [10]. However, conventional machine learning models often suffer from limited generalization capabilities due to their relatively low complexity. Furthermore, as sensor technologies advance and the volume of battery data continues to grow, traditional machine learning approaches struggle to effectively manage and process large datasets. In recent years, deep learning has emerged as a powerful solution to these challenges in battery SOH estimation, offering superior nonlinear fitting capabilities and the advantage of end-to-end feature learning [11].

Deep learning is a subfield of machine learning that emphasizes learning through multilayered neural networks which inherently involve a large number of parameters [12]. By leveraging multiple layers of abstraction, deep learning is capable of directly extracting features from raw data, enabling more complex pattern recognition. This ability positions deep learning as particularly advantageous for processing large datasets and achieving higher levels of abstraction [13,14]. With continuous advancements in computational power, limitations on computational resources no longer represent a barrier to the implementation of more sophisticated SOH estimation methods. Recent years have seen a surge in research applying deep learning techniques to SOH estimation, with models such as convolutional neural networks (CNN), long short-term memory (LSTM), and transformers increasingly being utilized and yielding promising results. These studies have significantly contributed to enhancing the robustness and accuracy of SOH estimation [15].

There have been numerous reviews on battery SOH estimation. For instance, Berecibar et al. provided a systematic review of SOH estimation methods for lithium-ion batteries [6]. Similarly, Yao et al. conducted a comprehensive review of various SOH estimation techniques [7]. Wang et al. offered an in-depth analysis of online SOH estimation methods [16], while Ren et al. focused on machine learning algorithms for SOH estimation [17]. Additionally, Vidial et al. primarily reviewed the application of machine learning in SOH estimation for electric vehicle batteries [10], whereas Guirguis et al. focused on transformer-based deep learning models for SOH estimation [18]. While these systematic reviews have made significant contributions to advancing research in this field, most existing reviews have concentrated primarily on overall SOH estimation methods or machine learning applications in SOH estimation, leaving the smaller body of work focusing on deep learning relatively underexplored. Therefore, this review aims to provide a comprehensive overview of deep learning techniques applied to SOH estimation by systematically summarizing the current research, identifying challenges, and outlining potential future directions to guide further investigation in this field.

The structure of this paper is as follows: Section 2 provides an overview of deep learning techniques and details the models often used in SOH estimation; Section 3 systematically describes the research status of deep learning in SOH estimation; Section 4 discusses current challenges and possible future research directions; and Section 5 summarizes and concludes the paper.

2. Overview of Deep Learning

Deep learning is a subset of machine learning based on artificial neural networks and focused on learning data representations. It typically involves multilayer neural network models with a vast number of parameters. Due to the intricate working mechanisms of lithium-ion batteries, the aging process has highly complex nonlinear characteristics. Additionally, the heterogeneity of batteries, including variations across different battery models as well as within the same model, can result in diverse aging behaviors. This variability poses significant challenges for accurate SOH estimation and knowledge transfer. With their enhanced capacity for nonlinear representation, deep learning models offer significant advantages in addressing the complexities inherent in SOH estimation. This section provides an overview of the deep learning algorithms commonly employed in SOH estimation, highlighting their strengths and potential applications. Deep learning models can be broadly categorized into CNNs, sequential models, and hybrid models.

2.1. Convolutional Neural Networks

A CNN is a biomimetic computational model within a deep learning framework, and is specifically designed to process data that possess a network-like structure. The fundamental CNN concept is inspired by the hierarchical perception mechanism of the biological visual system, enabling efficient feature extraction and abstraction from the input data [19]. Through an end-to-end learning paradigm, CNN progressively maps the original input to the high-level semantic features in a layer-by-layer fashion. This approach has been widely applied in fields such as computer vision, medical image analysis, and natural language processing. Considering that lithium-ion battery data can be represented in a grid-like structure formed by regular time–axis sampling, CNN-based approaches have become increasingly relevant for SOH estimation in recent years. The ability of CNNs to automatically extract features from such structured data makes them particularly suitable for the task of SOH estimation, allowing for effective and accurate prediction of battery health over time [20,21,22].

The hierarchical architecture of a CNN is shown in Figure 1.

The core modules of a CNN include the following:

(1)

Convolutional Layer

The convolutional layer serves as the core component of a CNN. The convolutional kernel performs a sliding window operation on the input data’s local receptive field to generate a feature map, which is primarily used for extracting key features. Through this convolution operation, the CNN learns the features of the input data by focusing on local regions, enabling local perception and feature extraction. In essence, the convolutional layer scans the input data using the kernel in order to create a feature map that highlights essential characteristics such as edges, textures, or temporal patterns (in the case of time series data) [24]. These extracted features are crucial for subsequent layers, where they are used to build the higher-level abstractions which are key for making accurate predictions such as SOH estimation in lithium-ion batteries.

(2)

Pooling Layer

The pooling layer employs pooling operations to perform spatial downsampling of feature maps, effectively retaining the most salient feature information. Common pooling operations include max pooling and average pooling. The pooling layer enhances the robustness of the model by providing invariance against small spatial translations and distortions, and also helps to reduce computational complexity.

