Method and system for lithium battery state of health estimation based on adversarial defense strategy

Abstract

The application discloses a lithium battery state of health estimation method and system based on an adversarial defense strategy, and the method comprises the following steps: acquiring the voltage and current of a lithium battery to be detected, integrating the current of partial segments to obtain the voltage and capacity curves of the partial segments; selecting the capacity corresponding to several voltages on the curve based on the voltage and capacity curves of the partial segments; and inputting the capacity corresponding to the several voltage values of the selected partial segments of the curve into a preset lithium battery state of health estimation model to obtain a lithium battery state of health estimation value, wherein the preset lithium battery state of health estimation model is obtained by training based on a convolution residual network and an adversarial attack defense strategy. The system comprises a data acquisition module, a data selection module and a data processing module. The application can accurately estimate the SOH of normal input samples, and still has accurate SOH estimation for input samples subjected to adversarial attacks.

Description

Technical Field

[0001] This invention belongs to the field of lithium battery health state estimation technology, specifically relating to a method and system for lithium battery health state estimation based on adversarial defense strategies. Background Technology

[0002] Lithium-ion batteries have become an indispensable part of the electrochemical energy storage field, widely used in electric vehicles and battery energy storage systems. However, the challenge of battery aging remains unavoidable, stemming from their complex and multifaceted degradation mechanisms, including lithium inventory loss, active material loss, and conductivity loss. To ensure the safe operation and power capability of lithium-ion batteries, precise monitoring of the battery's state of health (SOH) is necessary.

[0003] Deep learning methods, as a "black box" model, can extract and analyze underlying patterns from recorded historical datasets. This eliminates the need for in-depth understanding of the underlying electrochemical reactions within lithium batteries, offering a significant advantage in terms of applicability.

[0004] The following three existing implementations are closest to the patent application:

[0005] CN117741484A discloses a method for estimating the state of health (SOH) of a lithium battery based on the characteristics of its charging voltage curve. This invention relates to the field of lithium battery state prediction technology. The method includes: S1 acquiring charging voltage data sequences and health status data sequences for each charge-discharge cycle of the lithium battery; S2 determining the constant current charging time period and calculating the difference ΔV-i between the charging voltage curve and the average charging voltage during the constant current charging time period; S3 analyzing the correlation between ΔV-i obtained in S2 and the current maximum charging capacity of the battery using the Pearson correlation coefficient; S4 collecting training samples to train a convolutional neural network-long short-term memory network to obtain a lithium battery SOH estimation model; and S5 estimating the SOH of the lithium battery using the lithium battery SOH estimation model. This invention can extract effective characteristics representing lithium battery degradation from the voltage curve of the relatively fixed charging process of the lithium battery, improving the accuracy of the lithium battery SOH prediction model.

[0006] CN116774086B discloses a method for estimating the health status of lithium batteries based on multi-sensor data fusion. This invention discloses a method for estimating the health status of lithium batteries based on multi-sensor data fusion. It involves collecting lithium battery data, using detrended cross-correlation analysis to extract features from the dataset and partition the dataset; constructing a lithium battery pack health status model based on STGCN and Pyraformer models; extracting local battery capacity change features from a spatial parameter sequence composed of multi-dimensional parameter data of the lithium battery pack using STGCN, and inputting the obtained features into Pyraformer to establish a relationship between the spatial parameter sequence and the global lithium battery pack health status; initializing the SDO using uniform initialization, and introducing improved multivariate learning into the SDO to obtain an improved UMSDO algorithm; optimizing the hyperparameters in the STGCN-Pyraformer model using UMSDO to obtain the corresponding optimal parameters, and predicting the battery health status to obtain the predicted battery health status. This invention can be applied to the modeling process of lithium battery health status prediction, ensuring the accuracy of lithium battery health status prediction.

