A method and device for early warning positioning thermal runaway of lithium ion battery pack

Abstract

The present application relates to the technical field of lithium ion battery safety monitoring, and discloses a method and device for early warning and positioning thermal runaway of lithium ion battery pack; first, real-time voltage, temperature, current and number data of each cell are collected to construct a cell state observation tensor with a space-time label; then, a cell thermal-electric state extended Kalman estimation algorithm is run based on the tensor to generate a cell health state partial order lattice; then, a thermal runaway risk propagation phase space grid is constructed, and finally, local attractor region identification and early warning decision based on manifold analysis are executed under grid constraints to output high-risk source cell number and early warning level. The present application solves the problems of existing technology, such as lagging early warning caused by dependence on late physical signs, lack of dynamic evaluation of cell group state consistency, and inability to accurately locate faulty cells, and improves the timeliness and positioning accuracy of thermal runaway early warning.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery safety monitoring technology, and in particular to a method and apparatus for early warning and location of thermal runaway in lithium-ion battery packs. Background Technology

[0002] As a core energy storage component of new energy equipment, the thermal runaway safety monitoring of lithium-ion battery packs is a crucial aspect of ensuring reliable system operation. Existing thermal runaway detection technologies primarily involve deploying sensors to collect physical parameters such as voltage, temperature, and current of each cell. Some solutions additionally monitor gas concentration or casing pressure within the battery pack, and then process the collected data in conjunction with the battery management system (BMS). The core idea behind these technologies is to set fixed thresholds; when a parameter exceeds the threshold, a potential thermal runaway event is identified, triggering an alert. In conventional processing methods, the BMS focuses on the time-series changes of parameters within individual cells, using parameter fluctuation trends to aid in judgment and identify thermal runaway events, thus preventing safety accidents.

[0003] Existing thermal runaway detection technologies suffer from a delayed warning system. This problem arises because existing technologies rely on physical indicators such as sudden voltage drops, rapid temperature increases, gas releases, or pressure surges, which typically appear only in the mid-to-late stages of the thermal runaway process. Weak abnormal signals generated when a cell experiences an internal short circuit or early failure, such as slow voltage shifts or slight temperature increases, cannot be effectively detected because they do not reach set thresholds and lack dynamic assessment of the consistency of the cell group's state. This warning delay directly reduces occupants' emergency escape time. Furthermore, because it is impossible to accurately locate individual cells in a thermal runaway state, maintenance personnel must inspect multiple modules or even all cells within the battery pack one by one. This not only increases maintenance time and costs but also risks the spread of thermal runaway risk to adjacent cells due to the failure to isolate the at-risk cell in time, potentially leading to a wider safety incident. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, the present invention provides a method for early warning and location of thermal runaway in lithium-ion battery packs, comprising:

[0005] Collect real-time voltage, temperature, current and serial number data of each cell, and construct a cell state observation tensor with spatiotemporal labels;

[0006] Based on the cell state observation tensor with spatiotemporal labels, the extended Kalman estimation algorithm for cell thermoelectric state is run to generate a partial order lattice of cell health state.

[0007] Based on the partial order lattice of cell health status, a phase space grid for thermal runaway risk propagation is constructed.

[0008] Under the constraint of phase space grid for thermal runaway risk propagation, local attractor region identification and early warning decision-making based on manifold analysis are performed.

[0009] Furthermore, the steps for constructing the spatiotemporally labeled cell state observation tensor include:

[0010] Collect real-time voltage, real-time temperature, real-time current and serial number data of each cell to obtain the original data set of the cells;

[0011] Based on the original battery cell dataset, determine the sliding time window parameters;

[0012] Based on the cell numbering data and sliding time window parameters in the original cell dataset, the cell index, voltage sequence, temperature sequence and current sequence are obtained through index encoding and generation of physical quantity sequences.

[0013] Based on the numbering data in the original battery cell dataset, the Laplacian feature mapping method is used to perform position topology coding to obtain the position topology code of each battery cell;

[0014] Based on cell index, voltage sequence, temperature sequence, current sequence, location topology encoding, and sliding time window parameters, a cell state observation tensor with spatiotemporal labels is constructed.

[0015] Furthermore, the steps for performing location topological coding using the Laplacian feature mapping method include:

[0016] Based on the row and column coordinates contained in the serial number data in the original battery cell dataset, a spatial adjacency matrix is ​​constructed.

[0017] Calculate the degree matrix of the spatial adjacency matrix;

[0018] The Laplacian matrix is ​​calculated using the degree matrix and the spatial adjacency matrix.

[0019] Solve for the eigenvalues ​​and eigenvectors of the Laplace matrix, and select the four eigenvectors with the smallest eigenvalues; combine the eigenvectors to form the cell's location topology code.

[0020] Furthermore, the steps of running the extended Kalman estimation algorithm for the thermoelectric state of the battery cell and generating the partial order lattice of the battery cell health state include:

[0021] Based on the cell state observation tensor with spatiotemporal labels, the algorithm input data is extracted to obtain the algorithm input dataset;

[0022] Based on the algorithm input dataset and the electrochemical-thermal coupling model, the extended Kalman estimation algorithm for the thermoelectric state of the battery cell is run to obtain the binary pairs of each battery cell;

[0023] Based on the binary pairs of each cell, a partial-order lattice of cell health status is constructed.

[0024] Furthermore, the steps for running the extended Kalman estimation algorithm for the thermoelectric state of the battery cell include:

[0025] The state equation and observation equation of the extended Kalman estimation algorithm for the thermoelectric state of a battery cell are constructed. The state equation uses the state of charge and DC internal resistance of the battery cell as state variables. The input of the observation equation is the estimated value of the state variables, and the output is the predicted value of voltage and temperature.

[0026] Initialize the filter parameters, specifically including: for each cell index in the algorithm input dataset, initialize the initial values ​​of the state variables and the initial values ​​of the error covariance matrix respectively;

[0027] Perform the filtering prediction step;

[0028] The filtering update step includes: extracting the state of charge and DC internal resistance from the state variable update values ​​at each time step, taking the average value over the time steps, and using it as the final state of charge and DC internal resistance of the cell. The two are combined to form a binary tuple of the cell.

[0029] Furthermore, the steps for constructing the phase space grid for thermal runaway risk propagation include:

[0030] Based on the partially ordered lattice of cell health status and the spatiotemporally labeled cell status observation tensor, the associated data are extracted and integrated to obtain the basic dataset;

[0031] Based on the basic dataset, the SOC deviation, internal resistance deviation, voltage change rate and temperature change rate of each cell are calculated to obtain the four-dimensional state parameter set of the cell.

[0032] Based on the four-dimensional state parameter set of the battery cell and the basic dataset, the phase space points of the battery cell are constructed and associated with physical locations to obtain a set of battery cell phase space points with physical locations. The steps of constructing the battery cell phase space points and associating them with physical locations are as follows: the SOC deviation, internal resistance deviation, voltage change rate, and temperature change rate of each battery cell are used as four-dimensional coordinate parameters and combined to form the coordinate points of the battery cell in the four-dimensional phase space, i.e., the battery cell phase space points; the physical location number of each battery cell phase space point is extracted from the basic dataset and bound to the battery cell phase space points to obtain a set of battery cell phase space points with physical locations.

[0033] Based on the physical location distribution of battery cells in the basic dataset, a three-dimensional mesh cell set is obtained by dividing the data into three-dimensional mesh cells.

[0034] The set of phase space points of the battery cell with physical location is embedded into a three-dimensional mesh cell set, and a joint risk density function is constructed for each mesh cell to obtain a mesh cell set with joint risk density function.

[0035] Based on a set of grid cells with a joint risk density function, directed edges are established between adjacent grid cells to obtain a grid structure with initial directed edges.

[0036] Based on the partial order lattice of cell health status and the mesh structure with initial directed edges, the spatial distance and dominance strength are calculated, and the edge weights of the directed edges are modulated to obtain the mesh structure with weighted directed edges.

[0037] By integrating a set of grid cells with a joint risk density function with a grid structure with weighted directed edges, a phase space grid for thermal runaway risk propagation is constructed.

[0038] Furthermore, the steps of embedding the set of cell phase space points with physical locations into a set of three-dimensional mesh cells and constructing a joint risk density function include:

[0039] Based on the physical location number of the cell phase space point with physical location, the corresponding three-dimensional grid coordinates are matched, and the cell phase space point is assigned to the corresponding three-dimensional grid cell to form a set of grid cells containing cell phase space points.

[0040] For each grid cell containing a cell phase space point, a joint risk density function is constructed using a kernel density estimation method. The kernel density estimation method selects a Gaussian kernel function and determines the bandwidth parameter through cross-validation. With each cell phase space point as the center, the Gaussian kernel function is applied, and the joint risk density function values ​​at each location are summed to form a complete joint risk density function.

