A Simulation Method and System for Energy Storage Battery Modules Based on Machine Learning

Abstract

This application discloses a machine learning-based simulation method and system for energy storage battery modules. The method integrates physical mechanisms and data-driven approaches, identifies the root cause type of fault based on causal inference algorithms, determines the three-dimensional coordinates of the faulty cell based on solving the inverse heat conduction problem, generates a selective bypass strategy based on multi-objective optimization, and achieves dynamic model calibration and generalization based on virtual-real closed-loop verification and cross-module knowledge transfer. It realizes a technological leap from symptom identification to root cause discrimination, from module-level fuzzy early warning to cell-level precise positioning, and from overall throttling to differentiated intervention, solving the technical problems of model drift with aging and the limitations of single-module customization.

Description

Technical Field

[0001] This application relates to the field of new energy technology, and in particular to a simulation method and system for energy storage battery modules based on machine learning. Background Technology

[0002] Energy storage system safety management is a core technology direction for ensuring the stable operation of power systems in the new energy field. Among them, battery module thermal runaway early warning and fault diagnosis are key links in energy storage safety control, directly affecting the operational reliability of large-scale energy storage power stations and electric vehicles. As lithium-ion battery modules develop towards high energy density and compact structures, the complex electrothermal coupling effect between multiple cells within the module makes the fault propagation mechanism increasingly complex. Traditional fault detection methods based on threshold judgment are no longer sufficient to meet the needs of refined safety management. Therefore, developing causal inference methods that integrate physical mechanisms and data-driven approaches to achieve intelligent identification and precise spatial tracing of module-level fault root causes has become an important technical support for the evolution of energy storage battery management systems towards predictive maintenance.

[0003] Existing fault diagnosis technologies for energy storage battery modules primarily rely on correlation analysis, triggering fault alarms by monitoring abnormal fluctuations in external characteristic parameters such as voltage and temperature. However, this approach has inherent limitations: First, it lacks causal inference capabilities, failing to distinguish the root cause of abnormal temperature increases—that is, it cannot determine whether the heat dissipation failure is due to blockage in external thermal management pipes or self-generated heat caused by micro-short circuits within the cell, leading to blind fault handling strategies. Second, it lacks spatial positioning accuracy; existing methods can only determine overall module anomalies and cannot determine the precise coordinates of the faulty cell in three-dimensional space through thermal conduction inversion calculations, making modular and precise maintenance difficult. Third, existing simulation systems are disconnected from fault diagnosis, failing to automatically generate targeted degradation operation strategies (such as selectively bypassing faulty cells rather than overall derating) within milliseconds of detecting anomalies, resulting in unnecessary loss of system usable capacity. These shortcomings severely restrict the technological leap from "passive alarm" to "active immunity" in energy storage systems. Summary of the Invention

[0004] The main objective of this application is to provide a machine learning-based simulation method and system for energy storage battery modules to solve the problems mentioned in the background.

[0005] To achieve the above objectives, this application provides the following technical solution:

[0006] A machine learning-based simulation method for energy storage battery modules, characterized by the following specific steps:

[0007] S1. Data Acquisition and Fusion: Real-time acquisition of module voltage, current, temperature and pressure signals, construction of spatiotemporal observation matrix, calculation of residuals through physical baseline model and triggering diagnosis;

[0008] S2, Feature Decoupling: Project the residuals onto an orthogonal subspace to separate mechanical, electrochemical, and thermal management fault features and generate type-specific feature vectors;

[0009] S3. Causal Root Cause Identification: Based on physical priors, construct a structural causal model, perform counterfactual intervention reasoning, determine the dominant fault root cause type, and output the heat source hypothesis;

[0010] S4. Spatial coordinate inversion: Select the heat source model according to the root cause type, solve the inverse heat conduction problem, and invert the three-dimensional coordinates of the fault and the heat source intensity within the module.

[0011] S5. Propagation and Evolution Prediction: Multiphysics simulation is performed by embedding three-dimensional coordinates and heat source intensity to predict the spatiotemporal expansion range of the fault impact domain and the capacity loss trend.

[0012] S6. Degradation Strategy Optimization: Construct a multi-objective optimization function with the influence domain as a constraint to generate differentiated degradation control strategies with selective bypass and hot buffer;

[0013] S7. Virtual-Real Closed-Loop Correction: Execute differentiated degradation control strategies and compare actual responses with predicted results to correct structural causal model parameters and inversion weights online, achieving incremental updates;

[0014] S8. Knowledge Transfer and Generalization: Extract common fault features across modules, adapt to new modules with different structures through transfer learning, and expand the coverage of the physical baseline model through feedback.

[0015] Preferably, in step S1, the specific method of data acquisition and fusion is as follows:

[0016] S1.1 Real-time acquisition of voltage, current and surface temperature time-series signals of each individual cell in the module, fusion of multi-source heterogeneous data of mechanical expansion pressure and cooling system flow, and construction of original observation matrix with spatiotemporal consistency through multi-rate timestamp alignment and spatial domain interpolation registration.

[0017] S1.2 Input the original observation matrix into the physical baseline model of electrochemical thermomechanical coupling, calculate the theoretical expected response value under the current operating condition, extract the residual signal between the actual observation and the physical expectation through differential operation, and adaptively adjust the judgment threshold according to the battery aging state. When the cumulative residual value exceeds the judgment threshold, the fault diagnosis process is triggered.

[0018] Preferably, in step S2, the specific method of feature decoupling is as follows:

[0019] S2.1 Construct a contrastive learning architecture that integrates prior physical knowledge. Using the residual signal output from step S1 as the input sample, design differentiated contrastive loss functions for three types of faults: mechanical, electrochemical, and thermal management. Forcefully separate the overlapping components in the mixed features and establish a fault type-specific discrimination boundary so that similar fault features are clustered and dissimilar fault features are excluded.

[0020] S2.2 Map the aggregated feature components to three orthogonal feature subspaces, introduce physical orthogonality constraints to ensure decoupling of mechanical fault features from pressure change gradients and electrochemical fault features from internal resistance evolution trends, generate pure feature vectors corresponding to different fault mechanisms, and output them as observation nodes of the structural causal model to the causal root cause identification step.

[0021] Preferably, in step S3, the specific method for identifying the root cause is as follows:

[0022] S3.1. Take the pure feature vector output in step S2 as the set of observed variables, integrate the law of conservation of energy and the heat transfer mechanism to construct a priori topological constraint library, and use the causal discovery algorithm to deduce the directed causal edge between the root cause of the fault and the observed symptoms from the time series data of abnormal working conditions, establish a structural causal model that can explain the physical evolution mechanism and eliminate the interference of pseudo-correlation between variables.

[0023] S3.2. Based on the structural causal model, perform counterfactual intervention simulation. Through do-calculus, simulate virtual control operations to keep the cooling circuit constant and the charging and discharging current constant respectively. Calculate the posterior probability distribution of the temperature anomaly under each intervention scenario. Determine the dominant fault root cause type through the probability comparison mechanism and select the corresponding heat source model hypothesis to output to the spatial coordinate inversion step.

[0024] Preferably, in step S4, the spatial coordinate inversion is performed in the following way:

[0025] S4.1. Based on the root cause type determined in step S3, the corresponding heat source model hypothesis is adaptively selected. When the fault is determined to be internal electrochemical, the volume heat source model is selected. When the fault is determined to be contact interface, the surface heat source model is selected. A heat conduction inverse problem solution framework with boundary condition constraints is constructed. The surface temperature field distribution collected in step S1 is embedded into the solution domain as a boundary constraint and a regularization term is established.

[0026] S4.2. Based on the constructed inverse problem framework, a regularized numerical method is used to solve the ill-posed inverse problem of heat conduction. The three-dimensional spatial coordinates and heat source intensity of the internal heat source in the module are inferred from the surface temperature field distribution. The non-uniqueness of the solution is suppressed by physical information constraints. The three-dimensional spatial coordinates and heat source intensity obtained by inversion are used as initial conditions and output to the propagation evolution prediction step.

[0027] Preferably, in step S5, the specific method for propagation evolution prediction is as follows:

[0028] S5.1 Receive the three-dimensional spatial coordinate position and heat source intensity parameters obtained from the inversion in step S4, embed them as initial boundary conditions into the physical baseline model, activate the multi-physics simulation engine of electrochemical thermomechanical coupling, and perform forward propagation calculation to deduce the energy transfer and thermal mass diffusion evolution process of the fault in the module.

[0029] S5.2 Based on the multiphysics simulation results, predict the evolution trend of the three-dimensional isothermal surface of the temperature field and the temperature rise trajectory of adjacent units in the future period, dynamically calculate the spatiotemporal expansion range of the fault influence domain and quantify the overall capacity loss ratio of the module, and output the fault influence domain range and capacity loss trend as constraints to the degradation strategy optimization step.

[0030] Preferably, in step S6, the specific method for optimizing the degradation strategy is as follows:

[0031] S6.1 Construct a constrained boundary domain based on the spatiotemporal evolution range of the fault impact domain and the capacity decay trend predicted in step S5. Integrate the fault root cause types determined in step S4 to establish a multi-objective collaborative optimization function, taking into account the three objectives of maximizing the remaining available capacity, minimizing the thermal runaway risk index, and minimizing the impact of strategy switching as the optimization criteria.

