State-of-Health Prediction of Lithium-Ion Batteries Based on Diffusion Model with Transfer Learning

Abstract

An accurate state-of-health (SOH) prediction of lithium-ion batteries (LIBs) is crucial to their safe and reliable. Although recently the data-driven methods have drawn great attention, owe to its efficient deep learning, it is worthwhile to continue devoting many efforts to prediction performance. In practice, fast charging mode has been widely applied in battery replenishing, which poses challenges for SOH prediction due to the diversity of charging conditions and electrochemical properties of LIBs; although, the process is stable and detectable. Furthermore, most previous data-driven prediction methods based discriminative model cannot describe the whole picture of the problem though sample data, affecting robustness of model in real-life applications. In this study, it is presented a SOH prediction model based on diffusion model, as an efficient new family of deep generative model, with time series information tackled through Bi-LSTM and the features derived from the voltage profiles in multi-stage charging process, which can identify distribution characteristics of training data accurately. The model is further refined by means of transfer learning, by adding a featured transformation from the base model for SOH prediction of different type LIBs. Two different types of LIBs datasets are used to evaluate the proposed model and the verified results revealed its better performance than those of other methods, reducing efforts required to collect data cycles of new battery types with the generality and robustness.

1. Introduction

In recent years, considerable attentions have been paid to lithium-ion batteries (LIBs) as dominant energy storage devices and power sources for electric vehicles (EVs), which facilitate the saving of fossil energy and environmentally friendly transportation [1]. However, the battery capacity degradation, affected by some factors, such as temperature, current rate, and battery historical aging path, are raising concerns of LIBs safety and feasible application in EVs, which could cause underuse or overuse of battery pack. Particularly, when the EVs are driving under the random and complex operation instead of in the laboratory cycle, leading to resource waste or some potential disasters [2].

Hence, to achieve timely maintenance and prevent malfunction, accurate predicting the state-of-health (SOH) of LIBs can provide a reference for battery control strategy and protection mechanism, which has drastically altered and pushed forward the sustainable development of EVs [3]. SOH was first proposed to evaluate the aging degree of a battery relative to a new battery in real time and used as an important indicator defined as the ratio of the available capacity in current operation to the rated capacity offered by the manufacture. Due to the different emphasis of power capability and energy capability in hybrid EVs (HEV) and pure EVs (PEV), the resistance and capacity are used as the ratio of SOH for HEV and PEV, respectively [4]. This study focused on the capacity-based SOH prediction.

Many previous research have proposed numerous SOH prediction methods, which are usually divided into two categories, model-driven and data-driven methods. The model-driven methods principally include electrochemical model (EM) and equivalent circuit model (ECM). EM describes the reactions during charging and discharging by electrochemical equations, concentrating on intrinsic aging mechanism in accordance with electrochemical principles. Li et al. [5] analyzed the influence of 26 parameters on an electrochemical model under both the charging process and real-life driving cycles. Randall et al. [6] introduced the side reactions to model degradation mechanisms. However, it is difficult to describe the whole internal parameters of battery accurately by equations due to complicated reaction process. ECM transforms the battery to circuit combined by some electrical components, to simulate the charging and discharging behaviors of LIBs by mathematical equations [7]. Additionally, model parameters are identified through empirical assumptions and mathematical algorithms, such as extend Kalman filter (EKF) [8,9], recursive least squares (RLS) [10], and particle filter (PF) [11]. Nonetheless, it is also difficult to predict precisely due to the discrepancy between battery internal state and model hypothesis.

More recently, the data-driven methods have been spotlighted in SOH prediction of LIBs because of easy acquirement for data and the development of artificial intelligence (AI). The methods neither require the prior knowledge about battery mechanism nor make prediction with human subjective experience or intervention, linking the latent relationships between inputs and outputs. Even more impressively, successful application in natural language processing (NLP), computer vision (CV), and speech processing have enabled these techniques to accelerate progress in research of battery field.