[0041] Each grid cell is associated with its joint risk density function to obtain a set of grid cells with the joint risk density function.

[0042] Furthermore, the steps for performing local attractor region identification and early warning decision-making based on manifold analysis include:

[0043] Based on the phase space grid of thermal runaway risk propagation, the spatiotemporally labeled cell state observation tensor, and the partial order grid of cell health status, the associated data are extracted and integrated to obtain the analysis dataset.

[0044] Based on the analysis dataset, a global manifold analysis is performed on the vector field in the thermal runaway risk propagation phase space grid to obtain a manifold feature map. The global manifold analysis includes: dividing the vector field into local vector subfields according to the three-dimensional coordinates of the thermal runaway risk propagation phase space grid, with each local vector subfield containing vector information from a preset number of adjacent grid cells; calculating the continuity index of each local vector subfield, and selecting local vector subfields with continuity indices higher than a preset threshold; merging the selected local vector subfields to form continuous manifold segments, calculating the consistency of vector directions within each segment, identifying segments with vector directions that tend to be consistent and point towards the core region as potential local attractor region candidates, and integrating them to form a manifold feature map.

[0045] Based on manifold feature maps and analysis datasets, potential local attractor region candidates are extracted and verified to obtain effective local attractor regions.

[0046] Based on the effective local attractor region and the analysis dataset, candidate cells for high-risk thermal runaway sources are located.

[0047] Based on the effective local attractor region and the analysis dataset, the stability index of the effective local attractor region is calculated.

[0048] Based on the stability index and candidate cells with high risk of thermal runaway, the warning level is determined and the results are output.

[0049] Furthermore, the steps for calculating the stability index of this effective local attractor region include:

[0050] Extract the vector field data of all grid cells within the effective local attractor region from the analysis dataset, calculate the vector field divergence of each grid cell, and take the average value as the overall divergence of the region.

[0051] Extract the vector field data of all grid cells within the effective local attractor region from the analysis dataset, calculate the curl of the vector field of each grid cell, and take the average value as the overall curl of the region;

[0052] The global divergence and global curl are normalized separately.

[0053] The stability index is obtained by weighting and summing the normalized global divergence and global curl according to preset weights.

[0054] An apparatus for early warning and location of thermal runaway in a lithium-ion battery pack, used to implement the aforementioned method for early warning and location of thermal runaway in a lithium-ion battery pack, the apparatus comprising:

[0055] Cell data acquisition and tensor construction unit: used to acquire real-time voltage, temperature, current and number data of each cell, and construct cell state observation tensor with spatiotemporal labels;

[0056] Thermoelectric state estimation and partial order lattice generation unit: used to run the extended Kalman estimation algorithm for the thermoelectric state of the battery cell based on the cell state observation tensor with spatiotemporal labels, and generate a partial order lattice of the battery cell health state;

[0057] Risk propagation phase space grid construction unit: used to construct a phase space grid for thermal runaway risk propagation based on the partial order lattice of cell health state;

[0058] Location and early warning decision unit: used to perform local attractor region identification and early warning decision based on manifold analysis under the constraints of the phase space grid of thermal runaway risk propagation.

[0059] Compared to existing technologies, the advantages of this invention are as follows: By constructing a cell state observation tensor with spatiotemporal labels and incorporating cell location topological encoding as an independent dimension, this invention effectively solves the problem of traditional BMS systems ignoring cell spatial layout and misjudging normal heat conduction as internal short-circuit risk. It can accurately distinguish whether local overheating originates from thermal diffusion of neighboring cells or from internal faults within a single cell, eliminating misjudgments caused by missing spatial correlations at the data structure level. Combining the extended Kalman estimation algorithm of the electrochemical-thermal coupling model with the partial-order lattice of cell health status, it overcomes the limitation of traditional scalar indicators in being unable to identify hidden abnormal cells that are "not at the lowest state of charge and not at the highest internal resistance." By capturing the asymmetric degradation characteristics of cells through dominance relationships, such weakly abnormal cells are included in a minimal element set, achieving accurate screening of early potential hazards. The construction of a phase-space grid for thermal runaway risk propagation overcomes the limitations of traditional heatmaps, which can only display risk distribution but not propagation trends. Its non-uniform grid with vector fields visualizes the propagation path of risk from low-risk to high-risk areas, allowing for early identification of risk convergence centers. Local attractor region identification based on manifold analysis eliminates reliance on traditional static thresholds. By capturing dynamic risk convergence trends, it provides early warnings even when voltage and temperature do not reach thresholds, allowing for a longer emergency window for maintenance. The final location stage combines dual verification of lowest voltage and highest temperature with health status data to reduce false alarms caused by sensor interference. The output high-risk cell IDs can be directly associated with physical locations, avoiding tedious checking and significantly improving the timeliness and accuracy of thermal runaway warnings, ensuring the safe operation of lithium-ion battery packs. Attached Figure Description

[0060] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart of a method for early warning and location of thermal runaway in a lithium-ion battery pack according to the present invention;

[0062] Figure 2 This is a flowchart illustrating the partial order lattice construction and minimum element screening of the cell health status in an embodiment of the present invention.

[0063] Figure 3 This is a flowchart illustrating the verification process for effective local attractor regions in an embodiment of the present invention.

[0064] Figure 4 This is a functional block diagram of a device for early warning and location of thermal runaway of lithium-ion battery packs according to the present invention. Detailed Implementation

[0065] 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.

[0066] Example 1:

[0067] Please see Figure 1 As shown, this embodiment provides a method for early warning and location of thermal runaway in lithium-ion battery packs, including:

[0068] Step S1: Collect real-time voltage, temperature, current and number data of each cell, and construct a cell state observation tensor with spatiotemporal labels.

[0069] This step focuses on the need for structured data collection and integration of battery pack cell status. First, it collects real-time voltage, temperature, current, and cell number data for each cell, laying the foundation for subsequent tensor construction.

[0070] Specifically, the steps for collecting data from each battery cell and constructing a spatiotemporally labeled battery cell state observation tensor are as follows:

[0071] Step S11: Collect real-time voltage, real-time temperature, real-time current and serial number data of each cell to obtain the original data set of the cells.

[0072] The real-time voltage refers to the instantaneous potential difference across a single battery cell, which is synchronously acquired by a single-cell voltage acquisition module deployed on the positive and negative terminals of each cell. The sampling frequency can be dynamically adjusted according to the battery pack's operating conditions; for example, the sampling frequency can be increased during fast charging or high-rate discharging to capture rapid voltage fluctuations. The real-time temperature refers to the temperature value on or inside the battery cell, acquired through a combination of two methods: first, attaching an NTC thermistor to the surface of the cell casing to acquire the surface temperature; second, embedding a fiber optic temperature sensor near the cell's tabs to acquire the temperature of the internal core area, ensuring coverage of thermal runaway. The key areas prone to failure are controlled; the real-time current refers to the charging and discharging current flowing through a single cell, which is collected by connecting a micro shunt in series on each cell branch, and the current direction is recorded to distinguish between charging current and discharging current, avoiding misjudgment of the state due to confusion of current direction; the number data refers to the unique identification information of the cell in the battery pack, including the module number to which the cell belongs, the horizontal coordinate and the vertical coordinate within the module. This data is determined by topology planning during the battery pack production and assembly stage and is pre-written into the storage unit of the battery management system (BMS) as the basis for subsequent location of faulty cells.

[0073] Specifically, the process for collecting real-time data from each battery cell is as follows:

[0074] Step S111: By deploying a single-cell voltage acquisition module on the positive and negative terminals of each cell, the instantaneous potential difference across each cell is synchronously acquired, i.e., the real-time voltage; the sampling frequency is dynamically adjusted according to the battery pack's operating conditions (fast charging, high-rate discharging, or normal operating conditions) to obtain the real-time voltage data of each cell.

[0075] Step S112: Real-time temperature is collected using a combination of two methods. One method is to attach an NTC thermistor to the surface of the cell casing to collect the surface temperature. The other method is to embed an optical fiber temperature sensor near the cell tab to collect the temperature of the internal core area. The results of the two methods are integrated to obtain the real-time temperature data of each cell.

[0076] Step S113: Connect a miniature shunt in series on each cell branch to collect the charging and discharging current flowing through a single cell, and record the current direction to distinguish between charging and discharging current, so as to obtain the real-time current data of each cell.

[0077] Step S114: Read the pre-written cell number data from the storage unit of the battery management system (BMS). This data includes the module number to which the cell belongs, the horizontal coordinate and the vertical coordinate within the module, and obtain the cell number data.

[0078] By linking and integrating the above real-time voltage, real-time temperature, real-time current, and serial number data, the original dataset of the battery cell is obtained.