[0032] S6.2 Solve the multi-objective collaborative optimization function to obtain the Pareto optimal solution set, generate a differentiated degradation control strategy that includes selectively bypassing faulty cells, adjusting the charge and discharge rates of adjacent healthy cells to build a thermal buffer barrier, and activating the backup thermal management circuit, and demap the differentiated degradation control strategy into a sequence of control instructions that can be executed by the battery management system.

[0033] Preferably, in step S7, the specific method for virtual-real closed-loop correction is as follows:

[0034] S7.1 Map the differentiated degradation control strategy generated in step S6 to the physical energy storage entity system for execution and establish a virtual-real bidirectional data channel. Synchronously collect the actual voltage and temperature response time series data after the strategy is implemented, and calculate the virtual-real consistency deviation with the theoretical evolution trajectory predicted in step S5. Quantify the prediction error and determine whether it exceeds the preset tolerance threshold.

[0035] S7.2 When the virtual-real consistency verification shows that the prediction error exceeds the preset tolerance threshold, the causal edge weight coefficients of the structural causal model in step S3 and the regularization constraint parameters of the heat conduction inverse problem solution framework in step S4 are incrementally corrected using actual response time series data to realize the adaptive evolution and update of the digital twin model, and the evolved model parameters are fed back to step S1 and step S3 to improve the accuracy of subsequent diagnosis.

[0036] Preferably, in step S8, the specific method of knowledge transfer generalization is as follows:

[0037] S8.1 Mining historical case time series data accumulated in the virtual-real closed loop of step S7, extracting common fault evolution laws and causal topological patterns across geometric configurations and chemical systems, condensing them into a meta-knowledge base containing inversion structure and root cause discrimination knowledge and storing it.

[0038] S8.2 For new modules with heterogeneous geometric configurations or chemical systems, the causal reasoning structure and inverse problem solving in the transfer meta-knowledge base are adapted across domains. Based on a small number of target domain samples, the new module is fine-tuned to converge to usable diagnostic accuracy. The adapted model parameters are then fed back to step S1 to expand the coverage boundary of the physical baseline model.

[0039] This application also provides a machine learning-based energy storage battery module simulation system, including a data acquisition and fusion module, which acquires module voltage, current, temperature, pressure and flow signals, constructs an observation matrix through spatiotemporal registration, inputs a physical baseline model to calculate residuals and adaptively determines and triggers diagnosis based on thresholds;

[0040] The fault feature decoupling module receives residual input and compares with the learning architecture to separate mechanical, electrochemical, and thermal management fault features and maps them to an orthogonal subspace to generate pure feature vectors.

[0041] The causal root cause identification module constructs a structural causal model using features as observation nodes, deduces the causal edge between root causes and symptoms, and determines the root cause type by calculating the probability through counterfactual intervention.

[0042] The spatial coordinate inversion module selects the heat source model based on the root cause type, uses the temperature field as the boundary constraint, and regularizes and inversely calculates the three-dimensional coordinates of the fault and the intensity of the heat source.

[0043] The propagation and evolution prediction module embeds the coordinate heat source into the physical model, simulates and extrapolates energy transfer through multiphysics, and predicts the evolution of the temperature field and the range of its influence domain.

[0044] The degradation strategy optimization module constructs a multi-objective optimization function with the influence domain as a constraint, and solves the degradation strategy to generate selective bypass faulty cells and adjust the ratio of neighboring cells.

[0045] The virtual-real closed-loop correction module executes the strategy to collect response data, calculates the deviation from the predicted trajectory, and corrects the causal model weights and inverse problem parameters to update the model.

[0046] The cross-module knowledge transfer module mines historical cases to extract common patterns and builds a meta-knowledge base. It then adapts the causal structure of the new module for cross-domain transfer and fine-tunes convergence, expanding the baseline coverage.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] 1. This method achieves a fundamental leap from "symptom identification" to "root cause determination" by introducing causal inference; it achieves a breakthrough in spatial accuracy from "module-level fuzzy early warning" to "unit-level precise positioning" by solving the inverse problem of heat conduction; and it improves treatment efficiency from "overall conservative reduction" to "differentiated precise intervention" through multi-objective optimization strategy generation. Simultaneously, it solves the problem of static model accuracy drifting with battery aging through virtual-real closed-loop verification and incremental update mechanisms, and overcomes the application limitations of single-module customized models through cross-module knowledge transfer.

[0049] 2. This method advances the early warning time window for fault diagnosis of energy storage battery modules from the existing minute level to the hour level, and improves the spatial positioning accuracy to the single-unit level resolution, thereby enhancing the safety, economy and intelligent operation and maintenance level of energy storage systems, and providing reliable technical support for predictive maintenance of large-scale energy storage power stations and electric vehicles. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort:

[0051] Figure 1 This is a flowchart illustrating the steps of the method described in this application.

[0052] Figure 2 This is the system flowchart for this application. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Example 1: Please refer to Figure 1 A machine learning-based simulation method for energy storage battery modules, with the following specific steps:

[0055] S1. Data Acquisition and Fusion: Real-time acquisition of module voltage, current, temperature and pressure signals, construction of spatiotemporal observation matrix, calculation of residuals through physical baseline model and triggering diagnosis;

[0056] In the data acquisition and fusion process, the physical baseline model and data-driven algorithms are integrated in the following way: Real-time acquisition of voltage, current, and surface temperature time-series signals of each individual cell within the module, along with simultaneous acquisition of multi-source heterogeneous data on mechanical expansion pressure and cooling system flow. A spatiotemporally consistent original observation matrix is ​​constructed through multi-rate timestamp alignment and spatial domain interpolation registration. This original observation matrix is ​​then input into the electrochemical-thermomechanical coupled physical baseline model. This model consists of a bidirectional coupling of an electrochemical sub-model, a thermal sub-model, and a mechanical sub-model. The electrochemical sub-model uses a P2D pseudo-two-dimensional model to solve the Butler-Volmer electrochemical reaction kinetic equations, outputting the theoretical terminal voltage and heat generation power density. The thermal sub-model, based on the transient heat conduction equations, solves for the theoretical distribution of the three-dimensional temperature field. The mechanical sub-model, based on the thermoelastic mechanics equations, solves for the theoretical distribution of expansion displacement and pressure. The three sub-models achieve bidirectional closed-loop transmission through heat generation rate, temperature field, and expansion displacement. The theoretical terminal voltage, theoretical surface temperature, and theoretical mechanical expansion pressure output from the three sub-models are differentially analyzed with the corresponding measured values ​​in the original observation matrix to extract residual signals. The judgment threshold is adaptively adjusted based on the capacity decay coefficient and internal resistance growth coefficient corresponding to the battery aging state. When the cumulative residual value exceeds the adaptive judgment threshold, the subsequent fault diagnosis process is triggered.

[0057] S2, Feature Decoupling: Project the residuals onto an orthogonal subspace to separate mechanical, electrochemical, and thermal management fault features and generate type-specific feature vectors;

[0058] In the feature decoupling step, the construction and training of the artificial intelligence model are as follows: A contrastive learning architecture integrating prior physical knowledge is constructed. This architecture includes a one-dimensional convolutional neural network layer, a bidirectional long short-term memory network layer, and a projection head layer. The one-dimensional convolutional neural network layer has three convolutional layers with kernel sizes of 7, 5, and 3, and channel numbers of 32, 64, and 128, respectively. Each layer is followed by a batch normalization layer and a modified linear unit activation function. The bidirectional long short-term memory network layer has 128 and 64 hidden units, respectively. The projection head layer consists of two fully connected layers, mapping features to a 32-dimensional contrastive space. The residual signal output from step S1 is used as the input sample. The input data is a 512-point residual sequence normalized by Z-score. For three types of faults—mechanical, electrochemical, and thermal management—normalized temperature-scale contrastive loss functions with temperature coefficients of 0.1, 0.2, and 0.15 are set, respectively. Training was performed using a batch size of 64, an initial learning rate of 0.001, and the Adam optimizer. The training epochs were set to 200, and an early stopping mechanism was introduced. Overlapping components in the mixed features were forcibly separated, and a fault type-specific discrimination boundary was established, causing similar fault features to cluster and dissimilar fault features to exclude each other. The clustered feature components were mapped to three orthogonal feature subspaces. Physical orthogonality constraints were introduced to ensure decoupling of mechanical fault features from pressure change gradients and electrochemical fault features from internal resistance evolution trends. Pure feature vectors corresponding to different fault mechanisms were generated and served as the observation node outputs of the structural causal model.

