A lithium ion battery state of health estimation method and system based on feature fusion

Abstract

The application provides a lithium ion battery health state estimation method and system based on feature fusion, and belongs to the field of lithium ion battery state estimation, and comprises the following steps: obtaining steady-state data and transient data of a target domain lithium ion battery and a source domain lithium ion battery, performing feature enhancement on the transient data and the steady-state data of the target domain by using a denoising diffusion model to obtain transient data and steady-state data of an enhanced target domain; performing feature extraction on the transient data and the steady-state data of the source domain and the enhanced target domain respectively to obtain transient features and steady-state features; fusing the transient features and the steady-state features by using a dynamic weight matching mechanism and a conditional attention mechanism to obtain source domain fusion features and target domain fusion features; and realizing target domain lithium ion battery health state estimation by using an adversarial transfer network based on the source domain fusion features and the target domain fusion features. The application realizes accurate and robust SOH estimation of the target domain.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery state estimation, and particularly relates to a method and system for estimating the health status of lithium-ion batteries based on feature fusion. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Lithium-ion batteries are widely used in electric vehicles, energy storage systems, and portable electronic devices. As usage time increases, batteries gradually exhibit aging phenomena such as capacity decay and increased internal resistance. Accurately estimating their state of health (SOH) is crucial for ensuring system performance and safety.

[0004] Existing SOH estimation methods mainly include model-driven and data-driven methods. Among them, data-driven methods, which construct SOH estimation models based on historical battery characteristic data and machine learning algorithms, have attracted widespread attention from researchers because they do not require in-depth understanding of the inherently complex processes such as battery aging mechanisms. However, these methods still have certain drawbacks: First, current feature selection is still insufficient. Most studies rely on only one or more single features, which is insufficient to fully characterize the performance changes of batteries at different aging stages. In addition, rate fluctuations and temperature changes significantly increase the uncertainty of the degradation path, further weakening the correspondence between features and state of harm (SOH).

[0005] Second, the model's generalization ability is limited. In real-world operating conditions, data and SOH labels in the target domain are generally insufficient, and the performance of the source domain model degrades significantly when transferring across battery types or different operating conditions. At the same time, the distribution differences between the source and target domains make transfer learning susceptible to "negative transfer," making it difficult to guarantee reliability.

[0006] To address the first drawback, existing methods mostly employ feature fusion strategies, but these rely heavily on steady-state features. However, the steady-state phase occurs infrequently in actual operation, making the acquisition of steady-state features difficult. Furthermore, most fusion mechanisms are based on fixed weights or correlation allocations, making it difficult to adapt to different aging stages and variations in temperature and operating rates. Regarding the second drawback, while existing generative augmentation methods can alleviate the feature deficiency problem to some extent, they typically assume SOH label data reconstruction. When both labels and features are missing, relying solely on pre-trained parameters for direct transfer is highly prone to failure. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this invention proposes a feature fusion-based method and system for estimating the state of health (SOH) of lithium-ion batteries. This addresses the problem that existing single-feature or multi-feature fusion mechanisms are insufficient to fully characterize the degradation characteristics of lithium-ion batteries under different aging stages, temperatures, and rate conditions. Furthermore, in actual operating conditions, the limited feature data and the scarcity or absence of SOH labels often simultaneously affect the accuracy of model estimation. By fusing transient and steady-state features and employing adversarial transfer, different weighting coefficients are dynamically matched to different aging stages. The fusion term is modified according to different temperatures and rate conditions, enabling the fused features to adaptively optimize under multiple operating conditions. Finally, by utilizing the rich fused features and SOH label data in the source domain, accurate and robust SOH estimation in the target domain is achieved.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, this invention discloses a method for estimating the state of health of lithium-ion batteries based on feature fusion, comprising: Obtain steady-state and transient data of lithium-ion batteries in the target domain and the source domain. The steady-state data includes the main peak of the charging curve and the constant voltage charging time. The transient data includes the pulse response voltage difference and the charging relaxation voltage. A denoising diffusion model is used to enhance the features of transient and steady-state data in the target domain. By progressively adding noise in the forward direction and denoising through reverse diffusion, the missing feature data in the target domain is generated and completed, resulting in enhanced transient and steady-state data of the target domain. Feature extraction is performed on transient and steady-state data of the source domain and the enhanced target domain respectively to obtain transient and steady-state features of the source domain and the target domain. The transient and steady-state features are fused using a dynamic weight matching mechanism, and the fused features are corrected according to temperature and magnification conditional attention mechanism to obtain fused source domain features and fused target domain features. An adversarial transfer network is constructed based on source domain fusion features and their SOH label data. The constructed adversarial transfer network is used to estimate the target domain lithium-ion battery health status by applying the target domain fusion features.

