A method and system for short circuit detection within battery modules based on magnetic field detection

Abstract

This invention belongs to the field of battery testing technology, and particularly relates to a method and system for detecting short circuits within battery modules based on magnetic field detection. The method includes constructing a three-dimensional magnetic field sensing array; selecting an effective observation window based on quasi-static logic, and triggering a magnetic sensor to collect magnetic induction intensity data within the effective observation window; decoupling the background magnetic field and the topological current-carrying contribution magnetic field from the raw magnetic induction intensity data to obtain residual magnetic field characteristics; inputting the residual magnetic field characteristics into a constructed physical constraint neural network model to reconstruct the current density based on physical driving; and employing a multi-level fault response strategy based on the reconstructed current density to classify the battery state into different safety levels, thereby completing the detection of internal short circuits within the battery module. This invention, based on a three-dimensional magnetic field array sensing combined with a physical constraint deep learning model, achieves internal short circuit detection within battery modules, providing an optional implementation path for the early and accurate diagnosis of short circuit faults within battery modules.

Description

Technical Field

[0001] This invention belongs to the field of battery testing technology, and particularly relates to a method and system for detecting short circuits within battery modules based on magnetic field detection. Background Technology

[0002] Lithium-ion batteries, with their advantages of high energy density, long cycle life, and low self-discharge rate, have become the most mainstream power and energy storage carrier for new energy vehicles. However, as an energy storage device involving complex electrochemical reactions, lithium-ion batteries pose significant safety risks. On the one hand, due to the complex electrochemical and thermo-electric coupling processes involved in lithium-ion batteries, factors such as overcharging, over-discharging, overheating, mechanical stress, and manufacturing defects can lead to conditions such as separator damage, dendrite growth, or foreign object puncture, thereby inducing or accelerating the formation of internal short circuits (ISCs) and other safety hazards. On the other hand, to meet the demands of high-energy and high-power applications, multiple individual cells are typically integrated into battery modules through series or parallel connections. At the module level, electromagnetic interference introduced by current paths and structural components (such as busbars, connectors, and casings), as well as issues such as inconsistent individual cell parameters, operating current fluctuations, and series-parallel coupling effects, can cause weak abnormal current signals generated in the early stages of ISCs to be easily masked by normal operating currents and structural currents, thus increasing the difficulty of early fault identification and localization. In its early stages, an intermittent short circuit (ISC) often manifests as a weak leakage current or localized abnormal heating, making it difficult to detect directly. As the short circuit worsens, it can lead to accumulated localized temperature rise and further induce thermal runaway risks. Therefore, early and accurate identification of potential faults such as ISCs is crucial.

[0003] Existing research methods for lithium-ion battery fault diagnosis can be mainly divided into model-based methods, signal processing-based methods, and data-driven methods. These are analyzed in detail below:

[0004] (1) Model-based approach:

[0005] Model-based methods (such as equivalent circuit models and electrochemical models) typically characterize changes in the internal state of a battery through parameter identification (such as internal resistance, polarization-related parameters, etc.), and have certain advantages in estimating the state of charge (SOC) and state of health (SOH).

[0006] However, in early diagnosis of internal short circuits (ISC), such methods often rely heavily on the accuracy of model structure and parameters. When there are many cells in the module and there are differences in aging, temperature changes and fluctuations in operating conditions, the battery parameters exhibit nonlinear and time-varying characteristics, making it difficult to establish a high-fidelity model that balances accuracy and real-time performance, thus affecting the stability and robustness of the diagnostic results.

[0007] (2) Signal processing-based methods and data-driven methods:

[0008] Signal processing-based methods are the most commonly used approach in current battery management systems (BMS), primarily utilizing macroscopic external characteristic parameters such as voltage, current, and temperature for signal processing. For example, regarding voltage signals, BMS can identify faults through indicators such as terminal voltage fluctuations and differential voltage, facilitating engineering deployment. However, in the early stages of a micro-short circuit, the abnormal current may be in the milliampere range or lower, resulting in a small change in terminal voltage, which is easily masked by sensor noise, sampling errors, or SOC estimation errors, making early identification difficult. Regarding temperature signals, temperature sensors are mostly placed on the battery surface. Limited by heat diffusion time and environmental disturbances, the conduction of internal hot spots to the surface is delayed, making rapid early warning difficult.

[0009] In addition, gas-sensitive and acoustic detection methods often only show obvious characteristics after the fault has developed to a certain extent, which are relatively lagging alarm methods.

[0010] Given the limitations of the aforementioned methods in early ISC fault diagnosis, magnetic field detection technology based on the current magnetic effect has become an important direction for next-generation smart battery diagnostics. As a physical quantity generated by current, the spatial distribution of the magnetic field can reflect changes in the loop current and current distribution. Under quasi-static conditions, disturbances in the current distribution will cause corresponding changes in the external magnetic field distribution. Compared to measurements such as voltage and temperature, magnetic field detection has advantages in response speed and sensitivity to changes in deep current. However, its effectiveness is affected by sensor bandwidth, arrangement, and module structure conditions, and usually requires baseline modeling or differential strategies to suppress the background contribution of normal operating current.

[0011] Although the magnetic field method has a theoretical basis, its application to complex battery modules still faces many challenges: the high-voltage busbars and connectors inside the module generate strong background magnetic field interference; the metal shell and conductive structure may introduce induced eddy current effects under dynamic conditions, causing magnetic field amplitude attenuation and spatial distortion; at the same time, inferring the internal current distribution from the external magnetic field is an inverse magnetic field problem, which has ill-posedness and computational complexity constraints, and traditional high-fidelity modeling (such as finite element method) faces pressure from computing power and modeling costs in online real-time deployment. Summary of the Invention

[0012] To overcome the shortcomings of the prior art, this invention provides a method and system for short circuit detection within battery modules based on magnetic field detection. It achieves short circuit detection within battery modules by combining a three-dimensional magnetic field array sensing with a physical constraint deep learning model. The three-dimensional detection overcomes the shielding effect, and the physical constraints make up for the data shortcomings, providing an optional implementation path for the early and accurate diagnosis of short circuit faults within battery modules.

[0013] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0014] The first aspect of this invention provides a method for detecting short circuits within a battery module based on magnetic field detection.

[0015] A short-circuit detection method for battery modules based on magnetic field detection includes the following steps:

[0016] Magnetic sensors are deployed at designated locations on the battery module to construct a three-dimensional magnetic field sensing array;

[0017] Based on quasi-static logic, a valid observation window is selected, and the magnetic sensor is triggered to collect magnetic induction intensity data within the valid observation window to obtain the raw magnetic induction intensity data.

[0018] The background magnetic field and the topological current-carrying contribution magnetic field are decoupled from the original magnetic induction intensity data to obtain the residual magnetic field characteristics;

[0019] The residual magnetic field characteristics are input into the constructed physical constraint neural network model, and the reconstructed current density is derived based on physical driving.