[0007] CN116908692A discloses a robust joint estimation method for the state of charge (SOH) and state of health (SOH) of lithium-ion batteries. This invention relates to a robust joint estimation method for the SOH and SOH of lithium-ion batteries. It replaces the minimum mean square error (MSE) in traditional algorithms with minimum error entropy. Secondly, it designs a benchmark point, MEEF, by combining maximum correlation entropy with MEE. MEEF can adjust the average error to zero, further effectively handling non-Gaussian noise in complex environments. Finally, it derives a minimum error entropy bidirectional square root capacitive Kalman filter (MEEF-DSRCKF) algorithm based on a benchmark point. This patent uses a second-order equivalent circuit model as an example, using internal resistance to measure SOH, and achieves joint estimation of the lithium-ion battery through mutual updates of two MEEF-DSRCKF filters. Finally, under non-Gaussian noise conditions, the MEEF-DSRCKF algorithm, the MEE-DSRCKF algorithm, and the traditional DSRCKF algorithm are compared. The results show that this invention further improves the accuracy and robustness of the estimation of the battery's SOH and SOH under non-Gaussian noise conditions.

[0008] Inventions one through three described above are all for battery SOH assessment, and all three provide methods for evaluating battery SOH. They estimate battery SOH based on different battery characteristics, such as charging voltage curve characteristics, partial battery capacity change characteristics, and internal resistance characteristics, using different technical means and algorithms. The inventions aim to improve the accuracy and practicality of battery SOH assessment; they employ different data processing, model building, and algorithm frameworks to more accurately estimate battery SOH. Invention one calculates the difference ΔV-i between the charging voltage curve and the average charging voltage during the constant current charging period as a feature. Invention two extracts local battery capacity change characteristics from a spatial parameter sequence composed of multi-dimensional parameter data of the lithium battery pack. These two inventions can achieve accurate estimation of battery SOH, but lack consideration for model robustness. Invention three uses a second-order equivalent circuit model as an example, using internal resistance to measure SOH. Although this method can handle non-Gaussian noise in complex environments, its practicality and generalization are poor. The purpose of these inventions is to provide technical support for the reliable operation of energy storage batteries. Accurate assessment of battery SOH (State of Health) can optimize battery usage strategies, extend battery life, reduce failure risks, and provide a basis for the design and management of energy storage systems. The invention can be applied to the management and operation of energy storage systems. Through accurate assessment and online monitoring of energy storage battery SOH, system managers can optimize scheduling, capacity planning, and fault diagnosis based on the actual battery condition to ensure the stable operation of the energy storage system. Assessment and monitoring of energy storage battery SOH can optimize battery usage and maintenance strategies, reducing overcharging, over-discharging, and adverse operations, thereby extending battery life and cycle life. These inventions can be used to detect battery faults and abnormal conditions early. Online monitoring of battery SOH can identify potential fault characteristics and abnormal behaviors, providing early warning information so that appropriate maintenance and replacement measures can be taken to avoid the impact of battery failures on the energy storage system.

[0009] However, in practical deployment, the input sample data of the aforementioned invention is not only affected by sensor measurements, but may also be tampered with by malicious attackers after being uploaded to the cloud-based battery management system (BMS), leading to incorrect SOH estimation results. Adversarial attacks present attackers with new opportunities. These attacks slightly perturb the input sample data to alter the SOH estimation results. Deep learning models are highly susceptible to adversarial attacks, significantly reducing their robustness. In the event of a cyber-physical attack, the battery state may be modified to an incorrect value, potentially increasing the safety risks of energy storage systems with thousands of batteries. Attackers can modify the original input, guiding the trained deep learning model to an erroneous state, leading to dangerous health management decisions. For example, if an attacker sets the battery SOH to 0.9, while the actual SOH has reached 0.8, users might perceive the lithium battery as safe, potentially triggering thermal runaway or even serious safety incidents. If the SOH estimation model using deep learning lacks sufficient security measures to counter adversarial attacks, it is vulnerable to malicious attacks, leading to incorrect SOH estimations and exacerbating the complexity of maintaining the health and safety of battery storage systems or electric vehicles. Therefore, adversarial attacks pose new challenges to deep learning models, requiring them not only to have accurate SOH estimation but also to be robust to adversarial attacks. Summary of the Invention

[0010] The purpose of this invention is to provide a method and system for estimating the state of health (SOH) of lithium batteries based on adversarial defense strategies. This method can accurately estimate the SOH for normal input samples, while also providing relatively accurate SOH estimation for input samples subjected to adversarial attacks. This allows the model to simultaneously possess both the accuracy of SOH estimation and robustness against adversarial attacks.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] Methods for estimating the state of health of lithium batteries based on adversarial defense strategies include:

[0013] The voltage and current of the lithium battery under test are obtained, and the current of a portion of the segment is integrated in ampere-hours to obtain the voltage and capacity curves of the portion of the segment.

