A multi-source feature fusion double-branch lithium battery health estimation method

Abstract

The application discloses a kind of double-branch lithium battery health estimation methods of multi-source feature fusion, including the key operation data of lithium battery in the process of charging and discharging is collected by battery management system or test platform;The key operation data of lithium battery collected is extracted to multi-source health factor;The data preprocessing is carried out to the multi-source health factor extracted;Using the particle swarm optimization algorithm based on the principle of quantum mechanics is improved, the optimal section is searched in the reasonable range of charging voltage;In the determined optimal voltage section, the corresponding multi-source characteristic quantity is trained, the parallel double-branch network CMNN is established, the parameters are optimized by back propagation algorithm, and the QPSO-CMNN lithium battery state estimation model is obtained;The final model is used to estimate test set lithium ion battery SOH, the application uses QPSO to search optimal voltage section, adapts actual data missing and user battery actual use characteristics, with higher SOH prediction accuracy and robustness.

Description

Technical Field

[0001] This invention relates to the field of new energy battery management and intelligent prediction technology, specifically to a dual-branch lithium battery health estimation method that integrates multi-source features. Background Technology

[0002] During long-term cycling, the capacity and State of Health (SOH) of lithium-ion batteries gradually decline over time. Accurate state of health estimation is crucial for the safety control and lifespan prediction of battery management systems. Accurate SOH estimation is essential for extending battery life, avoiding overcharging and over-discharging, and enabling predictive maintenance. Existing battery health estimation methods mainly fall into two categories: model-driven and data-driven. Model-based methods primarily include equivalent circuit models and electrochemical models, possessing a clear theoretical foundation and good interpretability. However, model parameters are difficult to update in real time and are greatly affected by ambient temperature and aging conditions. Data-driven methods utilize deep learning or statistical models to learn degradation patterns from data such as voltage, current, and capacity, but typically require complete charge-discharge data. For example, many studies use full charge-discharge curves to train neural networks to predict capacity or SOH.

[0003] However, in practical applications, vehicles and energy storage systems often employ partial charging strategies, making it difficult to obtain the complete charging curve. Furthermore, single-source data or single-channel network models are susceptible to noise interference and struggle to simultaneously capture short-term fluctuations and long-term degradation information during battery degradation. Some studies have attempted to improve model performance using convolutional networks, temporal convolutional networks, or transfer learning, but these still suffer from dependence on the complete charging curve, making them unsuitable for partial charging scenarios; and lacking adaptive mechanisms in feature selection and data segment determination; additionally, the network model structure is too simplistic to simultaneously capture local details and global degradation trends. Therefore, there is an urgent need for a battery SOH estimation method that can automatically extract key voltage segment features under incomplete charging conditions, compensate for the shortcomings of single models, improve prediction accuracy, and adapt across operating conditions. Summary of the Invention

[0004] The purpose of this invention is to provide a health estimation method for dual-branch lithium batteries based on multi-source feature fusion, so as to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a multi-source feature fusion method for estimating the health of a dual-branch lithium battery, comprising the following operational steps:

[0006] Step S1: Collect key operating data during the charging and discharging process of the lithium battery through the battery management system or test platform.

[0007] Preferably, the key operating data includes battery terminal voltage, current, temperature, real-time capacity and time, complete charging curve, cumulative charged ampere-hour capacity, and recorded temperature changes to compensate for polarization effects, forming the original input matrix as the basis for subsequent feature extraction.

[0008] Step S2: Extract multi-source health factors from the collected key operating data of lithium batteries.

[0009] Preferably, multi-source feature extraction includes: extracting multi-source health features from the collected key operational data in both the data domain and the model domain to obtain data domain features and model domain features; the data domain features include four indicators: segment capacity, voltage rise rate, ... Peak position and peak; Model domain features include: ohmic resistance, polarization resistance and polarization capacitance, which are obtained by fitting a first-order RC circuit model to the charging curve and estimating the parameters using the least squares method; By obtaining the data domain features and model domain features, a multidimensional description of the battery degradation state, including battery capacity decay, electrochemical polarization and internal resistance growth, is constructed.

[0010] Step S3: Perform data preprocessing on the extracted multi-source health factors.