[0059] S3. Causal Root Cause Identification: Based on physical priors, construct a structural causal model, perform counterfactual intervention reasoning, determine the dominant fault root cause type, and output the heat source hypothesis;

[0060] In the causal root cause identification step, the construction and deduction of the artificial intelligence model are as follows: The pure feature vector output from step S2 is used as the set of observed variables. The set of observed variables includes at least mechanical fault feature nodes, electrochemical fault feature nodes, thermal management fault feature nodes, and temperature anomaly symptom nodes. A priori topological constraint library is constructed by integrating the law of conservation of energy and heat transfer mechanisms. This constraint library stipulates that: electrochemical heat generation feature nodes can only point to temperature anomaly symptom nodes, and reverse connections are prohibited; mechanical deformation heat generation feature nodes can only point to temperature anomaly symptom nodes, and reverse connections are prohibited; cooling flow feature nodes can only point to temperature anomaly symptom nodes, and reverse connections are prohibited. A constraint-based PC causal discovery algorithm is adopted. The Fisher-Z transform is used as the conditional independence test method, with a significance level set to 0.05. Within the directional constraints defined by the priori topological constraint library, a directed acyclic graph structure satisfying conditional independence is searched. The directed causal edges between fault root cause nodes and observed symptom nodes are deduced. A structural causal model that can explain the physical evolution mechanism is established, and spurious interference edges between variables are eliminated. Counterfactual intervention simulation is performed based on a structural causal model. A do-calculus intervention operation is performed on candidate root cause variables, severing all incoming causal edges of the intervention variable in the model and fixing its value to a constant value. The first intervention scenario executes do (cooling loop flow rate = rated design value). The second intervention scenario executes do (charging / discharging current = current constant value). Using Bayesian network inference, the posterior probability of the persistence of abnormal temperature symptoms is calculated for each of the two intervention scenarios. The probability comparison mechanism calculates the absolute value of the difference between the two posterior probability values. When the absolute value of this difference exceeds a preset threshold of 0.3, the variable category corresponding to the intervention scenario with a significantly reduced posterior probability is determined as the dominant root cause of the fault. Based on this, either a volumetric heat source model or a surface heat source model is adaptively selected for the output hypothesis.

[0061] S4. Spatial coordinate inversion: Select the heat source model according to the root cause type, solve the inverse heat conduction problem, and invert the three-dimensional coordinates of the fault and the heat source intensity within the module.

[0062] In the spatial coordinate inversion step, the physical model and data-driven domain are combined as follows: Based on the root cause type determined in step S3, the corresponding heat source model assumption is adaptively selected. When the fault is identified as internal electrochemical, a volume heat source model is selected, and the heat source term is embedded into the heat conduction equation in the form of heat generation rate density. When the fault is identified as a contact interface, a surface heat source model is selected, and the heat source term is embedded into the boundary conditions in the form of contact thermal resistance interface heat flux density. A framework for solving the inverse heat conduction problem with boundary condition constraints is constructed. The heat conduction governing equation is a transient heat conduction equation described by density, specific heat capacity, and thermal conductivity. The boundary conditions adopt the third type of convective heat transfer boundary and are coupled with the surface temperature field distribution measurement values ​​collected in step S1, which are embedded into the solution domain as boundary constraints. At the same time, a Tikhonov regularization stability term is established. Based on this inverse problem framework, a numerical method combining Tikhonov regularization and truncated singular value decomposition is used to solve the ill-posed inverse heat conduction problem. The regularization parameter is adaptively selected using the L-curve criterion. The three-dimensional spatial coordinates and heat source intensity parameters of the internal heat source within the module are inferred from the surface temperature field distribution. Physical information constraints are used to suppress the non-uniqueness of the solution. The inverted three-dimensional spatial coordinates and heat source intensity are output as initial conditions.

[0063] S5. Propagation and Evolution Prediction: Multiphysics simulation is performed by embedding three-dimensional coordinates and heat source intensity to predict the spatiotemporal expansion range of the fault impact domain and the capacity loss trend.

[0064] In the propagation evolution prediction step, the construction and calculation method of the multiphysics simulation engine are as follows: The three-dimensional spatial coordinates and heat source intensity parameters obtained from the inversion in step S4 are received and embedded into the physical baseline model as initial boundary conditions. The electrochemical-thermomechanical coupling multiphysics simulation engine is activated. This engine includes an electrochemical field solution unit, a thermal field solution unit, and a mechanical field solution unit. The electrochemical field solution unit uses a P2D pseudo-two-dimensional model to solve the solid-phase lithium-ion diffusion equation, the liquid-phase lithium-ion migration equation, and the Butler-Volmer electrochemical reaction kinetic equation. The thermal field solution unit, based on the transient heat conduction equation, solves the distribution data of the three-dimensional temperature field evolution over time within the module. The mechanical field solution unit, based on the thermoelastic mechanics equation, solves the thermal expansion displacement field and thermal stress field distribution. The three solution units achieve bidirectional closed-loop transmission through heat generation rate, temperature field, and expansion displacement. The finite element method or finite volume method is used to perform spatiotemporal discretization solutions on the coupled equations, deduce the energy transfer and thermal-mass diffusion evolution process of the fault within the module. Based on multiphysics simulation results, isosurface data with a preset temperature threshold are extracted in a time-series manner to deduce the evolution trend of the three-dimensional isothermal surface of the temperature field. Temperature time-series response data at the center coordinates of each individual unit within the vicinity of the faulty unit are extracted to construct the temperature rise trajectory of the adjacent units. A safe operating temperature threshold is set. The continuous spatial region where the temperature exceeds this threshold is defined as the fault influence domain. The spatial volume of the fault influence domain and its rate of expansion over time are calculated to obtain spatiotemporal boundary data. The capacity decay rate is calculated based on the theoretical terminal voltage response and temperature data of each unit, combined with a temperature-accelerated aging model. The capacity decay rate of the individual units located within the fault influence domain is weighted and summed with the rated total capacity of the module to quantify the overall capacity loss ratio of the module. The spatiotemporal boundary data of the fault influence domain and the overall capacity loss ratio of the module are output as constraints.

[0065] S6. Degradation Strategy Optimization: Construct a multi-objective optimization function with the influence domain as a constraint to generate differentiated degradation control strategies with selective bypass and hot buffer;

[0066] In the degradation strategy optimization step, the artificial intelligence algorithm is constructed and solved as follows: A constraint boundary domain is constructed based on the spatiotemporal evolution range of the fault influence domain and the capacity decay trend predicted in step S5. The fault root cause types determined in step S4 are integrated to establish a multi-objective collaborative optimization function. This function aims to maximize remaining available capacity, minimize the thermal runaway risk index, and minimize the strategy switching impact. The first component is the negative value of the remaining available capacity. The second component is the thermal runaway risk index. The third component is the norm of the strategy switching power impact. Constraints include the upper limit of the influence domain current and the upper limit of the temperature. The NSGA-II multi-objective genetic algorithm is used to solve this optimization function. The population size is set to 100. The crossover probability is set to 0.9. The mutation probability is set to 0.1. The number of iterations is set to 500 generations. The Pareto optimal solution set is obtained. A differentiated degradation control strategy is generated, including selectively bypassing faulty cells, adjusting the charge / discharge rate of neighboring healthy cells to construct a thermal buffer barrier, and activating the backup thermal management loop. The strategy solution is mapped into a sequence of control instructions executable by the battery management system.

[0067] S7. Virtual-Real Closed-Loop Correction: Execute differentiated degradation control strategies and compare actual responses with predicted results to correct structural causal model parameters and inversion weights online, achieving incremental updates;

[0068] In the virtual-real closed-loop correction step, the online training and parameter correction methods of the artificial intelligence model are as follows: The differentiated degradation control strategy generated in step S6 is mapped to the physical energy storage entity system for execution, establishing a virtual-real bidirectional data channel. The actual voltage and temperature response time-series data after the strategy implementation are synchronously collected. The virtual-real consistency deviation is calculated by comparing it with the theoretical evolution trajectory predicted in step S5. The root mean square error is used to quantify the prediction error, and it is determined whether it exceeds a preset tolerance threshold. When the virtual-real consistency verification shows that the prediction error exceeds the tolerance range, the causal edge weight coefficients of the structural causal model in step S3 and the regularization constraint parameters of the heat conduction inverse problem solution framework in step S4 are incrementally corrected using the actual response time-series data. The causal edge weight coefficients are corrected using an online gradient descent algorithm with a learning rate set to 0.01. The regularization parameters are adaptively adjusted according to the magnitude of the prediction error. When the error increases, the Tikhonov regularization strength is increased. Adaptive evolution and update of the digital twin model are achieved. The evolved model parameters are fed back to steps S1 and S3 to improve the accuracy of subsequent diagnostics.

[0069] S8. Knowledge Transfer and Generalization: Extract common fault features across modules, adapt to new modules with different structures through transfer learning, and expand the coverage of the physical baseline model through feedback.

[0070] In the knowledge transfer and generalization step, the construction and training methods of the artificial intelligence transfer learning model are as follows: First, mine the historical case time-series data accumulated in step S7 (virtual and real closed loop). Extract common fault evolution patterns and causal topological patterns across geometric configurations and chemical systems. Condense this into a meta-knowledge base containing inversion structure and root cause discrimination knowledge. This meta-knowledge base stores the weight parameters of the trained causal inference network, the regularization parameters of the inverse problem-solving network, and the common causal topological structure. For new modules with heterogeneous geometric configurations or chemical systems, the causal inference structure and inverse problem-solving network in the transfer meta-knowledge base are adapted across domains. A feature extraction-based transfer learning framework is adopted. Freeze the feature extraction layer consisting of a one-dimensional convolutional layer and a bidirectional long short-term memory network in the source domain model. Freeze the causal topological structure layer. Fine-tune the fully connected classification and regression layers on the target domain data of the new module. Set the fine-tuning learning rate to 0.0001. Set the batch size to 32. Set the number of training epochs to 50. Based on a small number of target domain samples, converge the new module to usable diagnostic accuracy. The adapted model parameters are fed back to step S1 to expand the coverage boundary of the physical baseline model.