[0020] A multi-level fault response strategy based on reconfigurable current density is adopted to classify the battery state into different safety levels and complete the internal short circuit detection of the battery module.

[0021] A second aspect of the present invention provides a short-circuit detection system for battery modules based on magnetic field detection.

[0022] A short-circuit detection system within a battery module based on magnetic field detection includes:

[0023] The magnetic field sensing module is configured to: deploy magnetic sensors at designated locations on the battery module to construct a three-dimensional magnetic field sensing array;

[0024] The raw data acquisition module is configured to: filter the effective observation window based on quasi-static logic, trigger the magnetic sensor to collect magnetic induction intensity data within the effective observation window, and obtain the raw magnetic induction intensity data.

[0025] The decoupling module is configured to decouple the background magnetic field from the topological current-carrying contribution magnetic field of the original magnetic induction intensity data to obtain the residual magnetic field characteristics.

[0026] The model inversion module is configured to input the residual magnetic field characteristics into the constructed physical constraint neural network model and invert the reconstructed current density based on physical drive.

[0027] The multi-level response module is configured to: adopt a multi-level fault response strategy based on reconfigured current density to classify the battery state into different safety levels and complete the internal short circuit detection of the battery module.

[0028] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the short-circuit detection method for a battery module based on magnetic field detection as described in the first aspect of the present invention.

[0029] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the short-circuit detection method for a battery module based on magnetic field detection as described in the first aspect of the present invention.

[0030] The above one or more technical solutions have the following beneficial effects:

[0031] This invention provides a method and system for detecting short circuits within battery modules based on magnetic field detection. It overcomes the dual bottlenecks of computing power and data in traditional magnetic field diagnostic techniques by innovatively introducing a Physical Information Neural Network (PINN) to handle the inverse magnetic field problem of battery modules. Compared to traditional finite element analysis (FEM) methods, this invention significantly reduces the time complexity of inversion calculations, enabling real-time monitoring. Compared to purely data-driven deep learning methods, this invention effectively solves the problem of poor generalization ability caused by the scarcity of short-circuit fault samples (the small sample dilemma) by introducing Maxwell's equations as strong physical constraints, ensuring extremely high diagnostic robustness even in industrial settings lacking massive amounts of fault label data.

[0032] This invention constructs a surface-gap combined three-dimensional sensing system to solve the signal shielding problem under complex grouped topologies. For the compact grouped structure and metal casing of high-voltage, high-capacity battery modules, this invention utilizes the gaps between battery rows to insert magnetic induction units, combined with a surface array to form a three-dimensional surrounding detection network, significantly improving the signal-to-noise ratio for deep leakage magnetic fields inside the battery. Combined with a dynamic filtering mechanism based on the rate of change of load current with time and a background magnetic field modeling and subtraction system based on three-dimensional geometric topology, it is equivalent to constructing a virtual magnetic cleanroom at the software level, effectively eliminating interference from eddy current effects and strong background magnetic fields from the busbars, achieving high-sensitivity non-destructive detection under all operating conditions.

[0033] This invention proposes a hierarchical computing architecture adapted for industrial applications, balancing the conflict between high-precision physical inversion and limited on-site computing power. Employing a "local lightweight feature extraction + backend high-precision physical reconstruction" architecture, the complex tasks of physical model training and fine-tuning are transferred to high-computing platforms (such as the cloud or central computing units), while local processing focuses solely on low-power data acquisition and initial screening. This design enables high-precision physical information model diagnostic technology to be deployed in computing-constrained BMS or edge controllers, significantly reducing system hardware costs and deployment barriers.

[0034] The multi-level early warning strategy proposed in this invention can accurately quantify short-circuit current, distinguish between early-evolutionary soft short circuits and sudden hard short circuits, and provide quantitative data support for BMS to formulate differentiated safety control strategies (such as early warning observation, active balancing or emergency disconnection), truly realizing the prevention of thermal runaway risks before they occur.

[0035] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0036] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0037] Figure 1 This is a flowchart of the method in Example 1.

[0038] Figure 2 This is a schematic diagram of the three-dimensional magnetic sensing array arrangement combining the surface and gap of the battery array module in Embodiment 1.

[0039] The attached diagram lists the components represented by each number as follows:

[0040] 1. Battery cell; 2. Heat dissipation channel gap; 3. Negative terminal; 4. Positive terminal; 5. Busbar; 6. Module surface magnetic sensor; 7. Module gap magnetic sensor; 8. Power system. Detailed Implementation

[0041] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0042] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0043] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0044] Example 1

[0045] Under complex working environments and conditions, lithium-ion batteries may experience internal short-circuit faults in individual cells due to issues such as lithium plating, leading to serious safety accidents. In particular, for battery modules composed of multiple cells connected in series, existing magnetic field inspection technologies face technical bottlenecks such as spatial blind spots caused by complex pack topology, significant eddy current shielding effects of metal casings under dynamic operating conditions, and the inability of limited computing power on the vehicle side to support high-precision physical inversion.

[0046] In view of the aforementioned safety risks and technical limitations, this invention provides a method for detecting short circuits within battery modules based on magnetic field detection. This method uses a magnetic sensor combined with PINN to diagnose short circuit faults within lithium-ion battery packs. It aims to overcome the problem of insufficient computing power on the vehicle side by leveraging cloud computing capabilities. Simultaneously, it addresses the technical challenges of existing technologies being unable to penetrate square dense array structures and traditional static inversion methods being unable to handle dynamic eddy current shielding and busbar interference. This enables accurate capture of early "soft short circuits" and sudden severe "hard short circuits" under all operating conditions, significantly improving the detection sensitivity and location accuracy of early internal short circuit faults, thereby enhancing the safety of lithium-ion battery systems.

[0047] like Figure 1 As shown, the short-circuit detection method within a battery module based on magnetic field detection includes the following steps:

[0048] Magnetic sensors are deployed at designated locations on the battery module to construct a three-dimensional magnetic field sensing array;

[0049] Based on quasi-static logic, a valid observation window is selected, and the magnetic sensor is triggered to collect magnetic induction intensity data within the valid observation window to obtain the raw magnetic induction intensity data.

[0050] The background magnetic field and the topological current-carrying contribution magnetic field are decoupled from the original magnetic induction intensity data to obtain the residual magnetic field characteristics;

[0051] The residual magnetic field characteristics are input into the constructed physical constraint neural network model, and the reconstructed current density is derived based on physical driving.

[0052] A multi-level fault response strategy based on reconfigurable current density is adopted to classify the battery state into different safety levels and complete the internal short circuit detection of the battery module.

[0053] The purpose of this embodiment is to overcome the shortcomings of existing technologies in diagnosing internal short circuits under complex grouping and dynamic operating conditions, solve the problems of hidden and undetectable internal short circuit faults, inaccurate location and inability to quantify them, and improve the safety and reliability of lithium-ion battery systems during operation.

[0054] The technical solution of this embodiment will be explained in detail below. In general, the execution steps of this embodiment include:

[0055] (i) Construct and calibrate a three-dimensional magnetic field sensing array, and arrange magnetic sensors in the accessible areas outside and inside the battery module.