[0011] Preferably, data preprocessing includes filtering, normalization, and time synchronization processing, and a standardized data matrix that can be directly used for feature extraction is generated after filtering and normalization.

[0012] Step S4: Use a particle swarm optimization algorithm based on quantum mechanics principles to search for the optimal segment within a reasonable range of charging voltage.

[0013] Preferably, the operation steps of the particle swarm optimization algorithm improved based on quantum mechanics principles include:

[0014] Step S41: Select the main charging segment of the charging curve ;

[0015] Step S42: Simulate partial charging of an actual vehicle in the main charging section. Select some charging segments ;

[0016] Step S43: Introduce a quantum behavior model so that the particle's position update no longer depends on velocity, but is obtained through the random collapse of the quantum wave function, in a simplified form:

[0017]

[0018] In the formula, This indicates the number of iterations, i.e., the current iteration of the algorithm. Indicates the first The particle in the first The position of the generation, For the first The particle position is at the... The position of the generation, Indicates the first The best position found by the particle so far It is the group's average optimal position. To obey A uniformly distributed random variable over an interval. This is the contraction coefficient, a global adjustment parameter used to control the particle search range;

[0019] Step S44: Initialize the particle swarm, with each particle representing a candidate segment. Calculate the Pearson correlation coefficient between each multi-source health factor and capacity within each particle segment. Calculate the objective function and update the particle positions. Output the optimal segment when the number of iterations or the error meets the conditions.

[0020] Step S5: Within the determined optimal voltage range, train the corresponding multi-source features to establish a parallel dual-branch network CMNN, optimize the parameters through the backpropagation algorithm, and obtain the QPSO-CMNN lithium battery state estimation model.

[0021] Preferably, the QPSO-CMNN lithium battery state estimation model includes:

[0022] Step S51: Within the range of the charging voltage, the multi-source health feature sequence extracted from the optimal segment obtained by the particle swarm optimization algorithm based on the improved quantum mechanics principle is randomly divided into a test set, a validation set, and a training set.

[0023] Step S52: Train the dual-branch deep learning estimation model CMNN using the training set data;

[0024] Step S53: Optimize the model parameters using the loss function and output the final QPSO-CMNN battery state estimation model for lithium battery SOH estimation.

[0025] Step S6: Use the final model to estimate the SOH of the test set lithium-ion batteries.

[0026] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention utilizes QPSO to search for the optimal voltage range, adapting to the lack of actual data and the actual usage characteristics of user batteries; it introduces multimodal fusion features, where each multi-source health factor provides degradation information from different physical perspectives, and the network automatically determines which features are more important at the current stage through weight learning, obtaining most of the useful information from the original lithium battery operating data, and finally establishing a QPSO-CMNN lithium battery state estimation model, which has higher SOH prediction accuracy and robustness than other ordinary models. Attached Figure Description

[0027] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0028] Figure 1 The flowchart illustrates a dual-branch lithium battery health estimation method based on multi-source feature fusion, as provided in this embodiment of the invention.

[0029] Figure 2 The flowchart for selecting the optimal voltage segment using the quantum particle swarm optimization algorithm is provided for embodiments of the present invention.

[0030] Figure 3 This is a network structure diagram of a dual-branch lithium battery health estimation method based on multi-source feature fusion, provided in an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Combination Figures 1-3 As shown, the present invention provides a technical solution: a multi-source feature fusion method for estimating the health of a dual-branch lithium battery, comprising the following steps:

[0033] Step S1: Collect lithium battery operating data.

[0034] In this embodiment, during battery operation or experimental testing, key operational data during the battery charging and discharging process are collected through the battery management system or testing platform. This includes battery terminal voltage, current, temperature, real-time capacity and time, as well as a complete charging curve, cumulative ampere-hour capacity, and recorded temperature changes to compensate for polarization effects. This data can comprehensively describe the charging and discharging behavior and thermoelectric characteristics of the lithium-ion battery, forming the original input matrix, which is the basis for subsequent feature extraction and health status estimation.

[0035] Step S2: Extract multi-source health factors from the collected lithium battery operation data.