[0071] In this embodiment: Step S1 completes the synchronous registration task of high-frequency voltage and current signals and low-frequency temperature and pressure signals by constructing a spatiotemporal alignment mechanism for multi-source heterogeneous data, and establishes an observation benchmark with physical consistency, thereby eliminating the timing misalignment error caused by heterogeneous sensor sampling, providing a highly reliable data input basis for subsequent fault diagnosis, and achieving the purpose of multimodal information fusion and noise suppression.

[0072] Step S2 constructs a physical constraint-guided contrastive learning framework to complete the decoupling and separation of three types of fault features: mechanical, electrochemical, and thermal management. It projects the aliased residual signals onto an orthogonalized feature subspace, thereby breaking through the bottleneck of mutual interference of fault mechanism signals in traditional methods. This achieves the purpose of purifying and characterizing fault type-specific features, providing a pollution-free input node for causal inference.

[0073] Step S3 completes the paradigm shift from correlation analysis to causal inference by constructing a structural causal model that integrates physical priors. It uses counterfactual intervention to infer and determine the type of root cause of the failure, thereby overcoming the limitation of existing technologies that can only identify symptoms but cannot trace the root cause. This achieves the goal of accurately determining the dominant mechanism of failure evolution and outputting the corresponding heat source hypothesis.

[0074] Step S4 completes the inversion calculation task from module-level surface temperature to field-source level three-dimensional coordinates by constructing a framework for solving the inverse heat conduction problem. Based on the root cause type, the heat source model is adaptively selected, thereby solving the problem that existing diagnostic methods can only determine the overall abnormality of the module but cannot locate the specific faulty unit. This achieves the purpose of accurate spatial coordinate positioning and heat source intensity identification.

[0075] Step S5 completes the task of dynamically predicting the fault evolution trend by performing forward propagation simulation with multi-physics coupling. Based on spatial coordinates and heat source parameters, it infers the temperature field expansion range and capacity decay trajectory, thereby transforming post-event alarm into pre-event prediction, achieving the purpose of early warning of fault propagation path and quantifying capacity loss trend.

[0076] Step S6 constructs a multi-objective collaborative optimization function to complete the intelligent generation task of differentiated degradation strategy. It coordinates the three objectives of remaining capacity, security risks and switching impact to generate selective bypass and hot buffer strategies, thereby overcoming the capacity waste problem caused by the overall degradation of existing technologies and achieving the goal of precise handling and maximizing available capacity.

[0077] Step S7 establishes a virtual-real bidirectional data channel to complete the incremental update task of the digital twin model. It uses the deviation between the actual response and the prediction results to correct the causal model and inversion framework parameters, thereby solving the problem of the static model's accuracy decaying with battery aging and achieving the goal of adaptive model evolution and maintaining long-term diagnostic stability.

[0078] Step S8 completes the generalization and adaptation task of diagnostic capabilities by constructing a meta-knowledge base and cross-domain transfer mechanism, extracts common fault evolution laws and transfers them to new modules with heterogeneous geometric configurations or chemical systems, thereby avoiding the resource consumption of repeated long-term aging experiments for each new structure, achieving the goal of rapid cold start and knowledge reuse.

[0079] Compared to existing threshold-based correlation alarm mechanisms, this method achieves a fundamental leap from "symptom recognition" to "root cause identification" by introducing causal inference. It achieves a breakthrough in spatial accuracy, moving from "module-level fuzzy early warning" to "individual-level precise positioning," through solving the inverse problem of heat conduction. Furthermore, it improves efficiency by shifting from "overall derating and conservative handling" to "differentiated and precise intervention" through multi-objective optimization strategies. Simultaneously, it solves the problem of static model accuracy drifting with battery aging through virtual-real closed-loop verification and incremental update mechanisms, and overcomes the application limitations of single-module customized models through cross-module knowledge transfer. Overall, this method advances the early warning time window for energy storage battery module fault diagnosis from the current minute level to the hour level, improves spatial positioning accuracy to individual-level resolution, and enhances the safety, economy, and intelligent operation and maintenance level of energy storage systems. It provides reliable technical support for predictive maintenance of large-scale energy storage power stations and electric vehicles.

[0080] Example 2: Please refer to Figure 1 In step S1, the specific method of data acquisition and fusion is as follows:

[0081] S1.1 Real-time acquisition of voltage, current and surface temperature time-series signals of each individual cell in the module, fusion of multi-source heterogeneous data of mechanical expansion pressure and cooling system flow, and construction of original observation matrix with spatiotemporal consistency through multi-rate timestamp alignment and spatial domain interpolation registration.

[0082] S1.2 Input the original observation matrix into the physical baseline model of electrochemical thermomechanical coupling, calculate the theoretical expected response value under the current operating condition, extract the residual signal between the actual observation and the physical expectation through differential operation, and adaptively adjust the judgment threshold according to the battery aging state. When the cumulative residual value exceeds the judgment threshold, the fault diagnosis process is triggered.

[0083] The physical baseline model for electrochemical-thermomechanical coupling consists of three bidirectional coupling levels: an electrochemical sub-model, a thermal sub-model, and a mechanical sub-model. The electrochemical sub-model employs a P2D pseudo-two-dimensional model based on porous electrode theory. Using the charge and discharge current of each individual cell as input, it solves for the concentration distribution of lithium ions in the solid-phase active particles and the liquid-phase electrolyte. Based on the Butler-Volmer electrochemical reaction kinetic equation, it calculates the local current density and reaction overpotential, and then outputs the theoretical terminal voltage and electrochemical heat generation power density of each individual cell.

[0084] The thermal sub-model uses the heat generation power density output by the electrochemical sub-model as the volume heat source term, embeds the transient heat conduction control equation based on the law of energy conservation, and combines the cooling system flow rate and inlet temperature to determine the third type of convective heat transfer boundary conditions, and solves the theoretical response value of the three-dimensional temperature field distribution inside the module.

[0085] The mechanical sub-model uses the three-dimensional temperature field distribution output by the thermal sub-model as the thermal load input, and calculates the theoretical distribution values ​​of the thermal expansion displacement of the battery cell caused by the temperature gradient and the mechanical expansion pressure inside the module based on the thermoelastic mechanics equations. The above three sub-models form a closed loop through bidirectional transmission of coupling variables.

[0086] The electrochemical sub-model inputs the heat generation power density into the thermal sub-model, which in turn feeds back the temperature field to the electrochemical sub-model to correct the electrochemical reaction kinetic parameters. Simultaneously, the temperature field is input into the mechanical sub-model as the thermal load boundary condition. The theoretical terminal voltage, theoretical surface temperature, and theoretical mechanical expansion pressure output from the electrochemical, thermal, and mechanical sub-models are compared with the measured values ​​of the corresponding physical quantities in the actual observation matrix to obtain a residual signal sequence. The judgment threshold is adaptively adjusted based on the capacity decay coefficient and internal resistance growth coefficient corresponding to the battery aging state. When the cumulative residual value exceeds this adaptive judgment threshold, the subsequent fault diagnosis process is triggered.

[0087] In this embodiment: Step S1.1 effectively solves the problem of observation distortion caused by time misalignment and uneven spatial distribution between high-frequency sampling of voltage and current and low-frequency acquisition of temperature and pressure through multi-rate timestamp alignment and spatial domain interpolation registration. It constructs a high-dimensional observation matrix with spatiotemporal consistency, ensuring that different physical quantities are comparable and calculable under a unified coordinate reference, and provides a high-fidelity data foundation for subsequent physical baseline comparison.

[0088] Step S1.2 uses an electrochemical-thermomechanical coupling model to deduce the theoretical expected response under the current operating condition, uses differential operations to extract the dynamic residual between the actual observation and the physical expectation, and combines the battery aging state to adaptively adjust the judgment threshold, effectively separating normal operating condition fluctuations from abnormal fault signals, avoiding the accumulation of false alarms caused by battery degradation, and realizing the accurate triggering and reliable transmission of fault-sensitive information.

[0089] Compared to existing technologies that rely on fixed thresholds for fault detection based on raw signals, the above steps establish a dynamic predictive benchmark by introducing a physical baseline model. This transforms fault detection from absolute value comparison to deviation analysis, effectively suppressing operating condition fluctuations and sensor noise interference. Simultaneously, multi-source data spatiotemporal registration eliminates phase misalignment caused by sampling heterogeneity, making the states of each unit within the module comparable and traceable within a unified spatiotemporal framework. Overall, this achieves precise purification from raw data to fault-sensitive information, reducing false alarm rates and providing a high signal-to-noise ratio input foundation for subsequent causal inference.

