A battery pack identification method and system based on a physical-probabilistic dual-track architecture

Abstract

This invention relates to a battery pack identification method and system based on a physics-probabilistic dual-track architecture, belonging to the field of data processing technology. The method includes: generating a virtual cell group containing differentiated initial parameters based on basic battery pack information; generating a battery pack system model based on the virtual cell group; acquiring dynamic parameter information output by the battery pack system model running in the virtual environment; inputting the dynamic parameter information output by the battery pack system model into a trained battery health state core predictor to obtain an estimated health state point value for each virtual cell; and obtaining a state estimation result based on the real-time parameter information of the actually running battery pack and the estimated health state points of each virtual cell. This invention enables accurate simulation and prediction of the dynamic external characteristics of the battery pack, the evolution of internal inconsistencies, and the probability distribution of the health state of each cell.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology and relates to a battery pack identification method and system based on a physical-probabilistic dual-track architecture. Background Technology

[0002] With the development of electric vehicles and energy storage systems, extremely high demands are placed on the accurate sensing, lifespan prediction, and safety management of battery pack status. Current mainstream battery pack modeling methods mainly suffer from the following limitations:

[0003] Accuracy and efficiency are difficult to balance: Models based on detailed electrochemical principles are highly accurate but computationally complex and difficult to run in real time; while simplified equivalent circuit models cannot accurately reflect the internal state of the battery, especially the inconsistencies between cells.

[0004] Model decoupling from reality: Traditional model parameters are usually fixed after offline calibration, making it impossible to track aging degradation and parameter drift throughout the entire battery life cycle. This causes the model to gradually deviate from the real battery, and the prediction accuracy decreases over time.

[0005] Crude health status assessment: Existing methods often treat the entire battery pack as a whole, or assess the health status based only on the worst cell, lacking a refined and probabilistic description of the health status and uncertainties of each cell in the pack, resulting in inaccurate risk warnings and lifespan predictions.

[0006] Heavy computational burden: If you try to build and run complex models for each of the hundreds of cells in the pack, it will generate catastrophic computational overhead, making it difficult to deploy in a real battery management system (BMS).

[0007] Therefore, there is an urgent need for a new physical-probabilistic dual-track modeling method that can achieve high-fidelity simulation, online updates, take into account computational efficiency, and provide a detailed assessment of the health status of battery packs. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to provide a battery pack identification method and system based on a physical-probabilistic dual-track architecture.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] One aspect of the present invention provides a battery pack identification method based on a physical-probabilistic dual-track architecture, the battery pack identification method based on the physical-probabilistic dual-track architecture comprising:

[0011] Obtain basic information about the battery pack;

[0012] A virtual cell group containing differentiated initial parameters is generated based on the basic information of the battery pack;

[0013] A battery pack system model is generated based on a virtual cell group containing differentiated initial parameters;

[0014] Obtain dynamic parameter information output by the battery pack system model running in a virtual environment;

[0015] Obtain the trained battery health state core predictor;

[0016] The dynamic parameter information output by the battery pack system model is input into the trained battery health state core predictor to obtain the estimated health state point value of each virtual cell.

[0017] Obtain real-time parameter information of the battery pack in actual operation;

[0018] The state estimation results are obtained based on the real-time parameter information of the actual operating battery pack and the estimated health state points of each virtual cell.

[0019] Furthermore, the basic information of the battery pack includes battery pack topology information, basic cell parameter information, historical cell test data information, and operating condition and system configuration information.

[0020] Furthermore, the step of generating a virtual cell group containing differentiated initial parameters based on the basic information of the battery pack includes:

[0021] The total number of virtual cells is obtained based on the battery pack topology information;

[0022] Based on the battery cell basic parameter information, obtain an initial parameter type list and a set of baseline values ​​for each parameter;

[0023] Each virtual cell has its own unique initial parameters based on the total number of virtual cells, the list of initial parameter types, and the set of baseline values ​​for each parameter. The unique initial parameters of each virtual cell form a virtual cell group containing differentiated initial parameters.

[0024] Furthermore, the generation of the battery pack system model based on the virtual cell group containing differentiated initial parameters includes:

[0025] Generate a set of individual ECM model instances for each virtual cell based on the virtual cell group containing differentiated initial parameters;

[0026] A cluster of battery pack individual model instances with complete electrical connections is generated based on the set of individual ECM model instances and the battery pack topology information.

[0027] An intermediate battery pack model with electro-thermal coupling is generated from the cluster of individual battery pack models that have completed electrical connections.

[0028] A prototype model of a battery pack with integrated electrothermal equalization function is generated based on the intermediate model of the electro-thermal coupled battery pack.

[0029] Define the input and output interfaces for the prototype model of the battery pack with integrated electrothermal equalization function, thereby obtaining the battery pack system model.

[0030] Furthermore, the core predictor for obtaining the trained battery health status includes:

[0031] Obtain a standardized battery health status training dataset;

[0032] Obtain the DKA-IC-SVGP model;

[0033] The DKA-IC-SVGP model is trained using a standardized battery health status training dataset to obtain a trained core predictor of battery health status.

[0034] Furthermore, the step of inputting the dynamic parameter information output by the battery pack system model into the trained battery health state core predictor to obtain the estimated health state point value of each virtual cell includes:

[0035] According to a preset cycle, the dynamic parameters of each virtual cell output by the model are collected in real time to obtain the dynamic parameter set of each virtual cell.

[0036] Feature extraction is performed on the dynamic parameter set of each virtual cell to obtain the standardized feature matrix of all virtual cells;

[0037] The standardized feature matrix of all virtual cells is input into the trained battery health state core predictor to obtain the health state point estimate of each virtual cell.

[0038] Furthermore, the step of obtaining the state estimation result based on the real-time parameter information of the actual operating battery pack and the estimated health state points of each virtual cell includes:

[0039] Generate a standardized real-time parameter sequence with timestamps based on the real-time parameter information of the battery pack in actual operation;

[0040] A time-aligned standardized real-time-virtual data set is generated based on the timestamped standardized real-time parameter sequence and dynamic parameter information;

[0041] An initialized SMC particle swarm is generated based on a time-stamped standardized real-time parameter sequence and the estimated health state points of each virtual cell.

[0042] Based on the time-aligned standardized real-time-virtual data set and the initialized SMC particle swarm, a structured state estimation result set is generated using the SMC-EM algorithm, and the structured state estimation result set is used as the state estimation result.

[0043] Furthermore, the generation of a structured state estimation result set using the SMC-EM algorithm based on the time-aligned standardized real-time-virtual data set and the initialized SMC particle swarm includes:

[0044] Step 11: Based on the initialized SMC particle swarm and the time-aligned standardized real-time-virtual data set, define a dual-scale particle update equation, dividing the particle filtering into cell-level microscale and battery pack-level macroscale. The microscale particle update formula is as follows:

[0045] =

[0046] in, This represents the updated microscopic particle state of the i-th battery cell in step t+1 and the k-th iteration. Let i be the microscopic particle state of the i-th cell in step t and iteration k. This represents the measured terminal voltage of the i-th cell in the k-th iteration. For the virtual model prediction of the terminal voltage of the i-th cell in step t and iteration k, Let be the set of particle states of all cells except the i-th cell in the t-th step and the k-th iteration; , These are data-driven coefficients and global constraint coefficients;

[0047] Step 12: Based on the updated particle swarm at the microscale, construct the macroscale parameter fusion formula:

[0048]

[0049] in, These are the model parameters updated in step t+1 and the kth iteration. This is a Bayesian prior based on the global parameters from the previous step; Let be the measured terminal voltage of the i-th cell in the k-th iteration; Let be the microscopic particle state of the i-th cell in step t+1 and the k-th iteration; These are the package-level global parameters for the (t+1)th step and the (k-1)th iteration; These are the model parameters to be optimized; It is a conditional probability distribution;

[0050] Step 13: In the expectation step of the expectation-maximization algorithm, the posterior expectation of the hidden state is calculated by weighted expectation of dual-scale particles. This posterior expectation is the hidden state estimation result.

[0051] Step 14: In the maximization step of the expectation-maximization algorithm, a Bayesian variational lower bound is introduced to construct the objective function:

[0052]

[0053] The updated model parameters are obtained by solving; where, These are the model parameters updated at step t+1; These are the current model parameters at step t; Let Y be the variational posterior distribution of the hidden states based on the model parameters at step t; Y is the set of measured data; X is the set of hidden states; and KL is the KL divergence. Let be the set of package-level macroscopic parameters at step t; This refers to the hidden state of a single battery cell.

[0054] Determine whether the model parameter update magnitude is less than the preset convergence threshold. If convergence is not achieved, feed back the updated model parameters to step 11 and repeat steps 11 to 14 until the convergence condition is met.

[0055] Step 15: Based on the converged model parameters and hidden state estimation results, integrate them according to the format of cell number, hidden state type, point estimate, probability distribution parameter, and timestamp to form a structured state estimation result set.

[0056] Furthermore, the step of training the DKA-IC-SVGP model using a standardized battery health status training dataset to obtain a trained core predictor of battery health status includes:

[0057] Step 21: Based on the standardized battery health status training dataset, construct the kernel space reconstruction criterion, map the original feature space to the adaptive kernel space through nonlinear transformation, and calculate the initial solution of the kernel space weight matrix Λ;

[0058] Step 22: Obtain the objective function for joint optimization of the induction point and kernel space;

[0059] Step 23: Optimize the initial solution of the kernel space weight matrix Λ by co-optimizing the objective function of the induced points and the kernel space, thereby obtaining the optimized kernel space weight matrix and the set of induced points;

[0060] Step 24: Substitute the optimized kernel space weight matrix and the set of induced points into the DKA-IC-SVGP model. Use an iterative training method, with the feature matrix of the training dataset as input and the real SOH as the label for model training. Calculate the prediction error every 10 iterations until the prediction error is less than a preset threshold or the number of iterations reaches a preset maximum value. Output the variational parameter update sequence during the training process and the final converged variational parameter set.