[0014] Based on the voltage and capacity curves of some segments, select the capacity corresponding to several voltages on the curve;

[0015] The capacity corresponding to several voltage values ​​of the selected curve segment is input into the preset lithium battery health state estimation model to obtain the lithium battery health state estimate. The preset lithium battery health state estimation model is obtained by training based on a convolutional residual network and an adversarial attack defense strategy.

[0016] A further improvement of this invention lies in acquiring the voltage and current of the lithium battery under test, performing ampere-hour integration on the current of a portion of the battery, and obtaining the voltage and capacity curves of that portion of the battery, including:

[0017] Among them, some segments V c The formula used for calculating the ampere-hour integral of the current is:

[0018] (1)

[0019] in Q ( V ci )for V c The i-th voltage selected above V ci The corresponding capacity, t V ci , t V start They are voltages V ci and V c Starting voltage V start The corresponding time, I ( t () represents the electric current.

[0020] A further improvement of this invention lies in selecting the capacity corresponding to several voltages on the curve based on the voltage and capacity curves of a portion of the segment, including:

[0021] Partial fragments V c Divided into N = ( V end - V start ) / Δ V + 1 segment, Δ V This indicates the voltage interval of each segment. V end for V c The end voltage is used to obtain the voltage sequence. V c1 , V c2 , V c3 , ..., V cN ] = [ V start , Vstart + Δ V ,..., V start + ( N -1) × Δ V ], and the capacity corresponding to the voltage sequence. Q ( V c ) = [ Q ( V c1 ), Q ( V c2 ), ..., Q ( V cN )).

[0022] A further improvement of this invention lies in that the preset lithium battery health state estimation model is obtained by training a convolutional residual network and an adversarial attack defense strategy, wherein the convolutional residual network includes M The input is processed through convolutional blocks, global average pooling layers, and fully connected layers to form a feature map. The global average pooling layer is used to reduce the dimensionality of the feature map. Finally, a fully connected layer is used to map the reduced feature map onto the battery SOH.

[0023] A further improvement of this invention is that the preset lithium battery health state estimation model is obtained by training a convolutional residual network and a defense strategy against adversarial attacks, wherein the defense strategy against adversarial attacks includes:

[0024] In each round of training a convolutional residual network, generated adversarial examples are added to the training dataset, enabling the model to learn the relationship between the adversarial examples and their corresponding outputs. The trained model adapts to changes from the original input to the adversarial examples. The optimization objective is expressed as follows:

[0025] (4)

[0026] The internal maximization stage seeks to maximize the input of a given model. x 0 The adversarial example that results in the highest loss function is identified in certain cases; the outer minimization problem aims to determine the parameters that minimize the loss of the internal normal samples and the found adversarial examples. This enhances the overall robustness of the model; L It is a loss function. y For true SOH, For parameters θ Convolutional residual network, The radius is the disturbance radius.

[0027] Systems for estimating the state of health of lithium batteries based on adversarial defense strategies include:

[0028] The data acquisition module acquires the voltage and current of the lithium battery under test, performs ampere-hour integration on the current of a portion of the battery, and obtains the voltage and capacity curves of that portion of the battery.

[0029] The data selection module selects the capacity corresponding to several voltages on the curve based on the voltage and capacity curves of a portion of the data segment.

[0030] The data processing module inputs the capacity corresponding to several voltage values ​​of the selected curve segment into a preset lithium battery health state estimation model to obtain the lithium battery health state estimate. The preset lithium battery health state estimation model is obtained by training based on a convolutional residual network and an adversarial attack defense strategy.

[0031] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method for estimating the state of health of a lithium battery based on an adversarial defense strategy.