[0056] To achieve accurate inversion of the spatial field, this method pre-establishes a physical coordinate system for the module. Preferably, a Cartesian coordinate system is established with the geometric center of the battery module as the origin. All sensor position coordinates and acquired vector magnetic field components are mapped to this unified coordinate system through a calibration matrix.

[0057] Figure 2 This is a schematic diagram of the three-dimensional magnetic sensing array arrangement combining surface and gap in the battery array module of this embodiment. Figure 2 As shown, multiple battery cells 1 are arranged side-by-side in two rows, each row forming a battery module. The two rows of battery modules together form a battery array module, with a heat dissipation channel gap 2 between them. Adjacent battery cells 1 are connected by busbars 5, and one side of the busbar 5 of the battery array module is connected to the power system 8. Module surface magnetic sensors 6 are positioned on the surface of the battery cells 1, and module gap magnetic sensors 7 are positioned within the heat dissipation channel gap 2 between the two rows of battery modules. It can be understood that the module gap magnetic sensors 7 can be selectively positioned on one side of the battery cell 1 forming the heat dissipation channel gap 2. A negative terminal 3 and a positive terminal 4 are provided on each battery cell 1.

[0058] During the construction of the three-dimensional magnetic field sensing array:

[0059] A flexible circuit board integrating a magnetic sensor array is inserted into the heat dissipation channel gap of the square array module, and a planar sensor array is arranged above / to the side of the module to form a "surface-gap" multi-directional embedded detection structure.

[0060] The sensors are calibrated for zero bias, proportional coefficient, installation attitude and position. The mapping relationship between the sensor coordinate system and the battery module geometric coordinate system is established, and the calibration parameters of each sensor are recorded for subsequent compensation.

[0061] Overall, this step constructed a three-dimensional magnetic field sensing array, realizing the acquisition and transmission of raw three-dimensional magnetic field signals from the lithium-ion battery module. Considering the spatial distribution characteristics of the square battery module, an embedded magnetic sensor array with a "surface + gap" arrangement was adopted. A flexible printed circuit board (FPC) integrating the magnetic sensor array was inserted into the large cooling channel gap between the two rows of batteries to detect and acquire deep leakage magnetic fields at close range. Simultaneously, under no-load conditions, the ambient magnetic field was acquired as a zero-bias reference magnetic field for background magnetic field subtraction to ensure the accuracy of detecting magnetic field changes caused by current variations due to short circuits within the battery pack.

[0062] (II) Construct a data filtering and synchronous acquisition system under operating conditions.

[0063] The operating current of the battery module is collected using a data filtering and synchronous acquisition system, and its current change rate index (or equivalent current increment index) is calculated within a sliding window to determine the quasi-static acquisition window.

[0064] Specifically, the determination is made when the current change rate does not exceed the first preset threshold or the current increment does not exceed the second preset threshold within the preset time window.

[0065] In this embodiment, to better align with the physical laws in subsequent calculations, the "rate of change of current" is specifically preferred as the core criterion. When the judgment condition meets the first preset threshold, the current time segment is identified as a data acquisition window that satisfies the quasi-static condition of the operating current, i.e., a valid observation window.

[0066] Then, within this window, the magnetic sensor array is triggered to synchronously acquire magnetic induction intensity data within the same time segment. During the acquisition process, the data undergoes timestamp alignment, noise reduction, and outlier removal.

[0067] It should be noted that the quasi-static logic screening aims to avoid eddy current interference in the casing structure caused by drastic changes in large current, but does not exclude high-frequency quirky magnetic field fluctuation signals caused by the internal short-circuit fault itself.

[0068] This embodiment constructs a data filtering and synchronous acquisition system, and applies a data filtering mechanism under dynamic operating conditions. This is achieved by acquiring the load current. And calculate the rate of change of load current with respect to time t. The system uses this as a clear trigger criterion to determine the appropriate acquisition time window for inversion using a quasi-static electromagnetic model.

[0069] When the system detects the absolute value of the rate of change of current, which is the primary criterion When the current moment meets the quasi-static condition, all magnetic and current sensors are triggered to synchronously acquire data and align their timestamps. ε This is the first preset threshold.

[0070] The physical basis for this step is:

[0071] According to Faraday's law of electromagnetic induction, the induced eddy currents generated in a metal shell in an alternating magnetic field are positively correlated with the rate of change of the magnetic field. When the current changes slowly and the magnetic field fluctuates rapidly and weakly, the induced eddy currents caused by the change of magnetic flux and their additional magnetic field distortion in the metal shell are usually significantly reduced. Therefore, it is more suitable to use the quasi-static / steady approximation electromagnetic model for subsequent inversion calculations, thereby improving the stability and reliability of the inversion results.

[0072] (III) Decoupling the background magnetic field from the topological current-carrying contribution and constructing residual magnetic field characteristics.

[0073] The magnetic field measured by the sensor simultaneously includes the magnetic field generated by the current-carrying cells and abnormal current channels, the background magnetic field generated by the current-carrying bus / connector, the ambient magnetic field, and the sensor's zero bias. To suppress background interference and enhance fault observability, three decoupling steps are performed:

[0074] (1) Environment and bias subtraction: Under no-load or reference conditions, the ambient magnetic field and sensor zero bias are collected, and the measurement data are compensated or subtracted in combination with the calibration parameters;

[0075] (2) Bus background magnetic field subtraction: Based on the three-dimensional geometric topology of the module, the bus and connectors are discretized into several equivalent current-carrying segments or current elements, and the real-time load current is used to subtract the background magnetic field. The contribution of the background magnetic field at each measuring point is calculated and subtracted based on the geometric parameters;

[0076] (3) Common mode suppression: In view of the common mode component in the magnetic field of each individual measurement point under normal conditions in the series module, the robust statistical method is used to estimate the common mode component, and the difference is performed on each individual measurement point to obtain the residual magnetic field data for inversion. At the same time, in order to avoid the influence of faulty individual on the common mode estimation, the stability is improved by “first roughly finding anomalies, then removing anomalies and recalculating the common mode”.

[0077] In this embodiment, a background magnetic field modeling and subtraction system based on three-dimensional geometric topology is constructed. Based on the three-dimensional CAD geometric topology of the module and the real-time load current, the spatial distribution contribution of the background magnetic field is calculated and suppressed. The background component and common-mode component in the observation data are suppressed to obtain the residual magnetic field characteristics that can better characterize the abnormal current.

[0078] Furthermore, in this embodiment, the busbar and connector are discretized into several equivalent current-carrying segments or current elements, according to... The background magnetic field contribution at each sensor location is calculated and subtracted. Robust common-mode estimation and differential methods are then used to suppress common-mode components. Decoupling stability is improved by first roughly identifying anomalous cells, then removing them, and finally re-estimating the common-mode. Through these software-level background and common-mode suppression measures, the masking effect of structural electromagnetic interference on weak fault characteristics can be effectively reduced, providing a pathway for extracting early-stage internal short-circuit features under strong noise conditions.