[0036] In this embodiment, multi-source health features are extracted from the data domain and model domain respectively from the collected operational data to construct a multi-dimensional description of the battery degradation state. The features from the data domain and model domain together reflect the multi-dimensional degradation characteristics of the battery, such as capacity decay, electrochemical polarization, and internal resistance growth. By fusing health factors from the data domain and model domain, the model's dependence on data is reduced, and the robustness of model predictions is improved.

[0037] For example, the extracted data domain features include four indicators: segment capacity. Voltage rise rate , Peak position With Peak The model domain features are obtained by fitting a first-order RC circuit model to the charging curve and estimating the parameters using the least squares method to obtain the ohmic resistance. Polarization resistance Polarized capacitors The calculation formulas for each indicator are as follows:

[0038]

[0039]

[0040]

[0041]

[0042]

[0043]

[0044]

[0045] In the formula, For the battery at all times The charging current, , These represent the starting and ending voltages defined in the battery charging curve, respectively. and These correspond to the times when the starting and ending voltages are reached, respectively. Indicates in The amount of charge introduced into the interval. It represents the amount of voltage change. Indicates that Maximum voltage value, Indicates the internal resistance of the ohm. Polarization resistance, This indicates a polarized capacitor.

[0046] Step S3: Perform data preprocessing on the extracted features;

[0047] In this embodiment, to improve the stability of subsequent optimization and modeling, the multi-source feature matrix is ​​preprocessed as follows: a Savitzky-Golay smoothing filter is used to eliminate sampling noise, and outliers are replaced with their medians; for each Normalization to The formula is as follows:

[0048]

[0049] In the formula, Indicates the first The normalized values ​​of each health factor Indicates the first Original values ​​of each health factor Indicates the first The minimum value of each health factor in all samples Indicates the first The maximum value of each health factor across all samples.

[0050] Step S4: Use a particle swarm optimization algorithm based on quantum mechanics principles to search for the optimal section within a reasonable range of charging voltage;

[0051] In this embodiment, the quantum mechanics-based improved particle swarm optimization algorithm (QPSO) is a type of particle swarm optimization algorithm (PSO) based on quantum mechanics principles. Compared with traditional PSO, QPSO is generally superior in global search capability and convergence performance. The flowchart of QPSO for selecting the optimal voltage segment is as follows. Figure 2 As shown, within the charging voltage range, the QPSO algorithm is used to automatically search for the voltage segment with the most information. The length of the voltage section needs to be... Within the range, by maximizing the objective function Seven multidimensional features within a selected segment were identified as most sensitive to battery state of harmonic equilibrium (SOH). This involves searching for the maximum value of the objective function. The expression is:

[0052]

[0053] In the formula, For the first indivual Pearson correlation coefficient with actual capacity The larger the value, the richer the health information contained in that segment.

[0054] For example, the specific steps of the quantum mechanics-based improved particle swarm optimization (QPSO) algorithm in searching for the optimal segment within a reasonable charging voltage range include:

[0055] Step S41: Research shows that a SOC range of 10%-80% is the charging range chosen by most users, so the main charging segment of the charging curve is selected. The data is most stable and representative within this range;

[0056] Step S42: Simulate partial charging of a real vehicle, where voltage often changes during charging. The left and right sides will not cover the entire process, only the main charging section. Select some charging segments ;

[0057] Step S43: QPSO introduces a quantum behavior model, so that the position update of particles no longer depends on velocity, but is obtained through the random collapse of the quantum wave function. It has the advantages of stronger global search capability and more stable convergence. The simplified form is:

[0058]

[0059] In the formula, This indicates the number of iterations, i.e., the current iteration of the algorithm. Indicates the first The particle in the first The position of the generation, For the first The particle positions (representing the start-end point of a voltage segment) are at the [number]th [position]. The position of the generation, Indicates the first The best position found by the particle so far It is the group's average optimal position. To obey A uniformly distributed random variable over an interval. This is the contraction coefficient, a global adjustment parameter used to control the particle search range.

[0060] Step S44: Initialize the particle swarm (each particle represents a candidate segment), calculate the Pearson correlation coefficient between each HF and capacity within each particle segment, calculate the objective function, update the particle positions, and output the optimal segment when the number of iterations or the error meets the conditions. This segment is the voltage range most sensitive to capacity degradation.