[0061] This application also provides a battery pack identification system based on a physical-probabilistic dual-track architecture, the battery pack identification system based on the physical-probabilistic dual-track architecture comprising:

[0062] A battery pack basic information acquisition module, which is used to acquire basic information about the battery pack;

[0063] A virtual cell cluster generation module is used to generate a virtual cell cluster containing differentiated initial parameters based on the basic information of the battery pack.

[0064] A battery pack system model generation module is used to generate a battery pack system model based on a virtual cell group containing differentiated initial parameters.

[0065] A dynamic parameter information acquisition module is used to acquire the dynamic parameter information output by the battery pack system model running in a virtual environment.

[0066] A battery health state core predictor acquisition module is used to acquire a trained battery health state core predictor.

[0067] The health state point estimation value acquisition module is used to input the dynamic parameter information output by the battery pack system model into the trained battery health state core predictor, thereby obtaining the health state point estimation value of each virtual cell.

[0068] A real-time parameter information acquisition module is used to acquire real-time parameter information of the battery pack in actual operation.

[0069] The state estimation result acquisition module is used to acquire state estimation results based on the real-time parameter information of the battery pack in actual operation and the estimated health state points of each virtual cell.

[0070] The present application enables accurate simulation and prediction of the dynamic external characteristics of the battery pack, the evolution of internal inconsistencies, and the probability distribution of the health status of each cell.

[0071] The beneficial effects of this invention are as follows:

[0072] (1) This invention handles fast-changing dynamics such as voltage and current response through the physical layer, while focusing on slow-changing health states through the probabilistic layer. The dual-track parallel architecture significantly reduces computing power requirements while ensuring simulation accuracy. The physical layer model uses vectorized parallel computing, supporting real-time calculation of hundreds of cells, while the probabilistic layer avoids redundant calculations through periodic batch updates, resolving the contradiction that high-precision models cannot run in real time in traditional methods. Technical support includes the generation of virtual cell groups and the design of second-order RC equivalent circuits for the battery pack system model, ensuring accurate dynamic response of external characteristics. Verification data shows that the simulation results are highly consistent with real operating conditions. For example, the total voltage of the battery pack decreases steadily with the discharge current without any sudden abnormalities, confirming the high fidelity of the model.

[0073] (2) By injecting manufacturing tolerances such as a 1.5% standard deviation in capacity and dynamic stress differences, the model can automatically reproduce the differentiation of SOC and voltage between cells and dynamically track the evolution of inconsistencies, providing a virtual experimental platform for balancing strategies. The core technology is an integrated electrothermal balancing model, which combines dual-mode balancing logic such as active and passive balancing to adjust parameter differentiation in real time. In the verification data, the cell SOC is continuously distributed, ranging from 0.64 to 0.74, reflecting the effectiveness of the manufacturing tolerances. The voltage range fluctuation is controlled within 0.126 to 0.13 volts, proving the model's accurate simulation and suppression capabilities for inconsistencies.

[0074] (3) Breaking through the limitations of traditional overall assessment, this invention outputs an estimated health status point and probability distribution (e.g., a 95% confidence interval) for each cell, quantifying the degree of aging and uncertainty, and helping to accurately locate the weakest cells. The technical support comes from the DKAIC-SVGP model, which enhances the distinguishability of aging features through kernel space reconstruction and induced point co-optimization, with a prediction error of less than 0.01. Validation data shows that the predicted SOH values ​​of the first 20 cells range from 0.85 to 0.95, with different probability distribution intervals, reflecting the refined assessment capability and providing a reliable basis for predictive maintenance.

[0075] (4) Through the SMCEM online joint identification framework, the model parameters can be adjusted in real time as the battery ages, avoiding the disconnect problem of traditional fixed parameter models and realizing the leap from static modeling to dynamic tracking. The key technology is the dual-scale particle update mechanism, such as micro-level cell and macro-level battery pack, combined with Bayesian variational lower bound optimization, which has a fast convergence speed and is completed within 50 iterations. In the validation data, the worst cell index changes dynamically, reflecting the model's real-time tracking ability of aging evolution, and confirming the effectiveness of parameter adaptive update.

[0076] (5) The dual-track architecture decouples the fast and slow conversion processes, and the probabilistic layer batch processing replaces the serial calculation of a single cell, reducing the computing power requirement by 30% compared to the traditional method, and is compatible with BMS hardware deployment. Technical support includes structured data interfaces such as API functions and vectorized operation optimization, supporting plug-and-play. Implementation cases show that the model runs stably with 273 cells in a 3-parallel 91-string topology, the calculation frequency matches the physical layer simulation step size of 1 second per step, there is no computing power overload, and the feasibility and economy of practical applications are enhanced.

[0077] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0078] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0079] Figure 1 This is a flowchart illustrating a battery pack identification method based on a physical-probabilistic dual-track architecture according to an embodiment of this application.

[0080] Figure 2 This is a flowchart illustrating the operation of SMC-EM according to an embodiment of this application;

[0081] Figure 3 This is a schematic diagram of the variation curves of pack voltage and current within a battery pack according to an embodiment of this application;

[0082] Figure 4 This is a schematic diagram of the cell SOC distribution at the end of the simulation of an embodiment of this application;

[0083] Figure 5 This is a schematic diagram of the dynamic changes in voltage inconsistency according to an embodiment of this application;

[0084] Figure 6 This is a schematic diagram illustrating the predicted SOH value and uncertainty according to an embodiment of this application;

[0085] Figure 7 This is a schematic diagram of the worst-case dynamic changes of a battery cell according to an embodiment of this application. Detailed Implementation

[0086] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0087] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0088] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0089] like Figure 1 The battery pack identification method based on the physical-probabilistic dual-track architecture shown includes:

[0090] Obtain basic information about the battery pack;

[0091] In this embodiment, the basic information of the battery pack includes battery pack topology information, basic cell parameter information, historical cell test data information, and operating condition and system configuration information.

[0092] Specifically, the battery pack topology information is obtained in the following ways: by consulting the battery pack design drawings, product specifications or BMS configuration files, the series and parallel topology of the battery pack (such as N series M parallel) is determined, and the total number of cells is determined (total number of cells = number of series N × number of parallel M); system-level parameters such as the battery pack nominal voltage and rated capacity (such as 91.5Ah in the example) are recorded simultaneously.

[0093] The basic parameters of the battery cell are obtained as follows: extract the nominal capacity (e.g., 30.5Ah in this example), nominal internal resistance (e.g., 0.2mΩ in this example), and open-circuit voltage-state-of-charge-temperature (OCV(SOC,T)) basic curves of a single battery cell from the technical manual or factory test report provided by the battery cell supplier; at the same time, collect the statistical laws of battery cell manufacturing tolerances, including key tolerance data such as capacity standard deviation (e.g., 1.5%), internal resistance coefficient standard deviation (e.g., 5%), and allowable range of OCV curve deviation.

[0094] Historical test data of battery cells are obtained in the following ways: through a laboratory charge and discharge test platform, cyclic aging data (such as changes in internal resistance and capacity decay under different cycles) and dynamic operating condition response data (such as voltage changes under different currents and temperatures) of the same batch of battery cells are collected, or historical operating data of BMS (including voltage, current, temperature and SOC records) of similar battery packs that have been in service are retrieved to form a battery cell performance database.

[0095] Operating condition and system configuration information are obtained in the following ways: basic operating condition data of the battery pack in actual application scenarios are collected, including typical charging and discharging current range, ambient temperature fluctuation range, and cycle working period; at the same time, the battery pack balancing logic type (active / passive) and data sampling related parameters (such as time step of 1.0 second and SOH prediction interval of 300 seconds in the embodiment) are specified to provide configuration basis for subsequent model construction and online identification.

[0096] A virtual cell group containing differentiated initial parameters is generated based on the basic information of the battery pack;

[0097] In this embodiment, generating a virtual cell group containing differentiated initial parameters based on the basic information of the battery pack includes:

[0098] The process of generating a virtual cell group containing differentiated initial parameters based on the basic information of the battery pack includes:

[0099] The total number of virtual cells is obtained based on the battery pack topology information. Specifically, based on the series and parallel topology, the total number of virtual cells is calculated as: total number of virtual cells = number of series cells N × number of parallel cells M, ensuring consistency with the actual number of cells in the battery pack.

[0100] Based on the cell's basic parameter information, an initial parameter type list and a set of reference values ​​for each parameter are obtained. Specifically, the types of initial parameters and their corresponding reference values ​​are extracted and determined from the cell technical manual, factory test report, or historical test data in the battery pack's basic information. The parameter types include at least: the basic curve of the three-dimensional lookup table OCV(SOC,T) for open circuit voltage-state of charge-temperature, the nominal value of the ohmic internal resistance R0, the nominal values ​​of the polarization resistance R1 and capacitor C1, the nominal values ​​of the polarization resistance R2 and capacitor C2, and the nominal value of the rated capacity of a single cell (e.g., a nominal capacity of 30.5Ah and a nominal internal resistance of 0.2mΩ for a single cell).

[0101] Each virtual cell has its own unique initial parameters based on the total number of virtual cells, the list of initial parameter types, and the set of baseline values ​​for each parameter. The unique initial parameters of each virtual cell form a virtual cell group containing differentiated initial parameters.