Some advanced data-driven methods based on traditional machine learning (ML) have been applied in SOH prediction, including artificial neural network (ANN) [12], support vector machine (SVM) [13,14], automated machine learning (AutoML) [15], and Gaussian process regression (GPR) [16,17], to capture the complicated relationship between input features and SOH. Additionally, Deep learning (DL) has become research focus recently by virtue of its remarkable performance in large-scale data processing. Some complex neural network models evolved from ANN, such as convolutional neural network (CNN) [18], recurrent neural network (RNN) [19] with its variants the long short-term memory (LSTM) [20] and gated recurrent unit (GRU) network [21], directly approximate high-dimensional nonlinear functions according to batteries life characteristics and have achieved more efficient prediction in complicated working condition. Moreover, the outstanding time series model Transformer [22] has been developed to process the characteristics among time sequence in battery degradation.

In fact, most of the data-driven methods for SOH estimation and prediction of LIBs characterize the discriminative model, which do not require prior knowledge of samples, but directly establish the mapping relationship between the input and output, that is, LIBs cycles (or other health indexes) and capacity. Then, the model will be established after adjusting the weights of the network with training through loss function, which cannot reflect the characteristics of the training data itself and further on the LIBs whole working scenario, endowed better fitting only with the quality and scale of the training data.

The generative model has gained considerable attention for learning joint probability distributions P(X,Y) over complex data, especially with the emergence of ChatGPT [23]. Unlike the discriminative model, its advantage is to model the observed sample data with probability density functions, add prior probabilities P(X), then obtain the posterior probability P(Y|X) by Bayesian formula, which is P(Y|X) = P(X,Y)/P(X) [24]. This modelling idea can identify distribution characteristics of training data accurately and better describe the whole picture of the problem though samples.

The extraction of battery life characteristics can be affected by several time series points involving high noise, which are generated as the asynchronism of message transmission, interference of signal sensor and cumulative error of battery management system (BMS) calculation. By generative model, it can avoid causing a serious deviation for the whole distribution of training data, to achieve robust modeling instead of isolated feature representations [25]. More importantly, as the development of BMS chips and cloud data storage have been breaking new ground in terms of sampling speed and data fidelity, battery operation process can be stored in a detailed large-scale datasets, and SOH prediction is more appropriate to generative model, which means that the prior probabilities P(X) can describe clearly the true distribution of the actual battery degradation process, so as to have a faster learned network learning convergence to the real model with increased sample size [26].

Deep generative model is a class of generative models that train deep neural networks to model the distribution of training samples, and it has noteworthy application in time series field, including time series anomaly detection [27], time series imputation [28], and time series forecasting [29]. Recently, there are several studies in which the deep generative model is used to estimate and predict the battery state, and the published studies based on the deep generative model show that both generative adversarial network (GAN) and variational autoencoders (VAE) with satisfactory battery state data generation results. Kim et al. [30] proposed a fully unsupervised methodology for LIBs capacity estimation, in which latent variables are extracted from electrochemical impedance spectroscopy (EIS) data by using information maximizing GAN. He et al. [31] pointed out a new algorithm named Dynamic-VAE to retrieve partial dimensions of the raw battery-charging data, detecting abnormal data points in the sequential information, so as to identify battery failure accurately. Ardeshiri et al. [32] developed a novel prognostic architecture based on a least-squares GAN with GRU for LIBs cycle life prediction as the generator and multi-layer perceptron as the discriminator, to obtain high prediction accuracy from time-domain features.

In recent years, following the emergence of GAN and VAE, diffusion model as an efficient new family of deep generative model has emerged in many fields with record-breaking performance, including image synthesis and video generation [33]. The diffusion model aims to transform the prior data distribution into random noise before revising the transformations step by step to rebuild a brand-new sample with the same distribution as the prior probabilities [34]. However, as far as it is known, this study is the first attempt to develop SOH prediction approach based on diffusion model with transfer learning, to accomplish an obvious enhancement in SOH prediction of different LIBs. The main contributions of this study are summarized as follows:

• The fast-charging mode, consisting of multiple stages matching the maximum charging rate under limited conditions, is widely popularized in real-life EV application. The statistical features are extracted from the voltage profile of each stage as the model latent variables. The results prove that the selected features can distinctly characterize the battery degradation trend in the multi-stage charging process.
• Based on diffusion model, the prediction model is established with time series information tackled via BiLSTM, a versatile multivariate probabilistic time series forecasting method that leverages estimating its gradient to learn and sample from the data distribution at each time step, autoregressively. The transfer learning method is applied on different batteries, enabling the generalizability of proposed approach.
• The proposed method is verified on two LIBs datasets with different electrochemical properties and charging conditions, and the verification results reveal the higher accuracy compared with other methods, accomplishing an obvious performance improvement in generalizability and robustness.

