Method for predicting early life of lithium ion battery under different aging conditions

Abstract

The invention relates to a method for predicting the early life of a lithium ion battery under different aging conditions, and the method comprises the steps: obtaining the cycle data of the battery under different aging conditions, carrying out the normalization capacity index interpolation resampling and signal normalization processing, and constructing the difference characteristics between cycles and the difference characteristics between the batteries; extracting and fusing multi-scale degradation information through two Kolmogorov-Arnold Networks (KANs) with the same structure; and finally, outputting a prediction result through a shared parameter linear layer, and carrying out weighted average to obtain a final life value. According to the method, high-precision prediction can be achieved only through early data of first 100 cycles, the method is suitable for different cycle temperatures, charging and discharging protocols and material systems, the service life test period is remarkably shortened, the research and development cost is reduced, and a reliable service life evaluation scheme is provided for a battery management system.

Description

Technical Field

[0001] This invention relates to a method for predicting the early lifespan of lithium-ion batteries under different aging conditions. It can be widely applied to the lifespan assessment of lithium-ion batteries used in new energy vehicle power batteries, energy storage systems, and portable electronic devices, and belongs to the field of lithium-ion battery lifespan prediction technology. Background Technology

[0002] Lithium-ion batteries, with their high energy density, long cycle life, and excellent power performance, have become the core power source for new energy vehicles, energy storage systems, and portable electronic devices. However, in actual use, the performance of lithium-ion batteries gradually declines with the number of cycles and usage time; capacity, internal resistance, and energy efficiency all degrade, eventually leading to the end of their lifespan.

[0003] Factors influencing the lifespan degradation of lithium-ion batteries are extremely complex, including ambient temperature, charge / discharge protocols, formation protocols, and electrode material systems. Different aging conditions lead to significant differences in battery degradation mechanisms and rates, resulting in highly uncertain and nonlinear lifespan predictions, posing a significant challenge to accurate forecasting.

[0004] Existing lithium-ion battery life prediction methods are mainly divided into three categories: First, empirical models, which rely on fitting a large amount of experimental data, such as the Arrhenius model, but have poor generalization ability and are difficult to adapt to various aging conditions; second, equivalent circuit models, such as the Thevenin model, which can reflect the physical meaning of electrochemical processes, but have complex parameter identification and cannot adapt to diverse aging conditions; and third, data-driven models, such as LSTM and CNN, which have advantages in capturing complex nonlinear relationships, but their performance is heavily dependent on a large amount of labeled full-life aging data.

[0005] Long-cycle aging experiments are costly and time-consuming, making it difficult to obtain sufficient degradation information in the early stages of battery life when capacity decay is not yet significant, thus limiting the practicality of existing models. Furthermore, the data distribution varies significantly under different aging conditions, and directly using this data for model training leads to a decline in predictive performance. Currently, the industry lacks a lifetime prediction model that can generalize across operating conditions and samples using limited cycle data in the early stages when battery capacity has not significantly decayed. There is an urgent need for a universal prediction method that takes into account the differences in aging conditions and the differences between batteries to overcome the technical bottleneck of high-precision, strong generalization ability prediction in the early stages. Summary of the Invention

[0006] To overcome the shortcomings of existing lithium-ion battery life prediction methods, such as insufficient generalization ability under various aging conditions, strong dependence on a large amount of full life data, and difficulty in achieving accurate early life prediction, this invention provides a method for predicting the early life of lithium-ion batteries under different aging conditions. By constructing inter-cycle difference features and inter-battery difference features, and introducing KANs for modeling, this method achieves high-precision prediction of the life of batteries in the early stages of life, thereby significantly reducing life test time and cost, and improving the versatility and scalability of battery life modeling.