(3)

Activation Function

Following the convolutional and fully connected layers, the model typically incorporates a nonlinear activation function to introduce nonlinearity, thereby enhancing the network’s capacity to model complex relationships. The most commonly used activation function is Rectified Linear Unit (ReLU), defined as $ReLU=max0,x$, where x is the input [25].

(4)

Fully Connected Layer

The final stage of a CNN typically employs one or more fully connected layers in which the high-level feature maps are flattened and fed into a multilayer perceptron for classification or regression, enabling global semantic understanding.

(5)

Output Layer

The output layer is typically configured as a softmax or linear activation layer, which is responsible for generating the final classification results and their corresponding probability distribution.

With the advancement of CNNs, a number of variants have been widely adopted in fields such as image classification, semantic segmentation, object detection, and video analysis. The key design elements of CNN, namely, local perception and parameter sharing, can greatly enhance model generalization ability and mitigate the risk of overfitting. This is especially beneficial for addressing battery heterogeneity, as CNN models can effectively adapt to data variations.

The end-to-end learning approach of CNN eliminates the need for manual feature engineering by automating both feature extraction and the classifier’s performance integration through backpropagation, thereby reducing reliance on expert knowledge of battery aging processes. Thanks to these advantages, CNNs have emerged as a prominent tool for SOH estimation in lithium-ion batteries, and their application has been extensively studied in recent years. The ability of CNNs to automatically learn relevant features from raw data makes them a powerful and efficient method for predicting battery health, especially in complex battery systems [26].

2.2. Sequence Models

In the field of deep learning, sequence models are designed to process sequential data, making them particularly effective for handling time-dependent information. These models are adept at capturing temporal dependencies within the data, which makes them well suited for applications involving sequential or time series data [27,28]. Given the strong time series characteristics of the aging process in lithium-ion batteries, many recent studies have focused on applying sequence models to estimate the SOH of lithium-ion batteries. These approaches have yielded promising results, demonstrating the effectiveness of sequence models in predicting battery health over time. By leveraging the temporal patterns inherent in the aging process, sequence models can provide more accurate and reliable SOH predictions, which is crucial for the safe and efficient management of battery-powered systems. Common sequence models used in SOH estimation include RNNs and their variants, transformer-based models, and hybrid models.

2.2.1. Recurrent Neural Networks and Variants

• Recurrent Neural Networks
  Recurrent Neural Networks (RNN) are designed for processing sequential data. Unlike traditional feedforward networks, RNNs maintain memory of previous inputs through cyclic connections in the hidden layers. This allows the network’s output at each time step to be influenced by both the current input and the previous hidden state. RNN are well suited for data processing tasks involving time series, text, and other sequential data structures [29,30]. A typical RNN architecture is illustrated in Figure 2.

  Figure 2. A simple recurrent neural network. The hidden state propagates through the input sequence; the weight matrices U, V, and W correspond to the linear operations on the input vector u, output vector v, and internal variables h, respectively. At each time step t, the operations depend on the internal variables evaluated at the previous time step t−1 [31].

  Figure 2. A simple recurrent neural network. The hidden state propagates through the input sequence; the weight matrices U, V, and W correspond to the linear operations on the input vector u, output vector v, and internal variables h, respectively. At each time step t, the operations depend on the internal variables evaluated at the previous time step t−1 [31].
  The core RNN concept is the incorporation of memory through recurrent connections in the hidden layers. Each neuron in the RNN receives inputs from both the current input and the previous hidden state, which jointly determine the current state. This mechanism allows RNNs to capture contextual information and model temporal dependencies in time series data. A typical RNN architecture includes three layers: input, hidden, and output.
  RNNs excel at processing and modeling time series data by passing information from previous time steps to the current state via updates to the hidden state at each step, effectively capturing temporal dependencies. Additionally, RNNs share the same set of weights across all time steps, enabling them to process data at different time points using a common weight set, which significantly reduces the model’s parameter count. However, standard RNNs face challenges such as the vanishing and exploding gradient problems when dealing with long time series. These issues hinder their ability to capture long-term dependencies within the sequence, resulting in limited performance on longer sequences. For applications such as lithium-ion battery modeling which involve long working cycles, standard RNNs may struggle to maintain accurate predictions across the entire cycle [28,32].
  To address the issue of long-term dependencies, several RNN variants have been developed to mitigate the shortcomings of standard RNNs in capturing long-term temporal dependencies, with notable examples including LSTM networks and gated recurrent units (GRU) [32].
• Long Short-Term Memory
  LSTM is a specialized RNN architecture designed to address the vanishing gradient problem that traditional RNNs face when processing long sequences [33]. LSTM effectively regulates the flow and updating of information through the integration of a gating mechanism, allowing for the preservation of long-term dependencies and selective retention or forgetting of historical information. Alongside the cell state, the core of the LSTM model consists of three gate structures: the forget gate, input gate, and output gate. These components enable LSTM to efficiently store, update, and manage information across extended sequences, ensuring robust modeling of temporal dependencies [34,35].
• Gated Recurrent Units
  GRU is a variant of LSTM with a simplified structure that makes model training more efficient [36]. GRU models consist of two main components, namely, the reset gate and update gate. This design is more streamlined compared to that of LSTM, and has a significantly reduced number of parameters while achieving similar performance. In the context of lithium-ion battery SOH estimation, GRU models can provide an effective solution by alleviating the computational burden of LSTM, particularly in practical EV systems where computational resources are limited [37].