The remainder of this paper is organized as: Section 2 introduces the methodology of diffusion model applied to SOH prediction and transfer learning. In Section 3, the feature extraction is proposed and the relationship to battery capacity is presented. Section 4 evaluated the experimental results of the proposed method for source and target datasets, and its valid comparison with other methods. Finally, major conclusions drawn from this study and outlines for future research scope are given in Section 5.

2. Methodology

2.1. Diffusion Model Applied to SOH Prediction

Diffusion model, short for the diffusion probabilistic model, emerges as the new advanced class of deep generative models that gradually convert data into noise, then generate new samples by learning the noise removed process, through the forward process and reverse process, respectively [35]. Figure 1 presents the intuition of diffusion model.

The denoising diffusion probabilistic model (DDPM) [34] applied in this study implements the operation via two Markov chains: the forward chain is designed with the goal to progressively add Gaussian noise to the data until it is transformed into a simple prior Gaussian distribution. The reverse chain aims at eliminating noise from data via a deep neural network, which can parameterize the transition by learning the reverse process [36].

In the forward process, the calculation of approximate posterior q(x_1:N|x₀) is fixed instead of trainable, as Equations (1) and (2),

$$q\left(x_{1:N}\left|x_0\right.\right)={\prod }_{n=1}^{N}q\left(x_n\left|x_{n-1}\right.\right)$$

(1)

$$q\left(x_n\left|x_{n-1}\right.\right)~N\left(x_n;\sqrt{1-{\beta }_n}{x}_{n-1},{\beta }_{n}I\right)$$

(2)

where β_t ∈ (0,1) represents the gradually added Gaussian noise to the signal according to a sequence of variance scales. N denotes the total number of samples, and I represents the unit matrix. The joint distribution p_θ(x_n−1|x_n) signified the reverse process, defined as a Markov chain with trainable instead of fixed Gaussian transitions starting with p(x_n) as Equations (3) and (4).

$$p_{\theta }\left(x_{0:N}\right)=p\left(x_N\right){\prod }_{n=N}^N{p}_{\theta }\left(x_{n-1}\left|x_n\right.\right)$$

(3)

$$p_{\theta }\left(x_{n-1}\left|x_n\right.\right)~N\left(x_{n-1},{\mu }_{\theta }\left(x_n,n\right),{\Sigma }_{\theta }\left(x_n,n\right)I\right)$$

(4)

where the mean μ_θ, and variance Σ_θ are parameterized by deep neural networks with shared parameters θ.

The goal of reverse Markov chain, that is the calculation of p_θ(x_n−1|x_n), is to eliminate the Gaussian noise added in the forward process. In the reverse process, x₀ can be sampled by a noise vector p(x_n), and iteratively sample from the learnable transition p_θ(x_n−1|x_n) until n = 1. To sample accurately, the reverse Markov chain is trained to match the forward Markov chain; thus, parameter θ needed to be adjusted in order that the posteriori distribution p_θ(x_n−1|x_n) of the reverse Markov chain closely approximates posterior distribution q(x_n−1|x_n,x₀) of the forward process given x₀.

Kullback–Leibler (KL) Divergence is the index used to measure the similarity of probability distribution, which can denote the resemblance between the approximate estimated probability distribution and the true probability distribution of the data as a whole. The smaller the divergence value means the two probability distributions are closer to each other, and it can be applied as an objective function to find the optimal value for any approximate distribution. The match of two Markov chains can be achieved by minimizing the Kullback–Leibler (KL) divergence between p_θ(x_n−1|x_n) and q(x_n−1|x_n,x₀) [34]. Furthermore, the KL-divergence between these two is transferred by minimizing the negative log-likelihood using Jensen’s inequality as Equation (5),

$$D_{KL}\left(q\left(x_{n-1}\left|x_n,x_0\right.\right)\Vert p_{\theta }\left(x_{n-1}\left|x_n\right.\right)\right)=E_q\left[\frac{1}{2{\Sigma }_{\theta }}\Vert {\tilde{\mu }}_n\left(x_n,x_0\right)-{\mu }_{\theta }{(x_n,n)\Vert }^2\right]+C$$