[0090] Example 3: Please refer to Figure 1 In step S2, the specific method for feature decoupling is as follows:

[0091] S2.1 Construct a contrastive learning architecture that integrates prior physical knowledge. Using the residual signal output from step S1 as the input sample, design differentiated contrastive loss functions for three types of faults: mechanical, electrochemical, and thermal management. Forcefully separate the overlapping components in the mixed features and establish a fault type-specific discrimination boundary so that similar fault features are clustered and dissimilar fault features are excluded.

[0092] S2.2 Map the aggregated feature components to three orthogonal feature subspaces, introduce physical orthogonality constraints to ensure decoupling of mechanical fault features from pressure change gradients and electrochemical fault features from internal resistance evolution trends, generate pure feature vectors corresponding to different fault mechanisms, and output them as observation nodes of the structural causal model to the causal root cause identification step.

[0093] In this embodiment: Step S2.1 constructs a contrastive learning architecture that integrates physical priors, takes the residual signal as input, and designs differentiated contrastive losses for three types of faults: mechanical, electrochemical, and thermal management. It forcibly separates the overlapping components in the mixed features and establishes type-specific discrimination boundaries, which promotes the clustering of similar fault features and the mutual exclusion of dissimilar features. This completes the preliminary clustering of fault modes and the construction of discrimination boundaries, effectively breaking through the identification bottleneck of mutual interference between mechanical vibration and electrochemical heat generation signals in traditional methods, and achieving the purpose of preliminary classification of fault features.

[0094] Step S2 maps the aggregated feature components to three orthogonal subspaces and introduces physical orthogonality constraints to ensure the decoupling of mechanical fault features from pressure change gradients and electrochemical fault features from internal resistance evolution trends. This generates pure feature vectors corresponding to different fault mechanisms, thereby completing the deep decoupling of fault mechanisms and feature purification. This effectively eliminates the coupling interference between fault signals from different physical fields, achieving the goal of providing high-purity observation nodes for subsequent causal inference.

[0095] Compared to the root cause misjudgment caused by direct correlation analysis based on mixed signals in existing technologies, this step uses a progressive process of contrastive clustering and orthogonal decoupling. First, it uses contrastive loss to force the separation of aliased features and establish a discrimination boundary. Then, it uses physical orthogonal constraints to ensure that various fault features are expressed in independent subspaces, thus achieving a fundamental leap from signal aliasing to mechanism separation.

[0096] This process improves the purity and separability of the pre-features for fault root cause identification, effectively solving the problem of misdiagnosis caused by similar symptoms but different mechanisms in the background technology. It lays a high signal-to-noise ratio data foundation for the accurate deduction of the fault evolution mechanism by the subsequent structural causal model, enabling the diagnostic system to have the essential ability to distinguish between mechanical damage, electrochemical failure and thermal management abnormalities.

[0097] Example 4: Please refer to Figure 1 In step S3, the specific method for identifying the root cause is as follows:

[0098] S3.1. Take the pure feature vector output in step S2 as the set of observed variables, integrate the law of conservation of energy and the heat transfer mechanism to construct a priori topological constraint library, and use the causal discovery algorithm to deduce the directed causal edge between the root cause of the fault and the observed symptoms from the time series data of abnormal working conditions, establish a structural causal model that can explain the physical evolution mechanism and eliminate the interference of pseudo-correlation between variables.

[0099] S3.2. Based on the structural causal model, perform counterfactual intervention simulation. Through do-calculus, simulate virtual control operations to keep the cooling circuit constant and the charging and discharging current constant respectively. Calculate the posterior probability distribution of the temperature anomaly under each intervention scenario. Determine the dominant fault root cause type through the probability comparison mechanism and select the corresponding heat source model hypothesis to output to the subsequent spatial coordinate inversion step.

[0100] In this embodiment: Step S3.1 uses the feature vector as the observation node, integrates energy conservation and heat transfer mechanism to construct topological constraints, and uses the causal discovery algorithm to deduce the causal relationship between the root cause of the fault and the observed symptoms, establishes a structural causal model to characterize the evolution law and eliminates the interference of spurious correlation between variables, completes the model construction task from correlation to causation, and achieves the purpose of establishing a cognitive framework for the fault evolution mechanism.

[0101] Step S3.2 Based on the structural causal model, perform counterfactual intervention simulation to simulate control scenarios of keeping the cooling circuit constant and keeping the charging and discharging current constant. Calculate the posterior probability distribution of the continuous temperature anomaly under each intervention scenario. Determine the type of root cause of the fault by comparing the probabilities and select the heat source model hypothesis to output to the subsequent spatial inversion step. Complete the root cause discrimination task under the intervention simulation to achieve the purpose of distinguishing between external thermal management failure and internal electrochemical fault and outputting the heat source model.

[0102] The pure feature vector output in step S2 is mapped to a set of observation nodes for the structural causal model. The set of observation nodes includes at least mechanical fault feature nodes, electrochemical fault feature nodes, thermal management fault feature nodes, and temperature anomaly symptom nodes. A priori topological constraint library is constructed by integrating the law of conservation of energy and the heat transfer mechanism. The priori topological constraint library stipulates that electrochemical heat generation feature nodes point to temperature anomaly symptom nodes and prohibits the establishment of reverse connections, mechanical deformation heat generation feature nodes point to temperature anomaly symptom nodes and prohibits the establishment of reverse connections, and cooling flow feature nodes point to temperature anomaly symptom nodes and prohibit the establishment of reverse connections. A constraint-based causal discovery algorithm is used to determine the statistical dependency between nodes by using conditional independence test. Within the directional constraint range defined by the prior topological constraint library, a directed acyclic graph structure that satisfies conditional independence is searched. The direction of the remaining undirected edges is determined by the directional propagation rule. The directed causal edges between the root cause node of the fault and the observed symptom node are deduced, and pseudo-correlation interference edges caused by common causes or confounding factors are eliminated.

[0103] When performing counterfactual intervention inference based on a structural causal model, a do-calculus intervention operation is performed on the candidate root cause variable, cutting off all incoming causal edges of the intervention variable in the model and fixing its value to a constant value. In the first intervention scenario, do(cooling loop flow rate = rated design value) is executed, forcibly fixing the cooling loop flow rate to the rated design value while allowing other variables to evolve naturally. The posterior probability value P(temperature anomaly | do(cooling flow rate = rated value)) of the persistence of temperature abnormality symptoms under this intervention is calculated based on Bayesian network inference. In the second intervention scenario, do(charging and discharging current = current constant value) is executed, forcibly fixing the charging and discharging current to the current constant value. The temperature under this intervention is calculated based on Bayesian network inference. The posterior probability value P(temperature anomaly | do(charging / discharging current = current value)) for the persistence of abnormal symptoms is calculated. The probability comparison mechanism is to calculate the difference between the posterior probabilities of the persistence of temperature anomaly under two intervention scenarios. When the absolute value of the difference between the two posterior probability values ​​exceeds a preset threshold, the variable category corresponding to the intervention scenario with a significantly reduced posterior probability is determined to be the dominant root cause of the fault. If the temperature anomaly probability after cooling flow intervention is significantly lower than the temperature anomaly probability after charging / discharging current intervention, and the difference between the two exceeds a preset threshold, then the external thermal management failure is determined to be the dominant root cause of the fault. Otherwise, the internal electrochemical failure is determined to be the dominant root cause of the fault, and the volume heat source model or surface heat source model is adaptively selected and output to the subsequent spatial coordinate inversion step accordingly.

[0104] Existing correlation analysis based on threshold judgment can only identify symptoms but cannot trace the root cause. This step constructs a structural causal model incorporating conservation law constraints to establish a causal link between the root cause of the fault and the observed symptoms. By using counterfactual intervention to simulate control scenarios, it can distinguish between root cause mechanisms such as external heat dissipation failure and internal micro-short circuits. This process overcomes the bottleneck of being unable to attribute abnormal temperature rises, enabling the diagnostic system to identify the dominant mechanism of fault evolution. This provides a basis for determining the root cause type for subsequent accurate location and differentiated treatment, achieving a leap from correlation alarm to causal diagnosis.

[0105] Example 5: Please refer to Figure 1 In step S4, the spatial coordinate inversion is performed as follows:

[0106] S4.1. Based on the root cause type determined in step S3, the corresponding heat source model hypothesis is adaptively selected. When the fault is determined to be internal electrochemical, the volume heat source model is selected. When the fault is determined to be contact interface, the surface heat source model is selected. A heat conduction inverse problem solution framework with boundary condition constraints is constructed. The surface temperature field distribution collected in step S1 is embedded into the solution domain as a boundary constraint and a regularization term is established.

[0107] S4.2. Based on the constructed inverse problem framework, a regularized numerical method is used to solve the ill-posed inverse problem of heat conduction. The three-dimensional spatial coordinates and heat source intensity of the internal heat source in the module are inferred from the surface temperature field distribution. The non-uniqueness of the solution is suppressed by physical information constraints. The three-dimensional spatial coordinates and heat source intensity obtained by inversion are used as initial conditions and output to the propagation evolution prediction step.