[0102] In this embodiment, each virtual cell's unique initial parameters are generated based on the total number of virtual cells, the initial parameter type list, and the set of baseline values ​​for each parameter. The unique initial parameters of each virtual cell constitute a virtual cell group containing differentiated initial parameters, including:

[0103] Based on the input tolerance rules or historical data, set differentiated distribution types and fluctuation ranges for each type of initial parameter, and prioritize the use of normal distribution (such as capacity standard deviation of 1.5% and internal resistance coefficient standard deviation of 5%). Clarify the allowable range of small deviations in the OCV curve to ensure that the distribution rules conform to the actual cell manufacturing differences, thereby obtaining the differentiated distribution rules (including distribution type, standard deviation / fluctuation range) of each initial parameter.

[0104] Using the Monte Carlo sampling method, and based on a predefined distribution rule, each virtual cell is randomly sampled individually. Differential fluctuations are then assigned to each virtual cell based on baseline parameter values, generating a unique first... i A dedicated initial parameter set for each virtual battery cell (Each cell's parameter set includes all preset parameters such as OCV(SOC,T), R0, R1 / C1, R2 / C2, and rated capacity).

[0105] The initial parameter set of each virtual cell is organized and archived according to its serial number and stored as a structure array. This array completely records the initial state of all virtual cells, forming a virtual cell group containing differentiated initial parameters.

[0106] A battery pack system model is generated based on a virtual cell group containing differentiated initial parameters;

[0107] In this embodiment, a virtual cell group containing differentiated initial parameters (stored in the form of a structured array) is used. The parameters of each cell at least include: open circuit voltage-state of charge-temperature three-dimensional lookup table OCV(SOC,T) (including calibration values ​​at different SOC and temperatures), ohmic internal resistance R0 (unit: mΩ), polarization resistor R1 and polarization capacitor C1, polarization resistor R2 and polarization capacitor C2, and rated capacity of a single cell. Q n (Unit: Ah) All parameters vary based on manufacturing tolerances or historical data (e.g., capacity standard deviation 1.5%, internal resistance coefficient standard deviation 5%).

[0108] In this real-world example, generating a battery pack system model from a virtual cell array containing differentiated initial parameters includes:

[0109] Define a second-order RC circuit topology: The model of each cell consists of an OCV voltage source + an ohmic internal resistance R0 + two series RC polarization branches (R1-C1, R2-C2), where the R1-C1 branch describes the fast polarization response and the R2-C2 branch describes the slow polarization response. The components form a complete circuit structure according to series logic.

[0110] Encapsulating Independent Atomic Subsystems: The differentiated initial parameters of each cell are bound to the aforementioned topology and encapsulated as independent atomic subsystems. Each subsystem has a built-in parameter storage module (which stores the corresponding cell's OCV(SOC,T) lookup table and R0, R1, R2, C1, C2, ...). Q n Only standardized input interfaces (receiving charge / discharge current, ambient temperature, initial SOC) and output interfaces (output voltage, real-time SOC, etc.) are exposed to ensure that the subsystem is reusable and easy to integrate later.

[0111] Determine the state equation and solution method: Based on the principle of second-order RC equivalent circuit, the polarization voltage differential equation is used to describe the dynamic characteristics of polarization response. The terminal voltage calculation equation integrates OCV, ohmic voltage drop and polarization voltage drop. The SOC update equation is derived based on the Coulomb counting method. The numerical solution adopts the well-known Euler method or Runge-Kutta method, and the discretization time step is set (e.g., 1.0 second) to balance the solution accuracy and real-time performance.

[0112] Supports batch parallel computing: Employing vectorized computational logic, the model solution process can accept batch cell input data in array format (current, temperature, and initial SOC of all cells), and perform state equation calculations for all cells in parallel. This avoids the waste of computing power caused by calculating each cell individually, adapting to the simultaneous operation requirements of hundreds of cells. Ultimately, it obtains a set of individual second-order RC equivalent circuit model (ECM) instances corresponding one-to-one with the virtual cell. Each instance contains a built-in set of differentiated parameters, standardized input / output interfaces, and complete state solution logic, which can be directly used for electrical connections and multi-scale coupling in subsequent system integration layers.

[0113] In this embodiment, generating a cluster of electrically connected battery pack unit models based on the set of individual ECM model instances and the battery pack topology information includes:

[0114] Digital modeling of topology: transforming series and parallel rules into a two-dimensional topological inclination matrix ConnMatrix [(N×M)×2], each row of the matrix stores the target parallel group number and the target series position number of a single cell, for example... ConnMatrix (k,1)=p、 ConnMatrix (k,2)=s represents that the k-th cell belongs to the t-th cell. p Parallel group, the first s Series connection position. ConnMatrix This represents a two-dimensional matrix that associates the series and parallel topologies of a battery pack. Each row stores the parallel group number and series position number of a single cell. Indicates the first The parallel group number to which each battery cell belongs; Indicates the first The serial position number of each battery cell.

[0115] The topology matrix mapping-port binding algorithm, combined with commonly used engineering modeling software, is used to implement programmatic connections. Specific steps are as follows:

[0116] Software selection: MATLAB / Simulink (via Simscape Electrical module library API), Python (NumPy+NetworkX network modeling library), or ANSYS Twin Builder (parametric modeling interface) are used, all of which support non-graphical programmatic calls to electrical interfaces.

[0117] Parallel connection implementation: Traverse the topological inclination matrix, group by parallel group number, and for each group of M cells, call the software's port binding function Port_Bind() to index the positive terminals of all cells. Bind to the same virtual public node Negative index Bind to the same virtual public node This enables electrical short-circuiting; simultaneously, through the software's built-in current distribution algorithm, it automatically distributes the total current of the parallel group. = / N ( The total current (total current) is evenly distributed to each cell in the group, i.e. = / M. Among them. Indicates the first Index of the positive electrical port of each battery cell; Indicates the first Index of the negative electrical port of each battery cell; Indicates the first The positive common electrical node of a parallel group; Indicates the first The negative common electrical node of a parallel group; This represents the total charging and discharging current of a single parallel group; This indicates the total charging and discharging current of the entire battery pack; Indicates the first The input charging and discharging current of each battery cell.

[0118] Series connection implementation: Sort all parallel groups by series position number, call the port binding function Port_Bind(), and bind the negative common node of the s-th parallel group. The positive common node of the (s+1)th parallel group Binding forms a series link; the software automatically superimposes the voltages of each parallel group. = (With consistent cell voltages within the group), the total package voltage is obtained. =Σ . Indicates the first i The output voltage of each battery cell. Indicates the first Each series position corresponds to the terminal voltage of the parallel group; Indicates the first The output voltage of each battery cell; This indicates the total terminal voltage of the entire battery pack; Indicates the first Each series position corresponds to the negative common electrical node of the parallel group; Indicates the first Each series position corresponds to the positive common electrical node of the parallel group; This represents a structure array that stores the differentiated initial parameters of 273 virtual battery cells.

[0119] Connection relationship verification and optimization: The electrical rule checking tools of the software (such as Simulink's ModelAdvisor and NetworkX's connectivity analysis function is_connected()) are called to verify that: all cells are connected to the backbone network (no isolated nodes); there are no open circuits in the series links and no short circuits in the parallel nodes; the deviation between the total voltage / total current calculation results and the theoretical values ​​of the series and parallel electrical formulas is ≤0.05%; if the requirements are not met, the node binding relationship is automatically adjusted by the algorithm until the verification is passed.

[0120] Generate reusable electrical network files: solidify the final node binding relationships, current / voltage transfer rules, and cell interface mapping tables into standardized files (such as XML and MAT files), embed them into the battery pack model, and support seamless integration and interface calls for subsequent thermal coupling models and balancing logic. Finally, obtain a cluster of battery pack individual model cells with complete electrical connections, including: a digital electrical topology network (including virtual common nodes and cell port binding relationships); automatic current / voltage allocation and superposition logic; standardized input interfaces (receiving total pack current and ambient temperature) and output interfaces (outputting total pack voltage, voltage of each parallel group, and terminal voltage / current / SOC of each cell); and an electrical connection verification report, which can be directly used for subsequent multi-physics coupling integration.

[0121] In this embodiment, the battery pack individual cell model cluster with complete electrical connections (including series and parallel electrical topology, current of each cell) I cell ( i Terminal voltage V cell ( i (SOC output interface, and cell physical arrangement coordinates).

[0122] Battery cell thermal characteristic parameters: thermal conductivity λ (Unit: W / (m)) K) Specific heat capacity c p (Unit: J / (kg)) K), cell density ρ (Unit: kg / m³) 3 ), polarization heat generation coefficient (related to charging and discharging current and polarization voltage);

[0123] Heat dissipation structure information: heat dissipation method (air cooling / liquid cooling / natural heat dissipation), heat dissipation surface area A (Unit: m) 2 ), heat dissipation medium parameters (air cooling air velocity) v Liquid cooling flow rate and specific heat capacity; environmental heat transfer coefficient for natural heat dissipation. h 0) Cell spacing d (Unit: mm) Thermal conductivity of the insulation material of the enclosure.

[0124] In this embodiment, generating an electro-thermal coupled intermediate battery pack model based on a cluster of battery pack individual model clusters with completed electrical connections includes:

[0125] Establish a cell instantaneous heat generation model: Using a composite model of Joule's law and polarization heat generation, based on the cell current, ohmic internal resistance and polarization voltage output by the electrical cluster, the instantaneous heat generation power of a single cell is quantitatively calculated to realize electrothermal mapping.

[0126] Constructing a heat conduction model between battery cells: Based on Fourier's law of heat conduction, a heat conduction matrix is ​​established by combining the physical arrangement coordinates of the battery cells. It can depict the heat transfer relationship between adjacent cells and reproduce the heat distribution evolution within the package.