(5)

where C is a constant and irrelevant to θ, and ${\tilde{\mu }}_n$ is mean of q(x_n−1|x_n,x₀). The objective can be simplified in Equation (6) by introducing a new noise network ε_θ, instead of KL divergence as training objective directly.

$$E_{x_0~q\left(x_0\right),\epsilon ~N\left(0,I\right)}\left[\lambda \left(n\right)\Vert \epsilon -{\epsilon }_{\theta }{\left(\sqrt{{\overline{\alpha }}_n}{x}_0+\sqrt{1-{\overline{\alpha }}_n}\epsilon ,n\right)\Vert }^2\right]$$

(6)

where λ(n) is a positive weighting function, and ε_θ is a deep neural network with parameter θ that predicts the noise vector ε given x₀.

In this study, a multivariable time series prediction method is pointed out. It combines diffusion model and autoregressive model, in which the multivariable means the multi-dimensional feature representing battery degradation trend. The conditional probability model of future time steps for a multivariable time series can be learned through the given value and covariate of the previous moment in the series, and further achieve the LIBs lifespan sequence prediction, and the conditional probability model is as Equation (7):

$$q_{\chi }\left(x_{t_0:T}^0\left|x_{1:t_0-1}^0,c_{1:T}\right.\right)={\Pi }_{t=t_0}^T{q}_{\chi }\left(x_t^0\left|x_{1:t-1}^0,c_{1:T}\right.\right)$$

(7)

where covariate c_i:T denotes characteristic of LIBs lifespan sequence, assumed to be known for all the time points, and $x_{i:t}^0\in {\mathbb{R}}^D$ is the multivariable time series data, which can be learned by the conditional diffusion model introduced above, where i ∈ {1,…,D}, and t means time step. The task is to predict the conditional probability distribution, calculated by diffusion model, of the whole time series based on the data sampled from the training interval, so as to accomplish the prediction of the future data in an autoregressive way.

BiLSTM [37] is a variant model composed of two LSTMs with different data transmission directions. With this structure of network, the important features and correlation of time sequence information before and after the current time point can be obtained in depth. BiLSTM is used as an appropriate autoregressive model to construct a time-dependent network as Equation (8), and it utilizes bidirectional time series information from the past and future derived from current output, mining the data connection of the whole time series deeply, improving utilization of the characteristics of long-term dependence in time series.

$$h_t=BiLST{M}_{\theta }\left(x_t^0,c_t,h_{t-1}\right)$$

(8)

where θ is the shared network weight with the conditional diffusion model, and h₀ = 0. Therefore, the expression of conditional probability model can be further simplified in the form of a conditional diffusion model as Equation (9),

$$q_{\chi }\left(x_{t_0:T}^0\left|x_{1:t_0-1}^0,c_{1:T}\right.\right)={\Pi }_{t=t_0}^T{p}_{\theta }\left(x_t^0\left|h_{t-1}\right.\right)$$

(9)

As depicted in Figure 2, both of x_t−1 and c_t−1 are used to generate h_t−1 after injected into BiLSTM as input, and x_t can be obtained by the conditional diffusion model with updated h_t−1 as condition. Therefore, a conditional probability model is obtained, which can predict the x_t with encoded h_t, given x_t−1 and c_t−1.

Combined with the previous introduction to DDPM, the fixed forward process of this model is still from x₀ to x_N, but the hidden status h_t−1 of each time step is added when building the noise network during the learned reverse process. The training process is performed in the pattern by randomly sampling training data for the acceptable robustness of the model, and the negative log-likelihood of the model with parameter θ is chosen as the optimized loss function as Equation (10), with the initially hidden state h_t−1 obtained by BiLSTM in prediction interval.

$${{\displaystyle \sum }}_{t=t_0}^T-\log p_{\theta }\left(x_t^0\left|h_{t-1}\right.\right)$$

(10)

Hence, a more optimized and detailed objective than Equation (6) can be obtained as Equation (11):

$$E_{x_0~q\left(x_0\right),\epsilon ~N\left(0,I\right)}\left[\Vert \epsilon -{\epsilon }_{\theta }{\left(\sqrt{{\overline{\alpha }}_n}{x}_t^0+\sqrt{1-{\overline{\alpha }}_n}\epsilon ,h_{t-1},n\right)\Vert }^2\right]$$

(11)

where ε_θ is a neural network conditioned on the hidden state h_t.