2.2.2. Self-Attention-Based Models

The attention mechanism is a technique in artificial neural networks that mimics cognitive attention. This mechanism enables the network to assign higher weights to certain parts of the input data while reducing the influence of others, thereby directing its focus toward the most critical segments. The core idea is to dynamically capture global dependencies at any position within the sequence through attention weights, replacing the recurrent structures found in traditional models. This allows the network to selectively emphasize relevant portions of the input data rather than treating all information uniformly when processing sequential data [38].

The transformer model is a neural network architecture based on the attention mechanism and primarily designed for sequence-to-sequence tasks [39]. Unlike RNNs, which rely on recurrent connections, the transformer architecture entirely replaces recurrence with self-attention mechanisms. This architectural shift significantly enhances parallelization during training, making such models well suited for large-scale datasets [40].

As research on transformer models has increased, their potential has been continuously explored and expanded. Compared to RNNs and their variants, transformer models exhibit superior ability to capture long-range dependencies, making them well suited for modeling the complex life cycle of lithium-ion batteries. Moreover, because transformers rely entirely on attention mechanisms, they can process all time steps in parallel during training, significantly accelerating the training process. This advantage facilitates the rapid development of SOH estimation tasks, enhancing both efficiency and scalability [18].

2.3. Hybrid Models

With the advancement of deep learning technologies, an increasing number of studies have begun to integrate various models to enhance sequence modeling capabilities. Hybrid models are particularly effective in this regard, as they combine the strengths of different types of models to further improve performance. The core design principle behind hybrid models is the fusion of key components from diverse neural network architectures with the aim of overcoming the limitations inherent in individual models. This design approach follows the principle of complementarity, utilizing fusion strategies such as cascade, parallel, and embedded structures to achieve functional synergy [41].

3. State of Health Estimation

The SOH of a battery is a critical indicator of the degradation in its performance, and is primarily used to assess the deviation of the battery’s current state from its original performance characteristics. SOH reflects the loss of battery performance over time, which may result from factors such as aging, usage patterns, or environmental influences. Temperature has a significant impact on the SOH of a battery. The internal resistance of the battery increases at low temperatures, leading to a decline in charging and discharging efficiency. Additionally, the slower chemical reaction kinetics at low temperatures result in a reduction in battery capacity. In contrast, high temperatures accelerate internal chemical reactions, which may temporarily enhance battery performance. However, prolonged exposure to elevated temperatures causes material degradation, particularly in the electrolyte and electrode materials, ultimately leading to capacity fading and an increase in internal resistance [42].

Moreover, charging and discharging currents also influence SOH. During high-current charging, the chemical reaction rate and current density within the battery increase, potentially causing overcharging or localized hotspots which can accelerate battery degradation. Similarly, excessive discharge currents may induce irreversible chemical reactions in the electrolyte and electrode materials, leading to capacity loss. Therefore, high-current charging and discharging significantly affect the accuracy of SOH estimation [43].

Commonly used SOH evaluation metrics include battery capacity and internal resistance. SOH is typically defined as the ratio of the battery’s current maximum available capacity to its rated capacity, expressed as a percentage. Equation (1) provides the most general health condition equation, with $Q_{max}$ representing the maximum battery charge and $C_n$ the battery’s nominal initial charge capacity [44].

$$\mathrm{SOH}={\displaystyle \frac{Q_{max}}{C_n}}$$

(1)

SOH is a critical metric in battery health management. Accurate and rapid SOH estimation plays a vital role in ensuring optimal performance and longevity of batteries. Currently, there are numerous research studies focusing on SOH estimation methods, which can be broadly categorized into two approaches: model-based and data-driven [45]. Model-based methods estimate SOH based on the degradation mechanisms of lithium-ion batteries. This approach has reached a relatively mature stage, with algorithms falling into two main categories, namely, electrochemical models and equivalent models. However, due to the complexity of real-world battery systems, model-based methods often exhibit poor robustness and are challenging to implement with high accuracy. In contrast, the rapid development of machine learning has led to data-driven methods garnering increasing attention. These methods bypass the need for complex chemical reaction models and instead directly construct SOH estimation models using historical or real-time data. By learning from large amounts of input–output data, such data-driven methods avoid the intricate physical modeling process and offer enhanced robustness in estimating SOH [46].

The deep learning approach for SOH estimation of lithium-ion batteries represents an advanced data-driven methodology. By leveraging black-box deep learning models, this approach enables end-to-end SOH estimation for complex lithium-ion batteries, eliminating the need for manual feature design that is typical in traditional data-driven methods. Furthermore, the robust nonlinear representation capabilities of deep learning models significantly enhance the accuracy of SOH estimation. The standard workflow for deep learning-based SOH estimation methods is illustrated in Figure 3. The process begins by collecting data from the battery, followed by selecting an appropriate deep learning model based on the characteristics of the data. Next, the model is trained using the collected data and the SOH is estimated. Finally, the performance of the estimation method is evaluated [47,48].