[0127] Establish a cell-environment heat dissipation model: Newton's law of cooling is used to calculate the convective heat dissipation power between the cell and the environment. The heat transfer coefficients of air cooling and liquid cooling are corrected by wind speed and flow rate, respectively. At the same time, the heat loss of the casing insulation is taken into account to correct the total heat dissipation power.

[0128] Achieving bidirectional electro-thermal coupling integration: By binding the thermal model to the electrical backbone network through a parameter interface, a closed-loop coupling of electrothermal generation → thermal influence on electricity is formed; among which, the SOH-temperature coupling internal resistance correction formula accurately reflects the synergistic effect of aging and temperature:

[0129]

[0130] In the formula, Reference temperature The initial internal resistance when SOH=1 This is the aging correction factor for the cell's internal resistance. k It is a non-linear decay exponent (with a value of 1.2 to 1.8). This represents the health status value of the i-th cell. This indicates that the i-th cell is at a temperature of T i Health status SOH i Real-time ohmic internal resistance; Temperature correction factor indicating cell internal resistance;

[0131] Coupled Simulation Solution and Verification: Engineering software such as ANSYS Icepak and MATLAB / Simulink Simscape Thermal were used to simultaneously solve the electrical state equations and heat conduction equations using the finite volume method (FVM). The time step was consistent with the electrical model (e.g., 1.0 s). The verification standard was: the deviation of the cell temperature rise rate from the theoretical value during high-current discharge was ≤5%, ensuring the coupling logic was error-free. The final intermediate model of the electro-thermal coupled battery pack was obtained, including: an electro-thermal bidirectional coupling logic module; and a real-time temperature output interface T for each cell. i Thermal conduction matrix and heat dissipation parameter configuration; standardized input interfaces (total current, ambient temperature, heat dissipation medium parameters) and output interfaces (total battery pack voltage, voltage / SOC / temperature of each cell, overall temperature distribution of the pack).

[0132] In this embodiment, generating a prototype battery pack model with integrated electrothermal equalization function based on the intermediate model of the electro-thermal coupled battery pack includes:

[0133] Establish a real-time monitoring module for equilibrium state: Collect data in batches from the output of the electro-thermal coupling intermediate model via a standardized interface. , , Calculate the voltage range of the cells within the package. Cell SOC range Simultaneously verify whether the temperature of each cell is within the safe equalization window, and output the equalization status flag and the target cell index (high-potential cell group). Low-potential battery pack This ensures that monitoring and physical layer data are updated synchronously.

[0134] Embedded dual-mode equalization control logic module:

[0135] Passive load balancing: when In the absence of an active balancing fault, the high-potential battery cell and the discharge resistor are controlled by a relay. When the circuit is turned on, excess power is released using a resistor-based energy dissipation balancing principle, with real-time feedback of discharge current and energy consumption data, until the voltage difference within the series is ≤ Automatic circuit disconnection. Among them, This represents the threshold value for the trigger voltage range in passive equalization. The threshold voltage range indicating the termination of the equalization function.

[0136] Active load balancing: when or At that time, based on the principle of inductive / capacitive energy coupling, the high-potential battery cell... Directed transfer of electrical energy to low-potential cells The transfer current does not exceed / ( (This refers to the average terminal voltage of the cells within the same series branch), achieving energy recovery-based balancing. This represents the threshold for the active equalization trigger SOC range; This represents the threshold value of the trigger voltage range for active balancing. This indicates the maximum transmission power of the active equalization module.

[0137] Achieving balanced-electrical-thermal synergistic coupling: The additional heat generated during the balancing process (heating from passive balancing resistors and losses from active balancing modules) is fed back to the thermal model in real time through the thermal coupling interface, participating in the calculation of the internal temperature field according to Fourier's law of heat conduction and Newton's law of cooling; simultaneously, the thermal model outputs T... i Inverse constraint equilibrium logic, when Automatically reduce equalization power when the temperature exceeds 45°C. Equalization should be paused when the temperature exceeds 50°C to prevent heat accumulation from accelerating cell aging.

[0138] Equalization logic validity verification: By comparing the changes in cell voltage / SOC range and temperature field stability before and after equalization (cell temperature rise ≤2℃ / min during equalization), the execution effect of the equalization strategy is verified. If the range still does not meet the requirements after equalization, a secondary equalization process is triggered until the termination threshold is reached. Finally, a prototype battery pack model with integrated electro-thermal equalization functionality is obtained, including: an equalization status monitoring module, a dual-mode equalization control module, and electro-thermal-equalization collaborative coupling logic; real-time output interfaces (voltage / SOC / temperature, equalization status flag, energy transfer / loss data for each cell); standardized input interfaces (total current, ambient temperature, heat dissipation medium parameters, equalization strategy configuration parameters); and an equalization log recording module (recording trigger time, equalization duration, and range improvement), which can be directly used for subsequent model input / output interface definition and system integration.

[0139] In this embodiment, a model input / output interface is defined for the prototype battery pack model with integrated electrothermal equalization function, thereby obtaining the battery pack system model, including:

[0140] Define the model's external input interface (receiving operating condition data such as total current and ambient temperature) and output interface (outputting the total voltage of the battery pack and detailed status of each cell, such as voltage, SOC, and temperature); optimize the model's execution efficiency through vectorized operations to ensure parallel computation of batch cells, and finally obtain a complete battery pack system model.

[0141] The range of parameters collected for the battery pack system model running in the virtual environment (based on a dual-track architecture data stream) includes:

[0142] Physical layer rapidly changing dynamic parameters (real-time output): total battery pack voltage and current; terminal voltage of each cell. SOC, real-time temperature Polarization voltage , No. i Real-time ohmic internal resistance of each cell Balance status indicators (whether balanced, balance mode, balance current). Indicates the first The polarization voltage of the first-order RC polarization branch of a battery cell; Indicates the first The polarization voltage of the second-order RC polarization branch of the battery cell.

[0143] Probabilistic layer slowly varying health parameters (periodic output): State of Health (SOH) point estimate for each cell, SOH probability distribution (mean / variance); SOH range of cells within the package, index of worst-performing cell.

[0144] Derivative parameters (calculated in real time): Voltage range of cells within the package SOC range Instantaneous heat generation power of a single cell and average temperature of the package.

[0145] In this embodiment, obtaining the dynamic parameter information output by the battery pack system model running in the virtual environment includes:

[0146] Configure a tiered data collection trigger mechanism:

[0147] Rapidly changing parameters: synchronized with the model simulation time step (e.g., 1 second / time), using simulation step trigger + batch reading mode to ensure real-time matching with the physical layer dynamic response.

[0148] Slow-varying parameters: triggered at a preset period (e.g., 300 seconds / time), consistent with the SOH prediction period of the probability layer, to avoid redundant acquisition occupying computing power.

[0149] Derived parameters: Calculated in real time based on rapidly changing parameters and output synchronously with the rapidly changing parameters.

[0150] Design a standardized data acquisition interface:

[0151] Using the model's built-in batch data output interface, all parameters are encapsulated into a matrix (dimension of "time step × parameter type × number of cells") in a structured format of timestamp + parameter type + cell index + value, supporting the reading of all parameters of 273 cells (such as a 3 parallel 91 string topology) at one time.

[0152] The interface is compatible with virtual simulation environments such as MATLAB / Simulink and Python. It can be directly called through API functions (such as get_model_output()) without manual extraction, ensuring data timing alignment.

[0153] Data transmission and caching:

[0154] Real-time caching: A circular buffer is used to store rapidly changing parameters of the most recent 1000 time steps, supporting real-time viewing and anomaly backtracking during model operation.

[0155] Persistent storage: The data is split into hourly segments and stored as CSV or HDF5 format files. The file name includes the simulation task ID and the start timestamp. The file contains a parameter description header and a data checksum to ensure traceability.

[0156] Data validity verification:

[0157] Range verification: Based on the rated parameters of the battery cell, set thresholds (such as voltage 1.5~4.2V, temperature -20~60℃), remove abnormal values ​​that exceed the reasonable range, mark them as "invalid data" and retain the timestamp.

[0158] Logical verification: Verify the consistency of parameters under series and parallel topologies (e.g., total series voltage = sum of voltages in each series, total parallel current = sum of currents in each cell). If the deviation exceeds 5%, trigger an alarm and record the verification log.

[0159] Data integration and correlation:

[0160] By associating fast-changing parameters, slow-changing parameters, and derived parameters with timestamps, a "single time step full parameter snapshot" is generated, which supports filtering and querying by cell index (e.g., the 100th cell) or parameter type (e.g., SOH).

[0161] Obtain the trained battery health state core predictor;

[0162] In this embodiment, obtaining the trained battery health state core predictor includes:

[0163] Obtain a standardized battery health status training dataset; specifically, use laboratory battery pack charge and discharge condition data (including current, voltage, temperature, SOC change curves, capacity decay records, etc. under different cycle numbers, different current intensities, and different ambient temperatures) to clean the input data (remove outliers and fill in missing data), extract key performance indicators for each cell under different operating conditions (such as average current, temperature rise slope, internal resistance change rate, cumulative charge and discharge throughput, etc. within the cycle), and construct a standardized model training dataset to ensure that the data covers the aging characteristics of the entire battery life cycle.

[0164] A DKA-IC-SVGP model is obtained; the DKA-IC-SVGP model is trained using a standardized battery health state training dataset to obtain a trained core predictor of battery health state. Specifically, a deep kernel approximation-induced point-sparse variational Gaussian process (DKA-IC-SVGP dynamic kernel adaptation and incremental compression sparse variational Gaussian process) model framework is constructed. Using cell features from the training data as input and actual SOH (battery health state) as output, the kernel matrix (characterizing feature correlation), induced points (reducing computational complexity), and variational parameters (optimizing probability prediction accuracy) of the model are determined through iterative optimization, enabling the model to learn the mapping relationship between cell features and health state degradation.