2.2. Transfer Learning

Most machine learning methods normally are applied under assumed preconditions that training data (source data) and validation data (target data) conform to independent and identical distribution. In practice, however, the distributions of source data and target data are usually diverse, and the inadequate label data can hinder model performance in task data.

Transfer learning, applied to solve this problem in this study, can obtain the knowledge from other tasks and transfer them to new tasks, which can expedite the utilization of models among different datasets [38]. It is not necessary that source data and target data are independent and identical distribution with transfer learning, so the learned model (source task) features and parameters can be transferred to a new analogous model (target task), to relieve the dependence on label data and to accelerate model convergence. The target task need not be trained completely due to the knowledge transferred from source task, which can improve the modeling efficiency and expand the scope of practical application.

For SOH prediction of different types of LIBs in diverse working conditions and compensating multiple data losses of prediction models and improving model generalization ability, a flexible estimation scheme integrating diffusion model and transfer learning was proposed. It comprises two steps: the first is that target networks pre-training through source dataset to obtain base model, and the second is that transfer learning with fine-tuning strategy to adjust the parameters of middle layer in target networks with other layers unchanged. As shown in Figure 3, in the source task, the diffusion model is finally established through training data and verification of test data. The model parameters and the weights, which represent the internal degradation mode of the battery in the source task, will be locked, retained, and transferred to the target task for an easier SOH prediction. Then, the full connect (FC) layer make a critical difference in fine-tuning SOH prediction for specific batteries or conditions, the weights of which are re-trained based on target battery. Finally, the model network of target task is established efficiently, through the knowledge transferred from source task with small data volumes.

3. Experiment Validation

3.1. Experimental Dataset

Generally, the battery capacity is obtained by a standard full discharge-charge process. For the EV driving scenario in real-life, however, the batteries are discharged under erratic conditions instead of controllable as in the laboratory, which depends on the user behavior with uncertainties in environmental and unpredictable load demand during driving. Conversely, the battery charge is often accomplished in the time of the EVs parking with a stable and detectable grid connection. Therefore, due to the regular charging strategy, a charge process rather than a discharge process is preferred for on-board battery health monitoring.

Fast charging in direct current (DC) mode, widely used in the charging of electric vehicles, have attracted extensive interests because of its efficiency in energy replenishing and quick relief for range anxiety. In order to ensure charging safety and battery life, the BMS can calculate the charging current rate in different stage under the maximum charging power limit, based on the real-time temperature and state-of-charge (SOC) of the batteries, and then transmit the results to charging grid to control the output current. Therefore, in fast charging mode, the current rate and constant current (CC) charging time are quite different due to the discrepancy of temperature and SOC among batteries, which can affect the consistency of data characteristics and demand higher requirements for generalization ability of prediction model.

In this study, for simulating the diversity of multi-stage CC charging conditions, the two different types of LIBs datasets with different electrochemical properties and charging conditions are considered. The first is the lithium iron phosphate (LFP) battery dataset from TOYOTA Research Institute [39], named Dataset 1 as source dataset; and the second is the nickel cobalt manganese (NCM) battery dataset from laboratory, named Dataset 2 as target dataset. All batteries underwent a multi-stage charge with three current rate profile and a constant current discharging profile. The specifications of these two datasets are tabulated in

Table 1. Summary of two LIBs datasets.

Spec.	Dataset 1	Dataset 2
Rated capacity (Ah)	1.1	73
Number of cells	4	2
Charging current (A)	2.64-1.43-1.1	29.2-21.9-3.65
Discharging current (A)	4.4	73
Max cut-off voltage (V)	3.6	4.3
Min cut-off voltage (V)	2.0	2.7
Number of cycles (Bn mm denotes the battery code)	B1-01:1062 B1-02:1266 B1-03:1114 B1-04:1047	B2-01:373 B2-02:373

, and the experimental steps of a cycle are shown in Figure 4, and the battery capacity degradation curves are described in Figure 5.