[0009] Secondly, this invention discloses a lithium-ion battery health state estimation system based on feature fusion, comprising: The data acquisition module is configured to acquire steady-state and transient data of lithium-ion batteries in the target domain and the source domain, wherein the steady-state data includes the main peak of the charging curve and the constant voltage charging time, and the transient data includes the pulse response voltage difference and the charging relaxation voltage. The feature enhancement module is configured to: use a denoising diffusion model to enhance the features of the transient and steady-state data of the target domain; and generate and complete the missing feature data of the target domain by gradually adding noise in the forward direction and denoising by reverse diffusion, thereby obtaining the enhanced transient and steady-state data of the target domain. The feature fusion module is configured to: extract features from transient and steady-state data of the source domain and the enhanced target domain respectively to obtain transient features and steady-state features, and fuse the transient features and steady-state features using a dynamic weight matching mechanism and a conditional attention mechanism to obtain source domain fusion features and target domain fusion features; The prediction and estimation module is configured to: construct an adversarial transfer network based on the source domain fusion features and its SOH label data, and use the constructed adversarial transfer network to estimate the target domain lithium-ion battery health status based on the target domain fusion features.

[0010] Thirdly, the present invention discloses an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when run by the processor, complete the above-mentioned steps of lithium-ion battery SOH based on the fusion of transient and steady-state characteristics and anti-migration.

[0011] Fourthly, the present invention discloses a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the above-mentioned lithium-ion battery health state estimation method based on feature fusion.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a method for estimating the health status of lithium-ion batteries based on feature fusion. The method introduces a dynamic weight matching mechanism and a conditional attention mechanism. The dynamic weight matching mechanism sets different weight coefficients according to different aging stages to achieve dynamic fusion of transient and steady-state features; the conditional attention mechanism modifies the fusion term according to different temperatures and power ratings, enabling the fused features to adaptively optimize under multiple operating conditions.

[0013] This invention addresses the problem of missing feature data in lithium-ion batteries in the target domain by employing a denoising diffusion model for feature enhancement. Through progressive noise addition and denoising reconstruction, missing features are generated, completed, and enhanced.

[0014] When the SOH labels of lithium-ion batteries in the target domain are limited or missing, an adversarial migration strategy is used to align the feature distributions of the source and target domains and learn domain-independent features. At the same time, the rich fusion features of the source domain and the SOH label data are fully utilized to achieve accurate and robust SOH estimation of the target domain.

[0015] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0016] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0017] Figure 1 This is a flowchart of the lithium-ion battery health status estimation based on feature fusion as described in Embodiment 1 of the present invention.

[0018] Figure 2 This is a diagram illustrating the overall framework for lithium-ion battery health status estimation based on feature fusion as described in Embodiment 1 of the present invention.

[0019] Figure 3 This is a schematic diagram of source and target domain SOH estimation for lithium-ion batteries as described in Embodiment 1 of the present invention.

[0020] Figure 4 This is a schematic diagram illustrating the distribution of domain and target domain feature data and label data as described in Embodiment 1 of the present invention.

[0021] Figure 5 This is a framework diagram of the dynamic weight matching mechanism described in Embodiment 1 of the present invention.

[0022] Figure 6 This is a schematic diagram of the experimental SOH estimation results described in Embodiment 1 of the present invention. Detailed Implementation

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0025] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0026] Example 1 In one or more embodiments, a method for estimating the state of health of lithium-ion batteries based on feature fusion is disclosed, such as... Figures 1-2 As shown, it includes the following steps: Step S1: Obtain steady-state and transient data of the target domain lithium-ion battery and the source domain lithium-ion battery, as well as tag information.