[0032] Compared with the prior art, the present invention has at least the following beneficial technical effects:

[0033] The present invention provides a method and system for estimating the state of health (SOH) of lithium batteries based on an adversarial defense strategy. The neural network using an adversarial training strategy can achieve accurate SOH estimation for normal samples. While neural networks using conventional training strategies can only achieve accurate SOH estimation for normal samples, their SOH estimation results are poor when encountering adversarial samples. In contrast, the neural network using the adversarial training strategy can achieve relatively accurate SOH estimation for adversarial samples. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the RCN structure;

[0035] Figure 2 A schematic diagram showing the SOH estimation results for normal samples and adversarial samples;

[0036] Figure 3 A comparison of normal and adversarial samples in the initial and final loops of NCM #1;

[0037] Figure 4 This is a schematic diagram showing the SOH estimation results for adversarial samples during normal training and adversarial training.

[0038] Figure 5 This is a block diagram of the system for estimating the health status of lithium batteries based on an adversarial defense strategy, as described in this invention. Detailed Implementation

[0039] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0040] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0041] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0042] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0043] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0044] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0045] Example 1

[0046] The present invention provides a method for estimating the state of health of lithium batteries based on adversarial defense strategies, comprising:

[0047] The voltage and current of the lithium battery under test are obtained, and the current of a portion of the segment is integrated in ampere-hours to obtain the voltage and capacity curves of the portion of the segment.

[0048] Based on the voltage and capacity curves of some segments, select the capacity corresponding to several voltages on the curve;

[0049] The capacity corresponding to several voltage values ​​of the selected curve segment is input into the preset lithium battery health state estimation model to obtain the lithium battery health state estimate. The preset lithium battery health state estimation model is obtained by training based on a convolutional residual network and an adversarial attack defense strategy.

[0050] In this embodiment, the preset lithium battery health state estimation model is obtained by training a convolutional residual network and an adversarial attack defense strategy. The convolutional residual network includes... M The input is processed through convolutional blocks, global average pooling layers, and fully connected layers to form a feature map. The global average pooling layer is used to reduce the dimensionality of the feature map. Finally, a fully connected layer is used to map the reduced feature map onto the battery SOH.

[0051] Defense strategies against attacks include:

[0052] In each round of training a convolutional residual network, generated adversarial examples are added to the training dataset, enabling the model to learn the relationship between the adversarial examples and their corresponding outputs. The trained model adapts to changes from the original input to the adversarial examples. The optimization objective is expressed as follows:

[0053] (4)

[0054] The internal maximization stage seeks to maximize the input of a given model. x 0 The adversarial example that results in the highest loss function is identified in certain cases; the outer minimization problem aims to determine the parameters that minimize the loss of the internal normal samples and the found adversarial examples. This enhances the overall robustness of the model; L It is a loss function. y For true SOH, For parameters θ Convolutional residual network, The radius is the disturbance radius.

[0055] Example 2

[0056] This invention provides a method for estimating the state of health (SOH) of lithium batteries based on an adversarial defense strategy. The invention uses coulomb counts across a portion of the voltage range as health features. The proposed convolutional residual network (RCN) model is employed to estimate the SOH of the battery. Figure 1 The basic elements of the RCN model include MThe system consists of convolutional blocks (Conv Blocks), global average pooling (GAP) layers, and fully connected (FC) layers. The input is processed through Conv Blocks to form a feature map. The Conv Blocks with residual connections can better extract the features behind the input. Then, GAP layers are used to reduce the dimensionality of the feature map. Finally, FC layers are used to map the reduced feature map to the battery state of energy (SOH).

[0057] Conv Blocks contain residual connections that preserve the initial input within each block, thus avoiding gradient vanishing or exploding as seen in traditional neural networks and ensuring the accuracy of battery SOH estimation. Figure 1 The Conv block in the network consists of three key components: convolutional (Conv) layers, batch normalization (BN) layers, and rectified linear unit (ReLU) layers. The Conv layers are responsible for extracting features from the input data using convolutional kernels. Considering the time-series nature of the input data, a one-dimensional convolutional method is chosen for this stage. The BN layers play a crucial role in maintaining a stable input distribution for each subsequent layer, which helps mitigate gradient vanishing and accelerates network convergence. ReLU acts as a non-linear activation function.