Figure 6 illustrates the voltage variation trends of B1-01, as an example, in the three charging CC stages with increasing cycle number. The voltage curves move leftward, and shape turns to be gentle, which shows that the charging capacity is decreasing, and internal resistance is increasing. Hence, the degradation characteristic can be identified from a slight shift in curve shape of CC stage.

3.2. Feature Extraction

Summarized statistics have been applied to effectively quantify the shape and position characteristics of 2D curve, such as the voltage curve [39] and the voltage relaxation curve [40]. As mentioned above, it is suitable for feature extraction in CC charging stages, due to its high correlation with battery degradation and easy access in real-life driving cycles. The voltage curves of three CC stages are converted to 18 statistical features in three group in total, that is, variance (Var), skewness (Ske), maximum (Max), minimum (Min), mean (Mean), and excess kurtosis (Kur) in every group, and the mathematical description of the six features refers to the calculations in [40].

The Var represents the distribution of the voltage value in one CC stage and decrease in it indicates that the voltages show a concentrated distribution as battery degradation, and vice versa. The Ske means the relationship between sampled voltage data and average voltage, where voltage rises more rapidly with more Ske, consistent with the shape of the voltage curve. Likewise, the lower Kur denotes the gentler distribution of the voltage comparing with a normal distribution. Both Ske and Kur can characterize the shape of the voltage curve. The relationship between battery capacity and the corresponding features extracted from CC1 stage of Dataset 1 is presented in Figure 7, and it is difficult to describe the relationships even only by some certain functions, although the Max, Min and Mean just present an obvious monotonous correlation between capacity and voltage throughout the battery life, respectively. To avoid missing correlative features, which have different correlations in different batteries, all of these 18 features are adopted as model input without feature reduction.

3.3. Implementation Details

The leave-one-out evaluation [7] is used in this experiment settings, instead of the data used for training and validation is acquired from one battery, which is more practical in online prediction for entire life cycle of LIBs without the dependence on other cycle of the same battery. The proposed model can be evaluated in source task as: three batteries are used as training set and another battery is validation set. Four experiments in total are performed, through which the model network and evaluating indexes are determined for source task.

The framework of proposed approach is depicted in Figure 8, combined with off-line modelling and online modelling. In off-line process, the model is established by source dataset, Dataset 1, and validated through leave-one-out evaluation. To promote the generalization in target task, the trained network is transferred and fine-tuned by task dataset, Dataset 2. In the target task, the network layers are frozen so that weights cannot be modified further, and only the FC layer is trained by the data from Dataset 2. It is noteworthy that transfer learning can be applied based on the cycle data reaching 80% SOH at least, to achieve the rapid establishment of prediction model in target task. In online process, the extracted features, from raw data of different LIBs and charging conditions, are directly applied to predict the battery SOH based on the proposed model with transfer learning.

4. Results and Discussion

4.1. Evaluation Criteria

In this study, as the performance of the proposed method is evaluated, the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) indexes are calculated as follows:

$$\mathrm{RMSE}=\sqrt{\frac{{{\displaystyle \sum }}_{i=1}^n{(C_{real}\left(n\right)-C_{prd}\left(n\right))}^2}{n}}$$

(12)

$$\mathrm{MAE}=\frac{1}{n}{{\displaystyle \sum }}_{i=1}^n\left|C_{real}\left(n\right)-C_{prd}\left(n\right)\right|$$

(13)

$$\mathrm{MAPE}=\frac{1}{n}{{\displaystyle \sum }}_{i=1}^n\frac{\left|C_{real}\left(n\right)-C_{prd}\left(n\right)\right|}{C_{real}\left(n\right)}\ast 100\%$$

(14)

where n is the number of cycles, C_real is the real capacity, and C_prd is the predicted capacity. The results of the evaluation may not be comparable due to the capacity range of different batteries varies greatly, and normalizing can remove the influence caused by different capacity range. The min-max normalization is adopted to map the feature S into uniform interval as follows:

$$x^{\prime }=a+\frac{\left(x-\min \left(S\right)\right)\left(b-a\right)}{\max \left(S\right)-\min \left(S\right)}$$

(15)

where x is true value and x′ is mapped value of interval [a, b].