[0079] Step S12: Based on the original battery cell dataset, determine the sliding time window parameters, including the window length and sliding step size.

[0080] Specifically, the process for determining the sliding time window parameters is as follows:

[0081] Step S121: Statistically analyze the average evolution time of thermal runaway from early weak anomalies to obvious signs in the historical charge and discharge cycle data of the battery pack; this statistical analysis is based on historical thermal runaway case data and battery pack charge and discharge cycle records to ensure coverage of the thermal runaway evolution pattern under different operating conditions.

[0082] Step S122: Set 1.2-1.5 times the average evolution time statistically obtained in step S121 as the length of the sliding time window to ensure that the window can fully cover the early state change process of thermal runaway; at the same time, set 1 / 5 of the window length as the sliding step size to avoid data omission due to excessive step size or data redundancy due to excessive step size, and finally obtain the sliding time window parameters.

[0083] Step S13: Based on the cell number data and sliding time window parameters in the original cell dataset, perform index encoding and physical quantity sequence generation operations to obtain the cell index, voltage sequence, temperature sequence and current sequence.

[0084] Specifically, the process of index encoding and physical quantity sequence generation is as follows:

[0085] Step S131: Based on the numbering data (including module number, horizontal coordinate, and vertical coordinate) in the original battery cell dataset, assign a unique battery cell index to each battery cell; the battery cell index corresponds one-to-one with the module number and row and column coordinates of the battery cell, and the physical location of the battery cell in the battery pack can be directly traced through the index to obtain the battery cell index.

[0086] Step S132: For each cell, arrange its real-time voltage data in chronological order within the sliding time window to generate the voltage sequence of that cell; each sliding time window corresponds to a set of voltage sequences, and the sequence data is updated when the window slides.

[0087] Step S133: Using the same time sequence as in step S132, arrange the real-time temperature data of each cell within the sliding time window to generate the temperature sequence of the cell.

[0088] Step S134: Extract the real-time instantaneous current value of each cell within the sliding time window according to the time node, arrange them in chronological order, and generate the current sequence of the cell.

[0089] Step S14: Based on the numbering data in the original battery cell dataset, the Laplace feature mapping method is used to perform position topology coding to obtain the position topology code of each battery cell.

[0090] Specifically, the generation process of location topology coding is as follows:

[0091] Step S141: Construct a spatial adjacency matrix based on the row and column coordinates contained in the serial number data in the original battery cell dataset; the elements in the spatial adjacency matrix take the value of 1 or 0. When two battery cells have a direct adjacent relationship in physical space (such as being adjacent horizontally or vertically), the corresponding element takes the value of 1, otherwise it takes the value of 0.

[0092] Step S142: Calculate the degree matrix of the spatial adjacency matrix; the degree matrix is ​​a diagonal matrix, and the value of the element on its diagonal is equal to the sum of all elements in the corresponding row of the spatial adjacency matrix, that is, the number of adjacent cells for each cell.

[0093] Step S143: Calculate the Laplacian matrix using the degree matrix and the spatial adjacency matrix. The calculation formula is: Laplacian matrix = degree matrix - spatial adjacency matrix.

[0094] Step S144: Solve for the eigenvalues ​​and eigenvectors of the Laplace matrix, and select the first 4 eigenvectors with smaller eigenvalues; combine these eigenvectors to form the cell's positional topology code, which can reflect the spatial relative position of the cell within the battery pack and its correlation with potential heat conduction paths.

[0095] Step S15: Based on the cell index, voltage sequence, temperature sequence, current sequence, location topology encoding, and sliding time window parameters, integrate and construct a cell state observation tensor with spatiotemporal labels.

[0096] Specifically, the process of constructing the cell state observation tensor with spatiotemporal labels is as follows:

[0097] Step S151: Determine the four dimensions of the tensor, namely: the first dimension (cell dimension) is the cell index, the second dimension (time dimension) is the time step of the sliding time window, the third dimension (physical quantity dimension) is the physical quantity type (including voltage, temperature and current), and the fourth dimension (spatial dimension) is the location topology encoding.

[0098] Step S152: Map the voltage sequence, temperature sequence, and current sequence of each cell to different physical quantity types in the third dimension to ensure that the physical quantity data of each time step is accurately associated with the cell index and time step.

[0099] Step S153: Embed the position topology code of each cell into the fourth dimension, so that each cell carries the corresponding spatial position code information in all time steps and all physical quantity types.

[0100] Step S154: Integrate four-dimensional data to form a cell state observation tensor with spatiotemporal labels; each four-dimensional coordinate unit of the tensor corresponds to a certain type of physical quantity data of a cell at a specific time step and the cell's position topology code, realizing the structured integration of the cell state's temporal evolution information, physical quantity monitoring information and spatial location information.

[0101] Traditional BMS systems, when processing battery cell data, often treat the voltage, temperature, and other data of each cell as isolated time series, focusing only on the physical quantity changes of a single cell and completely ignoring the impact of the spatial layout of cells within the battery pack on the propagation of heat anomalies. For example, when a cell generates heat during normal operation, causing a slight temperature increase in neighboring cells, a traditional system might mistakenly interpret this normal heat conduction as a risk of internal short circuit in a neighboring cell, leading to false alarms. This step addresses this by constructing a cell state observation tensor with spatiotemporal labels, incorporating location topology encoding as an independent dimension into the tensor structure. This allows the system to correlate the spatial relationships of cells in subsequent analysis, accurately distinguishing whether localized overheating originates from heat diffusion from neighboring cells or an internal fault within the individual cell. This solves the problem of misjudgment caused by the lack of spatial correlation information at the data structure level. Meanwhile, this tensor integrates scattered voltage, temperature, and current data with spatiotemporal labels into a structured input, providing a complete and relevant data source for the extended Kalman estimation algorithm for the thermoelectric state of the battery cells in step two. This ensures that subsequent algorithms can utilize the temporal evolution coupling and spatial thermal conduction path dependence between battery cells to more accurately estimate the health status of the cells, avoiding estimation bias caused by isolated data. Furthermore, the spatiotemporal correlation information preserved by the tensor provides a foundation for subsequent analysis of the propagation path of thermal runaway risk, helping to capture subtle but crucial abnormal evolution trends within the battery cell group in advance, providing key data support for overcoming the bottleneck of lagging early warning in existing technologies.

[0102] Step S2: Based on the cell state observation tensor with spatiotemporal labels, run the extended Kalman estimation algorithm for cell thermoelectric state to generate a partial order lattice of cell health state.

[0103] This step focuses on the requirements of accurate estimation of the thermoelectric state of the battery cells and ranking of the health status of the population. First, the battery cell state observation tensor with spatiotemporal labels generated in step S1 is called. Combined with the electrochemical-thermal coupling model, the extended Kalman estimation algorithm of the battery cell thermoelectric state is run, and finally the partial order lattice of the battery cell health status is constructed. The electrochemical-thermal coupling model refers to a mathematical model that integrates the electrochemical reaction process and the heat conduction process of the battery cell. It is used to describe the coupling relationship between the battery cell's state of charge, DC internal resistance, voltage, temperature, and current. The core parameters of this model include the number of moles of positive electrode active material, the number of moles of negative electrode active material, the electrolyte conductivity, the battery cell's equivalent heat capacity, and the battery cell's thermal conductivity coefficient. These parameters are obtained through battery cell material characterization experiments and thermal characteristic tests. For example, the number of moles of positive and negative electrode active materials is determined by X-ray diffraction analysis, the electrolyte conductivity is measured by the four-probe method, the battery cell's equivalent heat capacity is determined by differential scanning calorimetry, and the battery cell's thermal conductivity coefficient is tested by laser scintillation. Before practical application, the model parameters need to be calibrated based on standard charge-discharge cycle experiments to ensure that the deviation between the model output and the actual state of the battery cell is controlled within a preset range. The state of charge refers to the proportion of the current remaining charge of the battery cell to its rated capacity, which reflects the energy storage capacity of the battery cell, and its value ranges from 0 to 1; the DC internal resistance refers to the equivalent internal resistance exhibited by the battery cell under DC operating conditions, which reflects the sum of the resistance to electrochemical reaction and ohmic resistance inside the battery cell, and its value varies with the degree of aging of the battery cell and temperature, and is usually larger at low temperature or high aging.

[0104] Specifically, the steps for running the extended Kalman lattice algorithm for estimating the thermoelectric state of the battery cell and generating the partial order lattice of the battery cell's health state are as follows:

[0105] Step S21: Based on the spatiotemporally labeled cell state observation tensor generated in step S1, extract the algorithm input data to obtain the algorithm input dataset.

[0106] Specifically, the process for extracting the algorithm input data is as follows:

[0107] Step S211: Extract the real-time voltage sequence, real-time temperature sequence, and real-time current sequence corresponding to the third-dimensional physical quantity type from the spatiotemporally labeled cell state observation tensor. These sequences are the core observation data reflecting the thermoelectric state of the cell.