[0027] This embodiment involves steady-state and transient data of lithium-ion batteries. The steady-state data includes the main peak of the charging IC curve. Constant voltage charging time Transient data includes impulse response voltage difference. Charging relaxation voltage .like Figure 3 The diagram illustrates the background of the source-domain and target-domain SOH estimation problem for lithium-ion batteries, which is the focus of this invention. Source-domain lithium-ion batteries contain abundant feature data (including steady-state and transient features) and complete SOH labeling information, while target-domain lithium-ion batteries contain only limited feature data and SOH labels. For simplicity, the source-domain lithium-ion battery will be referred to as the source domain, and the target-domain lithium-ion battery as the target domain, respectively.

[0028] Specifically, transient and steady-state data are extracted. After CC-CV charging and dynamic discharge in the source and target domains, a pulsed current is applied to obtain steady-state characteristic data: the main peak of the charging curve. Constant voltage charging time Transient characteristic data: Impulse response voltage difference Charging relaxation voltage At this point, the transient and steady-state feature data and SOH labels of the source domain are relatively abundant, while the transient and steady-state feature data and SOH labels of the target domain are relatively limited, with some SOH labels even missing. For example... Figure 4 As shown, this illustrates the distribution of source and target domain features and label data.

[0029] The extracted transient and steady-state data, along with label information, are used to construct the dataset, which includes: Source domain steady-state feature dataset is The source domain transient feature dataset is The target domain steady-state feature dataset is The target domain transient feature dataset is , can be represented as: (1) In the formula, and The source domain is respectively The steady-state vector and transient vector of each cycle, where , ; and The first The steady-state vector and transient vector of each cycle, where , .

[0030] The source domain SOH label dataset is The target domain SOH label dataset is , can be represented as: (2) In the formula, and The source domain and the target domain are respectively the first The SOH label data for each cycle.

[0031] Step S2: Use a denoising diffusion model to enhance the features of the transient and steady-state data of the target domain. By gradually adding noise in the forward direction and denoising in the reverse direction, the missing feature data of the target domain is generated and completed, and the enhanced transient and steady-state data of the target domain are obtained.

[0032] The target domain generative diffusion feature enhancement utilizes a denoising diffusion model to enhance both transient and steady-state features of the target domain. Through progressive noise addition and denoising reconstruction, missing features are generated, completed, and enhanced. The following detailed description focuses on the transient features of the target domain; the enhancement of steady-state features follows the same process.

[0033] First, forward diffusion is performed on the transient feature data. Add Gaussian noise, single step ( Add noise as follows: (3) In the formula, For noise; For positive diffusion Noisy samples of the step; Representing the Step noise intensity (typical) ); It is the identity matrix; and These represent the mean and variance of the added Gaussian noise, respectively. It follows a Gaussian distribution.

[0034] closed loop( Add noise as follows: (4) (5) In the formula, This refers to the original feature data; To accumulate to The retention factor of the step; Representing the Step noise intensity.

[0035] correspond The increase in noise is expressed as: (6) In the formula, It is random Gaussian noise.

[0036] Secondly, reverse diffusion denoising filters out Gaussian noise. Single-step reverse ( The diffusion formula is: (7) (8) (9) In the formula, This is a reverse denoising diffusion model; and These are the mean and variance of the Gaussian noise after inverse denoising, respectively. This is the variance scheduling value for the current time step, used to control noise intensity; For U-net neural network pairs Prediction of noise components in the middle.

[0037] correspond The updated denoising formula is as follows: (10) In the formula, It is random Gaussian noise.

[0038] In the denoising process, the loss function for predicting noise components is the mean squared error (MSE), as shown in the following formula: (11) In the formula, This is the loss function.

[0039] After multiple iterations Ultimately, the enhanced features are obtained: (12) In the formula, For enhanced transient features of the target domain; For the first A transient vector of a cycle.

[0040] This completes the enhancement of the transient features of the target domain. Similarly, the enhanced steady-state features of the target domain can be obtained. .

[0041] Step S3: Extract transient and steady-state features from the transient and steady-state data of the source domain and the enhanced target domain, respectively. Use a dynamic weight matching mechanism and a conditional attention mechanism to fuse the transient and steady-state features to obtain source domain fusion features and target domain fusion features.