[0165] In this embodiment, the DKA-IC-SVGP model is obtained; the DKA-IC-SVGP model is trained using a standardized battery health status training dataset to obtain a trained battery health status core predictor, including:

[0166] Step 21: Based on the standardized battery health status training dataset, construct the kernel space reconstruction criterion, map the original feature space to the adaptive kernel space through nonlinear transformation, and calculate the initial solution of the kernel space weight matrix Λ;

[0167] Specifically, based on a standardized battery health status training dataset, a kernel space reconstruction criterion is constructed. The original feature space is mapped to an adaptive kernel space through a nonlinear transformation, and the formula for generating the reconstruction kernel function is defined:

[0168] ,in, The reconstructed kernel function; For the first i , j The original feature vector of each battery cell (including aging-related features such as changes in current, temperature, and internal resistance); For the first i The transpose of the eigenvectors mapped to individual battery cells; For nonlinear mapping functions ( , The angular frequency parameter of the nonlinear mapping function. for), The kernel space weight matrix is ​​used for weighted reconstruction of the high-dimensional feature space; it is obtained by solving eigenvalue decomposition. (U is the eigenvector matrix obtained from eigenvalue decomposition, and Σ is the eigenvalue diagonal matrix), output the reconstructed kernel function expression and the kernel space weight matrix. The initial solution; This represents the reconstructed kernel function, which characterizes the similarity of the feature vectors of the two battery cells.

[0169] Step 22: Obtain the objective function for joint optimization of the induction point and kernel space;

[0170] In this embodiment, based on the reconstruction kernel function output in step 21, a co-optimization objective function for the induced point and kernel space is designed:

[0171]

[0172] in, To predict the covariance matrix, K Z represents the full kernel matrix of the original feature space, and Z is the set of induced points. For noise variance, I Let λ be the identity matrix and λ be the regularization coefficient. Output the expression of the collaborative optimization objective function and the initial values ​​of each parameter. To minimize the objective function operation (the optimization objects are the set of induced points Z and the kernel space weight matrix) Y represents the set of measured data; This is the transpose of the measured dataset; λ The regularization coefficient is used. To predict the inverse of the covariance matrix; Kernel space weight matrix The trace (the sum of the elements on the main diagonal of the matrix, used to constrain matrix complexity); Cross kernel matrix transpose; It is the inverse matrix of the induced point kernel matrix; The kernel matrix between the induced points; This is the cross-core matrix between the induction point and the original characteristics of the battery cell;

[0173] Step 23: Optimize the objective function and kernel space weight matrix by co-optimizing the induced points and kernel space. The initial solution is optimized to obtain the optimized kernel space weight matrix and the set of induced points; specifically, the objective function of step 2 is used as the optimization object, and the closed-form solution of the collaborative optimization is solved by the Lagrange multiplier method:

[0174]

[0175]

[0176] in, The nonlinear mapping characteristic matrix for all battery cells ( , N (Total number of virtual battery cells). This is the transpose of the characteristic matrix of the nonlinear mapping; The regularization coefficient is used. It is the identity matrix; is the inverse matrix; Y is the set of measured health status labels (the actual SOH values ​​of all cells, used as label data for model training); This is the transpose of the measured set of health status labels. To find the set of induction points that minimize the objective function; The trace of the matrix (the sum of the elements on the main diagonal, which quantifies the overall energy of the matrix and is used here to measure the prediction error). z Represents the set of induced points (sparsely sampled points in a high-dimensional feature space); Represents the set of optimal induction points after collaborative optimization;

[0177] Output the optimized kernel space weight matrix With the set of induction points ;

[0178] Step 24: Substitute the optimized kernel space weight matrix and the set of induced points into the DKA-IC-SVGP model. Use an iterative training method, with the feature matrix of the training dataset as input and the real SOH as the label for model training. Calculate the prediction error every 10 iterations until the prediction error is less than a preset threshold or the number of iterations reaches a preset maximum value. Output the variational parameter update sequence during the training process and the final converged variational parameter set.

[0179] In this embodiment, the optimized kernel space weight matrix and the set of induced points are substituted into the DKA-IC-SVGP model. An iterative training method is used, with the feature matrix of the training dataset as input and the true SOH as the label. The prediction error is calculated every 10 iterations until the prediction error is less than a preset threshold or the number of iterations reaches a preset maximum. The output includes the variational parameter update sequence during the training process and the final converged variational parameter set:

[0180] Reconstruct the optimized kernel function output from step 23 (including...) ) and the set of induction points Substitute the DKA-IC-SVGP model and use an iterative training method (update the variational parameters in each iteration). Use the feature matrix of the training dataset as input and the real SOH as the label to train the model. Calculate the prediction error once every 10 iterations until the prediction error is less than a preset threshold (e.g., mean absolute error ≤ 0.01) or the number of iterations reaches a preset maximum value (e.g., 100 times). Output the variational parameter update sequence during the training process and the final converged set of variational parameters. This is the optimal kernel space weight matrix after co-optimization.

[0181] The converged core parameters (including the optimized kernel matrix and the set of induced points) are then used. Kernel space weight matrix The parameters (variable parameters) are stored in a structured file format and the trained battery health state core predictor file is output.

[0182] The training method for the DKA-IC-SVGP model proposed in this application has the following advantages over existing methods:

[0183] Traditional model training employs a step-by-step strategy: first fixing the kernel function and then optimizing the induced point positions, or first fixing the induced points and then optimizing the kernel function parameters. This approach presents a contradiction: the core function of the kernel function is to map the original features (such as changes in cell current, temperature rise, and internal resistance) to a high-dimensional feature space, while the role of the induced points is to sample key nodes in the high-dimensional space to reduce computational complexity. If the two are optimized independently, the effective feature region mapped by the kernel function may not coincide with the sparse region sampled by the induced points. This can lead to the model's inaccurate capture of the nonlinear characteristics of cell aging (such as the abrupt change in SOH from 0.95 to 0.85) and the complex correlation features of operating stress.

[0184] This application employs an integrated design of kernel space reconstruction and induced point co-optimization, simultaneously solving for the kernel space weight matrix Λ and the induced point set Z during training. The reconstructed kernel function actively enhances the discriminative power of key features such as cell aging and temperature stress through nonlinear mapping, while the co-optimized objective function ensures that the induced points are accurately sampled within the effective feature-dense region after kernel function mapping (such as the key interval of SOH nonlinear decay). The two complement each other, enabling the model to accurately capture the SOH evolution pattern of the entire lifecycle of the cell from "new cell → aged cell," solving the problem of large prediction bias for nonlinear features in traditional models.

[0185] Traditional high-precision health prediction models (such as complex neural networks and full Gaussian process models) either require a large number of parameter iterations, resulting in high computational power consumption (making them unusable on BMS hardware), or simplify the model to reduce computational power (such as fixing simple kernel functions and reducing the number of induced points), leading to a decrease in prediction accuracy. It is difficult to balance "accuracy" and "efficiency".

[0186] This application reconstructs the kernel function through multi-dimensional nonlinear mapping, which enhances the expressive power of the original features without increasing model complexity. It is better at distinguishing the differentiated aging characteristics of different battery cells than traditional single kernel functions (such as the RBF kernel); the collaboratively optimized closed-form solution ( , The method uses the Lagrange multiplier method to solve the problem in one step, avoiding the loss of parameter optimality caused by iterative trial and error, and ensuring that the prediction error is controlled at a low level (e.g., mean absolute error ≤ 0.01).

[0187] The collaborative optimization of the inducer points avoids redundant sampling (inducer points are only placed in the effective feature area), reducing the number of inducer points by more than 30% compared with traditional random sampling. At the same time, the batch parallel processing logic supports the input of feature matrices of hundreds of cells at one time and the parallel output of SOH prediction results. The calculation frequency matches the probability layer update cycle (e.g., 300 seconds / time), and will not occupy the real-time simulation computing power of the physical layer.

[0188] Traditional SOH prediction models are mostly trained based on data from a single battery cell. When extended to hundreds of cells in a battery pack, the model is prone to generalization bias due to feature differentiation caused by manufacturing tolerances (such as capacity standard deviation of 1.5% and internal resistance standard deviation of 5%) and differences in operating stress. It cannot accurately output the personalized health status of each cell.

[0189] The training process in this application is based on a batch dataset containing inconsistencies (simulating the differentiated characteristics of hundreds of battery cells). The co-optimized kernel space weight matrix Λ automatically adapts to the differences in feature distribution among different battery cells. For example, for battery cells with rapid capacity decay, the kernel function will strengthen the weights of features such as cumulative throughput and temperature rise slope; for battery cells with large internal resistance drift, it will strengthen the weight of the "internal resistance change rate" feature. At the same time, the distribution of the induced point set Z will cover the feature variation range of all battery cells, ensuring that the features of each battery cell can be effectively mapped and sampled during batch input, avoiding prediction bias caused by feature differences.

[0190] Traditional step-by-step optimization of models is prone to parameter oscillations (e.g., after optimizing the induction point, the kernel function parameters need to be readjusted and iterated repeatedly), and the convergence speed is slow (requiring more than 500 iterations); moreover, the kernel function, induction point, and variational parameters after training are stored in a scattered manner, and the interface needs to be manually adapted during deployment, resulting in high engineering application costs.

[0191] The co-optimization of the kernel space and the induced point in this application is solved by a closed-form solution, which does not require repeated iterative adjustments. The number of convergences is reduced from more than 500 to less than 100, and there is no parameter oscillation. The training results are reproducible.