[0108] Step S212: Extract the cell index corresponding to the first dimension and the time step corresponding to the second dimension from the cell state observation tensor with spatiotemporal labels, so as to associate the observation data of different cells at different time nodes.

[0109] Step S213: Integrate the extracted real-time voltage sequence, real-time temperature sequence, real-time current sequence, cell index, and time step to obtain the algorithm input dataset; the fourth-dimensional position topology encoding of the cell state observation tensor with spatiotemporal labels is not involved in the calculation of this step, but is only used for subsequent steps to associate cell spatial position information.

[0110] Step S22: Based on the algorithm input dataset generated in step S21 and the electrochemical-thermal coupling model, run the extended Kalman estimation algorithm for the thermoelectric state of the battery cell to obtain the (state of charge, DC internal resistance) binary for each battery cell.

[0111] The extended Kalman spectroscopy (EPS) estimation algorithm for battery cells refers to an algorithm that simultaneously estimates the state of charge (SOC) and DC internal resistance of battery cells based on the extended Kalman filter framework and incorporating an electrochemical-thermal coupling model. Specifically, the algorithm's execution flow is as follows:

[0112] Step S221: Construct the state equation and observation equation of the algorithm. Based on the electrochemical-thermal coupling model, the state of charge and DC internal resistance of the battery cell are used as state variables to construct the state equation. The input of the state equation is the real-time current sequence in the algorithm input dataset, and the output is the predicted value of the state variable at the next moment. Process noise is introduced into the state equation. The variance matrix of the process noise is determined by statistically analyzing the state fluctuation data of the battery cell under standard operating conditions. For example, under constant temperature and constant current charging and discharging conditions, the measured values ​​of the state of charge and DC internal resistance of the battery cell are continuously collected, and the deviation between the measured values ​​and the model prediction values ​​is calculated. The variance of this deviation is used as the diagonal element of the process noise variance matrix. Based on the real-time voltage and temperature sequences in the algorithm input dataset, an observation equation is constructed. The input of the observation equation is the estimated value of the state variables, and the output is the predicted value of voltage and temperature. Observation noise is introduced into the observation equation. The variance matrix of this observation noise is determined by statistically analyzing the measurement errors of the voltage and temperature sensors. For example, in a sensor static calibration experiment, the deviation between the measured value and the true value of the sensor at standard voltage and standard temperature is recorded, and the variance of this deviation is used as the diagonal element of the observation noise variance matrix.

[0113] Step S222: Initialize filter parameters. For each cell index in the algorithm input dataset, initialize the initial values ​​of the state variables and the initial values ​​of the error covariance matrix. The initial value of the state of charge (SCC) among the initial values ​​of the state variables is determined by the relationship curve between the open-circuit voltage and the SCC of the cell. That is, collect the open-circuit voltage of the cell at the current temperature, query the pre-calibrated open-circuit voltage-SCC curve, and obtain the corresponding SCC value as the initial value. The initial value of the DC internal resistance is determined by the initial internal resistance test data of the cell at the factory. If the factory data is lacking, it is calculated by a small current charge and discharge experiment at room temperature. For example, in an environment of 25℃, charge and discharge the cell at 0.1 times the rated current, record the voltage change and current value during the charge and discharge process, and calculate the DC internal resistance as the initial value using Ohm's law. The initial value of the error covariance matrix is ​​a 2×2 diagonal matrix. The diagonal elements are the variances of the initial estimation error of the state of charge and the initial estimation error of the DC internal resistance, respectively. These variances are determined by statistically analyzing the deviation data of multiple initial value calibration experiments. For example, if the same cell is calibrated 10 times for initial state of charge and DC internal resistance, the deviation between each calibration value and the average value is calculated, and the variance of this deviation is used as the value of the corresponding diagonal element.

[0114] Step S223: Perform the filtering prediction step. Based on the state equation constructed in step S221 and the initial values ​​of the state variables and the error covariance matrix initialized in step S222, combined with the real-time current sequence in the algorithm input dataset, calculate the predicted values ​​of the state variables and the predicted values ​​of the error covariance at each time step. For the k-th time step, the predicted value of the state variables is calculated by substituting the updated value of the state variables at the (k-1)-th time step into the state equation. The predicted value of the error covariance is calculated using the Jacobian matrix of the state equation, the updated value of the error covariance at the (k-1)-th time step, and the process noise variance matrix. The Jacobian matrix of the state equation is obtained by taking the partial derivative of the nonlinear function in the state equation, which is used to linearize the nonlinear relationship of the state equation and ensure that the iterative operation of the extended Kalman filter can proceed normally.

[0115] Step S224: Perform the filtering update step. First, based on the observation equation constructed in step S221 and the predicted values ​​of the state variables at the current time step, calculate the predicted values ​​of voltage and temperature. Then, compare the predicted values ​​with the real-time voltage and real-time temperature at the current time step in the algorithm input dataset to obtain the observation residual. Next, calculate the Kalman gain based on the predicted value of the error covariance, the Jacobian matrix of the observation equation, and the observation noise variance matrix. The value of the Kalman gain is used to balance the influence of the predicted values ​​of the state variables and the observation residual on the updated values ​​of the state variables. When the observation noise is small, the Kalman gain is larger and more dependent on the observation data for updating. Finally, use the Kalman gain and the observation residual to correct the predicted values ​​of the state variables and the predicted values ​​of the error covariance to obtain the updated values ​​of the state variables and the error covariance at the current time step. Extract the state of charge and DC internal resistance from the updated values ​​of the state variables at each time step, take the average value over the time steps, and use it as the final state of charge and DC internal resistance of the cell. The two are combined to form a (state of charge, DC internal resistance) binary for the cell.

[0116] Step S23: Based on the (state of charge, DC internal resistance) binary pairs of each cell generated in step S22, construct a partial order lattice of cell health status.

[0117] Among them, the partial-order lattice of cell health status refers to a partially ordered set used to describe the dominance relationships among the health statuses of individual cells in a cell group. Its structure includes a set of elements and dominance relationships; specifically, the construction process is as follows:

[0118] Step S231: Collect the (state of charge, DC internal resistance) binary pairs of all cells generated in step S22 to form the element set of the partial order lattice of cell health status.

[0119] Step S232: Define the dominance relationship. If the (state of charge, DC internal resistance) binary pair of cell A satisfies that the state of charge ≤ the state of charge of cell B, and the DC internal resistance ≥ the DC internal resistance of cell B, then the health state of cell A is said to be dominated by the health state of cell B, denoted as cell A ≤ cell B. The physical meaning of this definition is: a lower state of charge indicates a weaker energy storage capacity, and a higher DC internal resistance indicates greater internal losses. Together, they reflect a worse health state of the cell.

[0120] Step S233: For any two different pairs in the element set, compare them according to the dominance relationship defined in step S232 to determine whether a dominance relationship exists between them; draw directed edges for pairs with a dominance relationship in the direction of "dominated element → dominating element", remove redundant directed edges (i.e., if there exists cell A ≤ cell B and cell B ≤ cell C, then remove the direct directed edge from cell A to cell C, and only retain the indirect dominance relationship), forming a Hasse graph of the partial order lattice of cell health status; the nodes of the Hasse graph are the (charge state, DC internal resistance) pairs of the cells, the corresponding cell index is marked next to the node, and the edges are the directed edges of the dominance relationship.