[0061] Step S5: Within the determined optimal voltage range, train the corresponding multi-source features to establish a parallel dual-branch network CMNN, and optimize the parameters through the backpropagation algorithm (BP) to obtain the QPSO-CMNN lithium battery state estimation model.

[0062] In this embodiment, the specific implementation method includes:

[0063] Step S51: Within a reasonable charging voltage range, the battery will be charged... The optimal segment obtained from the search was extracted. The sequences are randomly divided into a test set, a validation set, and a training set;

[0064] Step S52: Train the dual-branch deep learning estimation model CMNN using the training set data;

[0065] In this embodiment, the model structure diagram is as follows: Figure 3 As shown, each module is explained below:

[0066] The local branch employs a convolutional neural network to extract highly sensitive local features. It uses a two-level one-dimensional convolutional structure, consisting of standard convolution and depthwise separable convolution, combined with ReLU activation, GroupNorm, and Dropout regularization to extract short-term features between adjacent loops. This branch focuses on extracting local change features between adjacent loops, such as sudden changes in internal resistance and temperature fluctuations.

[0067] For example, the steps for extracting local features from local branches are as follows:

[0068] Step a1: For the input feature matrix Standard one-dimensional convolution is used to extract preliminary local features. The formula is as follows:

[0069]

[0070] In the above formula, Represents the ReLU activation function. This indicates that the convolution kernel covers the convolution of the previous and next 5 loops, achieving local focusing of the receptive field and initially refining local behavioral patterns.

[0071] Step a2: Further extract fine-grained features using depthwise separable convolution. The formula is as follows:

[0072]

[0073] In the formula, This represents a depthwise convolution operation with a kernel size of 3. This represents a 1×1 pointwise convolution operation. This structure significantly reduces parameters and computational costs while still maintaining the ability to capture short-term abrupt signals at high resolution.

[0074] Step a3: Convert the features output by convolution... Applying normalization and activation functions yields the feature outputs of local branches. The formula is as follows:

[0075]

[0076] In the formula, GroupNorm (group number set to 1) is used to prevent scale differences in various battery health factors from affecting network training; the ReLU nonlinear activation function is introduced to enable the model to learn the nonlinear mapping relationship between battery features and capacity; Dropout represents the random deactivation of some neurons to prevent overfitting.

[0077] The global branch is based on a state-space model architecture, utilizing a state-space kernel and dynamic gating mechanisms to effectively model the nonlinear evolution patterns in long-sequence data. By combining state updates with selective reads and writes, this branch can effectively capture the slow degradation trends of lithium batteries over thousands of cycles (such as SEI film thickening and lithium loss) while maintaining linear reasoning complexity.

[0078] For example, the steps for extracting global features using the global branch are as follows:

[0079] Step b1: Input dynamic block feature matrix The original feature dimensions are transformed through linear mapping. Projected onto hidden dimensions The initial hidden state is represented as The formula is as follows:

[0080]

[0081] In the formula, and These are the learnable parameters of the projection layer.

[0082] Step b2: Stacking Each residual block is composed of a root mean square normalized (RMSNorm) block and a MambaBlock concatenated. The output expression for each layer is:

[0083]

[0084] In the formula, Represents a single , For the first Layer learnable parameters This indicates the output of the previous layer. Mean normalization is performed at each time step to enhance training stability, and residual connections ensure a stable flow of information and gradients.

[0085] Step b3: Within each MambaBlock, the input features are expanded by a factor of 2 along the channel dimension and then split into two sub-streams;

[0086] In this embodiment, one input state branch is responsible for the actual temporal modeling; the other is a gated branch used for adaptive modulation of the state branch's output. The state branch first undergoes a one-dimensional depthwise separable convolution, then passes through a Swish activation function before being fed into a Selective State Space Model (SSM) state space unit for dynamic modeling. Parallel scanning is used during the training phase to accelerate processing at all time steps; during the inference phase, only the previous state and some historical information need to be cached. Single-step computation and memory overhead are constant, making it suitable for sequences of arbitrary length. The gated branch passes through only one Swish layer, and its output is used as a gating weight to perform element-wise multiplication modulation on the state branch output, achieving a filtering effect similar to attention. Normalization is then performed using LayerNorm to obtain the output of the current layer.