[0058] Adversarial attacks refer to attackers introducing subtle perturbations into the original data to generate adversarial examples that can affect the results of deep learning models. When exposed to designed adversarial examples, these attacks can cause trained deep learning models to produce a large number of incorrect results. This manipulation can have a significant impact on battery SOH estimation. Projective gradient descent (PGD) is chosen as the adversarial attack method. It is a gradient-based attack that iteratively introduces perturbations into the input using the gradient of the loss function. PGD is considered one of the most powerful first-order adversarial attack techniques, posing a significant challenge to the robustness of deep learning models. By successfully defending against PGD attacks, the model can address other first-order adversarial attack methods, thereby enhancing the model's robustness. The goal of the deep learning model is to minimize the loss function using gradient descent. L The goal of a PGD adversarial attack is to manipulate the input, making the SOH estimate deviate as much as possible from the actual SOH. Therefore, the input for a PGD attack is an upward curve in the opposite direction of gradient descent, and it is calculated using the following formula:

[0059] (2)

[0060] in x Indicates input, ε Indicates step size, x t Indicates the first t The adversarial example input after the next iteration. x 0 This represents the original input, i.e., the normal sample. x Indicates the acceptable perturbation radius and will include adversarial samples. x t Projection Back x 0 + x of l ∞ Sphere.

[0061] Adversarial training is a primary strategy for defending against adversarial attacks, enhancing the robustness of deep learning models. For deep learning frameworks undergoing normal training, the core objective is to determine the optimal model parameters that minimize the loss function across the entire training dataset. (Equation (3)). This process aims to achieve both high accuracy and good generalization ability when the model is faced with normal samples.

[0062] (3)

[0063] Standard normal training has proven effective across various tasks, but it often lacks robustness against adversarial examples. This limitation necessitates improvements to standard normal training methods.

[0064] In each round of training, by adding generated adversarial examples to the training dataset, the model learns the relationship between the adversarial examples and their corresponding outputs. This allows the trained model to adapt to subtle changes from the original input to the adversarial examples, thereby enhancing its robustness. The optimization objective can be restated as follows:

[0065] (4)

[0066] The internal maximization stage seeks to maximize the input of a given model. x 0 The adversarial examples that result in the highest loss function are identified. Then, the outer minimization problem aims to determine the parameters that minimize the loss of both the internal normal samples and the found adversarial examples. This enhances the overall robustness of the model; y For true SOH, For parameters θ Convolutional residual network, The radius is the disturbance radius.

[0067] Example 3

[0068] This invention was validated on NCM batteries. 66 NCM batteries were divided into a training set and a test set at a ratio of 4:1. The test set contained 14 NCM batteries, numbered NCM #1-NCM #14.

[0069] First, extract specific features from the battery, specifically the charging voltage range. V c Used for SOH estimation, where V c The beginning and end are respectively V start and V end .Will V c Divided into N = ( V end - V start ) / Δ V + 1 segment, Δ V This represents the voltage interval of each segment. Therefore, we obtain the sequence [ V c1 , V c2 , V c3 , ..., V cN ] = [ V start , V start + Δ V , ..., V start + ( N -1) × Δ V Then, it is represented as Q ( V c ) = [ Q ( V c1 ), Q ( V c2 ), ..., Q ( V cN The input features of )] and Q ( V ci The result is determined by equation (1).

[0070] (1)

[0071] in, t V ci , t V start They are V ci and Vstart The corresponding time, I ( t ( ) represents current. It should be noted that... Q ( V c This can be extracted from partial charging cycles, making it easy to obtain directly from the BMS. Voltage range V c The range is V start =3.55V to V end = 3.74V, Δ V =0.01V, N =20.