[0121] Step S234: Extract the set of minimum elements from the Hasse diagram. The set of minimum elements refers to the set of elements in the set of elements where no other element can dominate the element. This set corresponds to the potential high-risk cell candidate set. The cell corresponding to the minimum element is not dominated by any other cell in the group, and its health status is worse than or equal to all other cells in the group. Please refer to the flowchart of the partial order lattice construction and minimum element screening of cell health status in S23 above. Figure 2 ;

[0122] This step effectively addresses the lack of dynamic assessment capabilities for the consistency of cell group states in existing technologies by running the extended Kalman lattice for cell thermoelectric state estimation and constructing a partial order lattice for cell health states. Its role and advantages are mainly reflected in three aspects: First, addressing the limitations of traditional methods that rely solely on a single physical quantity or scalar consistency index (such as mean deviation or standard deviation), this step integrates an electrochemical-thermal coupling model and extended Kalman filtering. It can simultaneously estimate the state of charge and DC internal resistance using multi-source data of voltage, temperature, and current, avoiding estimation bias caused by interference from a single sensor and improving the accuracy of cell health state assessment. Second, compared to traditional methods that can only reflect local deviations in cell state, the (state of charge, DC internal resistance) binary tuple obtained in this step can more comprehensively characterize the cell health state, providing richer feature dimensions for group consistency assessment. Secondly, traditional methods cannot identify hidden abnormal cells that are "neither at the lowest state of charge nor at the highest DC internal resistance, but significantly deviate from the group's dominant path." For example, a cell may have a state of charge near the group mean, but its DC internal resistance is slightly higher than the mean, and its state of charge is lower than all cells with lower DC internal resistance. In this case, traditional scalar indicators are unlikely to detect the abnormality. However, the partial order lattice of cell health status constructed in this step can capture this asymmetric degradation feature through the dominance relationship. This allows such hidden abnormal cells to be included in the set of minimal elements, enabling the identification of early and subtle anomalies and providing a key basis for early warning of thermal runaway. Finally, the partially ordered lattice of cell health status and its set of minimal elements generated in this step provide core input data for constructing the phase space grid for thermal runaway risk propagation in the subsequent step three. The candidate set of potentially high-risk cells in the minimal element set can narrow down the scope of subsequent risk location and improve location efficiency. At the same time, the distribution and dominance relationships of cell group health status contained in the partially ordered lattice can provide the basic laws of health status evolution for analyzing the direction of risk propagation in step three, avoiding subsequent risk propagation analysis from deviating from the actual health status of the cells and ensuring the logical coherence and reliability of the entire early warning and location process. In addition, by identifying potentially high-risk cells in advance, it can provide maintenance personnel with targets for early intervention and prevent risky cells from further deteriorating to the middle and late stages of thermal runaway.

[0123] For example, suppose a battery pack contains 100 cells. In step S21, the voltage, temperature, and current data of each cell are extracted from the cell state observation tensor with spatiotemporal labels. After running the algorithm in step S22, 100 (state of charge, DC internal resistance) binary pairs are obtained. In step S23, the dominance relationship of these binary pairs is judged. It is found that the binary pairs of 5 cells are not dominated by any other cells. The indices of these 5 cells are included in the minimum element set as high-risk candidates. The Hasse diagram shows that the state of charge of these 5 cells is in the lower range of the group, the DC internal resistance is in the higher range of the group, and there is no mutual dominance relationship between them. They belong to the group of cells with the worst health status.

[0124] Step S3: Construct a phase space grid for thermal runaway risk propagation based on the partial order lattice of cell health status.

[0125] This step addresses the need for joint modeling of the "state-space-propagation" of battery cell risks. First, it utilizes the partially ordered lattice of battery cell health states generated in step S2 and the spatiotemporally labeled battery cell state observation tensor generated in step S1 to extract correlation information from these two types of data and construct a basic dataset. Through multi-dimensional parameter calculation and grid modeling, a phase space grid for thermal runaway risk propagation is ultimately formed. This phase space grid for thermal runaway risk propagation refers to a grid structure that integrates battery cell state information, physical location information, and risk propagation trends. Its core features are non-uniformity and a vector field, used to represent the spatial distribution and dynamic propagation patterns of the risk.

[0126] Specifically, the steps for constructing the phase space grid for thermal runaway risk propagation are as follows:

[0127] Step S31: Based on the partial order lattice of cell health status generated in step S2 and the cell status observation tensor with spatiotemporal labels generated in step S1, extract the associated data and integrate them to obtain the basic dataset.

[0128] Specifically, the process for extracting and integrating related data is as follows:

[0129] Step S311: Extract the (state of charge, DC internal resistance) binary pairs and corresponding cell indices of all cells from the partial sequence grid of cell health status to obtain cell health status data.

[0130] Step S312: Extract the real-time voltage sequence, real-time temperature sequence, and physical location number (row and column coordinates) associated with the cell index from the spatiotemporally labeled cell state observation tensor. At the same time, verify the accuracy of the physical location number through the fourth-dimensional location topology encoding to obtain the cell state and location data.

[0131] Step S313: Associate the cell health status data with the cell status and location data according to the cell index, and integrate them to form a basic dataset.

[0132] Step S32: Based on the basic dataset generated in step S31, calculate the SOC deviation, internal resistance deviation, voltage change rate and temperature change rate of each cell to obtain the four-dimensional state parameter set of the cell.

[0133] Specifically, the process for calculating the four-dimensional state parameters of the battery cell is as follows:

[0134] Step S321: Calculate SOC deviation and internal resistance deviation. Based on the (state of charge, DC internal resistance) pairs in the basic dataset, after removing the cell data corresponding to the minimum element set, calculate the group SOC reference value and group DC internal resistance reference value of the cells by weighting the number of dominations. Subtract the group SOC reference value from the SOC of each cell to obtain the SOC deviation, and subtract the group DC internal resistance reference value from the DC internal resistance of each cell to obtain the internal resistance deviation. The group SOC reference value of the cells is calculated from the SOC of all cells in the partial order grid of cell health status. Specifically, it is the weighted average of the SOC of the remaining cells after removing the cells in the minimum element set of the partial order grid. The weight value is positively correlated with the number of dominations of the cell in the partial order grid, that is, the more times the cell dominates other cells, the greater its weight.

[0135] Step S322: Calculate the voltage change rate and temperature change rate. Extract real-time voltage and temperature sequences from the basic dataset. Using the same sliding time window as in step S1, perform linear fitting on the two types of sequences respectively. The slopes of the fitted lines are used as the voltage change rate and temperature change rate, respectively.

[0136] Step S323: Associate the SOC deviation, internal resistance deviation, voltage change rate, and temperature change rate of each cell according to the cell index to obtain the four-dimensional state parameter set of the cell.

[0137] Step S33: Based on the four-dimensional state parameter set and basic dataset of the battery cell generated in step S32, construct the phase space points of the battery cell and associate them with physical locations to obtain a set of phase space points of the battery cell with physical locations.

[0138] Specifically, the process of constructing the phase space points of the battery cell and associating them with their physical locations is as follows:

[0139] Step S331: Use the SOC deviation, internal resistance deviation, voltage change rate, and temperature change rate of each cell as four-dimensional coordinate parameters, and combine them to form the coordinate point of the cell in the four-dimensional phase space, i.e., the cell phase space point.

[0140] Step S332: Extract the physical location number (row and column coordinates) of each cell phase space point from the basic dataset, bind it with the cell phase space point, and integrate them to obtain a set of cell phase space points with physical location.

[0141] Step S34: Based on the distribution of the physical location of the battery cells in the basic dataset, divide the data into three-dimensional mesh cells to obtain a set of three-dimensional mesh cells.

[0142] Specifically, the process of dividing a 3D mesh into units is as follows:

[0143] Step S341: Using the physical layout of the battery pack as the base, establish a three-dimensional coordinate system, with the x-axis corresponding to the horizontal direction, the y-axis corresponding to the vertical direction, and the z-axis corresponding to the module stacking direction.

[0144] Step S342: Calculate the average distribution density of cells in the battery pack, and determine the preset threshold for the number of cells in each grid cell (preferably set to 15-25 cells in the range of the average distribution density of cells in the battery pack, to ensure that the grid cell contains enough samples to characterize the local group characteristics while maintaining spatial resolution, so that the risk source location accuracy is better than the spatial span of a single grid cell); divide the three-dimensional grid cells according to the physical location number (row and column coordinates) and the preset threshold for the number of cells, and assign a unique three-dimensional grid coordinate to each grid cell to form a set of three-dimensional grid cells.

[0145] Step S35: Embed the set of cell phase space points with physical locations generated in step S33 into the set of three-dimensional mesh cells generated in step S34, and construct a joint risk density function for each mesh cell to obtain a set of mesh cells with joint risk density function.

[0146] Specifically, the process of embedding and function construction is as follows:

[0147] Step S351: Based on the physical location number (row and column coordinates) of the cell phase space points with physical location, match the corresponding three-dimensional grid coordinates, and assign the cell phase space points to the corresponding three-dimensional grid cells to form a set of grid cells containing cell phase space points.

[0148] Step S352: For each grid cell containing cell phase space points, construct a joint risk density function using the kernel density estimation method. Select a Gaussian kernel function and determine the bandwidth parameter using cross-validation (divide the cell phase space points into training and validation sets, and select the bandwidth with the largest log-likelihood value); with each cell phase space point as the center, apply the Gaussian kernel function and sum the joint risk density function values ​​at each location to form a complete joint risk density function.

[0149] Step S353: Associate each grid cell with its joint risk density function to obtain a set of grid cells with the joint risk density function.

[0150] Step S36: Based on the set of mesh cells with joint risk density function generated in step S35, establish directed edges between adjacent mesh cells to obtain a mesh structure with initial directed edges.

[0151] Specifically, the process of establishing a directed edge is as follows:

[0152] Step S361: Determine adjacent mesh cells. The standard is that the difference between the three-dimensional mesh coordinates of two mesh cells in the horizontal, vertical or stacking direction is 1 (physical location is directly adjacent).