As illustrated in Figure 4, the application of prominent deep learning techniques in SOH estimation can be broadly categorized into CNN-based estimation methods, sequence model-based estimation methods, and hybrid model estimation methods [49,50].

3.1. Convolutional Neural Network-Based Methods

CNNs have become a prominent research model in the field of computer vision, and their application to SOH estimation of lithium-ion batteries has yielded promising results in recent years. Chemali et al. proposed a CNN-based framework to directly estimate the SOH from voltage, current, and temperature measurements taken during battery charging. Their research explored CNNs with varying numbers of layers and neurons, demonstrating that CNNs can accurately estimate SOH using partial charging data under two different temperature conditions [51]. Similarly, Chen et al. developed a novel SOH estimation method based on partial constant-voltage (CV) charging phase data and a CNN model. They proposed a method for estimating SOH with high precision by using partial CV charging data to analyze the trends of the CV charging phase as the battery ages [52]. Lu et al. introduced a CNN based on feature fusion for estimating battery health states from local periodic voltage measurements, resulting in significantly improved accuracy of battery SOH estimation [53]. Furthermore, Chen et al. developed a CNN-based estimation model for estimating SOH from constant current charge–discharge data. They found that the CNN method has limitations when predicting SOH for batteries below 80% health or when attempting to estimate SOH from curves with varying C-rates [54].

Bockrath et al. proposed an SOH estimation algorithm based on different segmented local discharge curves for lithium-ion batteries. In their approach, raw sensor data are directly input into a temporal convolutional network (TCN) without the need for feature engineering. The network is capable of processing raw sensor data and estimating battery SOH under various aging and degradation conditions. This work highlights the potential of CNNs in SOH estimation, particularly TCNs [55]. Likewise, Gao et al. proposed a novel SOH estimation framework combining mixers and bidirectional temporal convolutional networks (BTCNs). This framework not only maximizes the use of local and global features from the input data to estimate SOH but also reduces redundancy in temporal and channel information [56].

Zhang et al. focused on improving the generalization of SOH estimation methods, proposing an end-to-end SOH estimation approach based on relaxation voltage that does not rely on specific cycle conditions. They input the relaxation voltage curve at full charge into a 1D CNN to directly estimate SOH. To address the issue of limited data, they employed transfer learning to transfer knowledge from source domains to target domains [57]. Ma et al. proposed a novel battery personalization SOH estimation method based on transfer learning. Their method utilizes a CNN combined with an improved domain adaptation technique to construct the SOH estimation model. The CNN is employed to automatically extract features from the raw charging voltage trajectories, while the maximum mean discrepancy (MMD) domain adaptation method is applied to reduce the distribution discrepancy between the training and testing battery data. This approach extends MMD from classification tasks to regression tasks for SOH estimation [58].

3.2. Sequence Model-Based Methods

3.2.1. Methods Based on Recurrent Neural Networks and Variants

Ansari et al. proposed an improved hybrid model combining recurrent neural networks (RNNs) and jellyfish optimization (JFO) to estimate the SOH of lithium-ion batteries. The JFO algorithm was applied to obtain the most suitable hyperparameters for the RNN model, aiming to achieve accurate SOH estimation results [59].

Wu et al. proposed an efficient estimation method based on machine learning. First, voltage distribution and incremental capacity changes during the charge–discharge process were obtained through cyclic life testing to extract health features related to battery degradation. Next, gray relational analysis and entropy weight methods were used to analyze the health features. Finally, LSTM was employed to estimate the battery’s SOH. Ma et al. proposed an SOH estimation method based on an improved LSTM network and the extraction of health indicators (HIs) from the charging and discharging process [60]. Lin et al. proposed an SOH estimation method based on a multi-source feature attention LSTM network. This approach considers the analysis of eight health indicators extracted from incremental capacity (IC), differential temperature, and differential thermoelectric voltage curves. Experimental results showed that the LSTM model with a local attention mechanism outperformed traditional LSTM and LSTM models based on global attention mechanisms in terms of SOH estimation accuracy [61]. Liu et al. introduced an accurate SOH estimation method for rechargeable lithium-ion batteries based on the charging process and a long short-term memory recurrent neural network (LSTM-RNN). To learn the mapping function without relying on battery models or filtering techniques, their method first inputs health indicators (HIs) extracted from the charging process into the LSTM-RNN and trains it to encode the dependencies of the relevant data sequences. Subsequently, the trained LSTM-RNN can accurately estimate the online SOH of the lithium-ion battery using the extracted HIs [62]. Zhang et al. studied an SOH estimation method based on incremental capacity (IC) and LSTM networks. They proposed an improved IC curve acquisition method based on the reference voltage, which retains the important features of the IC curve while reducing computational complexity. This approach enhances the efficiency of SOH estimation by leveraging critical information from the IC curve while minimizing the processing load [63]. Gong et al. proposed an improved LSTM-based data-driven estimation method to address the issues of parameter determination and slow convergence in the classic LSTM approach. They utilized a particle swarm optimization (PSO) algorithm to optimize the key hyperparameters in the neural network. Their experimental results demonstrated improved prediction accuracy by at least 5% compared to the classic LSTM method [64].