[0072] The proposed RCN model is then trained using a standard training strategy, and corresponding adversarial attack samples are generated using PGD. The hyperparameter configuration of RCN is shown in the table below:

[0073] Table 1. Hyperparameter Configuration of RCN

[0074]

[0075] The parameters for PGD are shown in the table below:

[0076] Table 2 Hyperparameter Configuration of PGD

[0077]

[0078] The hyperparameters for normal training are shown in the table below:

[0079] Table 3 Hyperparameters for normal training

[0080]

[0081] Taking NCM #1 in the test set as an example, the SOH estimation results of the RCN model after normal training are as follows for normal samples and adversarial samples. Figure 2 As shown:

[0082] The figure shows that, without adversarial attacks, the estimated SOH is very accurate and close to the reference value, proving the effectiveness of the proposed RCN. However, when the system is subjected to adversarial attacks, the SOH estimate deviates significantly from the reference SOH. Furthermore, Figure 3 (a) and (b) show a comparison between normal input and input to a PGD adversarial attack at the initial and final lifetime cycles of NCM #1. Figure 3As shown, the adversarial examples obtained from PGD attacks are very similar to normal samples, meaning it is difficult to visually distinguish whether an input has been attacked. However, the SOH estimation results of adversarial examples and normal samples show significant differences, highlighting the far-reaching impact of PGD attacks.

[0083] Table 4 details the average numerical data for all 14 test cell sets. The MAE and RMSE of the SOH estimation for normal samples are very low, at 0.418% and 0.654%, respectively. These values ​​indicate that the accuracy of SOH estimation is very high in the absence of adversarial attacks. In contrast, adversarial examples generated by PGD exhibit significantly larger errors in SOH estimation. The MAE and RMSE generated by PGD adversarial examples are 12.6% and 14.3%, respectively. These error metrics are 30.14 times and 21.87 times that observed in normal samples, respectively, indicating a substantial decrease in SOH estimation accuracy. These large differences demonstrate the adverse impact of adversarial attacks on the accuracy of SOH estimation.

[0084] Table 4. SOH estimation results for adversarial examples and normal examples from standard training.

[0085]

[0086] The above results highlight that adversarial examples can cause errors in SOH estimation, thus requiring adversarial training as a defense strategy to improve the original model. Furthermore, adversarial examples generated with different parameter settings have a significant impact on battery SOH estimation; this section delves into two parameters: perturbation radius. x and the number of iterations.

[0087] right Figure 4 A comparative analysis of the RMSE and MAE of adversarial examples with the same parameter set in (a) / (c) and (b) / (d) reveals the effectiveness of the adversarial training defense strategy in reducing the SOH estimation error caused by adversarial examples. Even in scenarios where the adversarial examples produce a maximum RMSE of 0.2445 for the normally trained model (e.g., with a perturbation radius of 0.019 and 10 iteration steps), the SOH estimation error of the adversarially trained RCN model is significantly reduced to 0.037. This is approximately 15% of the RMSE under normal training conditions, significantly demonstrating the effectiveness of adversarial training defense in enhancing model robustness. Furthermore, for normal samples, the RMSE and MAE of the adversarially trained RCN model are 0.01890 and 0.01176, respectively, indicating that the model can achieve accurate SOH estimation for both normal and adversarial samples.

[0088] Example 4

[0089] like Figure 5 As shown, the system for estimating the health status of lithium batteries based on adversarial defense strategies provided by the present invention includes:

[0090] The data acquisition module acquires the voltage and current of the lithium battery under test, performs ampere-hour integration on the current of a portion of the battery, and obtains the voltage and capacity curves of that portion of the battery.

[0091] The data selection module selects the capacity corresponding to several voltages on the curve based on the voltage and capacity curves of a portion of the data segment.

[0092] The data processing module inputs the capacity corresponding to several voltage values ​​of the selected curve segment into a preset lithium battery health state estimation model to obtain the lithium battery health state estimate. The preset lithium battery health state estimation model is obtained by training based on a convolutional residual network and an adversarial attack defense strategy.

[0093] Example 5

[0094] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method for estimating the state of health of a lithium battery based on an adversarial defense strategy.