[0153] Step S362: Based on the maximum value of the joint risk density function, establish directed edges from the grid cells with lower maximum values ​​to the grid cells with higher maximum values, forming a grid structure with initial directed edges.

[0154] Step S37: Based on the partial order lattice of cell health status generated in step S2 and the mesh structure with initial directed edges generated in step S36, calculate the spatial distance and dominance strength, modulate the edge weights of the directed edges, and obtain a mesh structure with weighted directed edges.

[0155] Specifically, the process of modulating edge weights is as follows:

[0156] Step S371: Calculate the spatial distance, which is the straight-line distance between the geometric centers of adjacent grid cells. The true spatial distance is obtained by calculating the square root of the square of the coordinate difference through the three-dimensional grid coordinates. Then, the true spatial distance is dimensionless, that is, the ratio of the true spatial distance to the maximum geometric size of the battery pack is used to ensure the dimensionality consistency of the edge weight calculation results.

[0157] Step S372: Calculate the dominance strength by counting the number of dominance relationships between cells in the starting grid cell and cells in the ending grid cell, and dividing the result by the product of the total number of cells in the two grid cells. The value range is 0-1.

[0158] Step S373: Calculate the edge weights according to the formula "edge weight = (1 / spatial distance) × dominance strength × adjustment coefficient" (the adjustment coefficient is calibrated using historical thermal runaway case data); associate the edge weights with the corresponding directed edges to obtain a mesh structure with weighted directed edges.

[0159] Step S38: Integrate the set of mesh cells with joint risk density function generated in step S35 with the mesh structure with weighted directed edges generated in step S37 to construct a phase space mesh for thermal runaway risk propagation.

[0160] Specifically, the integration and construction process is as follows:

[0161] Step S381: Embed the joint risk density function into the corresponding grid cell, and connect the weighted directed edges to the adjacent grid cells to form the basic grid structure.

[0162] Step S382: Adjust the size of the grid cells according to the cell distribution density (small volume in dense areas and large volume in sparse areas) to give the grid non-uniformity; take the direction of the directed edge as the direction of the vector field, and the edge weight is positively correlated with the vector length to give the grid vector field characteristics, and finally obtain the phase space grid for thermal runaway risk propagation.

[0163] This step effectively solves the core problems of isolated and static risk assessment in existing technologies by constructing a phase space grid for thermal runaway risk propagation. Its role and advantages are mainly reflected in four aspects: First, compared with the limitations of traditional methods that treat battery cells as independent nodes and only assess the risk state of individual battery cells, this step deeply couples the four-dimensional state parameters of the battery cell (SOC deviation, internal resistance deviation, voltage change rate, and temperature change rate) with the physical location. It realizes the joint modeling of "state-space" through three-dimensional grid cells, so that risk assessment is no longer separated from the actual spatial distribution of the battery cell, laying the foundation for accurate location of risk sources. For example, when the joint risk density of grid cells in a certain area increases, the physical location of the area can be directly locked through grid coordinates, avoiding the ambiguity caused by the inability of traditional methods to associate spatial locations. Second, the introduction of the joint risk density function breaks through the static limitations of traditional threshold determination. By using kernel density estimation to characterize the continuity of risk distribution within a grid cell, it can capture the early trend of "risk continuously increasing even though the threshold has not been reached." For example, the temperature change rate of a cell in a certain grid cell may not have reached the dangerous threshold, but the joint risk density continues to rise due to the superposition of SOC deviation and internal resistance deviation. This change can be intuitively presented through the function curve, providing a basis for early warning. Third, the directed edge structure with vector field realizes the visualization of the thermal runaway risk propagation trend. Traditional heatmaps can only show the spatial distribution of risk and cannot reflect how the risk propagates. However, the vector field in this step can clearly define the propagation path of risk from low-risk areas to high-risk areas. For example, when a certain grid cell becomes the risk convergence center, the vectors of the surrounding grid cells all point to this center, which can identify the source and direction of risk diffusion in advance, providing accurate guidance for maintenance personnel to isolate risky units. Fourth, the grid structure provides structured input for the identification of local attractor regions in step four. The overall manifold characteristics of the vector field are the core basis for identifying attractor regions. Without the phase space grid constructed in this step, step four cannot determine whether the risk is self-reinforcing from the perspective of "dynamic evolution" and can only rely on the instantaneous state threshold. At the same time, the non-uniformity design of the grid takes into account both positioning accuracy and computational efficiency. In densely populated areas of cells, small grids improve the accuracy of risk positioning, while in sparse areas, large grids reduce the consumption of computational resources, ensuring that the system can still maintain stable operation under complex working conditions and reducing the risk of false alarms or missed alarms.

[0164] Step S4: Under the constraints of the phase space grid for thermal runaway risk propagation, perform local attractor region identification and early warning decision based on manifold analysis.

[0165] This step focuses on the dynamic trend prediction and precise location of thermal runaway risk. First, it utilizes the thermal runaway risk propagation phase space grid generated in step S3, the cell state observation tensor with spatiotemporal labels generated in step S1, and the cell health state partial order grid generated in step S2. Correlation information is extracted from these three types of data to construct an analysis dataset. Through manifold analysis, region verification, risk location, index calculation, and early warning determination, the high-risk source cell number and early warning level are finally output. The local attractor region refers to a region in the thermal runaway risk propagation phase space grid where the risk density gradient of several adjacent grid cells continuously points towards the same central grid cell. Its core characteristic is that the risk converges towards the center, and the trend strengthens over time. The determination requires meeting two conditions: "the gradient ratio of adjacent grid cells pointing towards the center exceeds a preset proportion" and "remains stable for multiple consecutive time steps." The proportion is determined through the gradient distribution characteristics of historical cases, and the time step matches the sliding time window length of step S1.

[0166] Specifically, the steps for performing local attractor region identification and early warning decision-making based on manifold analysis are as follows:

[0167] Step S41: Based on the thermal runaway risk propagation phase space grid generated in step S3, the cell state observation tensor with spatiotemporal labels generated in step S1, and the cell health state partial order grid generated in step S2, extract and integrate the associated data to obtain the analysis dataset.

[0168] Specifically, the process for extracting and integrating related data is as follows:

[0169] Step S411: Extract vector field information, joint risk density function of each grid cell, and grid cell-cell index mapping relationship from the thermal runaway risk propagation phase space grid to obtain grid core data.

[0170] Step S412: Extract the latest time step real-time voltage, real-time temperature and cell number corresponding to each cell index from the cell state observation tensor with spatiotemporal labels to obtain the cell real-time state data; extract the (state of charge, DC internal resistance) binary corresponding to each cell index from the cell health state partial order grid to obtain the cell health state data.

[0171] Step S413: Link the grid core data, real-time cell status data and cell health status data according to the cell index, and integrate them to form an analysis dataset.

[0172] Step S42: Based on the analysis dataset generated in step S41, perform global manifold analysis on the vector field in the phase space grid of thermal runaway risk propagation to obtain manifold feature map.

[0173] The global manifold analysis refers to a comprehensive analysis method for the vector field in the phase space grid of thermal runaway risk propagation, analyzing both the global topological structure and local variation trends, used to uncover risk convergence or diffusion patterns. Specifically, the process of performing global manifold analysis is as follows:

[0174] Step S421: Divide the vector field into local vector subfields according to the three-dimensional coordinates of the phase space grid for thermal runaway risk propagation. Each local vector subfield contains vector information of a preset number of adjacent grid cells. The preset number is determined by the average density of the grid cells.

[0175] Step S422: Calculate the continuity index of each local vector subfield (calculated by the average difference in direction between adjacent vectors within the subfield; the smaller the average, the better the continuity), and filter out local vector subfields with continuity indices higher than a preset threshold. This preset threshold is set to the 85th-90th quantile of the distribution of continuity indices of all local vector subfields under normal operating conditions, ensuring that the selected subfields have significantly better continuity than those in conventional thermal diffusion scenarios.

[0176] Step S423: Merge the filtered local vector subfields to form continuous manifold segments. Calculate the consistency of vector directions within each segment. Identify segments whose vector directions tend to be consistent and point towards a core region as potential local attractor region candidates, and integrate them to form a manifold feature map. Define vector directions as consistent: divide the magnitude of the sum of all unit vectors within the segment by the total number of vectors, with a value range of [0,1]. The closer the value is to 1, the higher the consistency. The criterion for potential local attractor region candidates is that the consistency value exceeds 0.75.

[0177] Step S43: Based on the manifold feature map generated in step S42 and the analysis dataset generated in step S41, extract potential local attractor region candidates and perform verification to obtain effective local attractor regions.