[0079] The following is a detailed explanation.

[0080] Based on the principle of superposition, the k-th sensor in the sensor array is located at... r k The total magnetic flux density vector measured at point and time t It can be represented as a vector sum of the following components:

[0081]

[0082] To ensure that the mathematical expression conforms to the laws of electromagnetic field physics, all parameters retain their spatial vector attributes, and their physical meanings are uniformly defined as follows: The vector of the magnetic field component generated by the internal short-circuit abnormal current to be detected; The structural background magnetic field vector generated by the current-carrying busbars and connectors inside the module; This is the common-mode magnetic field vector generated at each cell by the normal operating current of the series-connected battery pack. The static environment magnetic field and the sensor's zero bias vector; For measuring the noise vector, t represents time.

[0083] To extract weak fault features (i.e., extract the This embodiment performs the following three-level decoupling steps:

[0084] 1. Environment and zero bias deduction: The system is powered on and there is no load current ( Under the condition of ) to perform benchmark acquisition, obtain This is then subtracted from subsequent measurement data to obtain the net current-carrying magnetic field.

[0085] 2. Topology Background Subtraction: Based on the module's three-dimensional geometric topology, an equivalent current-carrying model of the busbar and connectors is established. Preferably, the complex busbar is discretized into a set of several finite-length linear current elements. Based on the real-time acquired load current... Using the Biot-Savart law, the structural background magnetic field at each observation point was calculated. r k Theoretical estimate at the location The calculation formula is as follows:

[0086]

[0087] The physical meaning and vector definition of each parameter are as follows:

[0088] Indicates the permeability of free space; This represents the real-time load current of the battery module (scalar) collected at time t. j This is the numbering of the equivalent discrete linear current element. l j For the first j The integration path of a linear current element; d l j For the first j A linear current element l j The vector line element on the line has the same direction as the current direction; r k is the spatial position vector of the kth magnetic sensor (i.e., the position of the observation point); r j Vector line element l j The spatial location vector of the location (i.e., the source point location); r k - r j () represents the spatial distance vector from the source point to the observation point. Let be the magnitude of the distance vector.

[0089] The theoretical estimate obtained by subtracting this physical law from the net current-carrying magnetic field. To eliminate strong structural background interference.

[0090] 3. Robust Common-Mode Suppression: Taking advantage of the consistent operating current of all cells in a series module, robust statistical methods are used to suppress the common-mode magnetic field generated by normal electrochemical reactions. To avoid contaminating the common-mode reference with abnormal cell data from internal short circuits, a two-step strategy of "first coarse identification, then elimination and re-evaluation" is adopted.

[0091] (1) Coarse identification: For sensor sets belonging to the same battery pack or electrical group The median of the magnetic field distribution after background subtraction is calculated as the instantaneous common-mode reference for the group, and the deviation of each sensor data from the initial reference is calculated. Using robust statistical indicators (such as absolute median difference MAD) or preset deviation thresholds, suspected abnormal individual sensors with excessive deviations are roughly identified.

[0092] (2) Re-evaluation by elimination: Data from suspected abnormal individual sensors are eliminated, and the mean is recalculated or pruned using the data from the remaining normal individual sensors. This recalculation is then used as the final instantaneous common-mode baseline for the group. .

[0093] After the above processing, the residual magnetic field, which mainly contains information about internal short-circuit faults, is obtained. :

[0094]

[0095] The physical meanings of each parameter are strictly as follows:

[0096] Indicates the location at the space observation point r k The residual magnetic field vectors extracted at point t, which mainly contain internal short-circuit anomaly characteristics; This indicates that the k-th sensor is at the observation point. r k The total magnetic flux density vector actually measured at point and time t; This represents the estimated vector of the static environmental magnetic field subtracted during calibration in step 1 above, and the zero bias of the sensor. This represents the estimated vector of the background magnetic field of the busbar structure calculated based on physical laws in step 2 above; This represents the instantaneous common-mode reference magnetic field vector generated by the normal electrochemical reaction, which is re-estimated after outliers are removed using the robust statistical method in step 3 above.

[0097] The residual magnetic field This data will be used as input to the subsequent physical inversion model to reconstruct the internal anomalous current distribution.

[0098] (iv) Construct a conditional inversion model with physical consistency constraints to estimate the equivalent current distribution inside the module under limited sensor and noise conditions.

[0099] The physically constrained field reconstruction model constructed in this embodiment is used to invert and reconstruct the internal current distribution from the residual magnetic field characteristics under the constraint of electromagnetic field physical laws. Furthermore, the model can be implemented based on a Physical Information Neural Network (PINN) or other numerical frameworks with function fitting and differential calculation capabilities, realizing the joint driving force of "data observation + physical laws".

[0100] The model uses residual magnetic field data acquired synchronously by all sensors within a quasi-static acquisition window. The main input is the magnetic vector potential of the internal space of the module. For output variables, where Let L be the total number of magnetic sensors deployed. The model solves for the optimal magnetic vector potential distribution by minimizing a hybrid objective function L that includes data residuals and physical constraints. .

[0101] The hybrid objective functionL The definition of is:

[0102]

[0103] The definitions of each item are as follows:

[0104] Data consistency constraints This term is used to characterize the model's predicted magnetic field (based on the curl relationship between the magnetic field and the magnetic vector potential). ,in, Indicates magnetic flux density. This represents the curl operator, used to describe the vortex characteristics of a vector field at a point in space. (representing magnetic vector potential) and input magnetic field The fitting error at the sensor location. Defined as the norm error at the observation point, its calculation formula is:

[0105]

[0106] Where k represents the k-th sensor; K is the total number of magnetic sensors deployed; r k Let be the spatial position vector of the k-th magnetic sensor. For model prediction at point r k The magnetic potential vector at point and time t; The L2 norm squared, representing the error between two three-dimensional spatial vectors (that is, the sum of the squared differences of the components of the two three-dimensional vectors in the three orthogonal directions X, Y, and Z), physically characterizes the squared deviation of the spatial Euclidean distance between the predicted magnetic field and the actual residual magnetic field at that point.

[0107] Physical specification constraints To eliminate gauge uncertainties in the magnetic vector potential inversion process and ensure the uniqueness and convergence stability of the inversion solution, Coulomb gauge physical conditions are introduced. Physical gauge constraints. Defined as the divergence integral term within the continuous space V of the module:

[0108]

[0109] in Let V denote the divergence operator, and V be the continuous volume to be solved inside the battery module.

[0110] Boundary and Sparse Regularization Constraints This design incorporates the actual physical boundary characteristics of the battery module, including normal current constraints on the battery insulation boundary and L1 norm sparse regularization constraints on the current density.