[0042] Feature fusion is performed on transient and steady-state data from the source domain and the enhanced target domain, respectively. This step includes three stages: feature extraction, dynamic weight matching, and conditional attention correction. This embodiment uses the enhanced target domain transient feature data. Steady-state characteristic data The following is a detailed description using an example. The feature fusion of the source domain adopts the same implementation process, which will not be repeated here.

[0043] Step S3-1: Extract features from the transient and steady-state data of the source domain and the enhanced target domain, respectively.

[0044] Using Temporal Convolutional Networks (TCNs) for transient features For feature extraction, TCN primarily employs two core mechanisms: dilated causal convolution and residual connections. The former utilizes the dilated receptive field to efficiently capture causal dependencies in long sequences, while the latter overcomes the gradient vanishing problem in deep networks through skip connections, ensuring the model is trainable and can effectively extract deep features. The simplified TCN extraction formula is: (13) In the formula, Features of TCN; For TCN dilated causal convolution and residual connection functions; For TCN expansion parameters; This represents the size of the TCN convolution kernel.

[0045] Using Multilayer Perceptron (MLP) to analyze steady-state features For feature extraction, MLP, based on hidden layers and activation functions, can automatically learn complex nonlinear feature combinations, achieving high-precision fitting of arbitrarily complex functional relationships. The simplified extraction formula of MLP is: (14) In the formula, Features of MLP; For MLP activation functions; and These are the MLP weight parameters and bias terms, respectively.

[0046] Fully connect the features extracted by TCN and MLP: (15) In the formula, It is a fully connected feature; For fully connected layer activation functions; and These are the weight parameters and bias terms of the fully connected layer, respectively.

[0047] SOH preliminary regression output: (16) In the formula, These are preliminary regression results; and These are the regression layer weight parameters and bias terms, respectively.

[0048] Step S3-2: Based on the preliminary regression results, determine the aging stage, dividing the aging process into three stages: mild aging, moderate aging, and severe aging. Obtain the correlation matrix using the Pearson correlation coefficient based on the aging stage. Generate dynamic weights based on the correlation matrix and perform feature weighting to obtain preliminary fused features. Design a dynamic weight matching mechanism to fuse the extracted transient and steady-state features. This mechanism sets dynamic weight matrices according to different aging stages to ensure that the fused feature data accurately matches the corresponding aging degree.

[0049] Specifically, such as Figure 5 As shown, based on the preliminary SOH regression results, the aging stage is determined, and the entire aging process is divided into three stages: mild aging (0.95≦SOH<1), moderate aging (0.8≦SOH<0.95), and severe aging (SOH<0.8). In the mild aging stage, steady-state characteristics take precedence; in the moderate aging stage, equilibrium weights are established; and in the severe aging stage, transient weights take precedence.

[0050] Using Pearson correlation coefficient To measure the correlation between the characteristics of each stage and SOH, a correlation matrix is ​​obtained. : (17) In the formula, The correlation coefficient represents the steady-state characteristics of the mild aging stage. The correlation coefficient represents the transient characteristics of the mild aging stage. Index for aging stages. 0 represents mild aging, 1 represents moderate aging, and 2 represents severe aging.

[0051] Initialize the correlation weight matrix : (18) In the formula, To initialize the correlation weight matrix; This represents the row normalization function for a matrix.

[0052] Initialize aging phase scaling factor : (19) In the formula, The current sample is in the aging stage. The probability distribution; For learnable aging phases - modal scaling matrix.

[0053] Generate dynamic weights: (20) Feature weighting: (twenty one) In the formula, Representative enhances the target domain. The steady-state data of each cycle (at which point the cycle is in a stable state). (Aging stage) Representative enhances the target domain. Transient data for each cycle; For the weighted enhanced target domain Steady-state data for each cycle; For the weighted enhanced target domain Transient data for each cycle.

[0054] Weighted fusion of transient and steady-state characteristics: (twenty two) In the formula, Preliminary fusion characteristics; The activation function is used for feature concatenation. Concatenate the weight vectors for the features; This is the feature splicing bias term.