[0007] A method for predicting the early lifespan of lithium-ion batteries under different aging conditions includes the following steps:

[0008] Step 1: Obtain current, voltage, capacity, and temperature cycling data of lithium-ion batteries under different charging protocols, discharging protocols, and cathode material systems;

[0009] Step 2: Since the test duration varies for different batteries and cycle numbers, normalized capacity Q∈[0,1] is used as a unified index variable. The capacity-current, capacity-voltage, and capacity-temperature curves are interpolated and resampled to obtain a charging capacity-current I with dimension 1×1000. c (Q), Discharge Capacity - Current I d (Q), Charging capacity - voltage V c (Q), Discharge Capacity-Voltage V d (Q), Charging capacity - Temperature T c (Q), discharge capacity-temperature T d (Q), the characteristic of the charge / discharge voltage difference ΔV(Q)=V c (Q)-V d (Q), equivalent internal resistance characteristic R(Q)=(V c (Q)-V d (Q)) / (I c (Q)-I d (Q));

[0010] Step 3: Normalize all feature parameters obtained in Step 2 to the interval [0,1] to obtain the feature vector X = [I c (Q),I d (Q),V c (Q),V d (Q),T c (Q),T d [(Q),ΔV(Q),R(Q)]∈R 8×1000 Achieve cross-cycle and cross-battery data alignment and standardization.

[0011] Step 4: Calculate the inter-cycle differential characteristic X tar,d =X tar,[1:100] -Xtar,10 , where X tar,[1:100] X is the feature vector of the target battery in the first 100 cycles. tar,10 The feature vector of the target battery in the 10th cycle; calculate the inter-battery difference feature X. ref,d =X tar,[1:100] -X ref,[1:100] , where X ref,[1:100] The feature vectors of the first 100 cycles of a reference battery with a known lifespan are used to capture the differences in degradation patterns between different operating conditions, thereby improving the model's ability to generalize lifespan predictions under different aging conditions.

[0012] Step 5: Extract inter-cycle differential features X using two structurally identical Kolmogorov-Arnold Networks (KANs). tar,f =KAN(X) tar,d ) and inter-cell differences X ref,f =KAN(X) ref,d The Kolmogorov-Arnold Networks uses a B-spline function based on gridded basis functions as the activation function to achieve deep fusion of multi-scale degradation information. Through function decomposition and adaptive nonlinear mapping mechanism, the KAN network can express complex nonlinear relationships with fewer parameters, significantly reducing the battery requirements of the training model and reducing the time and cost of lifetime testing.

[0013] Step 6: Output the predicted lifetime value of the prior target battery through two linear layers with shared parameters. Lifetime difference between the target cell and the reference cell The weighted average yields the predicted final lifespan of the target battery. Where α∈[0,1], and N is the number of reference batteries;

[0014] Step 7: Select M reference batteries with known lifetimes, construct target battery-reference battery pairs to generate training samples, and apply the loss function. The model parameters w, θ, and φ are updated using Adam as the optimizer, where λ is the balance coefficient. The target battery represents the difference in actual lifetime between itself and the j-th reference battery. The trained model is used to accurately predict lifetime under different aging conditions, with Adam as the optimizer and gradient backpropagation algorithm used to update the model's parameters. The trained model is then used to predict the lifetime of retired batteries.

[0015] Step 8: Input the first 100 cycles of data for the battery with unknown lifespan, execute steps 2 to 6, and output the remaining lifespan prediction results.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] This invention enables high-precision lifetime prediction in the early stages before battery capacity shows significant degradation, exhibiting strong versatility and generalization ability. By constructing inter-cycle and inter-cell differential features, it effectively utilizes limited early-cycle data, reducing reliance on full-lifetime test data. Employing a KAN structure, it efficiently models complex degradation patterns with fewer parameters, alleviating the problem of limited small-sample data. This invention operates stably under different cycling temperatures, charge / discharge protocols, and material systems, providing a universal and reliable prediction method for battery management system lifetime assessment, significantly shortening lifetime testing cycles and reducing the research and development costs of long-lifetime batteries.

[0018] Significantly improved generalization ability: By constructing inter-cycle difference features and inter-battery difference features, combined with the adaptive nonlinear mapping ability of KANs, the model can be adapted to different cycle temperatures, charge and discharge protocols, formation protocols and electrode material systems. It achieves high-precision prediction on 5 different feature datasets (MART, HUST, ISU_ILCC, MIX-1, MIX-2), verifying the strong generalization performance across operating conditions and samples.

[0019] Reduced data dependence and experimental costs: Prediction can be completed with only early data from the first 100 cycles of the battery, without the need for full-life aging data, which significantly shortens the life test cycle (by more than 60%), reduces experimental equipment and time costs, and accelerates the research and development process of long-life lithium batteries.