Chen et al. proposed a precise SOH estimation method based on temperature prediction and GRU neural networks. First, using randomly discontinuous short-term charging data, the extreme learning machine (ELM) method was employed to predict the temperature variations during the constant current charging process. Next, the finite difference method was used to calculate the raw temperature differential changes, which were smoothed using a Kalman filter. On this basis, multidimensional health features were extracted from the differential temperature curve to reflect the battery’s degradation from various angles, with six highly correlated features selected using the Pearson correlation coefficient method. With all relevant health features prepared, a GRU neural network was then used to predict the SOH. Experimental results demonstrated that this method effectively controlled the health state error to within 2.28% based solely on partial random and discontinuous charging data, thereby validating its prediction performance [65].

Hong et al. proposed an SOH prediction method for automotive batteries based on a GRU neural network. Their method first extracts and preprocesses actual vehicle operation data to improve data reliability, including data cleaning and segmentation. Then, Kalman filtering and recursive least squares are used to identify the Ohmic internal resistance (OIR). A GRU neural network is established to predict the battery SOH, providing a novel solution for the accurate estimation of SOH in real-world vehicle batteries [66]. Wang et al. proposed a lithium-ion battery SOH estimation method that combines empirical mode decomposition (EMD), random forests (RFs), and GRU. This method first extracts the time intervals of constant voltage rise and constant voltage drop as health indicators (HIs) and uses Pearson’s correlation coefficient to analyze the relationship between the health indicators and SOH. Next, the EMD method is used to decompose the battery SOH data and the variance contribution ratio (VCR) is introduced to measure the relationship between the intrinsic mode function (IMF) components and SOH. Finally, an SOH estimation model based on the EMD–VCR–GRU–RF framework is established, effectively balancing prediction accuracy and computational efficiency [67].

3.2.2. Self-Attention Mechanism-Based Methods

The transformer model and its variants based on the self-attention mechanism are among the most representative models in this field. Luo et al. proposed a novel SOH estimation method based on a data preprocessing approach and a CNN–transformer framework. In the data preprocessing phase, highly correlated features are selected using the Pearson correlation coefficient (PCC). Principal component analysis (PCA) is then employed to eliminate redundant feature information, reducing the computational complexity of the estimation model. Subsequently, all features are normalized using the min–max scaling method, which accelerates the training process and minimizes the cost function. After preprocessing, the features are fed into the CNN–transformer model. The NASA battery dataset was used as both the training and testing dataset for building the proposed model. Simulation results demonstrated that the absolute estimation error for each dataset was within 1%. The mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) were all maintained below 0.55%, validating the estimation performance [68].

Chen et al. innovatively proposed a new SOH estimation method based on vision transformer networks (ViTs). By analyzing the training speed and accuracy at different sampling points, they designed an adaptive algorithm that selects the most suitable sampling data for the ViT, thereby guiding battery data acquisition in practical systems and reducing the manual effort involved in the process of neural network training. Furthermore, they enhanced ViT by incorporating dimensional transformation layers, multilayer perceptrons (MLP), and trainable regression labels. Experiments conducted on two distinct datasets demonstrated that the proposed framework achieved a precision exceeding 0.01, outperforming other available techniques. This research represents an innovative application of ViT models to SOH estimation, offering a novel perspective [69].

Xu et al. proposed a lithium-ion battery SOH estimation method based on incremental capacity analysis (ICA) and transformer models. First, the ICA method was used to extract the battery’s original incremental capacity (IC) curves, which were then processed using a dual filtering approach combining moving average smoothing and Gaussian smoothing to enhance the curve and extract peak features. Subsequently, a transformer network model based on a multihead attention mechanism was established. The extracted IC curve peak features were input to the model to estimate the SOH of the lithium-ion battery using the transformer. This study utilized experimental data from three different lithium-ion battery sources and conducted experiments based on various input features, prediction start points, and ambient temperatures. A comparative analysis with common machine learning methods was also performed. The experimental results demonstrated that the proposed method outperforms traditional machine learning methods in terms of long-term prediction accuracy and temperature adaptability. This research innovatively combines ICA and transformer models and emphasizes the use of data processing techniques to refine raw data, thereby enhancing the ability of transformer models to accurately estimate battery SOH [70].

Shu et al. proposed an efficient SOH estimation algorithm based on customized voltage segments and the latest transformer models. They first investigated the voltage changes during the battery charging process at different states of health (SOH). By analyzing the trend of changes in the charging voltage slope, specific voltage segments were extracted as health indicators to characterize SOH variations. Building upon this, they proposed an integrated transformer model with a multihead self-attention mechanism to capture temporal information and extract useful features, ultimately achieving precise SOH estimation. This approach leverages both the voltage behavior during charging and the advanced capabilities of transformer models to provide accurate and efficient estimation of battery SOH [71].