[0055] Step S3-3: Based on the conditional attention mechanism, the initial fused features are corrected to obtain the final fused features. A temperature and rate-based conditional attention mechanism is designed to correct the fused features. This mechanism sets multi-dimensional modulation factors according to different ambient temperatures and charge / discharge rates of lithium-ion batteries, effectively weakening the interference of different operating conditions on feature fusion, ultimately making the corrected fused features more accurately adapt to the battery performance state under different operating conditions. Specifically: A conditional attention mechanism based on temperature and charge / discharge rate is introduced to correct the initial fused features. Conditional attention is an attention mechanism in deep learning designed to dynamically adjust the model's attention to input information based on specific conditions, thereby improving model performance and generalization ability. The temperature and charge / discharge rate of the lithium-ion battery serve as dynamic modulation factors, enabling the model to adaptively adjust the fused features under different operating conditions. Specifically, a conditional attention mechanism based on temperature and charge / discharge rate is designed to specifically correct the initially fused feature data. Multi-dimensional modulation factors are adaptively set according to the battery's ambient temperature and charge / discharge rate. By adjusting the weight allocation of feature dimensions in real time, the interference of different operating conditions on feature representation is effectively weakened, ultimately allowing the corrected fused features to more accurately adapt to the battery performance state under different operating conditions. The process includes the following steps: Temperature and magnification are encoded and combined into a condition vector. e : (twenty three) In the formula, For temperature; It is a multiplier; and These are activation functions for temperature and multiplier encoding, respectively.

[0056] Generate multi-head attention query: (twenty four) (25) In the formula, and These represent the query weight vector and bias term, respectively. To represent the first A query for each attention head; This represents the number of attention heads.

[0057] The initial fusion features are converted into channel codes and then projected into multi-head key and value: (26) (27) In the formula, and Representing the first The key and value of each attention head; and These are the learnable channel position codes for the Key and Value, respectively; and These represent linear mapping layers for Key and Value, respectively, used to map channel-level scalars to a high-dimensional vector space, and further extended to a multi-head representation.

[0058] Scaling the dot product per head attention: (28) In the formula, For the first The first one in the attention. One key; for and Match score; for The dimension; For the first Attention points allocated to Attention weights; For the first A person's attention The value vector; The feature dimension after fusion; For the first The output of each attention head.

[0059] The outputs of each attention point are concatenated, and the feature dimensions are fused through a linear transformation: (29) (30) In the formula, For feature dimensions; This is a tensor concatenation function; and These are the mapping weight vector and the bias term, respectively; This is the fusion bias term output by the temperature- and magnification-based conditional attention mechanism.

[0060] For initial fusion features Correction: (31) In the formula, For the final fusion feature; For layer normalization function; This is an adjustable parameter for the correction item.

[0061] Thus, the fusion of transient and steady-state features of the enhanced target domain has been completed to obtain the fused target domain features, denoted as... Similarly, the source domain transient and steady-state features are fused to obtain the source domain fused features, denoted as... .

[0062] Step S4: Construct an adversarial transfer network based on the source domain fusion features and its SOH label data, and use the constructed adversarial transfer network to achieve high-precision estimation of the target domain lithium-ion battery health status based on the target domain fusion features.

[0063] SOH adversarial migration prediction based on rich fusion feature data from the source domain Based on the corresponding SOH labels, an adversarial migration network is constructed to achieve high-precision estimation of the SOH in the target domain.

[0064] Adversarial transfer networks introduce a domain discriminator to guide the feature extractor in learning indistinguishable shared feature representations between the source and target domains, while preserving task-related information, thus achieving cross-domain knowledge transfer. Even with limited samples in the target domain, this method can leverage the rich data in the source domain to enhance feature representation capabilities, improving prediction accuracy and model generalization performance, without explicitly matching the feature distributions of the source and target domains. The specific implementation process is as follows: Define the feature distributions of the source and target domains: (32) In the formula, It is the set of source domain feature distributions; This is the target domain feature distribution set (target domain SOH label data is scarce or unlabeled).

[0065] Feature extraction and regression prediction: (33) In the formula, For feature extraction layer; Features of the source domain; Features of the target domain; For regression prediction layer; and These are the parameters for the feature extraction layer and the regression prediction layer, respectively. and The source domain and the target domain are respectively the first One cycle of SOH regression prediction values; and These are the source domain and target domain SOH regression prediction sets, respectively.