[0020] High early prediction accuracy: For scenarios where there is no obvious capacity decay in the early stage of battery life, weak degradation information is captured by multi-scale difference features, and the prediction error is significantly reduced. The average error of the MIX-1 dataset is only 57 cycles, and the average error of the MART dataset is 76 cycles, which meets the accuracy requirements of practical applications for early life assessment.

[0021] High model training efficiency: KANs optimize local parameters through the B-spline activation function, reducing the number of model parameters and the risk of overfitting. Combined with the target battery-reference battery pair sample expansion strategy, it can still train efficiently with small sample data, improving the convenience of model deployment. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1This is a flowchart of a method for predicting the early lifespan of lithium-ion batteries under different aging conditions according to the present invention.

[0024] Figure 2 This is a visualization of the life prediction results of a lithium-ion battery under different aging conditions (different fast charging protocols) based on the MART dataset.

[0025] Figure 3 This is a visualization of the life prediction results of a lithium-ion battery under different aging conditions (different discharge protocols) based on the HUST dataset, which is a visualization of the life prediction results of a lithium-ion battery under different aging conditions (different discharge protocols) according to the present invention.

[0026] Figure 4 This is a visualization of the life prediction results of a lithium-ion battery under different aging conditions (different charge and discharge protocols) based on the ISU_ILCC dataset, which is a visualization of the life prediction results of a lithium-ion battery under different aging conditions (different charge and discharge protocols) according to the present invention.

[0027] Figure 5 This is a visualization of the early life prediction results of lithium-ion batteries under different aging conditions (different formation protocols) according to the present invention on the MIX-1 dataset.

[0028] Figure 6 This is a visualization of the life prediction results of a lithium-ion battery early life prediction method under different aging conditions (different materials and temperature protocols) based on the MIX-2 dataset. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Reference Figure 1 This invention proposes a method for predicting the early lifespan of lithium-ion batteries under different aging conditions, comprising the following steps:

[0031] Step 1: Obtain current, voltage, capacity, and temperature cycling data of lithium-ion batteries under different charging protocols, discharging protocols, and cathode material systems;

[0032] Step 2: Since the test duration varies for different batteries and cycle numbers, normalized capacity Q∈[0,1] is used as a unified index variable. The capacity-current, capacity-voltage, and capacity-temperature curves are interpolated and resampled to obtain a charging capacity-current I with dimension 1×1000. c (Q), Discharge Capacity - Current I d(Q), Charging capacity - voltage V c (Q), Discharge Capacity-Voltage V d (Q), Charging capacity - Temperature T c (Q), discharge capacity-temperature T d (Q), the characteristic of the charge / discharge voltage difference ΔV(Q)=V c (Q)-V d (Q), equivalent internal resistance characteristic R(Q)=(V c (Q)-V d (Q)) / (I c (Q)-I d (Q));

[0033] Step 3: Normalize all the feature parameters obtained in Step 2 for all battery signals, mapping them to the [0,1] interval to eliminate differences caused by different battery systems and design parameters, resulting in X = [I c (Q),I d (Q),V c (Q),V d (Q),T c (Q),T d [(Q),ΔV(Q),R(Q)]∈R 8×1000 Achieve data alignment and standardization across cycles and batteries;

[0034] Step 4: Calculate the inter-cycle differential characteristic X tar,d =X tar,[1:100] -X tar,10 , where X tar,[1:100] X is the feature vector of the target battery in the first 100 cycles. tar,10 The feature vector of the target battery in the 10th cycle; calculate the inter-battery difference feature X. ref,d =X tar,[1:100] -X ref,[1:100] , where X ref,[1:100] The feature vectors of the first 100 cycles of a reference battery with a known lifespan are used to capture the differences in degradation patterns between different operating conditions, thereby improving the model's ability to generalize lifespan predictions under different aging conditions.

[0035] Specifically, the difference characteristics between the first 100 cycles and the 10th cycle of a target battery with an unknown lifetime are calculated as follows:

[0036] X tar,d =X tar,[1:100] -X tar,10

[0037] Where X tar,[1:100] X represents the characteristics of the target battery after 100 cycles. tar,10 X represents the characteristic of the tenth cycle.tar,d This represents the inter-cycle variation characteristics, characterizing the inter-cycle degradation trend of the same battery in the early aging stage; simultaneously, the variation characteristics between the target battery and a reference battery with a known lifespan are calculated in the first 100 cycles, as shown below.