[0192] After training, the kernel matrix and the optimized , The variational parameters are uniformly fixed into structured files, which can be directly called through API functions (such as load_soh_predictor()) without manual adaptation. At the same time, the model interface is compatible with mainstream virtual simulation environments and BMS hardware platforms such as MATLAB / Simulink and Python. During deployment, only the files need to be loaded to run, without the need for repeated training.

[0193] In this embodiment, the dynamic parameter information output by the battery pack system model is input into the trained battery health state core predictor to obtain the estimated health state point values ​​of each virtual cell, including:

[0194] According to a preset period (such as 300 seconds for SOH prediction in the example), the dynamic parameters of each virtual cell output by the model are collected in real time to obtain the dynamic parameter set of each virtual cell, including the current internal resistance (R0, R1, R2), real-time temperature, working stress history (such as the charging and discharging current sequence of the past N cycles), cumulative charging and discharging throughput, temperature rise slope, internal resistance change rate and other key parameters to ensure that the parameters cover the cell aging related characteristics.

[0195] Feature extraction is performed on the dynamic parameter set of each virtual cell to obtain a standardized feature matrix for all virtual cells. Specifically, for each virtual cell, the dynamic parameters are processed according to the feature extraction rules. Statistical features (such as average current and peak current) are calculated for time-series parameters, and trend features (such as internal resistance change rate and temperature rise slope) are calculated for variable parameters. All features are integrated into a standardized feature vector of fixed dimension. Finally, an N×K dimensional feature matrix is ​​formed by combining the cell numbers (N is the total number of virtual cells, and K is the feature dimension).

[0196] The standardized feature matrices of all virtual cells are input into a trained battery health state core predictor to obtain the health state point estimates for each virtual cell. Specifically, the feature matrices are input into the core predictor in batches, and the prediction function is called to perform batch forward propagation operations. Utilizing the probabilistic prediction capability of the DKA-IC-SVGP model, the health state (SOH) related results of all virtual cells are calculated in parallel at once, avoiding the computational waste caused by calculating each cell individually. The output is an N×1 dimensional SOH point estimate vector (the mean health state of each cell) and an N×1 dimensional SOH variance vector (the uncertainty quantification result of the health state of each cell). The SOH point estimates and corresponding variances are associated and matched according to the cell sequence number to form a combination of health state point estimate and probability distribution for each virtual cell. This result is stored in a structured data format (such as arrays or data tables) for easy online joint identification and retrieval.

[0197] In this embodiment, obtaining the state estimation result based on the real-time parameter information of the actual operating battery pack and the estimated health state points of each virtual cell includes:

[0198] A standardized real-time parameter sequence with timestamps is generated based on the real-time parameter information of the battery pack in actual operation. Specifically, real-time parameters are continuously collected through the data acquisition channel of the battery management system (BMS) and the data is aligned according to the timestamps. Outliers with voltage / current mutations are removed according to preprocessing rules, missing data is filled in using interpolation, and temperature data is converted into units consistent with the virtual model (such as °C), thus forming a standardized real-time parameter sequence.

[0199] In this embodiment, generating a timestamped standardized real-time parameter sequence based on the real-time parameter information of the actually operating battery pack includes:

[0200] A time-aligned standardized real-time-virtual data set is generated based on the standardized real-time parameter sequence with timestamps and dynamic parameter information. Two types of data are collected synchronously at a uniform time step (e.g., 1.0 second): real-time parameter information of the actual operating battery pack (bus voltage, total charging and discharging current, ambient temperature, and pack temperature collected by the BMS) and output data of the operating battery pack system model (voltage, SOC, temperature, internal resistance, etc. of each virtual cell, with timestamps). Abnormal abrupt values ​​in the real-time parameters are removed, and missing data is filled in using interpolation. The real-time parameters and the output data of the virtual model are matched one by one according to the timestamps to achieve time alignment and form a real-time-virtual paired data sequence, ensuring that the data dimensions and time nodes are completely consistent.

[0201] An initialized SMC particle swarm is generated based on the timestamped standardized real-time parameter sequence and the estimated health state points of each virtual cell. Specifically, the SOH variance is converted into the prior confidence of the model parameters (the smaller the SOH variance, the higher the prior weight of the corresponding cell parameters). The Sequential Monte Carlo (SMC) particle filter environment is initialized, and the particle swarm size is set (100~500 particles per cell). Based on the current model parameters and SOH prior information, an initial particle swarm containing hidden states such as SOC and polarization voltage (V1 / V2) is generated for each virtual cell.

[0202] Based on the time-aligned standardized real-time-virtual data set and the initialized SMC particle swarm, a structured state estimation result set is generated using the SMC-EM algorithm, and the structured state estimation result set is used as the state estimation result. Specifically, for each cell's particle swarm, forward propagation is performed based on the physical layer model parameters to calculate the virtual cell output (terminal voltage, temperature response) corresponding to each particle. Using the bus voltage and temperature in the real-time parameters as a benchmark, the likelihood probability between the virtual particle output and the measured data is calculated, and the particle weights are adjusted in conjunction with the parameter prior weight matrix. All particle weights are normalized, and low-confidence particles with weights below a threshold are removed. A multinomial resampling method is used to replicate high-weight particles and eliminate low-weight particles according to their weight ratios, maintaining particle swarm diversity. Using the particle weights as coefficients, the expected value E[X|Y,θ] of each virtual cell's hidden state (SOC, polarization voltage V1 / V2, true internal resistance) is calculated. The uncertainty of the hidden state is quantified through the particle distribution variance, thereby obtaining the point estimates of each virtual cell's hidden state (mean SOC, mean polarization voltage, mean true internal resistance) and the probability distribution of the hidden state (particle swarm variance / 95% confidence interval).

[0203] The results are integrated into a structured dataset in the format of cell serial number-hidden state type-point estimate-probability distribution parameter-time stamp to ensure that each data item is traceable and can be directly used for subsequent EM algorithm parameter updates. Finally, a structured state estimation result set (including the point estimate of the core hidden state of each virtual cell and its probability distribution) is obtained.

[0204] In another embodiment, the structured state estimation result set is generated using the SMC-EM algorithm based on a time-aligned normalized real-time-virtual data set and an initialized SMC particle swarm optimization dataset, including:

[0205] Step 11: Based on the initialized Sequential Monte Carlo (SMC) particle swarm optimization and the time-aligned normalized real-time-virtual dataset, define a dual-scale particle update equation. The particle filtering is divided into a cell-level microscale and a battery pack-level macroscale. The microscale focuses on estimating the hidden states (SOC, polarization voltage) of a single cell, while the macroscale constrains the consistency of pack-level parameters. The microscale particle update formula is:

[0206]

[0207] in, This represents the updated microscopic particle state of the i-th battery cell in step t+1 and the k-th iteration. Let i be the microscopic particle state of the i-th cell in step t and iteration k. This represents the measured terminal voltage of the i-th cell in the k-th iteration. For the virtual model prediction of the terminal voltage of the i-th cell in step t and iteration k, Let be the set of particle states of all cells except the i-th cell in the t-th step and the k-th iteration; These are data-driven coefficients and global constraint coefficients;

[0208] Step 12: Based on the updated particle swarm at the microscale, construct the macroscale parameter fusion formula:

[0209]

[0210] in, These are the model parameters updated in step t+1 and the kth iteration. This is a Bayesian prior based on the global parameters from the previous step; Let be the measured terminal voltage of the i-th cell in the k-th iteration; Let be the microscopic particle state of the i-th cell in step t+1 and the k-th iteration; These are the package-level global parameters for the (t+1)th step and the (k-1)th iteration; These are the model parameters to be optimized; It is a conditional probability distribution;

[0211] Step 13: In the expectation step (E step) of the expectation-maximization (EM) algorithm, the traditional single-point expectation is replaced by dual-scale particle weighted expectation to calculate the posterior expectation of the hidden state (SOC, polarization voltage) of each cell. This posterior expectation is the hidden state estimation result.

[0212] Step 14: In the maximization step (M-step) of the Expectation-Maximization (EM) algorithm, a Bayesian variational lower bound is introduced to construct the objective function:

[0213]

[0214] The updated model parameters are obtained by solving; where, These are the model parameters updated at step t+1; These are the current model parameters at step t; Let Y be the variational posterior distribution of the hidden states based on the model parameters at step t; Y is the set of measured data; X is the set of hidden states; and KL is the KL divergence. Let be the set of package-level macroscopic parameters at step t; This refers to the hidden state of a single battery cell.

[0215] Determine whether the update magnitude of the model parameters is less than the preset convergence threshold (e.g., parameter change rate ≤ 0.001). If convergence is not achieved, feed back the updated model parameters to step 11 and repeat steps 11 to 14 until the convergence condition is met (e.g., the number of iterations does not exceed the preset maximum value, such as 50 times).

[0216] Step 15: Based on the converged model parameters and hidden state estimation results, integrate them according to the format of cell number, hidden state type, point estimate, probability distribution parameter, and timestamp to form a structured state estimation result set.

[0217] The state estimation result acquisition method of this embodiment has the following advantages:

[0218] Traditional identification algorithms often use single-scale state modeling, treating the battery pack as a whole or focusing only on the local state of a single cell. They do not consider the hierarchical relationship between cells and the battery pack, which leads to the estimation of hidden states (such as single cell SOC and polarization voltage) being affected by pack-level dynamic interference. Furthermore, pack-level parameters (such as total cascade internal resistance) are difficult to reflect the differentiated characteristics of cells, ultimately resulting in a contradiction between inaccurate local states and distorted global parameters, which cannot accurately support high-fidelity simulation at the physical layer.