[0108] In this embodiment: Step S4.1 adaptively selects the heat source model according to the root cause type. The internal electrochemical fault adopts the volume heat source model, and the contact interface fault adopts the surface heat source model. Construct a solution framework for the inverse heat conduction problem with boundary condition constraints. Embed the surface temperature field distribution as a boundary constraint into the solution domain and add a stability constraint term to complete the heat source model adaptation and boundary constraint embedding task, so as to achieve the purpose of selecting the appropriate physical model and establishing a stable solution framework according to the difference of fault mechanism.

[0109] Step S4.2 uses a stabilization numerical method to solve the inverse problem of heat conduction. The three-dimensional spatial coordinates and heat source intensity parameters of the internal heat source within the module are inferred from the surface temperature field distribution. The non-uniqueness of the solution is suppressed by physical information constraints. The three-dimensional spatial coordinates and heat source intensity obtained by inversion are used as initial conditions to output to subsequent steps, completing the spatial coordinate inversion and heat source intensity identification tasks, and achieving the purpose of accurately locating the faulty unit and quantifying the heat source intensity.

[0110] Existing technologies can only determine overall module anomalies but cannot pinpoint specific faulty cells. This step adaptively selects a heat source model based on the root cause type, constructs a framework for solving the inverse heat conduction problem, embeds the surface temperature field as a boundary constraint into the solution domain, and uses a stabilized numerical method to inversely deduce the three-dimensional coordinates and intensity parameters of the internal heat source. This process overcomes the barrier of observing the surface to locate the internal source, achieving a leap from fuzzy early warning at the module level to precise location at the cell level. The spatial positioning error is controlled within the size range of a single cell, providing reliable coordinate guidance for subsequent accurate replacement of faulty cells, significantly improving operation and maintenance efficiency and system availability.

[0111] Example 6: Please refer to Figure 1 In step S5, the specific method for propagation evolution prediction is as follows:

[0112] S5.1 Receive the three-dimensional spatial coordinate position and heat source intensity parameters obtained from the inversion in step S4, embed them as initial boundary conditions into the physical baseline model, activate the multi-physics simulation engine of electrochemical thermomechanical coupling, and perform forward propagation calculation to deduce the energy transfer and thermal mass diffusion evolution process of the fault in the module.

[0113] The multiphysics simulation engine for electrochemical-thermomechanical coupling includes an electrochemical field solution unit, a thermal field solution unit, and a mechanical field solution unit. The electrochemical field solution unit adopts a P2D pseudo-two-dimensional model and solves the solid-phase lithium-ion diffusion equation, the liquid-phase lithium-ion migration equation, and the Butler-Volmer electrochemical reaction kinetic equation based on porous electrode theory. It takes the charge and discharge current of a single cell and the heat source intensity parameters obtained from step S4 as inputs, and outputs the distribution of electrochemical reaction heat generation rate and the theoretical terminal voltage response of each single cell. The thermal field solution unit receives the distribution of electrochemical reaction heat generation rate and heat source intensity parameters as heat source terms, embeds the transient heat conduction equation based on the law of energy conservation, and combines the convection heat dissipation boundary conditions of the module shell and the contact heat conduction boundary conditions between adjacent cells to solve the distribution of the three-dimensional temperature field inside the module over time. The data processing unit receives three-dimensional temperature field distribution data as thermal load input and solves the thermal expansion displacement field and thermal stress field distribution of the battery cell and module structure based on the thermoelastic mechanics equations. The electrochemical field solution unit, thermal field solution unit, and mechanical field solution unit achieve bidirectional closed-loop transmission through coupling variables. The electrochemical field solution unit inputs the heat generation rate to the thermal field solution unit, and the thermal field solution unit feeds back the temperature distribution data to the electrochemical field solution unit to correct the electrochemical reaction kinetic parameters. At the same time, the temperature distribution data is input to the mechanical field solution unit as thermal load boundary conditions, and the mechanical field solution unit feeds back the expansion displacement field to the thermal field solution unit to correct the contact thermal conduction boundary conditions. The finite element method or finite volume method is used to perform spatiotemporal discretization solution of the coupled equations to deduce the energy transfer and heat and mass diffusion evolution process of the fault within the module.

[0114] S5.2 Based on the multiphysics simulation results, predict the evolution trend of the three-dimensional isothermal surface of the temperature field and the temperature rise trajectory of adjacent units in the future period, dynamically calculate the spatiotemporal expansion range of the fault influence domain and quantify the overall capacity loss ratio of the module, and output the fault influence domain range and capacity loss trend as constraints to the degradation strategy optimization step.

[0115] Based on the three-dimensional temperature field time-series data output by the multiphysics simulation engine, isosurface data with preset temperature thresholds are extracted according to the time sequence to deduce the evolution trend of the three-dimensional isothermal surface of the temperature field; the temperature time-series response data of the center coordinates of each cell in the vicinity of the fault cell are extracted to construct the temperature rise trajectory of the vicinity cells; a safe operating temperature threshold is set, and the continuous spatial region in the three-dimensional temperature field time-series data where the temperature value exceeds the safe operating temperature threshold is determined as the fault influence domain, the spatial volume of the fault influence domain and its expansion rate over time are calculated, and the spatiotemporal boundary data of the fault influence domain are obtained; based on the theoretical terminal voltage response of each cell output by the electrochemical field solution unit and the temperature data of each cell output by the thermal field solution unit, the capacity decay rate of each cell is calculated in combination with the temperature accelerated aging model, and the capacity decay rate of the cells located in the fault influence domain is weighted and summed with the rated total capacity of the module to quantify the overall capacity loss ratio of the module; the spatiotemporal boundary data of the fault influence domain and the overall capacity loss ratio of the module are output as constraints to the subsequent degradation strategy optimization steps.

[0116] In this embodiment: Step S5.1 receives three-dimensional spatial coordinates and heat source intensity parameters as initial boundary conditions and embeds them into the physical baseline model. The multiphysics coupling simulation engine is activated to perform forward propagation calculations, deduce the energy transfer and thermal diffusion evolution process of the fault within the module, complete the calculation task of the fault dynamic evolution mechanism, and achieve the purpose of establishing a physical picture of fault propagation and providing a computational basis for subsequent predictions.

[0117] Step S5.2 Based on the multiphysics simulation results, predict the evolution trend of the isothermal surface of the temperature field and the temperature rise trajectory of adjacent units in the future period, dynamically calculate the spatiotemporal expansion range of the fault influence domain and quantify the overall capacity loss ratio of the module, and output the fault influence domain and capacity loss trend as constraints to complete the task of predicting the fault propagation trend and quantifying the impact, so as to achieve the purpose of grasping the fault evolution situation in advance and providing a quantitative basis for the disposal strategy.

[0118] Existing technologies mostly remain at the fault alarm stage, lacking the ability to predict evolution trends. This step, by embedding the fault location and intensity parameters obtained from inversion into the physical baseline model, activates multiphysics coupled simulation to deduce the fault energy transfer and heat and mass diffusion processes, thereby predicting the temperature field evolution trend and the expansion range of the influence domain, realizing the transformation from static alarm to dynamic prediction. This process shifts from post-event handling to pre-event intervention, providing early warning of fault propagation paths and capacity loss trends several hours in advance, allowing sufficient response time for the generation of differentiated degradation strategies, and significantly improving the proactive safety protection capabilities of energy storage systems.

[0119] Example 7: Please refer to Figure 1 In step S6, the specific optimization method of the degradation strategy is as follows:

[0120] S6.1 Construct a constrained boundary domain based on the spatiotemporal evolution range of the fault impact domain and the capacity decay trend predicted in step S5. Integrate the fault root cause types determined in step S4 to establish a multi-objective collaborative optimization function, taking into account the three objectives of maximizing the remaining available capacity, minimizing the thermal runaway risk index, and minimizing the impact of strategy switching as the optimization criteria.

[0121] S6.2 Solve the multi-objective collaborative optimization function to obtain the Pareto optimal solution set, generate a differentiated degradation control strategy that includes selectively bypassing faulty cells, adjusting the charge and discharge rates of adjacent healthy cells to build a thermal buffer barrier, and activating the backup thermal management circuit, and demap the differentiated degradation control strategy into a sequence of control instructions that can be executed by the battery management system.

[0122] In this embodiment: Step S6.1 constructs a constraint boundary based on the spatiotemporal range of the fault influence domain and the capacity decay trend, integrates the fault root cause types to establish a multi-objective collaborative optimization function, and takes the maximization of remaining available capacity, the minimization of thermal runaway risk index and the minimization of strategy switching impact as optimization criteria, thereby transforming the propagation prediction results and root cause discrimination information into quantifiable strategy optimization boundaries, completing the objective function construction task under multiple constraints, and achieving the goal of establishing a strategy optimization benchmark that takes into account safety, economy and stability.

[0123] Step S6.2 Solve the multi-objective collaborative optimization function to obtain the optimal strategy solution set, generate a differentiated degradation control strategy that includes selective bypass of faulty cells, adjustment of the charge and discharge rates of adjacent healthy cells to form a thermal buffer, activation of the backup thermal management circuit, and mapping the strategy solution into a sequence of control instructions that can be executed by the battery management system. This completes the transformation from optimization solution to control instruction generation, achieving the goal of generating accurate and executable disposal solutions and avoiding capacity waste caused by overall derating.