[0066] This embodiment thus solves the generalization difficulty caused by the scarcity of target domain SOH tags.

[0067] Preferably, an adversarial migration network loss function is constructed: (34) (35) (36) In the formula, and These are the regression loss term and the domain adversarial loss term, respectively. To balance the weights of regression loss and adversarial loss ( (often determined based on ablation experiments). For the MSE function; To balance the regression loss weights in the target domain ( The value is determined based on the sample size of the target domain SOH labels. It is a mean function; For domain discrimination layer; These are the parameters for the domain discrimination layer.

[0068] High-precision estimation of the target domain SOH is achieved by optimizing the loss function of the adversarial transfer learning network. Specifically, the parameters of the feature extraction layer and the regression prediction layer are jointly optimized. This allows it to minimize regression loss while maximizing domain adversarial loss; simultaneously, it optimizes the parameters of the domain discriminant layer. This method minimizes domain adversarial loss, thereby aligning feature distributions between the source and target domains and improving the prediction accuracy of SOH in the target domain. Thus, this method achieves high-precision SOH estimation under different aging stages, temperatures, and rate conditions, even with missing feature data for the target lithium-ion battery and scarce SOH tags.

[0069] Furthermore, the effectiveness of the above method was verified through experiments.

[0070] Charge-discharge experiments were conducted on the battery under different temperatures and rates, and the battery management system was used to collect the battery voltage and current data. Subsequently, the proposed method was applied to SOH estimation, and in the estimation process, some feature data of the target domain battery and SOH label data were intentionally removed to simulate the scenario of feature loss under actual working conditions.

[0071] The SOH estimation results are as follows: Figure 6 As shown in the figure, the experimental results demonstrate that the proposed method can accurately estimate the state of health (SOH) of the target lithium-ion battery even when feature data is missing and SOH labels are scarce. Through dynamic weighted feature fusion, conditional attention mechanism fusion correction, and denoising diffusion model enhancement, the estimation accuracy of SOH in the target domain is effectively improved. The adversarial transfer learning strategy enables the feature extractor to learn domain-independent features, achieving feature distribution alignment between the source and target domains and enhancing the robustness and generalization ability of the model under different battery types or operating conditions. Furthermore, the proposed feature fusion and transfer learning framework can be flexibly extended to other battery types or multi-condition health state prediction tasks, demonstrating good generalization and engineering application value.

[0072] Example 2 In one or more embodiments, a lithium-ion battery health state estimation system based on feature fusion is disclosed, specifically including: The data acquisition module is configured to acquire steady-state and transient data of lithium-ion batteries in the target domain and the source domain, wherein the steady-state data includes the main peak of the charging curve and the constant voltage charging time, and the transient data includes the pulse response voltage difference and the charging relaxation voltage. The feature enhancement module is configured to: use a denoising diffusion model to enhance the features of the transient and steady-state data of the target domain; and generate and complete the missing feature data of the target domain by gradually adding noise in the forward direction and denoising by reverse diffusion, thereby obtaining the enhanced transient and steady-state data of the target domain. The feature fusion module is configured to: extract features from transient and steady-state data of the source domain and the enhanced target domain respectively to obtain transient features and steady-state features, and fuse the transient features and steady-state features using a dynamic weight matching mechanism and a conditional attention mechanism to obtain source domain fusion features and target domain fusion features; The prediction and estimation module is configured to: construct an adversarial transfer network based on the source domain fusion features and its SOH label data, and use the constructed adversarial transfer network to estimate the target domain lithium-ion battery health status based on the target domain fusion features.

[0073] Example 3 This embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps of the above-described lithium-ion battery health state estimation method based on feature fusion.

[0074] Example 4 This embodiment provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the above-described lithium-ion battery health state estimation method based on feature fusion.

[0075] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0076] 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.

[0077] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed 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.