4.2. Performance on Source Dataset

The 18 features are extracted from raw voltage curve of three CC stages, as model inputs. B1-02, B1-03, and B1-04 are used as training set, and the results of B1-01, as a validation set, can be predicted by proposed model. According to this grouping, the results of other three batteries can be validated, respectively. The capacity predictions and relative errors of B1-01, B1-02, B1-03, and B1-04 are depicted in Figure 9. The red solid line denotes predicted capacity, and blue solid line is real capacity. The relative errors determined by the two capacity values are showed less than 4.5% in validation set, which indicates the predictions in SOH trends with high accuracy and robustness for the same category of LIBs.

The RMSE, MAE, and MAPE value are listed in

Table 2. Prediction errors of four batteries in Dataset 1.

Index	B1-01	B1-02	B1-03	B1-04
RMSE	0.0049	0.0187	0.0034	0.0166
MAE	0.0044	0.0183	0.0022	0.0164
MAPE	0.43%	1.81%	0.23%	1.55%

. Although the lowest prediction accuracy among the four batteries is B1-02, the RMSE, MAE, and MAPE of it are 0.0187, 0.0183, and 1.81%, respectively, the trained model can operate reliably without initial cycles provided and with more practical significance.

To verify the feasibility of the proposed approach, the prediction errors of approaches reported in previous studies were also listed in

Table 3. Comparison of prediction errors with other ML methods.

Index	RNN	LSTM	GRU	Transformer	CNN-Transformer	Proposed Approach
RMSE	0.0253	0.0167	0.0121	0.0364	0.0227	0.0109
MAE	0.0271	0.0134	0.0115	0.0278	0.0218	0.0103
MAPE	2.94%	1.32%	1.07%	3.03%	2.11%	1.01%

, and the same training and validation for Dataset 1 are conducted with RNN, LSTM, GRU, Transformer [22], and CNN-Transformer [41]. It is obvious that the proposed approach achieves a better performance than other methods.

4.3. Performance on Target Dataset

The performance was validated for different battery types in different charging conditions by applying transfer learning to the proposed model trained by the source dataset and re-training using a single battery from the target dataset. Figure 10 depicts the capacity predictions and relative errors of B2-01 and B2-02, illustrating that transfer learning also achieves the high performance of models in SOH prediction of different LIBs.

Furthermore, to confirm the effectiveness of transfer learning, the results obtained by using the proposed model with transfer learning were compared to the model without transfer learning; that is, the model trained by the source dataset is applied directly to the target dataset without transfer learning. These results which are normalized as the capacity range of Dataset 2 in

Table 4. Comparison of the predicted results in two cases of Dataset 2.

Case	Index	B2-01	B2-02
With transfer learning	RMSE	0.0031	0.0028
	MAE	0.0027	0.0023
	MAPE	0.28%	0.24%
Without transfer learning	RMSE	0.0411	0.0437
	MAE	0.0387	0.0388
	MAPE	4.03%	4.05%

manifest that applying transfer learning promotes the proposed model generalization in different situations.

5. Conclusions

This study, as far as is known, is the first to propose the diffusion model applied in the SOH prediction of LIBs and enhanced the SOH prediction with transfer learning. First, for multiple stages of battery charging, the statistical features of voltage profile, extracted as latent variables, are employed to characterize the LIBs degradation. Meanwhile, it is proposed a prediction method based on diffusion model where time series information is accumulated and transmitted through BiLSTM. It can relieve the influence of data noise by learning the probability distribution of training data. Additionally, transfer learning can implement the rapid convergence of proposed model in new battery type and new condition. Then, the experiments on two LIBs cycle life datasets based on different electrochemical properties and charging conditions verify the accuracy and robustness. It is confirmed that the diffusion model with transfer learning could effectively reduce the dependency of label data in target task and build corresponding networks based on less data and faster training. For charging safety, fast charging will be divided into multiple stages under different charging current according to the current state or type of LIBs. The proposed approach is verified generalization ability and prediction accuracy in the multi-stage charging process for different type of LIBs under different condition, which is more practical in fast charging condition. As a next step, it deserves to proceed with further studies on other deep generative model applied SOH prediction of LIBs and to investigate the effective feature optimization techniques in multi-stage charging.