[0095] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0096] This application is described with reference to flowchart illustrations and / or block diagrams of methods, systems, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0097] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0098] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0099] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0100] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that can be understood by those skilled in the art. The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A method for estimating the state of health of lithium batteries based on adversarial defense strategies, characterized in that, include: The voltage and current of the lithium battery under test are obtained, and the current of a portion of the segment is integrated in ampere-hours to obtain the voltage and capacity curves of the portion of the segment. Based on the voltage and capacity curves of a subset of segments, the capacities corresponding to several voltages on the curves are selected. include: Partial fragments V c Divided into N = ( V end - V start ) / Δ V + 1 segment, Δ V This indicates the voltage interval of each segment. V end for V c The end voltage is used to obtain the voltage sequence. V c1 , V c2 , V c3 , ..., V cN ] = [ V start , V start + Δ V , ..., V start + ( N -1) × Δ V ], and the capacity corresponding to the voltage sequence. Q ( V c ) = [ Q ( V c1 ), Q ( V c2 ), ..., Q ( V cN )]; The capacity corresponding to several voltage values ​​of the selected curve segment is input into the preset lithium battery health state estimation model to obtain the lithium battery health state estimate. The preset lithium battery health state estimation model is obtained by training based on a convolutional residual network and an adversarial attack defense strategy. The pre-defined lithium battery health state estimation model is obtained by training a convolutional residual network and an adversarial attack defense strategy. The convolutional residual network includes... M The input is processed by a convolutional block, a global average pooling layer, and a fully connected layer. The input is processed by the convolutional block to form a feature map. The global average pooling layer is used to reduce the dimension of the feature map. Finally, the fully connected layer is used to map the reduced feature map to the battery SOH. The pre-defined lithium battery health state estimation model is obtained by training a convolutional residual network and a defense strategy against adversarial attacks. The defense strategy against adversarial attacks includes: In each round of training a convolutional residual network, generated adversarial examples are added to the training dataset, enabling the model to learn the relationship between the adversarial examples and their corresponding outputs. The trained model adapts to changes from the original input to the adversarial examples. The optimization objective is expressed as follows: (4) The internal maximization stage seeks to maximize the input of a given model. x 0 The adversarial example that results in the highest loss function is identified in certain cases; the outer minimization problem aims to determine the parameters that minimize the loss of the internal normal samples and the found adversarial examples. This enhances the overall robustness of the model; L It is a loss function. y For true SOH, For parameters θ Convolutional residual network, The radius is the disturbance radius.

2. The method for estimating the state of health of a lithium battery based on an adversarial defense strategy according to claim 1, characterized in that, The voltage and current of the lithium battery under test are acquired, and the current of a portion of the battery is integrated in ampere-hours to obtain the voltage and capacity curves of that portion, including: Among them, some segments V c The formula used for calculating the ampere-hour integral of the current is: (1) in Q ( V ci )for V c The i-th voltage selected above V ci The corresponding capacity, t V ci , t V start They are voltages V ci and V c Starting voltage V start The corresponding time, I ( t () represents the electric current.

3. A system for estimating the state of health of a lithium battery based on an adversarial defense strategy, characterized in that, The system is based on the lithium battery health state estimation method based on adversarial defense strategy as described in claim 1, comprising: The data acquisition module acquires the voltage and current of the lithium battery under test, performs ampere-hour integration on the current of a portion of the battery, and obtains the voltage and capacity curves of that portion of the battery. The data selection module selects the capacity corresponding to several voltages on the curve based on the voltage and capacity curves of a portion of the data segment. The data processing module inputs the capacity corresponding to several voltage values ​​of the selected curve segment into a preset lithium battery health state estimation model to obtain the lithium battery health state estimate. The preset lithium battery health state estimation model is obtained by training based on a convolutional residual network and an adversarial attack defense strategy.

4. The system for estimating the state of health of a lithium battery based on an adversarial defense strategy according to claim 3, characterized in that, In the data acquisition module, the voltage and current of the lithium battery under test are acquired. The current of a portion of the battery is integrated in ampere-hours to obtain the voltage and capacity curves of that portion, including: Among them, some segments V c The formula used for calculating the ampere-hour integral of the current is: (1) in Q ( V ci )for V c The i-th voltage selected above V ci The corresponding capacity, t V ci , t V start They are voltages V ci and V c Starting voltage V start The corresponding time, I ( t () represents the electric current.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method for estimating the state of health of a lithium battery based on an adversarial defense strategy as described in claim 1 or 2.

Images