[0111]

[0112] Where S is the insulating shell boundary surface of the module, and n is the unit normal vector of the shell boundary surface. Indicates spatial location r k The internal current density vector at time t; To reflect the L1 norm of the prior of local clustering of internal short-circuit faults in space, the first term of the formula aims to achieve a strict physical constraint, namely that abnormal current cannot penetrate the insulating battery casing; the second term of the formula utilizes the sparse prior of highly local clustering of internal short-circuit faults in three-dimensional space to suppress spurious noise artifacts in non-fault regions caused by ill-posed inversion.

[0113] In the above formula, λ1, λ2, and λ3 are all weighting coefficients used to balance the different dimensions and numerical magnitudes among the various constraint terms.

[0114] Inversion and current distribution reconstruction are performed. Based on the residual magnetic field data obtained in the above steps, the internal current distribution of the module is output. To improve computational stability and anti-interference capability, the "abnormal intensity distribution" is first obtained at the individual unit level to quickly locate suspected faulty units. Then, a more refined current distribution reconstruction is performed on the suspected faulty unit region, thereby improving the reliability of positioning and reducing the computational burden on irrelevant areas.

[0115] The Physics-Constrained Field Reconstruction Model constructed in this embodiment is used to invert and reconstruct the internal current distribution from the residual magnetic field characteristics under the condition of satisfying the physical laws of electromagnetic fields.

[0116] Furthermore, the model can be implemented based on Physical Information Neural Network (PINN) or other numerical frameworks with function fitting and differential calculation capabilities, achieving a joint drive of "data observation + physical laws".

[0117] The model building and solution process follows the following physical logic:

[0118] First, establish the curl relationship between the magnetic field and the magnetic vector potential. Since the divergence of the curl field is always zero, that is... ,in, Indicates spatial location r k The magnetic flux density vector at time t is defined as follows. This definition automatically satisfies the flux continuity theorem (i.e., the magnetic field is source-free), thus avoiding the need for additional mandatory constraint calculations to ensure zero magnetic field divergence in the objective function. This naturally guarantees the physical self-consistency of the magnetic field in the model structure, eliminating the need for additional divergence constraints.

[0119] Secondly, under the quasi-static magnetic approximation, Ampere's circuital law is introduced as the core physical governing equation. The differential form of Ampere's circuital law is: ,in This represents the free magnetic permeability (a physical constant). The model does not directly differentiate with respect to noisy data; instead, it iteratively optimizes and minimizes the aforementioned loss function to find the optimal magnetic vector potential distribution. Subsequently, based on Ampere's circuital theorem and the partial differential equation, the corresponding optimal three-dimensional internal current density field was derived. The corresponding calculation formula is:

[0120]

[0121] in, This represents the spatial location obtained through physical constraint inversion. r k The optimal three-dimensional internal current density vector estimate at time t; This represents the predicted magnetic flux density vector calculated from the optimal magnetic vector potential, i.e. ; This represents the optimal magnetic vector potential distribution obtained by minimizing the mixed objective function L; This indicates that a second curl operation (i.e., the curl of the curl) is performed on the optimal magnetic vector potential field.

[0122] This solution process results in an output current density estimate. It can reproduce the observed residual magnetic field data and satisfy the physical constraints of the electromagnetic field partial differential equation.

[0123] To address the ill-posedness of the inversion problem and suppress noise, the model solves by iteratively optimizing the aforementioned hybrid objective function L. Specifically:

[0124] By minimizing data consistency constraints This enables the model to predict the magnetic field. At each sensor location r k The residual magnetic field that accurately approximates the input ;

[0125] Physical gauge constraints are constructed by introducing Coulomb gauge conditions. Constraints This effectively eliminates the gauge uncertainty of the magnetic vector potential and ensures the inversion solution. The uniqueness and stability of iterative convergence;

[0126] Meanwhile, by constructing boundary and sparse regularization constraints... The internal current density field derived from Ampere's circuital law Apply normal current constraint on the boundary of the insulating shell And the L1 norm constraint that reflects the prior of local spatial clustering of internal short-circuit faults. This greatly suppresses spurious numerical artifacts generated in non-faulty regions.

[0127] Through the above-mentioned physical constraint inversion, the present invention can achieve high-precision imaging and positioning of abnormal current distribution inside the module under limited measurement points and noise interference.

[0128] (v) Constructing an output strategy for fault identification, quantification, and hierarchical alarm.

[0129] Based on the reconstructed current distribution, abnormal regions are identified and the locations of corresponding faulty cells are determined. Subsequently, based on the reconstructed three-dimensional current density field... J It identifies abnormal connected regions within the module whose amplitude exceeds the background noise threshold. .

[0130] The equivalent short-circuit current index is calculated using a predefined integration method to quantify the abnormal current. The output is classified into alarm levels based on the magnitude of the equivalent short-circuit current index, the duration and growth trend of the abnormality, and the reliability of the inversion results.

[0131] Predefined integration methods can be:

[0132] Within the abnormally connected region, define a cross section perpendicular to the principal direction of the current. S isc The equivalent short-circuit current is calculated using surface integrals. ,in n isc cross section S isc The normal vector.

[0133] In this embodiment, a multi-level fault response strategy based on inverted current density is implemented. The system classifies the battery state into different safety levels and performs corresponding operations based on the equivalent internal short-circuit current and reconstructed current density distribution obtained from the inversion.

[0134] Level L0 (Normal State): The reconstructed current density field has no obvious accumulation or the equivalent internal short-circuit current is lower than the first preset threshold (i.e., background noise, such as 100mA). It is determined that there is no obvious abnormality, only the health log is recorded, and no alarm is triggered.

[0135] L1 level (early soft short circuit): When the equivalent internal short-circuit current is between the first and second preset thresholds (e.g., between 100mA and 1A) and the duration exceeds the preset window, it corresponds to micro dendrite growth or slight damage to the diaphragm. At this time, the system determines that there is an evolution trend but no immediate risk of thermal runaway, triggers an early warning signal, increases the data sampling and uploading frequency, and adds the unit to the "key monitoring list" to analyze the short-circuit current growth rate to predict the evolution trend.

[0136] Level L2 (Sudden Hard Short Circuit): When the equivalent internal short-circuit current exceeds the second preset threshold (e.g., 1A), or the local temperature rise rate predicted by the fused temperature model exceeds the safety threshold, the corresponding diaphragm ruptures or suffers severe physical failure. At this time, the system determines that there is an extremely high risk of thermal runaway, immediately triggers the highest level alarm, and sends an active control request (such as limiting power, starting the maximum power of liquid cooling, or disconnecting the high-voltage circuit), while simultaneously generating an emergency work order and remotely locking the charging function.

[0137] (vi) A hierarchical computing architecture is used to perform the magnetic field inversion and fault diagnosis.

[0138] The architecture comprises a local monitoring end (such as a BMS or edge controller) and a high-performance processing end (such as a cloud server or central computing platform) for communication connectivity. The local monitoring end is configured to perform data collection, preliminary screening, and lightweight feature extraction, and to perform rapid inference using locally stored lightweight models. The high-performance processing end is configured to receive data uploaded from the local monitoring end and perform high-precision inversion calculations and model training using the PINN model. This architecture effectively addresses the extremely high demand for local hardware computing power in complex physical model inversion through a collaborative approach of "edge-side lightweight inference + backend physical heavy computation," while ensuring the real-time performance and accuracy of diagnostics.