[0038] X ref,d =X tar,[1:100] -X ref,[1:100]

[0039] Where X ref,[1:100] X represents the characteristics of the reference battery during its first 100 cycles. ref,d It represents the characteristic differences between the target battery and the reference battery, captures the differences in degradation patterns between different operating conditions, and thus improves the generalization ability of the model to predict lifetime under different aging conditions.

[0040] Step 5: Extract inter-recurrent differential features X using two Kolmogorov-Arnold Networks with identical structures. tar,f =KAN(X) tar,d ) and inter-cell differences X ref,f =KAN(X) ref,d The Kolmogorov-Arnold Networks uses a B-spline function based on gridded basis functions as the activation function to achieve deep fusion of multi-scale degradation information. The KAN network, through function decomposition and adaptive nonlinear mapping mechanism, can express complex nonlinear relationships with fewer parameters, significantly reducing the battery requirements of the training model and reducing the time and cost of lifetime testing.

[0041] Specifically, two KANs,h with identical structures are used. θ and h φ Extract inter-cycle difference features and inter-cell difference features X respectively. tar,f and X ref,f To achieve deep fusion of multi-scale degradation information, the KAN network uses function decomposition and adaptive nonlinear mapping mechanisms, as shown below:

[0042]

[0043]

[0044] Where l represents the index of the network layer, φ represents the activation function, and nl and nl+1 represent the number of nodes in the l-th and l+1-th layers, respectively. The proposed method uses a B-spline function based on gridded basis functions to represent different forms of activation functions, as shown below.

[0045]

[0046] Among them, B k (x) is a B-spline basis function, c l,i,j,k For the corresponding basis function B k The trainable weights of (x) are given by K, which represents the number of grid points. The B-spline function operates independently in different local regions, optimizing only the affected local parameters, thus effectively reducing the risk of model overfitting when lifetime prediction data is limited. Inter-cycle differential features and inter-cell differential latent features X tar,f and X ref,f It can be represented as:

[0047] X tar,f =KAN(X) tar,d ), X ref,f =KAN(X) ref,d )

[0048] Step 6: Output the predicted lifetime value of the prior target battery through two linear layers with shared parameters. Lifetime difference between the target cell and the reference cell The weighted average yields the predicted final lifespan of the target battery. Where α∈[0,1], and N is the number of reference batteries;

[0049] Specifically, the latent features output the predicted lifetime value of the prior target battery through two linear layers with shared parameters. Lifetime difference between the target cell and the reference cell Finally, by taking a weighted average of the two results, the final predicted lifetime of the target battery is obtained.

[0050] Step 7: Select M reference batteries with known lifetimes, construct target battery-reference battery pairs to generate training samples, and apply the loss function. The model parameters w, θ, and φ are updated using Adam as the optimizer, where λ is the balance coefficient. Let $i$ be the difference in actual lifetime between the $i$-th target battery and the $j$-th reference battery. The trained model is used to accurately predict lifetime under different aging conditions, with Adam as the optimizer and gradient backpropagation algorithm used to update the model parameters. The trained model is then used to predict the lifetime of retired batteries.

[0051] Specifically, M batteries with known lifespans are used as reference batteries. The model is trained using the following loss function, and the trained model is used to accurately predict the lifespan under different aging conditions. Adam is used as the optimizer, and the gradient backpropagation algorithm updates the model parameters. By constructing target battery-reference battery pairs, a large number of samples can be obtained for model training. Finally, the trained model is used to predict the lifespan of retired batteries.

[0052] Step 8: Input the first 100 cycles of data for the battery with unknown lifespan, execute steps 2 to 6, and output the remaining lifespan prediction results.

[0053] In the experiment, 1068 battery samples were obtained from 13 selected datasets and divided into five subsets based on their degradation dynamics. The battery cycling data used were all from publicly available databases, including cycling aging data under various battery chemistry systems, charge / discharge protocols, and temperature conditions. These datasets are widely used in academia and are publicly available.