[0124] Existing technologies mostly focus on fault alarms or single-threshold protection, with responses often limited to overall current limiting or shutdown. They lack the ability to generate refined strategies based on fault evolution prediction, leading to excessive and highly impactful capacity loss. This step addresses this by constructing a multi-objective collaborative optimization framework, transforming fault impact domain prediction and root cause identification results into constraints for strategy optimization, thus shifting from overall capacity reduction to precise intervention. Through differentiated methods such as selective bypass and thermal buffering, it maximizes the retention of usable capacity while ensuring controllable thermal runaway risk, significantly improving the energy storage system's ability to balance economy and safety.

[0125] Example 8: Please refer to Figure 1 In step S7, the specific method for virtual-real closed-loop correction is as follows:

[0126] S7.1 Map the differentiated degradation control strategy generated in step S6 to the physical energy storage entity system for execution and establish a virtual-real bidirectional data channel. Synchronously collect the actual voltage and temperature response time series data after the strategy is implemented, and calculate the virtual-real consistency deviation with the theoretical evolution trajectory predicted in step S5. Quantify the prediction error and determine whether it exceeds the preset tolerance threshold.

[0127] S7.2 When the virtual-real consistency verification shows that the prediction error exceeds the preset tolerance threshold, the causal edge weight coefficients of the structural causal model in step S3 and the regularization constraint parameters of the heat conduction inverse problem solution framework in step S4 are incrementally corrected using actual response time series data to realize the adaptive evolution and update of the digital twin model, and the evolved model parameters are fed back to step S1 and step S3 to improve the accuracy of subsequent diagnosis.

[0128] In this embodiment: Step S7.1 maps the differentiated degradation control strategy to the physical energy storage entity system for execution, establishes a virtual and real bidirectional data channel to synchronously collect the actual voltage and temperature response time series data after the strategy is implemented, performs consistency deviation calculation with the theoretical prediction trajectory, quantifies the prediction error and determines whether it exceeds the preset tolerance threshold, thereby completing the closed-loop monitoring and prediction accuracy verification task of the strategy implementation effect, achieving the purpose of establishing a digital twin virtual and real verification mechanism and providing a measured data basis for model optimization.

[0129] Step S7.2 When the virtual-real consistency verification shows that the prediction error exceeds the tolerance range, the causal edge weight coefficients of the structural causal model and the parameters of the heat conduction inverse problem solution framework are incrementally corrected using the actual response time series data. This enables the digital twin model to adaptively evolve and update with the battery aging state. The corrected model parameters are then fed back to the data acquisition and causal inference stages, thereby completing the task of online compensation for model accuracy and long-term stability maintenance, achieving the goal of suppressing model drift and maintaining diagnostic accuracy without decay.

[0130] Existing technologies mostly employ static offline training models, which remain fixed once deployed. This makes it difficult to adapt to the accuracy degradation caused by the continuous drift of aging characteristics throughout the battery's lifespan, resulting in a significant decrease in diagnostic accuracy over time. This step establishes a two-way data channel between the physical system and the digital twin's predicted trajectory in real time, forming a closed-loop verification mechanism. When the deviation exceeds the tolerance, the model correction process is automatically triggered, using actual operating data to update the causal model weights and inverse problem solution parameters online, enabling the digital twin model to adaptively evolve with battery aging.

[0131] This process overcomes the limitations of static models, enabling the diagnostic system to have self-learning and self-correction capabilities. It can maintain a high level of diagnostic accuracy even under long-term operation, effectively solving the problem of model accuracy decaying over time and ensuring diagnostic reliability throughout the entire life cycle from early stages to retirement.

[0132] Example 9: Please refer to Figure 1 In step S8, the specific methods of knowledge transfer generalization are as follows:

[0133] S8.1 Mining historical case time series data accumulated in the virtual-real closed loop of step S7, extracting common fault evolution laws and causal topological patterns across geometric configurations and chemical systems, condensing them into a meta-knowledge base containing inversion structure and root cause discrimination knowledge and storing it.

[0134] S8.2 For new modules with heterogeneous geometric configurations or chemical systems, the causal reasoning structure and inverse problem solving in the transfer meta-knowledge base are adapted across domains. Based on a small number of target domain samples, the new module is fine-tuned to converge to usable diagnostic accuracy. The adapted model parameters are then fed back to step S1 to expand the coverage boundary of the physical baseline model.

[0135] In this embodiment: Step S8.1 mines historical case data accumulated during the virtual-real closed-loop verification process, extracts common fault evolution laws and causal topological patterns that do not depend on specific geometric configurations and chemical systems, and condenses the inversion structure and root cause discrimination knowledge into a reusable meta-knowledge base, thereby breaking through the limitations of single-module customized models, establishing a basic support system for cross-platform diagnostic knowledge, and avoiding the resource consumption of repeatedly conducting long-term aging experiments for each new type of battery.

[0136] Step S8.2 When facing new module scenarios with different geometric configurations or chemical systems, the causal reasoning structure and inverse problem solving ability fixed in the meta-knowledge base are directly transferred. Based on a small number of target domain samples, the new module is quickly fine-tuned to achieve usable diagnostic accuracy. The adapted model parameters are fed back to the data acquisition stage to expand the coverage boundary of the physical baseline model, thereby completing the cross-domain adaptation and generalization deployment of diagnostic capabilities.

[0137] Existing technologies are mostly customized for specific modules, lacking cross-platform reusability. This necessitates the collection of large amounts of data for long-term training for new battery models, resulting in lengthy development cycles and high costs. This step constructs a meta-knowledge base by mining common patterns from historical cases, enabling a leap from specific scenarios to general capabilities in diagnostic knowledge. Furthermore, through transfer learning mechanisms, new modules can quickly acquire diagnostic capabilities with only a small number of samples, effectively solving the cold start problem. This process significantly shortens the development and verification cycle of new energy storage systems, reduces data acquisition costs, and enables the method to be universally applicable across geometric configurations and chemical systems for industrial application.

[0138] This application also provides a machine learning-based energy storage battery module simulation system, including a data acquisition and fusion module, which acquires module voltage, current, temperature, pressure and flow signals, constructs an observation matrix through spatiotemporal registration, inputs a physical baseline model to calculate residuals and adaptively determines and triggers diagnosis based on thresholds;

[0139] The fault feature decoupling module receives residual input and compares with the learning architecture to separate mechanical, electrochemical, and thermal management fault features and maps them to an orthogonal subspace to generate pure feature vectors.

[0140] The causal root cause identification module constructs a structural causal model using features as observation nodes, deduces the causal edge between root causes and symptoms, and determines the root cause type by calculating the probability through counterfactual intervention.

[0141] The spatial coordinate inversion module selects the heat source model based on the root cause type, uses the temperature field as the boundary constraint, and regularizes and inversely calculates the three-dimensional coordinates of the fault and the intensity of the heat source.

[0142] The propagation and evolution prediction module embeds the coordinate heat source into the physical model, simulates and extrapolates energy transfer through multiphysics, and predicts the evolution of the temperature field and the range of its influence domain.

[0143] The degradation strategy optimization module constructs a multi-objective optimization function with the influence domain as a constraint, and solves the degradation strategy to generate selective bypass faulty cells and adjust the ratio of neighboring cells.

[0144] The virtual-real closed-loop correction module executes the strategy to collect response data, calculates the deviation from the predicted trajectory, and corrects the causal model weights and inverse problem parameters to update the model.

[0145] The cross-module knowledge transfer module mines historical cases to extract common patterns and builds a meta-knowledge base. It then adapts the causal structure of the new module for cross-domain transfer and fine-tunes convergence, expanding the baseline coverage.

[0146] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0147] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

[0148] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A machine learning-based simulation method for energy storage battery modules, characterized in that: The specific steps are as follows: S1. Data Acquisition and Fusion: Real-time acquisition of module voltage, current, temperature and pressure signals, construction of spatiotemporal observation matrix, calculation of residuals through physical baseline model and triggering diagnosis; S2, Feature Decoupling: Project the residuals onto an orthogonal subspace to separate mechanical, electrochemical, and thermal management fault features and generate type-specific feature vectors; S3. Causal Root Cause Identification: Based on physical priors, construct a structural causal model, perform counterfactual intervention reasoning, determine the dominant fault root cause type, and output the heat source hypothesis; S4. Spatial coordinate inversion: Select the heat source model according to the root cause type, solve the inverse heat conduction problem, and invert the three-dimensional coordinates of the fault and the heat source intensity within the module. S5. Propagation and Evolution Prediction: Multiphysics simulation is performed by embedding three-dimensional coordinates and heat source intensity to predict the spatiotemporal expansion range of the fault impact domain and the capacity loss trend. S6. Degradation Strategy Optimization: Construct a multi-objective optimization function with the influence domain as a constraint to generate differentiated degradation control strategies with selective bypass and hot buffer; S7. Virtual-Real Closed-Loop Correction: Execute differentiated degradation control strategies and compare actual responses with predicted results to correct structural causal model parameters and inversion weights online, achieving incremental updates; S8. Knowledge Transfer and Generalization: Extract common fault features across modules, adapt to new modules with different structures through transfer learning, and expand the coverage of the physical baseline model through feedback.