3.3. Hybrid Methods

Hybrid methods combine multiple models to leverage the strengths of different components, potentially improving the accuracy of SOH estimation. Gu et al. proposed a hybrid model utilizing cascaded CNN and transformer models. The Pearson correlation coefficient was used to select highly correlated features, then principal component analysis (PCA) was applied to reduce computational complexity [72]. This approach achieved favorable results on the NASA dataset, demonstrating the high precision and stability of hybrid CNN-transformer models in battery SOH estimation. Similarly, Bao et al. designed an end-to-end multi-battery shared hybrid neural network (NN) prediction framework. Their framework represents a fusion of multiple models for use in SOH estimation, integrating a CNN, variational long short-term memory (VLSTM) networks, and a dimensional attention mechanism (CNN-VLSTM-da) for estimating the SOH of lithium-ion batteries [73]. Hong et al. proposed an SOH estimation method based on a dual-stage attention recurrent neural network (DARNN) and time-varying charging process health feature (HF) extraction. Their proposed method was validated using the NASA battery dataset. The results showed that this method can accurately estimate the SOH of lithium-ion batteries [74].

Mazzi et al. proposed a hybrid model for real-time SOH estimation based on a 1D-CNN and bidirectional gated recurrent units (BiGRU). This deep learning framework extracts relevant features from input data using the 1D CNN layer, then performs sequential learning in both directions via the BiGRU layer. This method demonstrated higher accuracy in estimating SOH compared to other commonly used models [75]. Li et al. introduced an end-to-end prediction framework for SOH estimation consisting of a 1D-CNN and an active state-connected LSTM network, achieving high precision in SOH estimation [76]. A hybrid model proposed by Fan et al. combines the GRU and CNN deep learning techniques to capture shared information and temporal dependencies in charging curves. Their method leverages observed data such as voltage, current, and temperature from new charging curves to achieve more accurate SOH estimation [77]. Tang et al. developed a hybrid neural network model with an attention mechanism for SOH estimation of lithium-ion batteries. This model combines CNN, convolutional block attention modules (CBAM), and LSTM networks. Through transfer learning and fine-tuning strategies, it estimates SOH under different battery operating conditions, achieving good accuracy [78]. Xu et al. proposed a hybrid CNN-LSTM model with skip connections to address the network degradation problem caused by LSTM, providing improved accuracy and robustness [79].

Mao et al. introduced an advanced SOH prediction method combining a 2D-CNN and bidirectional long short-term memory (BiLSTM) with Gramian angular field (GAF) technology. This approach offers a new perspective for SOH estimation [80]. Zhao et al. proposed a hybrid attention and deep learning method for SOH prediction of lithium-ion batteries. The model integrates the advantages of CNNs, GRU networks, and attention mechanisms to predict the battery health state. In a comparison with eleven mainstream prediction methods, their approach demonstrated superior performance in predicting battery health [81]. He et al. developed a multiscale convolutional attention mechanism (MCA) to automatically capture and enhance the highly correlated features of SOH degradation in charging data. They combined this with a residual module to create a hybrid neural network model with improved SOH estimation performance. Jia et al. proposed a hybrid prediction model combining bidirectional gated recurrent unit (BiGRU) and transformer elements with multi-head attention to effectively address the challenges of long time series SOH estimation. This model leverages the ability of BiGRU to capture temporal dependencies in the data and the transformer attention mechanism to focus on key features, making the resulting model suitable for estimating SOH over extended periods. The integration of these components helps to improve the performance in terms of both accuracy and robustness when handling complex battery data [82].

Zheng et al. proposed an SOH estimation method for lithium-ion batteries during random charging processes based on a convolutional gated recurrent unit (CNN-GRU) framework. This method adaptively extracts key features from voltage, current, and temperature curve segments during charging, enabling end-to-end SOH estimation of lithium-ion batteries [83]. Yao et al. combined CNN, wavelet neural network (WNN), and wavelet long short-term memory (WLSTM) to design an SOH estimation method based on battery aging factors. Their proposed CNN–WNN–WLSTM estimation scheme inherits the fast convergence and robust stability of WNN and the ability of LSTM to extract temporal sequence features from data. This method is well suited for SOH estimation [84].

Liu et al. proposed a method to address the measurement errors in voltage and temperature sampling by designing reconstructed feature sequences (RFSs) to suppress signal noise. They then introduced a CNN–GRU network with an attention mechanism to estimate the SOH based on short-term RFS samples. To further improve accuracy, a parallel structure was designed to effectively integrate feature information from the flow, raw samples, and RFSs. Extensive experiments were conducted on the Oxford battery degradation dataset to validate the performance of the proposed method [85].

Huang et al. proposed an SOH estimation method based on the temporal pattern attention (TPA) mechanism and a CNN–LSTM model. First, health factors related to capacity degradation were selected based on the charging and discharging curves of lithium-ion batteries, while local anomalies in the health factor data were detected using the local outlier factor (LOF) method. These anomalies were then corrected through Lagrange interpolation to ensure the continuity of the health factor data. Second, the effectiveness of the health factors was evaluated using Pearson and Spearman correlation analyses and the valid health factors were compiled into a dataset of health factors (HFs). Third, the TPA mechanism was integrated into the CNN–LSTM model to form the CNN–LSTM–TPA model, enhancing its ability to capture critical information. The whale optimization algorithm (WOA) was employed to optimize the model’s hyperparameters [86].