[0054] First, the MATR dataset, containing 81 fast-charging protocols, was used to evaluate the adaptive capability of the proposed method under different fast-charging scenarios. Second, the HUST dataset, containing 77 discharge protocols, was used to verify the performance of the method under different discharge protocols. Furthermore, the ISU_ILCC dataset, containing 64 charge-discharge protocols, was used to evaluate the model's lifetime prediction capability under cross-charge-discharge protocol conditions. Further, the Stanford dataset, containing 62 formation protocols, and the MICH dataset, containing 2 formation protocols, were merged into MIX-1 to test the model's prediction performance under different formation protocols. Finally, 311 battery samples from the remaining 8 datasets were combined to form the MIX-2 dataset, which includes different electrode materials, cycling protocols, ambient temperatures, and depths of discharge. By evaluating the proposed method on these five datasets with different degradation dynamics and comparing it with baseline models, the effectiveness of the proposed method in lifetime prediction under various degradation conditions was verified.

[0055] Training process: Each dataset is divided into training and test sets in a 7:3 ratio. Target cell-reference cell pairs are constructed to generate training samples. Model parameters are updated by minimizing the loss function. Model performance is validated and the learning rate is adjusted every 50 rounds. Validation results: The prediction error on the test set is as follows:

[0056] MART dataset: Average error over 76 cycles;

[0057] HUST dataset: Average error 178 cycles;

[0058] ISU_ILCC dataset: Average error 89 cycles;

[0059] MIX-1 dataset: Average error over 57 cycles;

[0060] MIX-2 dataset: average error over 120 cycles.

[0061] Visualization of prediction results, such as Figure 2-6As shown, the model's predicted values ​​fit the actual lifetime values ​​well, verifying the effectiveness of the method. The results demonstrate that the model proposed in this invention can achieve high-precision lifetime prediction under different aging conditions, verifying the model's universality and accuracy.

[0062] For a target battery with an unknown lifespan, collect its current, voltage, capacity, and temperature data from the first 100 cycles, process them according to steps two to four above to obtain the difference features, input them into the trained model, and output the final predicted remaining lifespan value. This can be directly integrated into the battery management system (BMS) to achieve real-time lifespan assessment.

[0063] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A method for early life prediction of lithium-ion batteries under different aging conditions, characterized in that: The method comprises the following steps: Step one: obtaining the current, voltage, capacity, and temperature cycle data of lithium ion batteries under different charging protocols, discharging protocols, and positive electrode material systems; Step two: using normalized capacity Q ∈ [0, 1] as a unified index variable, the capacity-current, capacity-voltage and capacity-temperature curves are interpolated and resampled to obtain the charging capacity-current I c (Q), discharging capacity-current I d (Q), charging capacity-voltage V c (Q), discharging capacity-voltage V d (Q), charging capacity-temperature T c (Q), discharging capacity-temperature T d (Q), charging and discharging voltage difference characteristics ΔV(Q) = V c (Q)-V d (Q), equivalent internal resistance characteristics R(Q) = (V c (Q)-V d (Q)) / (I c (Q)-I d (Q)) Step three: normalize all the feature parameters obtained in step two to the interval [0, 1] to obtain the feature vector X = [I c (Q), I d (Q), V c (Q), V d (Q), T c (Q), T d (Q), ΔV(Q), R(Q)] ∈ R 8×1000 ; Step four: Calculate inter-cycle difference feature X tar,d = X tar,[1:100] - X tar,10 where X tar,[1:100] is the feature vector of the target battery for the first 100 cycles, X tar,10 is the feature vector of the target battery for the 10th cycle; Calculate inter-battery difference feature X ref,d = X tar,[1:100] - X ref,[1:100] where X ref,[1:100] is the feature vector of the known lifetime reference battery for the first 100 cycles; Step five: two identical Kolmogorov-Arnold Networks are used to extract the inter-cycle difference feature X tar,f = KAN(X tar,d ) and the inter-battery difference feature X ref,f = KAN(X ref,d ), and the Kolmogorov-Arnold Networks use B-spline functions based on gridding basis functions as activation functions; Step six: output the predicted lifetime value of the target battery through two shared parameter linear layers and the lifetime difference between the target battery and the reference battery Weighted average to get the final lifetime prediction value of the target battery Where α ∈ [0, 1], N is the number of reference batteries; Step seven: select M known life reference batteries, construct target battery-reference battery pairs to generate training samples, use loss function Update the model parameters w, θ, φ with Adam as the optimizer, where λ is the balance coefficient, is the true life difference between the ith target battery and the jth reference battery. 2.The method of claim 1, wherein: It also includes step eight: inputting the first 100 cycle data of the unknown battery life, executing steps two to six, and outputting the remaining service life prediction result.

Images