[0139] In addition, this embodiment constructs an adaptive sampling and triggering mechanism under dynamic operating conditions. The local monitoring terminal controls the magnetic induction array and the voltage sensor to synchronize with the clock at the microsecond level; it monitors the rate of change of the load current in real time, and when a window that meets the quasi-static conditions (i.e., no significant eddy current interference) is detected, it triggers high-frequency snapshot acquisition; it performs preliminary processing on the acquired data based on the locally pre-stored bus geometry parameters, extracts the residual magnetic field characteristics, and performs lossless compression and packaging.

[0140] If the initial value of the current density distribution output by the local lightweight model shows an anomaly or the confidence level is lower than the preset threshold, the data is marked as a "difficult sample" and a high-priority upload mechanism is immediately triggered to send it to the high-performance processing end for secondary confirmation.

[0141] After receiving the data, the high-performance processing unit performs spatiotemporal alignment and fusion with historical data, data from the same batch of battery cells, and synchronously uploaded voltage, temperature, and SOC data. For uploaded "difficult samples" or long-term monitoring data, a loss function containing residual terms and sparse regularization terms from the physical equations is constructed. Using automatic differentiation technology, incremental learning is performed based on the original model parameters to solve for the optimal network weight parameters. This mechanism avoids the time consumption of retraining with full data and ensures that the model can quickly adapt to the magnetic field baseline drift caused by battery aging, achieving adaptive diagnosis throughout the entire life cycle.

[0142] The following section will provide a specific application scenario for this embodiment.

[0143] In a typical implementation of this embodiment, a module is constructed using six large-capacity square aluminum-cased lithium-ion batteries (280Ah) arranged in a two-row, three-column (2×3) configuration. This module, when expanded, can be widely used in energy storage and electric vehicle fields. The specific implementation includes the following steps:

[0144] 1. Construct a three-dimensional magnetic field sensing array and establish a coordinate system.

[0145] Array Arrangement: A "surface-gap" composite array is constructed. At a predetermined distance (5mm in this embodiment) above the module's top cover, magnetic induction units (surface array, 6 in total) are arranged corresponding to the geometric center of each individual battery cell to capture the macroscopic magnetic field distribution on the battery surface. A flexible circuit board (FPC) probe (gap array) integrating 3 magnetic induction units is inserted into the cooling channel gap (approximately 3mm wide) between the two rows of batteries. The sensor on the FPC is located deep within the gap, close to the side of the battery, to capture the deep leakage magnetic field attenuated by the aluminum casing at close range.

[0146] Sensor selection: Select a high-sensitivity fluxgate sensor or magnetoresistive sensor with a bandwidth covering DC to kHz levels.

[0147] Coordinate Definition and Calibration: A global Cartesian coordinate system is established with the geometric center of the battery array module as the origin, the long side of the module as the x-axis, the arrangement direction as the y-axis, and the vertical height direction as the z-axis. All coordinates are obtained through high-precision calibration. K = Position coordinates of 9 magnetic sensors in the global coordinate system The attitude matrix maps the local magnetic field components output by each sensor to a magnetic induction vector in the global coordinate system. .

[0148] When the system is powered on and there is no load ( Under these conditions, the sensor readings are collected as the zero-bias reference for the ambient magnetic field. Store it.

[0149] 2. Construct a physical constraint neural network model.

[0150] On the data processing side, a deep neural network (PINN) for solving the inverse magnetic field problem is pre-built and initialized. This network serves as the main body of the PINN, and its input is the decoupled residual magnetic field. The output is the spatial magnetic vector potential. .

[0151] Physical constraint construction: Defining the hybrid loss function The solution to the PINN model is guided by incorporating electromagnetic physics laws as residual terms into the loss function, achieving a dual drive from both data and physics. The neural network's predicted magnetic field approximates the observed value, thus optimizing the solution. Introducing the Coulomb gauge To ensure the uniqueness of the inversion solution, a regularization term is constructed using the L1 sparsity of internal short circuits and the insulation boundary conditions of the module casing. And in solving, strictly in the differential form of Ampere's circuital law ( () serves as the core physical control equation.

[0152] 3. Data filtering and collection under dynamic operating conditions.

[0153] Quasi-static triggering: The acquisition module monitors the load current in real time at high frequencies (e.g., 1 kHz, kilohertz). and its rate of change Set the quasi-static trigger threshold. , amperes / second.

[0154] When detected If there is a sudden acceleration, it is determined that there is strong eddy current interference, and the magnetic field inversion is stopped; when it is detected that... When the time (such as uniform speed or static state) is determined as a valid observation window.

[0155] In particular, the static state ( The optimal observation window is the one with the least background magnetic field interference, and the system should maintain monitoring during this period.

[0156] Synchronous acquisition: Within the effective window, the magnetic sensor array and the current sensor are triggered to perform microsecond-level synchronous acquisition (synchronization error controlled within 100µs (micrometers)) to obtain the time. t Raw magnetic flux density data With current And transmit it to the computing module.

[0157] 4. Background magnetic field modeling and decoupling based on three-dimensional geometric topology.

[0158] After receiving the data, the calculation module performs three-level physical feature decoupling to extract the residual magnetic field containing only fault information. :

[0159] (1) Environmental bias correction: Subtract the environmental baseline calibrated in step S1. Thus, a net current-carrying magnetic field is obtained;

[0160] (2) Bus decoupling: Based on the stored 3D CAD model of the bus, it is equivalent to a finite-length collection of current-carrying elements. Utilizing the Biot-Savart law, based on the real-time current... Calculate the theoretical value of the structural background magnetic field at each measuring point. And deduct it (Note: if (If this term is zero, then the value of this term is zero).

[0161] (3) Common-mode suppression: Utilizing the principle of current consistency in series modules, the median of the magnetic field data of a set of sensors belonging to the same battery pack or electrical group is calculated as the instantaneous common-mode reference. And subtract it. The final residual magnetic field is obtained:

[0162]

[0163] 5. PINN inversion solution and hierarchical diagnosis.

[0164] Inversion and reconstruction: Input the PINN model, calculate the physical residuals using automatic differentiation, and minimize the loss function through iterative optimization. L Update the neural network weights and solve for the optimal magnetic vector potential distribution. And derive the estimated three-dimensional current density inside the module:

[0165]

[0166] Diagnostic criteria: Based on reconstructed current density distribution Identify abnormal accumulation regions. Identify connected regions within the module where the current density magnitude is significantly higher than the background noise as fault domains. .

[0167] The connected regions within the module where the current density magnitude is significantly higher than the background noise are identified as fault regions. Within the fault region, the principal current direction is determined, and a cross-section containing the fault center and perpendicular to this principal direction is taken. S isc The equivalent short-circuit current is calculated using surface integral quantization:

[0168]

[0169] in n isc cross section S isc The normal vector.