[0087] Step b4: Finally, obtain the hidden representation of the global branch. ,

[0088] Hidden representation The calculation is as follows:

[0089]

[0090] In the formula, This represents the feature vector obtained after all residual blocks have been completed. This refers to LayerNorm.

[0091] Step b5: Fuse the prediction results of the dual-branch network at the time step level. and They correspond one-to-one in time, requiring no interpolation, and the prediction results are output through a fully connected layer (FC).

[0092] Step S53: The mean squared error (MSE) is used as the loss function, and the network weights are adaptively updated using the Adam optimizer and backpropagation. After each training round, the model performance is evaluated using the validation set. The learning rate is dynamically adjusted based on changes in the validation error, and an early stopping strategy is implemented to ensure model convergence and stability. After training, the final QPSO-CMNN model is built for lithium battery SOH estimation.

[0093] Step S6: Prediction phase, using the final model to estimate the SOH of the test set lithium-ion batteries;

[0094] In this embodiment, the operating data of the lithium battery to be predicted is processed through the same steps S1–S3 to extract the filtered multi-source features, which are then input into the trained QPSO-CMNN model. The predicted capacity and corresponding health status are calculated using the following formulas:

[0095]

[0096] In the formula, This represents the battery's current maximum usable capacity, while This indicates the battery's nominal capacity.

[0097] This invention utilizes multi-source feature extraction and QPSO algorithm optimization of voltage segments and parallel dual-branch network fusion to construct an adaptive and interpretable high-efficiency lithium battery state estimation framework, achieving high-precision prediction of battery SOH under partial charging conditions.

[0098] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0099] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method of lithium battery health estimation with multi-source feature fusion and double branches, characterized in that: The method comprises the following operation steps: Step S1: collecting key operation data of lithium batteries in the charging and discharging process through a battery management system or a test platform; The key operation data includes battery terminal voltage, current, temperature, real-time capacity and time, complete charging curve, accumulated charging ampere-hour capacity, and recorded temperature change to compensate for polarization effect; Step S2: extracting multi-source health factors from the collected key operation data of lithium batteries; Step S3: data preprocessing of the extracted multi-source health factors; Step S4: searching for an optimal section within a reasonable range of charging voltage by using a particle swarm optimization algorithm improved based on the principle of quantum mechanics; Step S5: training corresponding multi-source characteristic quantities within the determined optimal voltage section, establishing a parallel double-branch network CMNN, optimizing parameters by a back propagation algorithm, and obtaining a QPSO-CMNN lithium battery state estimation model; Step S6: estimating the test set lithium ion battery SOH by using the final model.

2. The multi-source feature fusion dual-branch lithium battery health estimation method according to claim 1, characterized in that: The multi-source feature extraction in step S2 comprises: Respectively extracting multi-source health features from the collected key operation data in the data domain and the model domain to obtain data domain features and model domain features; The data field features include four indicators: segment capacity, voltage rise rate, Peak position with peak; The model domain features include: ohmic resistance, polarization resistance, and polarization capacitance, which are obtained by fitting a first-order RC circuit model to the charging curve and estimating the parameters by the least square method; A multi-dimensional description of the battery degradation state including battery capacity attenuation, electrochemical polarization, and internal resistance growth is constructed by obtaining the data domain features and the model domain features.

3. The multi-source feature fusion dual-branch lithium battery health estimation method of claim 2, wherein: The data preprocessing of the extracted multi-source health factors in step S3 comprises: filtering, normalization, and time synchronization processing, and generating a standardized data matrix that can be directly used for feature extraction after filtering and normalization.