[0178] For specific details on the verification process of effective local attractor regions, please refer to [link / reference]. Figure 3 ;

[0179] The extraction and verification process is as follows:

[0180] Step S431: Extract potential local attractor region candidates from the manifold feature map, and obtain grid cell association data and risk density gradient information in each candidate region; the risk density gradient refers to the ratio of the difference in joint risk density function between adjacent grid cells to the grid spacing, which is used to quantify the rate and direction of risk change, and a positive value indicates that the risk increases from the starting cell to the ending cell.

[0181] Step S432: Count the number of risk density gradients pointing from adjacent grid cells to the central cell within the candidate region, and calculate the proportion of this number to the total number of adjacent grid cell pairs within the candidate region. If the proportion exceeds a preset threshold, proceed to the next verification step. The central cell refers to the grid cell in the local attractor region that is jointly pointed to by surrounding gradients, typically the core of risk convergence. The preset threshold is preferably set to 0.75-0.85, determined by statistically analyzing the gradient pointing proportion distribution of confirmed local attractor regions in historical thermal runaway cases, and taking its 80th quantile to ensure that only regions with significant convergence trends are retained for verification.

[0182] Step S433: Check the stability of the risk density gradient pointing relationship over multiple consecutive time steps. If it remains stable within a preset number of time steps, it is determined to be a valid local attractor region; otherwise, the candidate region is removed.

[0183] Step S44: Based on the effective local attractor region generated in step S43 and the analysis dataset generated in step S41, locate candidate cells with high risk of thermal runaway.

[0184] Specifically, the location process is as follows:

[0185] Step S441: Extract the central cell of the effective local attractor region and obtain the cell index list corresponding to the central cell from the analysis dataset; at the same time, retrieve the high-risk cell candidate set from step S234, filter out the intersection of the central cell cell index list and the high-risk cell candidate set, and obtain the high-risk candidate cell list within the central cell.

[0186] Step S442: Extract the latest real-time voltage and temperature of two types of cells from the analysis dataset: First, extract the voltage and temperature of all cells in the high-risk candidate cell list within the central unit, sort them, and determine the low-voltage cells (lowest voltage) and high-temperature cells (highest temperature) within the list; if the high-risk candidate cell list within the central unit is empty, extract the voltage and temperature of cells in the cell index list of all cells in the central unit, sort them, and determine the low-voltage cells and high-temperature cells. Lowest voltage means that the latest real-time voltage of a cell in the corresponding list is lower than that of other cells in the list, and the voltage is extracted from the cell state observation tensor with spatiotemporal labels; highest temperature means that the latest real-time temperature of a cell in the same list is higher than that of other cells in the list, and the temperature is extracted from the cell state observation tensor with spatiotemporal labels.

[0187] Step S443: If the low-voltage cell and the high-temperature cell are the same cell, then directly identify it as a candidate cell for high-risk thermal runaway source (if the candidate cell for high-risk thermal runaway source belongs to the high-risk cell candidate set, its high-risk attribute can be confirmed first); if they are not the same cell, first determine whether they belong to the high-risk cell candidate set: if only one of them belongs, then select the cell in the candidate set as the candidate cell for high-risk thermal runaway source; if both belong or neither belongs, then extract the (state of charge, DC internal resistance) binary pair of the two from the analysis dataset, and select the cell with a lower state of charge and a higher DC internal resistance as the candidate cell for high-risk thermal runaway source.

[0188] Step S45: Based on the effective local attractor region generated in step S43 and the analysis dataset generated in step S41, calculate the stability index of the effective local attractor region.

[0189] Specifically, the calculation process is as follows:

[0190] Step S451: Extract vector field data of all grid cells within the effective local attractor region from the analysis dataset, calculate the vector field divergence of each grid cell, and take the average value as the overall divergence of the region. The vector field divergence is a physical quantity describing the degree of divergence or convergence of the vector field; a negative value indicates convergence. It needs to be combined with grid non-uniformity correction and is obtained by calculating the sum of the rates of change of each coordinate axis of the vector within the grid.

[0191] Step S452: Extract vector field data of all grid cells within the effective local attractor region from the analysis dataset, calculate the vector field curl of each grid cell, and take the average value as the overall curl of the region. The vector field curl is a physical quantity describing the rotation trend of the vector field. A value close to 0 indicates that the directions are consistent. It needs to be combined with grid non-uniformity correction and is obtained by calculating the ratio of the difference in direction to the difference in position between adjacent vectors.

[0192] Step S453: Normalize the global divergence and global curl respectively (map to the 0-1 interval; the smaller the divergence, the closer the normalization value is to 1; the closer the curl is to 0, the closer the normalization value is to 1).

[0193] Step S454: The normalized divergence and curl values ​​are summed by weighting according to preset weights to obtain the stability index; the preset weights are determined by training with historical thermal runaway case data and are set according to the degree of influence of divergence and curl on the probability of thermal runaway.

[0194] Step S46: Based on the stability index generated in step S45 and the candidate cells for high-risk thermal runaway generated in step S44, determine the warning level and output the results.

[0195] Specifically, the process of determining and outputting is as follows:

[0196] Step S461: Determine the warning level based on the stability index: trigger a Level 1 alarm when the stability index is significantly close to 1, trigger a Level 2 alarm when it is in the middle range, and trigger a Level 3 alarm when it is close to 0.

[0197] Step S462: Extract the cell numbers of candidate cells with high risk of thermal runaway from the analysis dataset, associate the cell numbers with the warning level and output them to provide maintenance personnel with a basis for location and repair.

[0198] This step, by performing local attractor region identification and early warning decision-making based on manifold analysis, fundamentally solves the core defect of traditional technologies that rely on static thresholds and cannot predict the evolution trend of risks. Its role and advantages are mainly reflected in four aspects: First, compared with traditional methods that judge risks solely by whether temperature and voltage exceed fixed thresholds, this step captures the dynamic convergence trend of risks by detecting local attractor regions. Even if the current temperature and voltage have not reached the danger threshold, as long as the risk exhibits self-reinforcing convergence characteristics (i.e., forming attractors), an early warning can be issued. This judgment method based on evolution trends can provide occupants with a longer emergency escape time, making up for the shortcomings of traditional early warning lag. Second, the introduction of a dual verification condition of "lowest voltage and highest temperature" during the positioning process, combined with the health status data in the partial sequence grid of cell health status in step two, significantly reduces the risk of false alarms caused by interference from a single sensor or instantaneous fluctuations. For example, if a cell only has a temporarily low voltage but a normal temperature, or only has a temporarily high temperature but a normal voltage, it will not be identified as a high-risk source. Traditional single threshold methods are prone to misjudging such interference as risk. The multi-condition verification in this step significantly improves the robustness of the system. Third, the introduction of the stability index enables dynamic classification of warning levels. Different levels correspond to different levels of risk urgency. Maintenance personnel can take differentiated measures according to the level (such as immediate shutdown and maintenance for a level one alarm, and enhanced monitoring for a level two alarm), avoiding the rigid problem of the traditional method of "either alarm or not alarm," and improving the flexibility and efficiency of maintenance. Fourth, the final output cell number is directly associated with the physical location within the battery pack, solving the problem that traditional technologies cannot accurately locate faulty cells. Maintenance personnel do not need to check each cell one by one; they can directly find the faulty cell based on the output number and quickly isolate it. This reduces maintenance costs, shortens the time window for risk spread, and further reduces the probability of safety accidents.

[0199] For example, in a phase space grid for thermal runaway risk propagation, S42 analysis yields a candidate region containing 5 adjacent grid cells, and S43 verifies it as an effective local attractor region; S44 extracts the central cell cell index list (cells 101, 102, 103), whose latest voltages are 3.2V, 3.0V, and 3.1V, and temperatures are 35℃, 38℃, and 36℃, respectively, identifying cell 102 as a high-risk source candidate; S45 calculates the stability index to be 0.95, and S46 triggers a level 1 alarm and outputs the cell 102 number, allowing maintenance personnel to quickly locate and isolate the source.

[0200] Example 2:

[0201] This embodiment, based on Embodiment 1, provides a device for early warning and location of thermal runaway in lithium-ion battery packs, such as... Figure 4 As shown, it includes:

[0202] Cell data acquisition and tensor construction unit: used to acquire real-time voltage, temperature, current and number data of each cell, and construct cell state observation tensor with spatiotemporal labels;

[0203] Thermoelectric state estimation and partial order lattice generation unit: used to run the extended Kalman estimation algorithm for the thermoelectric state of the battery cell based on the cell state observation tensor with spatiotemporal labels, and generate a partial order lattice of the battery cell health state;

[0204] Risk propagation phase space grid construction unit: used to construct a phase space grid for thermal runaway risk propagation based on the partial order lattice of cell health state;

[0205] Location and early warning decision unit: used to perform local attractor region identification and early warning decision based on manifold analysis under the constraints of the phase space grid of thermal runaway risk propagation.