Tian et al. developed a hybrid model based on CNN, bidirectional LSTM (BiLSTM), and an attention mechanism (AM) to predict the SOH of lithium-ion batteries. By analyzing the battery’s charging and discharging processes, they extracted indirect health indicators (HI) that were strongly correlated with capacity. These HIs were then used as the input for the CNN, which utilized convolution and pooling operations across various layers to extract features from the battery’s time series data. Building upon this, a BiLSTM deep model was constructed to collect both forward and backward dependency data from the CNN. An attention mechanism was then employed to further emphasize the relationships between the sequential data, resulting in accurate SOH estimation. This study highlights the potential of attention mechanisms within hybrid models [87].

3.4. Comparison of Deep Learning Methods for State of Health Estimation

With the growing application of deep learning technologies in SOH estimation of EV lithium-ion batteries, researchers have proposed various methods based on deep learning to accurately estimate SOH. These approaches exhibit distinct characteristics and are tailored for different application scenarios. This section aims to discuss and compare several representative SOH estimation methods while emphasizing the unique features of each. Table 1 presents the classification and key characteristics of several representative research methods.

Due to characteristics of CNNs such as local perception, parameter sharing, and translation invariance, CNN-based methods can effectively extract time–space features from battery data, especially for SOH data with distinct patterns. However, CNN struggle to capture long-term temporal dependencies, leading to suboptimal performance when dealing with long time series data.

Methods based on typical RNNs benefit from the structural design of these models, which are better suited for handling time series data and have generally achieved good results in SOH estimation [88]. Nevertheless, typical RNNs suffer from the vanishing gradient problem, making them ineffective at capturing long-term trends such as battery degradation. As a result, RNNs are rarely used alone for complex SOH estimation tasks, and are typically employed as baseline models for comparison [89].

On the other hand, LSTM addresses the vanishing gradient issue by incorporating a gating mechanism, enabling the capture of long-term cyclical features of changes in battery health state over time. Therefore, LSTM is a common and effective choice in scenarios where historical battery data must be considered for SOH prediction. However, LSTM is computationally complex, especially when dealing with long time series, leading to higher training and prediction costs. Consequently, its use requires a careful consideration of computational resources [90].

In contrast, GRU also utilizes a gating mechanism to control the flow of information, while its structure is simpler with fewer gates (only the update and reset gates). This results in higher computational efficiency and faster convergence. However, due to the simplification of the model, GRU-based models may not be as powerful as LSTM-based ones when handling long and complex time series. Nevertheless, GRU offers sufficient performance for most SOH estimation tasks [91,92].

Transformer-based models abandon the recurrent nature of RNNs and their variants, instead relying entirely on the self-attention mechanism. Unlike LSTM and GRU, self-attention allows models to capture long-range dependencies at any position within the sequence, limited by the sequence length. Additionally, the computation process in transformers can be parallelized, which is more efficient when dealing with large-scale data. However, the training process of transformers is relatively complex and requires significant computational resources.

Hybrid models are typically constructed to address specific issues and leverage the strengths of different models. When model fusion is conducted appropriately, it can significantly enhance the accuracy and robustness of SOH estimation. However, fusing models increases both overall complexity and computational load. Moreover, if model fusion is not properly executed, it may not lead to improved performance.

Table 1. Comparison of SOH estimation methods in deep learning.

Methods	Researches	Advantages	Disadvantage
CNN-based methods	[51,52,53,54,55,57]	Efficient extraction of time-space features.	Limited ability to capture long-term dependencies.
RNN-based methods	[59]	Ability to handle time dependence in time series data.	Gradient disappearance problem.
LSTM-based methods	[60,63,64,93]	Ability to capture long-term dependencies.	High computational cost.
GRU-based methods	[65,66,67]	Compared with LSTM, the calculation efficiency is higher.	Reduced ability to handle long time series compared to LSTM.
Transformer-based methods	[68,69,70,71]	Parallel computation.	Complex training process with high computational resource consumption.
Hybrid models methods	[72,73,74]	Ability to leverage the strengths of different models.	Increases model complexity and computational cost.

Table 1. Comparison of SOH estimation methods in deep learning.

Methods	Researches	Advantages	Disadvantage
CNN-based methods	[51,52,53,54,55,57]	Efficient extraction of time-space features.	Limited ability to capture long-term dependencies.
RNN-based methods	[59]	Ability to handle time dependence in time series data.	Gradient disappearance problem.
LSTM-based methods	[60,63,64,93]	Ability to capture long-term dependencies.	High computational cost.
GRU-based methods	[65,66,67]	Compared with LSTM, the calculation efficiency is higher.	Reduced ability to handle long time series compared to LSTM.
Transformer-based methods	[68,69,70,71]	Parallel computation.	Complex training process with high computational resource consumption.
Hybrid models methods	[72,73,74]	Ability to leverage the strengths of different models.	Increases model complexity and computational cost.

Currently, SOH estimation methods based on hybrid models are a key research focus. Effective model fusion design can enhance both the accuracy and robustness of the resulting model. The core idea behind this approach is to combine the strengths of different models for improved performance.