4. The multi-source feature fusion dual-branch lithium battery health estimation method of claim 3, wherein: The operation steps of the particle swarm optimization algorithm improved based on the principle of quantum mechanics comprise: Step S41: selecting the main charging section of the charging curve ; Step S42: simulate partial charging of the actual vehicle, select a partial charging section in the main charging section ;​ Step S43: introducing a quantum behavior model to make the position update of particles no longer depend on the speed, but be obtained by random collapse of a quantum wave function, and the simplified form is: ； In the formula, This indicates the number of iterations, i.e., the current iteration of the algorithm. Indicates the first The particle in the first The position of the generation, For the first The particle position is at the... The position of the generation, Indicates the first The best position found by the particle so far It is the group's average optimal position. To obey A uniformly distributed random variable over an interval. This is the contraction coefficient, a global adjustment parameter used to control the particle search range; Step S44: initializing the particle swarm, each particle representing a candidate section, calculating the Pearson correlation coefficient of each health factor and capacity in each particle section, calculating the objective function, updating the particle position, and outputting the optimal section when the iteration number or error meets the condition.

5. The multi-source feature fusion dual-branch lithium battery health estimation method according to claim 4, characterized in that: Objective function is represented as: ； wherein is the Pearson correlation coefficient between the health factor and the actual capacity, The larger the value is, the more health information the section contains.​ 6. The multi-source feature fusion dual-branch lithium battery health estimation method according to claim 5, characterized in that: The QPSO-CMNN lithium battery state estimation model obtained in step S5 comprises: Step S51: within the set range of charging voltage, the multi-source health feature sequence extracted from the optimal section searched by the particle swarm optimization algorithm improved based on the principle of quantum mechanics is randomly divided into a test set, a validation set and a training set; Step S52: training the double-branch deep learning estimation model CMNN by using the training set data; Step S53: outputting the final QPSO-CMNN battery state estimation model by using the loss function to optimize the model parameters, which is used for lithium battery SOH estimation.

7. The multi-source feature fusion dual-branch lithium battery health estimation method according to claim 6, characterized in that: The double-branch deep learning estimation model CMNN comprises: a local branch and a global branch; The local branch uses a convolutional neural network to extract highly sensitive local features. It adopts a two-level one-dimensional convolutional structure, consisting of standard convolution and depthwise separable convolution, combined with ReLU activation, GroupNorm and Dropout regularization to extract short-term features between adjacent loops. This branch focuses on extracting local change features between adjacent loops. The global branch adopts a state-space model architecture, which uses a state-space kernel and dynamic gating mechanism to model the nonlinear evolution law in long sequence data. By combining state updates and selective reads and writes, it can effectively capture the slow degradation trend of lithium batteries over thousands of cycles while maintaining linear reasoning complexity.

8. The multi-source feature fusion dual-branch lithium battery health estimation method according to claim 7, characterized in that: Local branch extraction of local features includes: Step a1: Apply standard one-dimensional convolution to the input feature matrix to extract preliminary local features; Step a2: Further extract fine-grained features using depthwise separable convolution; Step a3: Apply normalization and activation functions to the features output by convolution to obtain the feature output of the local branch.

9. The multi-source feature fusion dual-branch lithium battery health estimation method of claim 8, wherein: Global branch extraction of global features includes: Step b1: Input the dynamic block feature matrix and project the original feature dimensions to the hidden dimensions through linear mapping; Step b2: Stack residual blocks, each residual block consisting of root mean square normalization and MambaBlock concatenated; Step b3: Within each MambaBlock, the input features are expanded by a factor of 2 along the channel dimension and then split into two sub-streams, one of which enters the state branch and the other goes through the gated branch; Step b4: Finally, the hidden representation of the global branch is obtained; Step b5: Fuse the prediction results of the local branch and the global branch at the time step level, and output them through a fully connected layer.

10. The multi-source feature fusion dual-branch lithium battery health estimation method of claim 9, wherein: Step b3 also includes: One state branch: responsible for the actual temporal modeling; the state branch first goes through a one-dimensional depthwise separable convolution, then through the Swish activation function and is fed into the state space unit of the selective state space model for dynamic modeling. Another gated branch is used to adaptively modulate the result of the state branch. The gated branch only passes through one Swish layer, and its output is used as a gate weight to perform element-wise multiplication modulation on the output of the state branch, achieving a filtering effect similar to attention. It is then normalized through layer normalization to obtain the output of the current layer.

Images