[0219] This application employs a dual-scale particle update mechanism, combining cell-level micro-scale and battery pack-level macro-scale. The micro-scale focuses on the local hidden states of each cell, such as SOC and polarization voltage, using data-driven coefficients to match the dynamic response of a single cell. The macro-scale constrains global parameters at the pack level (e.g., total cascade resistance and total voltage), using Bayesian priors to correlate parameters from the previous cycle, achieving accurate estimation of local states and collaborative optimization of global parameters. This mutual verification and constraint between the two scales avoids the limitations of single-scale modeling, reducing hidden state estimation errors (e.g., SOC estimation errors) from over 3% to less than 1%.

[0220] Traditional online identification algorithms often separate state estimation (SMC step) and parameter update (EM step) for execution. Either they complete the full state estimation first and then update the parameters, causing the parameters to lag behind the state changes; or they have too many alternating iterations (such as hundreds of times), which consumes a lot of computing power and cannot meet the real-time operation requirements of BMS (usually requiring millisecond-level response), making it difficult to balance identification accuracy and real-time performance.

[0221] This application uses the particle swarm updated at the microscale directly as the input for macroscale parameter fusion. The expectation step (E-step) of the EM algorithm calculates the hidden state posterior through the weighted expectation of dual-scale particles, and the maximization step (M-step) optimizes the model parameters synchronously based on the posterior, realizing one-step linkage between state estimation and parameter update, and avoiding parameter lag.

[0222] By introducing a Bayesian variational lower bound to construct the objective function, the search space for parameter optimization is reduced. At the same time, a reasonable convergence threshold (such as parameter change rate ≤ 0.001) and a maximum number of iterations (50 times) are set, which reduces the computational cost of each identification cycle by more than 40% compared with traditional algorithms. Moreover, the convergence is oscillating and perfectly matches the simulation rhythm of 1 second / step in the physical layer without consuming additional computing power.

[0223] Traditional online identification algorithms do not have dedicated logic for the series-parallel topology of battery packs. They simply superimpose the identification results of individual cells, which can easily lead to logical contradictions in the series and parallel parameters. For example, the total series voltage may deviate from the sum of the voltages of each series by more than 10%, and the total parallel current may not match the sum of the currents of each cell. This results in the identification results failing to reflect the electrical characteristics of the actual battery pack and making it even more difficult to support the simulation of the inconsistency evolution between cells.

[0224] The dual-scale update and parameter fusion deeply embed series and parallel topology constraints. The macro-scale parameter fusion formula clearly defines the electrical rules that the total series voltage equals the sum of the voltages of each series and the total parallel current equals the sum of the currents of each cell. During the EM step parameter update, the consistency of this logic is verified, and if the deviation exceeds 5%, the particle weights and parameter optimization direction are automatically adjusted. At the same time, the dual-scale particle update retains the differentiated state characteristics of each cell, enabling the identification results to accurately reproduce phenomena such as voltage differences and SOC differences caused by manufacturing tolerances and aging differences among cells within the package.

[0225] The following examples further illustrate this application in detail. It is understood that these examples do not constitute any limitation on this application.

[0226] Take the construction of a 3-parallel 91-string (273 cells in total) battery pack with a rated voltage of 332.1V and a capacity of 91.5Ah as an example, which is a physical-probabilistic dual-track architecture.

[0227] S1: Virtual cell cluster construction.

[0228] The nominal capacity of a single battery cell is set to 30.5Ah (91.5Ah / 3), and the nominal internal resistance is set to 0.2mΩ. Using random sampling from a normal distribution, 273 cells are assigned differentiated parameters such as capacity (standard deviation 1.5%), internal resistance coefficient (standard deviation 5%), and minor OCV curve shifts, and stored as a structure array cell_params(1:273). This array represents the initial state of a single battery cell in the digital twin.

[0229] S2: Multiscale physical model procedural integration.

[0230] Create a reusable atomic subsystem package for the individual ECM. Specifically, write a cell equivalent circuit model function that can accept the cell_params generated in S1, the total current of the pack, the temperature vector of each cell, and the time step as input in a vectorized manner. It internally solves the second-order RC state equations and outputs the terminal voltages of all cells and the updated internal states in parallel. This function can then serve as the core execution engine of the battery pack's physical layer.

[0231] S3: Probabilistic health prediction layer integration.

[0232] First, a DKA-IC-SVGP model was trained using laboratory battery pack charge-discharge data, and the trained kernel matrix, induced points, variational parameters, etc., were permanently saved. Then, a single-cell SOH prediction function was written. During simulation, feature vectors for each cell (e.g., average current, temperature rise slope, internal resistance change rate, cumulative throughput, etc. over the past N cycles) were periodically extracted from the updated cell_params. This 273xN feature matrix was input into the prediction function, which internally called the fixed model for batch forward propagation, outputting a 273x1 SOH mean vector and variance vector, thus completing a full-cell health check.

[0233] S4: Online joint identification and deployment.

[0234] The SMC-EM algorithm is deployed on real-time hardware. The algorithm continuously receives measured bus data from a real battery pack. Within each identification cycle: E-step: The SMC particle filter estimates the probability distribution of hidden states such as SOC and polarization voltage for the 273 cells based on the current model parameters. M-step: The EM algorithm calculates and updates the maximum likelihood estimate of the time-varying parameters (such as R0, R1, R2, etc.) of each cell in the physical model based on the state estimates provided by SMC. The updated parameters are fed back to the models in S1 and S2, thus completing a closed loop of "perception-learning-update". This process enables the physical-probabilistic dual-track architecture model to track the characteristic changes of real batteries caused by aging, increased inconsistency, etc., in real time.

[0235] For verification and results, please refer to [link / reference]. Figures 3-7 , Figure 3 The simulation results are plotted based on a 3-parallel, 91-string (273 cells) battery pack. The horizontal axis represents simulation time (in minutes), the left vertical axis represents the total battery pack voltage (in V), and the right vertical axis represents the charging and discharging current (in A). The curve shows that during the simulation, the battery pack discharges with a stable negative current (approximately -100A), and the total voltage steadily decreases from approximately 365V to 320V over the discharge time, without any significant abrupt changes or abnormal fluctuations. This curve verifies the rationality of the second-order RC equivalent circuit model and the electro-thermal coupling logic at the physical layer, reflecting the high fidelity of the dynamic response of the battery pack's external characteristics, and is consistent with the voltage-current variation law under actual discharge conditions.

[0236] Figure 4 This is a histogram showing the SOC distribution of 273 virtual cells after simulation. The horizontal axis represents the SOC value (range 0.64~0.74), and the vertical axis represents the number of cells in the corresponding SOC range. Due to the inclusion of manufacturing tolerances in the initial parameters (capacity standard deviation 1.5%, internal resistance standard deviation 5%), and the inconsistent evolution among cells during the simulation, the final SOC exhibits a continuous distribution without concentrated overlap. This figure verifies the effectiveness of the differentiated design of the virtual cell group and the accurate simulation capability of the physical layer model for the SOC differentiation among cells, consistent with the internal state distribution of a real battery pack.

[0237] Figure 5 The horizontal axis represents simulation time (in minutes, focusing on the critical period of 28-30 minutes), and the vertical axis represents the voltage range of the cells within the battery pack (in V, ranging from 0.126 to 0.13 V). The curves show that the voltage range fluctuates slightly over time but remains generally controllable, without a continuous widening trend. This result is attributed to the synergistic effect of the integrated electro-thermal-equilibrium model: the equilibrium logic intervenes in real time to suppress voltage differentiation, and the thermal coupling model corrects internal resistance to avoid local voltage anomalies, verifying the integrated model's ability to regulate and simulate inconsistencies within the battery pack.

[0238] Figure 6 The table shows the SOH prediction results for the first 20 virtual cells, with the horizontal axis representing the cell number and the vertical axis representing the SOH value (range 0.85~0.95). Each point in the figure represents the estimated SOH value of a single cell, and the implicit probability distribution range quantifies the prediction uncertainty. Because the DKA-IC-SVGP model is used for batch prediction, the SOH of different cells exhibits a differentiated distribution due to aging differences, and the predicted values ​​are concentrated in a reasonable healthy range (above 0.85), verifying the probabilistic layer's ability to provide a refined and probabilistic assessment of the health status of each cell.

[0239] Figure 7 The horizontal axis represents simulation time (in minutes), and the vertical axis represents the index number of the worst-performing cell (range 0-250, corresponding to 273 cells). The curve shows that the index of the worst-performing cell (lowest SOH or worst voltage / SOC) changes dynamically over time and is not fixed to any single cell. This phenomenon stems from the inconsistent evolution of cells within the battery pack: differences in operating stress and thermal distribution at different times lead to different aging rates for each cell, resulting in dynamic replacement of weaker cells. This figure verifies the accuracy of the model in dynamically tracking the health status of the battery pack and provides data support for locating weaker cells in real time during predictive maintenance.

[0240] This application has the following advantages:

[0241] A balance between high fidelity and high efficiency: Through a dual-track architecture of physical layer (fast dynamic changes) and probabilistic layer (slow health changes), the system ensures the accuracy of key dynamic changes (physical layer) while achieving refined health prediction with lower computing power (probabilistic layer), thus resolving the contradiction between accuracy and efficiency in traditional methods.

[0242] Realistic simulation of inconsistency evolution: By injecting manufacturing differences and dynamic stress differences, the model can automatically evolve the parameter differentiation between cells and accurately reproduce macroscopic observable phenomena such as voltage differences, providing a reliable virtual experimental platform for the study of equalization strategies.

[0243] Online evolution capability of the model: Through the SMC-EM online joint identification framework, the model parameters can be adaptively updated as the battery ages, ensuring the long-term accuracy of the digital twin throughout its entire life cycle and achieving a leap from static to real-time recording and prediction.