[0170] Output of grading results: based on Value, execute L0-L2 level three response:

[0171] Level L0 (Normal State): Equivalent Internal Short-Circuit Current If the current is less than the first preset threshold (set to 100mA in this example), the vehicle will not issue a notification, but the cloud will record a health log.

[0172] L1 level (soft short circuit): equivalent internal short circuit current If the battery is located between the first preset threshold and the second preset threshold (set to 1A in this example), corresponding to abnormal conditions such as micro-lithium dendrite growth and slight damage to the separator, the data processing terminal generates an early warning report and adds the battery to the "key monitoring list". The local monitoring terminal automatically increases the sampling frequency.

[0173] Level L2 (hard short circuit): Equivalent internal short circuit current If the value exceeds the second preset threshold, it corresponds to a serious physical failure and poses an extremely high risk of thermal runaway. The local monitoring terminal immediately triggers an audible and visual alarm and sends a power limit or relay disconnection request via the bus; the data processing terminal simultaneously pushes an emergency work order.

[0174] Example 2

[0175] This embodiment discloses a short-circuit detection system for battery modules based on magnetic field detection.

[0176] A short-circuit detection system within a battery module based on magnetic field detection includes:

[0177] The magnetic field sensing module is configured to: deploy magnetic sensors at designated locations on the battery module to construct a three-dimensional magnetic field sensing array;

[0178] The raw data acquisition module is configured to: filter the effective observation window based on quasi-static logic, trigger the magnetic sensor to collect magnetic induction intensity data within the effective observation window, and obtain the raw magnetic induction intensity data.

[0179] The decoupling module is configured to decouple the background magnetic field from the topological current-carrying contribution magnetic field of the original magnetic induction intensity data to obtain the residual magnetic field characteristics.

[0180] The model inversion module is configured to input the residual magnetic field characteristics into the constructed physical constraint neural network model and invert the reconstructed current density based on physical drive.

[0181] The multi-level response module is configured to: adopt a multi-level fault response strategy based on reconfigured current density to classify the battery state into different safety levels and complete the internal short circuit detection of the battery module.

[0182] This embodiment provides a short-circuit detection system within a battery module based on magnetic field detection, employing a hierarchical or distributed hardware architecture design. Wherein:

[0183] 1. Magnetic field sensing module (M1):

[0184] Function: Responsible for sensing the spatial magnetic field distribution around the battery module and converting analog signals into digital signals.

[0185] Configuration: Includes a first magnetic sensor array arranged on the surface of the battery module, a second magnetic sensor array embedded in the gap between battery rows on a flexible substrate (such as an FPC), and a high-precision analog-to-digital converter (ADC).

[0186] Optimal configuration: The sensor uses a high-sensitivity magnetic induction unit. This module is equipped with a hardware synchronous trigger interface to ensure that all magnetic field data channels and current sampling channels are strictly aligned in the time domain to guarantee the simultaneity of "snapshot" data.

[0187] 2. Raw Data Acquisition Module (M2)

[0188] It is mainly used for collecting and acquiring raw data.

[0189] 3. Decoupling module (M3).

[0190] The decoupling module is mainly used for local preprocessing and filtering of raw data:

[0191] Function: As the "edge outpost" of the system, it is responsible for cleaning, filtering and preliminary feature extraction of data to reduce transmission bandwidth pressure.

[0192] Configuration: Integrated into a local monitoring terminal (such as a vehicle-mounted BMS main control chip or an energy storage station edge gateway), it has certain vector computing capabilities.

[0193] Logic unit:

[0194] The dynamic filtering unit calculates the load current change rate in real time and triggers the raw data acquisition module M2 to collect data only when the quasi-static threshold is met, thus avoiding eddy current interference from the source.

[0195] The topology decoupling unit stores the geometric parameters of the module busbar, performs background field calculation and elimination and differential operation on the busbar, and extracts residual magnetic field characteristics;

[0196] The initial screening and inference unit runs a lightweight model or rule-based algorithm to quickly evaluate the current state and identify suspected abnormal data (difficult samples).

[0197] 4. Model Inversion Module (M4):

[0198] Function: As the "computational brain" of the system, it is responsible for running highly complex PINN models.

[0199] Configuration: Deployed on a high-performance computing platform (such as a cloud server cluster, central domain controller, or high-performance workstation).

[0200] Logic unit:

[0201] The multi-source fusion unit aligns the uploaded magnetic field data with voltage, temperature, SOC, and historical data in a spatiotemporal manner.

[0202] The physical training unit constructs a loss function that includes the Ampere circuital law residuals and sparse regularization terms, and uses automatic differentiation technology to incrementally learn the PINN model.

[0203] The high-precision inversion unit, based on a trained physical constraint model, finely reconstructs the three-dimensional current density distribution field inside the battery.

[0204] 5. Multi-level response module (M5):

[0205] Function: Responsible for converting inversion results into hierarchical control commands.

[0206] Configuration: Distributed to local terminals (such as dashboard HMI, system controller interface) and remote terminals (such as monitoring screen, after-sales push interface).

[0207] Execution logic:

[0208] The system receives the local criteria from the decoupling module M3 or the diagnostic results from the model inversion module M4. When it is determined to be a level 2 hard short circuit, it directly sends a power limiting or high voltage power-down request to the system controller and drives an audible and visual alarm. When it is determined to be a level 1 soft short circuit, it generates a visual report containing the faulty unit number, coordinates, and resistance value, and sends it to maintenance personnel or the after-sales system through M3.

[0209] Also includes:

[0210] Data transmission and communication module:

[0211] Function: Establishes a data link between the decoupling module M3 and the model inversion module M4.

[0212] Configuration: Supports wired or wireless communication protocols.

[0213] Transmission strategy: Supports dynamic switching between "normal low frequency" and "abnormal high frequency" modes; features breakpoint resume capability, caching critical fault data during communication interruptions and automatically retransmitting it after connection restoration. Data stream includes residual magnetic field characteristics / raw waveforms of uplink transmission, and diagnostic results / model update weights of downlink transmission.

[0214] Example 3

[0215] The purpose of this embodiment is to provide a computer-readable storage medium.

[0216] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the battery module short-circuit detection method based on magnetic field detection as described in Embodiment 1 of this disclosure.

[0217] Example 4

[0218] The purpose of this embodiment is to provide an electronic device.

[0219] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the short-circuit detection method for a battery module based on magnetic field detection as described in Embodiment 1 of this disclosure.