2. The machine learning-based simulation method for energy storage battery modules according to claim 1, characterized in that, In step S1, the specific methods for data acquisition and fusion are as follows: S1.1 Real-time acquisition of voltage, current and surface temperature time-series signals of each individual cell in the module, fusion of multi-source heterogeneous data of mechanical expansion pressure and cooling system flow, and construction of original observation matrix with spatiotemporal consistency through multi-rate timestamp alignment and spatial domain interpolation registration. S1.2 Input the original observation matrix into the physical baseline model of electrochemical thermomechanical coupling, calculate the theoretical expected response value under the current operating condition, extract the residual signal between the actual observation and the physical expectation through differential operation, and adaptively adjust the judgment threshold according to the battery aging state. When the cumulative residual value exceeds the judgment threshold, the fault diagnosis process is triggered.

3. The machine learning-based simulation method for energy storage battery modules according to claim 2, characterized in that, In step S2, the specific method of feature decoupling is as follows: S2.1 Construct a contrastive learning architecture that integrates prior physical knowledge. Using the residual signal output from step S1 as the input sample, design differentiated contrastive loss functions for three types of faults: mechanical, electrochemical, and thermal management. Forcefully separate the overlapping components in the mixed features and establish a fault type-specific discrimination boundary so that similar fault features are clustered and dissimilar fault features are excluded. S2.2 Map the aggregated feature components to three orthogonal feature subspaces, introduce physical orthogonality constraints to ensure decoupling of mechanical fault features from pressure change gradients and electrochemical fault features from internal resistance evolution trends, generate pure feature vectors corresponding to different fault mechanisms, and output them as observation nodes of the structural causal model to the causal root cause identification step.

4. The machine learning-based simulation method for energy storage battery modules according to claim 3, characterized in that, In step S3, the specific method for identifying causal roots is as follows: S3.

1. Take the pure feature vector output in step S2 as the set of observed variables, integrate the law of conservation of energy and the heat transfer mechanism to construct a priori topological constraint library, and use the causal discovery algorithm to deduce the directed causal edge between the root cause of the fault and the observed symptoms from the time series data of abnormal working conditions, establish a structural causal model that can explain the physical evolution mechanism and eliminate the interference of pseudo-correlation between variables. S3.

2. Based on the structural causal model, perform counterfactual intervention simulation. Through do-calculus, simulate virtual control operations to keep the cooling circuit constant and the charging and discharging current constant respectively. Calculate the posterior probability distribution of the temperature anomaly under each intervention scenario. Determine the dominant fault root cause type through the probability comparison mechanism and select the corresponding heat source model hypothesis to output to the spatial coordinate inversion step.

5. The machine learning-based simulation method for energy storage battery modules according to claim 4, characterized in that, In step S4, the spatial coordinate inversion is performed as follows: S4.

1. Based on the root cause type determined in step S3, the corresponding heat source model hypothesis is adaptively selected. When the fault is determined to be internal electrochemical, the volume heat source model is selected. When the fault is determined to be contact interface, the surface heat source model is selected. A heat conduction inverse problem solution framework with boundary condition constraints is constructed. The surface temperature field distribution collected in step S1 is embedded into the solution domain as a boundary constraint and a regularization term is established. S4.

2. Based on the constructed inverse problem framework, a regularized numerical method is used to solve the ill-posed inverse problem of heat conduction. The three-dimensional spatial coordinates and heat source intensity of the internal heat source in the module are inferred from the surface temperature field distribution. The non-uniqueness of the solution is suppressed by physical information constraints. The three-dimensional spatial coordinates and heat source intensity obtained by inversion are used as initial conditions and output to the propagation evolution prediction step.

6. The machine learning-based simulation method for energy storage battery modules according to claim 5, characterized in that, In step S5, the specific method for predicting propagation and evolution is as follows: S5.1 Receive the three-dimensional spatial coordinate position and heat source intensity parameters obtained from the inversion in step S4, embed them as initial boundary conditions into the physical baseline model, activate the multi-physics simulation engine of electrochemical thermomechanical coupling, and perform forward propagation calculation to deduce the energy transfer and thermal mass diffusion evolution process of the fault in the module. S5.2 Based on the multiphysics simulation results, predict the evolution trend of the three-dimensional isothermal surface of the temperature field and the temperature rise trajectory of adjacent units in the future period, dynamically calculate the spatiotemporal expansion range of the fault influence domain and quantify the overall capacity loss ratio of the module, and output the fault influence domain range and capacity loss trend as constraints to the degradation strategy optimization step.

7. The machine learning-based simulation method for energy storage battery modules according to claim 6, characterized in that, In step S6, the specific optimization method of the degradation strategy is as follows: S6.1 Construct a constrained boundary domain based on the spatiotemporal evolution range of the fault impact domain and the capacity decay trend predicted in step S5. Integrate the fault root cause types determined in step S4 to establish a multi-objective collaborative optimization function, taking into account the three objectives of maximizing the remaining available capacity, minimizing the thermal runaway risk index, and minimizing the impact of strategy switching as the optimization criteria. S6.2 Solve the multi-objective collaborative optimization function to obtain the Pareto optimal solution set, generate a differentiated degradation control strategy that includes selectively bypassing faulty cells, adjusting the charge and discharge rates of adjacent healthy cells to build a thermal buffer barrier, and activating the backup thermal management circuit, and demap the differentiated degradation control strategy into a sequence of control instructions that can be executed by the battery management system.

8. The machine learning-based simulation method for energy storage battery modules according to claim 7, characterized in that, In step S7, the specific method for virtual-real closed-loop correction is as follows: S7.1 Map the differentiated degradation control strategy generated in step S6 to the physical energy storage entity system for execution and establish a virtual-real bidirectional data channel. Synchronously collect the actual voltage and temperature response time series data after the strategy is implemented, and calculate the virtual-real consistency deviation with the theoretical evolution trajectory predicted in step S5. Quantify the prediction error and determine whether it exceeds the preset tolerance threshold. S7.2 When the virtual-real consistency verification shows that the prediction error exceeds the preset tolerance threshold, the causal edge weight coefficients of the structural causal model in step S3 and the regularization constraint parameters of the heat conduction inverse problem solution framework in step S4 are incrementally corrected using actual response time series data to realize the adaptive evolution and update of the digital twin model, and the evolved model parameters are fed back to step S1 and step S3 to improve the accuracy of subsequent diagnosis.

9. The machine learning-based simulation method for energy storage battery modules according to claim 7, characterized in that, In step S8, the specific methods of knowledge transfer generalization are as follows: S8.1 Mining historical case time series data accumulated in the virtual-real closed loop of step S7, extracting common fault evolution laws and causal topological patterns across geometric configurations and chemical systems, condensing them into a meta-knowledge base containing inversion structure and root cause discrimination knowledge and storing it. S8.2 For new modules with heterogeneous geometric configurations or chemical systems, the causal reasoning structure and inverse problem solving in the transfer meta-knowledge base are adapted across domains. Based on a small number of target domain samples, the new module is fine-tuned to converge to usable diagnostic accuracy. The adapted model parameters are then fed back to step S1 to expand the coverage boundary of the physical baseline model.

10. A machine learning-based simulation system for energy storage battery modules, characterized in that, The machine learning-based energy storage battery module simulation system is used to execute the machine learning-based energy storage battery module simulation method described in any of claims 1 to 9. The system includes a data acquisition and fusion module, which acquires module voltage, current, temperature, pressure, and flow signals, constructs an observation matrix through spatiotemporal registration, inputs a physical baseline model to calculate residuals, and adaptively determines and triggers diagnostics based on thresholds. The fault feature decoupling module receives residual input and compares with the learning architecture to separate mechanical, electrochemical, and thermal management fault features and maps them to an orthogonal subspace to generate pure feature vectors. The causal root cause identification module constructs a structural causal model using features as observation nodes, deduces the causal edge between root causes and symptoms, and determines the root cause type by calculating the probability through counterfactual intervention. The spatial coordinate inversion module selects the heat source model based on the root cause type, uses the temperature field as the boundary constraint, and regularizes and inversely calculates the three-dimensional coordinates of the fault and the intensity of the heat source. The propagation and evolution prediction module embeds the coordinate heat source into the physical model, simulates and extrapolates energy transfer through multiphysics, and predicts the evolution of the temperature field and the range of its influence domain. The degradation strategy optimization module constructs a multi-objective optimization function with the influence domain as a constraint, and solves the degradation strategy to generate selective bypass faulty cells and adjust the ratio of neighboring cells. The virtual-real closed-loop correction module executes the strategy to collect response data, calculates the deviation from the predicted trajectory, and corrects the causal model weights and inverse problem parameters to update the model. The cross-module knowledge transfer module mines historical cases to extract common patterns and builds a meta-knowledge base. It then adapts the causal structure of the new module for cross-domain transfer and fine-tunes convergence, expanding the baseline coverage.

Images