[0244] Refined and probabilistic health management: It can output the probability distribution of the health status of each cell, rather than a single mean, realizing transparent and quantitative insights into the health status of the battery pack. It can accurately locate the weak cells and assess the overall risk, providing core support for predictive maintenance.

[0245] This application also provides a battery pack identification system based on a physical-probabilistic dual-track architecture, the battery pack identification system based on the physical-probabilistic dual-track architecture comprising:

[0246] A battery pack basic information acquisition module, which is used to acquire basic information about the battery pack;

[0247] A virtual cell cluster generation module is used to generate a virtual cell cluster containing differentiated initial parameters based on the basic information of the battery pack.

[0248] A battery pack system model generation module is used to generate a battery pack system model based on a virtual cell group containing differentiated initial parameters.

[0249] A dynamic parameter information acquisition module is used to acquire the dynamic parameter information output by the battery pack system model running in a virtual environment.

[0250] A battery health state core predictor acquisition module is used to acquire a trained battery health state core predictor.

[0251] The health state point estimation value acquisition module is used to input the dynamic parameter information output by the battery pack system model into the trained battery health state core predictor, thereby obtaining the health state point estimation value of each virtual cell.

[0252] A real-time parameter information acquisition module is used to acquire real-time parameter information of the battery pack in actual operation.

[0253] The state estimation result acquisition module is used to acquire state estimation results based on the real-time parameter information of the battery pack in actual operation and the estimated health state points of each virtual cell.

[0254] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A battery pack identification method based on a physical-probabilistic dual-track architecture, characterized in that: The battery pack identification method based on a physical-probabilistic dual-track architecture includes: Obtain basic information about the battery pack; A virtual cell group containing differentiated initial parameters is generated based on the basic information of the battery pack; A battery pack system model is generated based on a virtual cell group containing differentiated initial parameters; Obtain dynamic parameter information output by the battery pack system model running in a virtual environment; Obtain the trained battery health state core predictor; The dynamic parameter information output by the battery pack system model is input into the trained battery health state core predictor to obtain the estimated health state point value of each virtual cell. Obtain real-time parameter information of the battery pack in actual operation; The state estimation results are obtained based on the real-time parameter information of the actual operating battery pack and the estimated health state points of each virtual cell. The basic information of the battery pack includes battery pack topology information, basic cell parameter information, historical cell test data information, and operating condition and system configuration information; The process of generating a virtual cell group containing differentiated initial parameters based on the basic information of the battery pack includes: The total number of virtual cells is obtained based on the battery pack topology information; Based on the battery cell basic parameter information, obtain an initial parameter type list and a set of baseline values ​​for each parameter; The unique initial parameters for each virtual cell are generated based on the total number of virtual cells, the list of initial parameter types, and the set of baseline values ​​for each parameter. The unique initial parameters of each virtual cell form a virtual cell group containing differentiated initial parameters. The process of generating a battery pack system model based on a virtual cell group containing differentiated initial parameters includes: Generate a set of individual ECM model instances for each virtual cell based on the virtual cell group containing differentiated initial parameters; A cluster of battery pack individual model instances with complete electrical connections is generated based on the set of individual ECM model instances and the battery pack topology information. An intermediate battery pack model with electro-thermal coupling is generated from the cluster of individual battery pack models that have completed electrical connections. A prototype model of a battery pack with integrated electrothermal equalization function is generated based on the intermediate model of the electro-thermal coupled battery pack. Define the model input and output interfaces for the prototype model of the battery pack with integrated electrothermal equalization function, so as to obtain the battery pack system model; The core predictor for obtaining the trained battery health status includes: Obtain a standardized battery health status training dataset; Obtain the DKA-IC-SVGP model; The DKA-IC-SVGP model is trained using a standardized battery health status training dataset to obtain a trained core predictor of battery health status.

2. The battery pack identification method based on a physical-probabilistic dual-track architecture according to claim 1, characterized in that: The step of inputting the dynamic parameter information output by the battery pack system model into the trained battery health state core predictor to obtain the estimated health state point value of each virtual cell includes: According to a preset cycle, the dynamic parameters of each virtual cell output by the model are collected in real time to obtain the dynamic parameter set of each virtual cell. Feature extraction is performed on the dynamic parameter set of each virtual cell to obtain the standardized feature matrix of all virtual cells; The standardized feature matrix of all virtual cells is input into the trained battery health state core predictor to obtain the health state point estimate of each virtual cell.

3. The battery pack identification method based on a physical-probabilistic dual-track architecture according to claim 2, characterized in that: The process of obtaining the state estimation result based on the real-time parameter information of the actual operating battery pack and the estimated health state points of each virtual cell includes: Generate a standardized real-time parameter sequence with timestamps based on the real-time parameter information of the battery pack in actual operation; A time-aligned standardized real-time-virtual data set is generated based on the timestamped standardized real-time parameter sequence and dynamic parameter information; An initialized SMC particle swarm is generated based on a time-stamped standardized real-time parameter sequence and the estimated health state points of each virtual cell. Based on the time-aligned standardized real-time-virtual data set and the initialized SMC particle swarm, a structured state estimation result set is generated using the SMC-EM algorithm, and the structured state estimation result set is used as the state estimation result.

4. The battery pack identification method based on a physical-probabilistic dual-track architecture according to claim 3, characterized in that: The generation of a structured state estimation result set using the SMC-EM algorithm, based on the time-aligned standardized real-time-virtual data set and the initialized SMC particle swarm, includes: Step 11: Based on the initialized SMC particle swarm and the time-aligned standardized real-time-virtual data set, define a dual-scale particle update equation, dividing the particle filtering into cell-level microscale and battery pack-level macroscale. The microscale particle update formula is as follows: = in, This represents the updated microscopic particle state of the i-th battery cell in step t+1 and the k-th iteration. Let i be the microscopic particle state of the i-th cell in step t and iteration k. This represents the measured terminal voltage of the i-th cell in the k-th iteration. For the virtual model prediction of the terminal voltage of the i-th cell in step t and iteration k, Let be the set of particle states of all cells except the i-th cell in the t-th step and the k-th iteration; , These are data-driven coefficients and global constraint coefficients; Step 12: Based on the updated particle swarm at the microscale, construct the macroscale parameter fusion formula: in, These are the model parameters updated in step t+1 and the kth iteration. This is a Bayesian prior based on the global parameters from the previous step; Let be the measured terminal voltage of the i-th cell in the k-th iteration; Let be the microscopic particle state of the i-th cell in step t+1 and the k-th iteration; These are the package-level global parameters for the (t+1)th step and the (k-1)th iteration; These are the model parameters to be optimized; It is a conditional probability distribution; Step 13: In the expectation step of the expectation-maximization algorithm, the posterior expectation of the hidden state is calculated by weighted expectation of dual-scale particles. This posterior expectation is the hidden state estimation result. Step 14: In the maximization step of the expectation-maximization algorithm, a Bayesian variational lower bound is introduced to construct the objective function: The updated model parameters are obtained by solving; where, These are the model parameters updated at step t+1; These are the current model parameters at step t; Let Y be the variational posterior distribution of the hidden states based on the model parameters at step t; Y is the set of measured data; X is the set of hidden states; and KL is the KL divergence. Let be the set of package-level macroscopic parameters at step t; This refers to the hidden state of a single battery cell. Determine whether the model parameter update magnitude is less than the preset convergence threshold. If convergence is not achieved, feed back the updated model parameters to step 11 and repeat steps 11 to 14 until the convergence condition is met. Step 15: Based on the converged model parameters and hidden state estimation results, integrate them according to the format of cell number, hidden state type, point estimate, probability distribution parameter, and timestamp to form a structured state estimation result set.

5. The battery pack identification method based on a physical-probabilistic dual-track architecture according to claim 4, characterized in that: The step of training the DKA-IC-SVGP model using a standardized battery health status training dataset to obtain a trained core predictor of battery health status includes: Step 21: Based on the standardized battery health status training dataset, construct the kernel space reconstruction criterion, map the original feature space to the adaptive kernel space through nonlinear transformation, and calculate the initial solution of the kernel space weight matrix Λ; Step 22: Obtain the objective function for joint optimization of the induction point and kernel space; Step 23: Optimize the initial solution of the kernel space weight matrix Λ by co-optimizing the objective function of the induced points and the kernel space, thereby obtaining the optimized kernel space weight matrix and the set of induced points; Step 24: Substitute the optimized kernel space weight matrix and the set of induced points into the DKA-IC-SVGP model. Use an iterative training method, with the feature matrix of the training dataset as input and the real SOH as the label for model training. Calculate the prediction error every 10 iterations until the prediction error is less than a preset threshold or the number of iterations reaches a preset maximum value. Output the variational parameter update sequence during the training process and the final converged variational parameter set.

6. A battery pack identification system based on the identification method according to any one of claims 1 to 5, characterized in that: The battery pack identification system includes: A battery pack basic information acquisition module, which is used to acquire basic information about the battery pack; A virtual cell cluster generation module is used to generate a virtual cell cluster containing differentiated initial parameters based on the basic information of the battery pack. A battery pack system model generation module is used to generate a battery pack system model based on a virtual cell group containing differentiated initial parameters. A dynamic parameter information acquisition module is used to acquire the dynamic parameter information output by the battery pack system model running in a virtual environment. A battery health state core predictor acquisition module is used to acquire a trained battery health state core predictor. The health state point estimation value acquisition module is used to input the dynamic parameter information output by the battery pack system model into the trained battery health state core predictor, thereby obtaining the health state point estimation value of each virtual cell. A real-time parameter information acquisition module is used to acquire real-time parameter information of the battery pack in actual operation. The state estimation result acquisition module is used to acquire state estimation results based on the real-time parameter information of the battery pack in actual operation and the estimated health state points of each virtual cell.

Images