[0078] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for estimating the state of health of a lithium-ion battery based on feature fusion, characterized in that, include: Obtain steady-state and transient data of lithium-ion batteries in the target domain and the source domain. The steady-state data includes the main peak of the charging curve and the constant voltage charging time. The transient data includes the pulse response voltage difference and the charging relaxation voltage. A denoising diffusion model is used to enhance the features of transient and steady-state data in the target domain. By progressively adding noise in the forward direction and denoising through reverse diffusion, the missing feature data in the target domain is generated and completed, resulting in enhanced transient and steady-state data of the target domain. Feature extraction is performed on transient and steady-state data of the source domain and the enhanced target domain respectively to obtain transient and steady-state features of the source domain and the target domain. The transient and steady-state features are fused using a dynamic weight matching mechanism, and the fused features are corrected according to temperature and magnification conditional attention mechanism to obtain fused source domain features and fused target domain features. The transient and steady-state features obtained from feature extraction are as follows: Transient and steady-state features are extracted from the source and target domains respectively. Temporal convolutional networks are used to extract transient features, and multilayer perceptrons are used to extract steady-state features. Fully connected features are obtained by fully connecting the features extracted by the temporal convolutional network and the multilayer perceptron. A preliminary regression output is obtained based on the fully connected features; An adversarial transfer network is constructed based on source domain fusion features and their SOH label data. The constructed adversarial transfer network is used to estimate the target domain lithium-ion battery health status by applying the target domain fusion features. The dynamic weight matching mechanism is as follows: based on the preliminary regression results, the aging stage is judged, and the aging process is divided into three stages, including mild aging, moderate aging and severe aging. According to the aging stage, the correlation matrix is ​​obtained using the Pearson correlation coefficient. Based on the correlation matrix, dynamic weights are generated and feature weighting is performed to obtain preliminary fusion features. The expression for generating dynamic weights based on the correlation matrix and performing feature weighting is as follows: In the formula, Representative enhances the target domain. Steady-state data for each cycle; Representative enhances the target domain. Transient data for each cycle; For the weighted enhanced target domain Steady-state data for each cycle; For the weighted enhanced target domain Transient data for each cycle; The correlation coefficient represents the steady-state characteristics of the mild aging stage. The correlation coefficient represents the transient characteristics of the mild aging stage. Index for aging stages; Weighted fusion of transient and steady-state characteristics: In the formula, These are the initial fusion features; The activation function is used for feature concatenation. Concatenate the weight vectors for the features; For feature splicing bias term; The conditional attention mechanism specifically includes: Multi-head attention is generated using a conditional vector that combines temperature and magnification. In the formula, and These represent the weight vector and the bias term, respectively. e It is a conditional vector; Q For attention head queries; To represent the first A query for each attention head; Represents the number of heads of attention; Q For the attention of multiple parties; The multi-head attention is scaled and dot-productd to obtain the output of each attention head. The outputs of each attention head are concatenated and the feature dimensions are fused by linear transformation mapping. Correction of initial fusion features: In the formula, The final fusion features for the target domain; For layer normalization function; Adjustable parameters for correction items; It is a fusion bias term based on the output of the conditional attention mechanism of temperature and magnification; The adversarial migration network achieves target domain lithium-ion battery health state estimation specifically as follows: In the formula, and These are the regression loss term and the domain adversarial loss term, respectively. To balance the weights of regression loss and adversarial loss; For the MSE function; To balance the regression loss weights in the target domain; It is a mean function; For domain discrimination layer; For the domain discrimination layer parameters; For feature extraction layer; Features of the source domain; Features of the target domain; The source domain final fusion feature is obtained by fusing transient and steady-state features of the target domain; For regression prediction layer; and These are the parameters for the feature extraction layer and the regression prediction layer, respectively. and Source domain and target domain respectively One cycle of SOH regression prediction values; and These are the source domain and target domain SOH regression prediction sets, respectively.

2. The lithium-ion battery health state estimation method based on feature fusion as described in claim 1, characterized in that, The feature enhancement involves enhancing the generative diffusion features of the target domain. A denoising diffusion model is used to enhance the transient and steady-state features of the target domain respectively. By progressively adding noise and reconstructing by denoising, missing features are generated, completed, and enhanced, resulting in enhanced transient and steady-state features of the target domain.

3. The lithium-ion battery health state estimation method based on feature fusion as described in claim 1, characterized in that, The preliminary regression output is obtained based on the fully connected features: In the formula, It is a fully connected feature; These are preliminary regression results; and These are the regression layer weight parameters and bias terms, respectively.