[0220] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0221] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0222] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A short-circuit detection method for battery modules based on magnetic field detection, characterized in that, Includes the following steps: Magnetic sensors are deployed at designated locations on the battery module to construct a three-dimensional magnetic field sensing array; Based on quasi-static logic, a valid observation window is selected, and the magnetic sensor is triggered to collect magnetic induction intensity data within the valid observation window to obtain the raw magnetic induction intensity data. The background magnetic field and the topological current-carrying contribution magnetic field are decoupled from the original magnetic induction intensity data to obtain the residual magnetic field characteristics; The residual magnetic field characteristics are input into the constructed physical constraint neural network model, and the reconstructed current density is derived based on physical driving. A multi-level fault response strategy based on reconfigurable current density is adopted to classify the battery state into different safety levels and complete the internal short circuit detection of the battery module. Loss function of physically constrained neural network models L Specifically: ； in: This is a data consistency constraint used to constrain the consistency between the reconstructed magnetic field and the measured residual magnetic field, and is defined as the norm error at the observation point. For physical gauge constraints, the Coulomb gauge condition is adopted, defined as follows: To eliminate gauge uncertainties in the magnetic vector potential and ensure the uniqueness of the inversion solution, where denoted by ; V is the continuous volume to be solved inside the battery module; For model prediction at point r k The magnetic potential vector at point and time t, where k represents the k-th sensor. r k Let K be the spatial position vector of the k-th magnetic sensor; The boundary and sparsity regularization constraints are combined with the physical characteristics of the battery module, including insulation boundary constraints and current sparsity constraints. , , These are all weighting coefficients used to balance the dimensions and numerical magnitudes of various quantities.

2. The short-circuit detection method within a battery module based on magnetic field detection as described in claim 1, characterized in that, Magnetic sensors are deployed at designated locations on the battery module to construct a three-dimensional magnetic field sensing array, specifically including: Magnetic sensors are placed in the gaps between the battery rows of the battery module, and a magnetic sensor array is arranged at a set position on the surface of the battery module to form a three-dimensional magnetic field sensing array. The magnetic sensor is calibrated for zero bias, proportional coefficient, installation attitude and position, and the mapping relationship between the magnetic sensor coordinate system and the battery module geometric coordinate system is established.

3. The short-circuit detection method within a battery module based on magnetic field detection as described in claim 1, characterized in that, Valid observation windows are selected based on quasi-static logic, specifically including: Define the length of the time window; Collect the current of the battery module within the current sliding window; If the current change rate does not exceed the first preset threshold, or the current increment does not exceed the second preset threshold, the quasi-static logic is satisfied, and the current sliding window becomes a valid observation window.

4. The short-circuit detection method within a battery module based on magnetic field detection as described in claim 1, characterized in that, The background magnetic field and the topological current-carrying contribution magnetic field are decoupled from the original magnetic flux density data to obtain the residual magnetic field characteristics, specifically including: Environment and bias subtraction: Under no-load or reference conditions, the ambient magnetic field and sensor zero bias are collected, and the original magnetic induction intensity data are compensated or subtracted in combination with calibration parameters. Bus background magnetic field subtraction: Based on the three-dimensional geometric topology of the battery module, the bus and connectors are discretized into several equivalent current-carrying line segments or current elements. The background magnetic field contribution at each measuring point is calculated and subtracted using real-time current and geometric parameters. Common-mode suppression: Robust common-mode estimation and differential methods are used to suppress common-mode components. The decoupling stability is improved by first roughly identifying abnormal cells, then removing abnormal cells and re-estimating the common mode, thus obtaining the residual magnetic field characteristics.

5. The short-circuit detection method within a battery module based on magnetic field detection as described in claim 1, characterized in that, The residual magnetic field characteristics are input into the constructed physical constraint neural network model, and the reconstructed current density is derived based on physical driving forces. Specifically, this includes: First, establish the curl relationship between the magnetic field and the magnetic vector potential; Secondly, under the quasi-static magnetic approximation, Ampere's circuital law is introduced as the physical governing equation; Construct an objective function that includes physical constraints, and find the optimal magnetic vector potential distribution by minimizing the loss function; Based on the optimal magnetic vector potential distribution, the internal current density is reconstructed to obtain the reconstructed current density. or, The physical constraint neural network model takes the residual magnetic field dataset of all magnetic sensors within the effective observation window as the conditional input, and the magnetic vector potential field in the continuous space of the battery module as the hidden state variable to be solved.

6. The short-circuit detection method within a battery module based on magnetic field detection as described in claim 1, characterized in that, A multi-level fault response strategy based on reconfigurable current density is adopted to classify the battery state into different safety levels and complete the internal short-circuit detection of the battery module, specifically including: Based on the reconstructed current density distribution, abnormal regions are identified and the corresponding faulty cell locations are determined. Identify and reconstruct abnormal connected regions within a faulty cell where the current density amplitude exceeds the background noise threshold. The equivalent short-circuit current in the abnormally connected region is calculated using a predefined integration method; Based on the equivalent short-circuit current magnitude and combined with preset safety level judgment thresholds, the battery status is classified into safety levels.

7. The short-circuit detection method for battery modules based on magnetic field detection as described in claim 1, characterized in that, Also includes: Build a hierarchical computing architecture that includes local monitoring terminals and high-performance processing terminals; The local monitoring terminal is used to perform the acquisition, decoupling, and residual magnetic field feature extraction of raw magnetic induction intensity data; High-performance processing equipment is used for inversion calculations and model training of physically constrained neural network models.

8. A short-circuit detection system within a battery module based on magnetic field detection, characterized in that, include: The magnetic field sensing module is configured to: deploy magnetic sensors at designated locations on the battery module to construct a three-dimensional magnetic field sensing array; The raw data acquisition module is configured to: filter the effective observation window based on quasi-static logic, trigger the magnetic sensor to collect magnetic induction intensity data within the effective observation window, and obtain the raw magnetic induction intensity data. The decoupling module is configured to decouple the background magnetic field from the topological current-carrying contribution magnetic field of the original magnetic induction intensity data to obtain the residual magnetic field characteristics. The model inversion module is configured to input the residual magnetic field characteristics into the constructed physical constraint neural network model and invert the reconstructed current density based on physical drive. The multi-level response module is configured to: adopt a multi-level fault response strategy based on reconfigured current density to classify the battery state into different safety levels and complete the internal short circuit detection of the battery module; Loss function of physically constrained neural network models L Specifically: ； in: This is a data consistency constraint used to constrain the consistency between the reconstructed magnetic field and the measured residual magnetic field, and is defined as the norm error at the observation point. For physical gauge constraints, the Coulomb gauge condition is adopted, defined as follows: To eliminate gauge uncertainties in the magnetic vector potential and ensure the uniqueness of the inversion solution, where denoted by ; V is the continuous volume to be solved inside the battery module; For model prediction at point r k The magnetic potential vector at point and time t, where k represents the k-th sensor. r k Let K be the spatial position vector of the k-th magnetic sensor; The boundary and sparsity regularization constraints are combined with the physical characteristics of the battery module, including insulation boundary constraints and current sparsity constraints. , , These are all weighting coefficients used to balance the dimensions and numerical magnitudes of various quantities.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the short-circuit detection method for battery modules based on magnetic field detection as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the short-circuit detection method for battery modules based on magnetic field detection as described in any one of claims 1-7.

Images