4. Challenges and Future Directions

The application of deep learning to the of SOH estimation of lithium-ion batteries in EVs has shown promising results in terms of both accuracy and performance. However, due to the intrinsic characteristics of deep learning technologies and the evolving nature of lithium-ion batteries, several challenges remain in applying deep learning to SOH estimation. This section aims to explore these challenges and outline potential future developments in this field. As detailed in Figure 5, the main challenges of deep learning-based SOH estimation methods are as follows:

• Generalization ability.
• Computational load.
• Battery pack limitations.

4.1. Generalization Ability

A substantial body of research has demonstrated that lithium-ion batteries exhibit inherent inconsistencies, manifesting as varying aging patterns across different batteries or even within cells of the same battery type [94]. This variability arises from multiple factors; for instance, production inconsistencies during manufacturing can lead to differences in battery performance, while the long lifecycle of battery systems amplifies these inconsistencies through cumulative effects, resulting in complex aging behaviors under real-world conditions. Moreover, factors such as diverse operating environments and different charging–discharging modes further contribute to the diversity of aging patterns in lithium-ion batteries. These inconsistencies undermine the generalization ability of SOH estimation models, necessitating the development of new models for each new battery type or operational condition, which in turn incurs higher training costs and requires greater resource investment. This situation presents a significant obstacle to the widespread application of deep learning methods for SOH estimation [95].

To enhance the generalization capability of these models, one straightforward approach is to increase the number of model parameters. Studies have indicated that a larger number of parameters improves the model’s learning capacity, facilitating the discovery of underlying commonalities across different aging modes of batteries [96]. However, a more promising direction for research lies in the exploration of more efficient model architectures. For instance, incremental learning represents an attractive approach. Incremental learning allows the model to continually adapt and learn from new data as new battery are obtained, avoiding the need to retrain the entire model [97]. This approach enables the model to respond to the dynamic and evolving conditions encountered during actual operation. Additionally, transfer learning offers a powerful alternative by enabling the transfer of knowledge (i.e., model parameters) from an existing model to a new one. This helps to expedite the training of the new model and avoids the need to restart training from scratch. Together, the paradigms of incremental learning and transfer learning promise to significantly enhance the efficiency of model training, reduce the time required to deploy deep learning models, and improve the overall generalization ability of SOH estimation techniques [98,99].

4.2. Computational Load

As deep learning models continue to scale, the number of parameters in these models increases, leading to higher demands on computational resource; however, due to the limited computing power of onboard chips in most EVs systems, deploying these deep learning models directly in real vehicles becomes challenging [100]. Additionally, the long inference times required by these models pose a significant obstacle to real-time SOH estimation. To address these issues, advanced computing architectures provide a promising solution. Studies have suggested leveraging cloud computing to manage the state estimation of lithium-ion batteries in EVs. This approach involves offloading the complex model inference process to cloud servers with greater computational capabilities and using advanced communication technologies to relay the results back to the vehicle, which can alleviate the computational burden on EVs’ onboard systems [101]. Building on this, the integration of edge computing alongside cutting-edge sensor and communication technologies represents a potential future direction for the development of efficient real-time SOH estimation systems for EVs [102].

4.3. Battery Pack Limitation

Although considerable research has been conducted on using deep learning for estimating the SOH of lithium-ion batteries, most studies have focused on individual battery cells. In contrast, EV systems utilize power battery packs composed of numerous battery cells, which makes the SOH estimation task more complex. SOH estimation models designed for individual battery cells are often ineffective when applied to battery packs. Additionally, the interactions between the different cells within a battery pack complicate estimation of the pack’s overall SOH [103]. Consequently, advancing research on SOH estimation for battery packs will be a critical direction for future development. While much of the existing research has focused on applying machine learning to battery cell SOH estimation, future efforts will likely involve transferring the knowledge gained from cell-based models to battery pack SOH estimation. Ultimately, this approach aims to enable accurate SOH estimation for real-world batteries used for vehicle power under actual operating conditions [104].

5. Conclusions

With the rapid development of EVs, estimating the SOH of lithium-ion batteries has become a significant research focus. Accurate SOH estimation is fundamental to ensuring the safety and efficiency of EV systems. While traditional machine learning algorithms such as artificial neural networks and support vector machines have demonstrated promising results, the complex operating conditions, varied operating modes, temperature fluctuations, and inherent heterogeneity of lithium-ion batteries pose substantial challenges for accurate SOH estimation.

The integration of deep learning techniques into SOH estimation has led to a growing body of research, demonstrating significant advantages over traditional machine learning methods. The primary contribution of this study is to summarize the latest advancements in deep learning techniques for estimating the SOH of lithium-ion batteries in EVs, offering a comprehensive discussion and comparison of various studies and highlighting key methodologies and developments in this field.

As the adoption of lithium-ion batteries continues to rise, accurate SOH estimation remains crucial to ensuring the reliability and robustness of EVs and other energy storage systems. Concurrently, the evolution of deep learning has introduced new algorithms and models with enhanced nonlinear representation capabilities. Future research on the application of deep learning to SOH estimation will likely focus on deploying models with larger parameter spaces, improving model generalization through innovative architectures as well as on designing more effective application frameworks to address the limitations of current technologies. These advancements will enable faster and more accurate SOH estimation, thereby ensuring the safety and performance of EVs.