4. A lithium-ion battery health state estimation system based on feature fusion, characterized in that, include: The data acquisition module is configured to acquire steady-state and transient data of lithium-ion batteries in the target domain and the source domain, wherein the steady-state data includes the main peak of the charging curve and the constant voltage charging time, and the transient data includes the pulse response voltage difference and the charging relaxation voltage. The feature enhancement module is configured to: use a denoising diffusion model to enhance the features of the transient and steady-state data of the target domain; and generate and complete the missing feature data of the target domain by gradually adding noise in the forward direction and denoising by reverse diffusion, thereby obtaining the enhanced transient and steady-state data of the target domain. The feature fusion module is configured to: extract features from transient and steady-state data of the source domain and the enhanced target domain respectively to obtain transient features and steady-state features, and fuse the transient features and steady-state features using a dynamic weight matching mechanism and a conditional attention mechanism to obtain source domain fusion features and target domain fusion features; The transient and steady-state features obtained from feature extraction are as follows: Transient and steady-state features are extracted from the source and target domains respectively. Temporal convolutional networks are used to extract transient features, and multilayer perceptrons are used to extract steady-state features. Fully connected features are obtained by fully connecting the features extracted by the temporal convolutional network and the multilayer perceptron. A preliminary regression output is obtained based on the fully connected features; The prediction and estimation module is configured to: construct an adversarial transfer network based on the source domain fusion features and its SOH label data, and use the constructed adversarial transfer network to estimate the target domain lithium-ion battery health status based on the target domain fusion features; The dynamic weight matching mechanism is as follows: based on the preliminary regression results, the aging stage is judged, and the aging process is divided into three stages, including mild aging, moderate aging and severe aging. According to the aging stage, the correlation matrix is ​​obtained using the Pearson correlation coefficient. Based on the correlation matrix, dynamic weights are generated and feature weighting is performed to obtain preliminary fusion features. The expression for generating dynamic weights based on the correlation matrix and performing feature weighting is as follows: In the formula, Representative enhances the target domain. Steady-state data for each cycle; Representative enhances the target domain. Transient data for each cycle; For the weighted enhanced target domain Steady-state data for each cycle; For the weighted enhanced target domain Transient data for each cycle; The correlation coefficient represents the steady-state characteristics of the mild aging stage. The correlation coefficient represents the transient characteristics of the mild aging stage. Index for aging stages; Weighted fusion of transient and steady-state characteristics: In the formula, These are the initial fusion features; The activation function is used for feature concatenation. Concatenate the weight vectors for the features; For feature splicing bias term; The conditional attention mechanism specifically includes: Multi-head attention is generated using a conditional vector that combines temperature and magnification. In the formula, and These represent the weight vector and the bias term, respectively. e It is a conditional vector; Q For attention head queries; To represent the first A query for each attention head; Represents the number of heads of attention; Q For the attention of multiple parties; The multi-head attention is scaled and dot-productd to obtain the output of each attention head. The outputs of each attention head are concatenated and the feature dimensions are fused by linear transformation mapping. Correction of initial fusion features: In the formula, The final fusion features for the target domain; For layer normalization function; Adjustable parameters for correction items; It is a fusion bias term based on the output of the conditional attention mechanism of temperature and magnification; The adversarial migration network achieves target domain lithium-ion battery health state estimation specifically as follows: In the formula, and These are the regression loss term and the domain adversarial loss term, respectively. To balance the weights of regression loss and adversarial loss; For the MSE function; To balance the regression loss weights in the target domain; It is a mean function; For domain discrimination layer; For the domain discrimination layer parameters; For feature extraction layer; Features of the source domain; Features of the target domain; The source domain final fusion feature is obtained by fusing transient and steady-state features of the target domain; For regression prediction layer; and These are the parameters for the feature extraction layer and the regression prediction layer, respectively. and Source domain and target domain respectively One cycle of SOH regression prediction values; and These are the source domain and target domain SOH regression prediction sets, respectively.

5. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the feature fusion-based lithium-ion battery health state estimation method according to any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete the lithium-ion battery health state estimation method based on feature fusion as described in any one of claims 1-3.

Images