Claims

1. A method of early warning positioning thermal runaway of a lithium ion battery pack, characterized in that, The method comprises: S1: Collecting real-time voltage, temperature, current and number data of each battery cell, and constructing a battery cell state observation tensor with space-time labels; S2: Based on the battery cell state observation tensor with space-time labels, running the battery cell thermal-electric state extended Kalman estimation algorithm to generate a battery cell health state partial order lattice; S3: Based on the battery cell health state partial order lattice, constructing a thermal runaway risk propagation phase space grid; S4: Under the constraint of the thermal runaway risk propagation phase space grid, performing local attractor region identification and early warning decision based on manifold analysis; The step of constructing the thermal runaway risk propagation phase space grid comprises: Based on the battery cell health state partial order lattice and the battery cell state observation tensor with space-time labels, extracting associated data and integrating to obtain a basic data set; Based on the basic data set, calculating the SOC deviation, internal resistance deviation, voltage change rate and temperature change rate of each battery cell to obtain a four-dimensional state parameter set of the battery cell; Based on the four-dimensional state parameter set of the battery cell and the basic data set, constructing a battery cell phase space point and associating a physical location to obtain a battery cell phase space point set with a physical location; the step of constructing a battery cell phase space point and associating a physical location is: taking the SOC deviation, internal resistance deviation, voltage change rate and temperature change rate of each battery cell as four-dimensional coordinate parameters, combining to form the coordinate point of the battery cell in the four-dimensional phase space, that is, the battery cell phase space point; extracting the physical location number of the battery cell corresponding to each battery cell phase space point from the basic data set, binding with the battery cell phase space point, and integrating to obtain a battery cell phase space point set with a physical location; Based on the distribution of the physical location of the battery cell in the basic data set, dividing three-dimensional grid units to obtain a three-dimensional grid unit set; Embedding the battery cell phase space point set with a physical location into the three-dimensional grid unit set, constructing a joint risk density function for each grid unit to obtain a grid unit set with a joint risk density function; Based on the grid unit set with a joint risk density function, establishing a directed edge between adjacent grid units to obtain a grid structure with an initial directed edge; Based on the battery cell health state partial order lattice and the grid structure with an initial directed edge, calculating the spatial distance and dominance intensity, modulating the edge weight of the directed edge to obtain a grid structure with a weighted directed edge; Integrating the grid unit set with a joint risk density function and the grid structure with a weighted directed edge to construct a thermal runaway risk propagation phase space grid.

2. The method of Claim 1, wherein, The step of constructing the battery cell state observation tensor with space-time labels comprises: S11: Collecting real-time voltage, real-time temperature, real-time current and number data of each battery cell to obtain a battery cell original data set; S12: Determining a sliding time window parameter based on the battery cell original data set; S13: Based on the number data in the battery cell original data set and the sliding time window parameter, generating physical quantity sequences through index encoding to obtain battery cell index, voltage sequence, temperature sequence and current sequence; S14: Based on the number data in the battery cell original data set, performing position topology coding using Laplace eigenmap method to obtain the position topology coding of each battery cell; S15: Integrating and constructing the cell state observation tensor with space-time labels based on the cell index, the voltage sequence, the temperature sequence, the current sequence, the position topology code and the sliding time window parameter.

3. The method of Claim 2, wherein the method further comprises: The step of performing position topology coding using Laplace eigenmap method includes: S141: Constructing a space adjacency matrix based on the row and column coordinates contained in the numbered data in the cell raw data set; S142: Calculating the degree matrix of the space adjacency matrix; S143: Calculating the Laplace matrix through the degree matrix and the space adjacency matrix; S144: Solving the eigenvalues and eigenvectors of the Laplace matrix, and selecting the first four eigenvectors with the smallest eigenvalues; combining the eigenvectors to form the position topology code of the cell.

4. The method of Claim 3, wherein the method further comprises: The steps of running the cell thermal-electric state extended Kalman estimation algorithm and generating the cell health state partial order lattice include: S21: Extracting algorithm input data based on the cell state observation tensor with space-time labels to obtain an algorithm input data set; S22: Running the cell thermal-electric state extended Kalman estimation algorithm based on the algorithm input data set and the electrochemical-thermal coupling model to obtain the binary tuple of each cell; S23: Constructing the cell health state partial order lattice based on the binary tuple of each cell.

5. The method of Claim 4, wherein the method further comprises: The steps of running the cell thermal-electric state extended Kalman estimation algorithm include: S221: Constructing the state equation and observation equation of the cell thermal-electric state extended Kalman estimation algorithm; the state equation takes the state variables of the state of charge and the direct current resistance of the cell as the state variables; the input of the observation equation is the estimated value of the state variable, and the output is the predicted value of the voltage and the temperature; S222: Initializing the filtering parameters, specifically including: initializing the initial value of the state variable and the initial value of the error covariance matrix for each cell index corresponding to the cell in the algorithm input data set; S223: Performing the filtering prediction step; S224: Performing the filtering update step, specifically including: extracting the state of charge and the direct current resistance in the state variable update value of each time step, taking the average value as the final state of charge and direct current resistance of the cell, and combining the two to form the binary tuple of the cell.

6. The method of Claim 5, wherein the method further comprises: The steps of embedding the cell phase space point set with physical positions into a three-dimensional grid element set and constructing a joint risk density function include: S351: Matching the corresponding three-dimensional grid coordinates according to the physical position number in the cell phase space point with physical positions, and assigning the cell phase space point to the corresponding three-dimensional grid element to form a grid element set containing cell phase space points; S352: For each grid element containing a cell phase space point, a kernel density estimation method is used to construct a joint risk density function; the kernel density estimation method selects a Gaussian kernel function, and the bandwidth parameter is determined by a cross-validation method; taking each cell phase space point as the center, applying the Gaussian kernel function, and summing to obtain the joint risk density function value of each position to form a complete joint risk density function; S353: Associating each grid element with its joint risk density function to obtain a grid element set with joint risk density functions.

7. The method of Claim 6, wherein the method further comprises: The steps of performing local attractor region identification and early warning decision based on manifold analysis include: S41: Based on the thermal runaway risk propagation phase space grid, the cell state observation tensor with a time and space label, and the cell health state partial order lattice, extract and integrate the relevant data to obtain an analysis dataset; S42: Based on the analysis dataset, perform overall manifold analysis on the vector field in the thermal runaway risk propagation phase space grid to obtain a manifold feature atlas; the overall manifold analysis includes: dividing the vector field into local vector subfields according to the three-dimensional coordinates of the thermal runaway risk propagation phase space grid, each local vector subfield containing vector information of a preset number of adjacent grid units; calculating the continuity index of each local vector subfield, and screening out local vector subfields with a continuity index higher than a preset threshold; merging the screened local vector subfields to form continuous manifold segments, calculating the vector direction consistency in each segment, and identifying the segments with consistent vector directions and pointing to the core region as potential local attractor region candidates to form the manifold feature atlas; S43: Based on the manifold feature atlas and the analysis dataset, extract the potential local attractor region candidates and perform verification to obtain effective local attractor regions; S44: Based on the effective local attractor regions and the analysis dataset, locate the high-risk source candidate cells of thermal runaway; S45: Based on the effective local attractor regions and the analysis dataset, calculate the stability index of the effective local attractor regions; S46: Based on the stability index and the high-risk source candidate cells of thermal runaway, determine the warning level and output the results.

8. The method of Claim 7, wherein the method further comprises: The step of calculating the stability index of the effective local attractor region includes: S451: Extract the vector field data of all grid units in the effective local attractor region from the analysis dataset, calculate the divergence of the vector field of each grid unit, and take the average value as the overall divergence of the region; S452: Extract the vector field data of all grid units in the effective local attractor region from the analysis dataset, calculate the curl of the vector field of each grid unit, and take the average value as the overall curl of the region; S453: Normalize the overall divergence and overall curl respectively; S454: Weighted sum the normalized overall divergence and overall curl according to the preset weight to obtain the stability index.

9. A device for early warning and locating thermal runaway of a lithium-ion battery pack, for implementing the method of any one of claims 1-8, characterized in that, The device includes: A cell data acquisition and tensor construction unit for acquiring real-time voltage, temperature, current and number data of each cell and constructing a cell state observation tensor with a time and space label; A thermal-electric state estimation and partial order lattice generation unit for generating a cell health state partial order lattice based on the cell state observation tensor with a time and space label and running a cell thermal-electric state extended Kalman estimation algorithm; A risk propagation phase space grid construction unit for constructing a thermal runaway risk propagation phase space grid based on the cell health state partial order lattice; A positioning and early warning decision unit for performing local attractor region identification and early warning decision based on manifold analysis under the constraint of the thermal runaway risk propagation phase space grid.

Images