A battery thermal runaway isolation method and device based on risk prediction, electronic equipment and storage medium

Abstract

The application discloses a battery thermal runaway isolation method and device based on risk prediction, electronic equipment and storage medium, and belongs to the technical field of battery thermal safety management. The method comprises the following steps: acquiring temperature time series data, internal resistance time series data and gas concentration time series data of battery cells in a battery module; performing alignment processing on the multi-source data to generate a multi-modal alignment feature matrix; determining a thermal runaway risk probability of the battery cells based on a thermal state evolution prediction model; constructing a weighted undirected graph according to the heat conduction paths and the charge-discharge loops among the battery cells, and performing graph cut minimum cut partition calculation to determine an optimal isolation edge set; and cutting the heat conduction paths and the charge-discharge loops corresponding to the optimal isolation edge set. Therefore, by implementing the application, the problems that the existing technology has, i.e., the deviation of the preposed risk prediction caused by the time sequence dislocation of the sensing signals and the difficulty in quantifying the real physical topology connection relationship so as to fail to accurately calculate the appropriate isolation boundary to implement physical cutting, can be solved.

Description

Technical Field

[0001] This invention relates to the field of battery thermal safety management technology, specifically to a battery thermal runaway isolation method, device, electronic device, and storage medium based on risk prediction. Background Technology

[0002] Thermal safety management of electrochemical energy storage systems is a crucial foundation for their large-scale application. Battery thermal runaway isolation technology based on risk prediction can proactively interrupt the physical propagation path of high-risk cells in the early stages of thermal runaway evolution. This proactive isolation method not only effectively prevents chain-like thermal propagation and ensures the overall safety of the energy storage system, but also confines the fault to a localized area, avoiding large-scale system failure. Therefore, it has significant application value in battery safety protection systems.

[0003] However, current risk-prediction-based thermal runaway isolation technologies generally suffer from inaccurate isolation boundary calculations and susceptibility to deviations in blocking actions. This deficiency arises because existing technologies typically ignore the temporal misalignment caused by differences in sampling frequencies between different modal sensing signals such as temperature, internal resistance, and gas, leading to errors in the initial risk probability prediction. Furthermore, after identifying high-risk targets, existing systems often employ fixed physical partitions or empirical power-off strategies, failing to mathematically quantify and model the actual physical topology and thermal conductivity of the individual cells within the battery module. This makes it difficult for the system to calculate appropriate partition boundaries to cut off heat and charge propagation within complex thermal conduction networks, thus hindering effective physical disconnection of specific heat conduction paths and charging / discharging circuits. Summary of the Invention

[0004] This invention provides a battery thermal runaway isolation method, device, electronic device, and storage medium based on risk prediction, which can solve the problems in the prior art where the timing misalignment of sensing signals leads to deviations in the pre-risk prediction, and the difficulty in quantifying the real physical topology connection relationship, thus making it impossible to accurately calculate the appropriate isolation boundary for physical disconnection.

[0005] One embodiment of the present invention provides a battery thermal runaway isolation method based on risk prediction, comprising: Acquire the time-series data of temperature, internal resistance, and gas concentration of the cells in the battery module; Using the timestamps of the internal resistance time series data as a reference, interpolation and alignment operations are performed on the temperature time series data and the gas concentration time series data, and then standardized to generate a multimodal alignment feature matrix. The multimodal aligned feature matrix is ​​input into a pre-trained thermal state evolution prediction model, which outputs the probability of thermal runaway risk of the battery cell. Identify the heat conduction paths and charging / discharging circuits between cells in a battery module, and construct a weighted undirected graph with cells as nodes, heat conduction paths and charging / discharging circuits between cells as edges, and parameters characterizing the thermal isolation priority between adjacent cells as edge weights; wherein the thermal isolation priority is positively correlated with the thermal conduction risk between cells, and the edge weights are negatively correlated with the thermal conduction risk; Using cells whose thermal runaway risk probability exceeds a preset threshold as trigger nodes, graph cut minimum cut partitioning calculation is performed in a weighted undirected graph to output the optimal isolation edge set; Cut off the heat conduction path and charging / discharging circuit corresponding to the optimal isolation edge set.

[0006] Furthermore, each cell in the battery module is equipped with a distributed temperature sensor on its surface, a microelectromechanical system gas sensor array is arranged inside the battery module, a high-frequency Hall current sensor is independently deployed on the series branch of each cell, and a voltage sampling channel is deployed at both ends of each cell. Acquire time-series data of temperature, internal resistance, and gas concentration of the cells in the battery module, including: Receive temperature time-series data collected by the distributed temperature sensors; Receive gas concentration time-series data collected by the microelectromechanical system gas sensor array; With an AC excitation signal superimposed on the main bus of the battery module, the current data of each series branch collected by the high-frequency Hall current sensor and the voltage response data of each cell collected by the voltage sampling channel are received respectively. According to the preset frequency domain signal separation algorithm, the AC response current component of each cell branch is extracted from the current data. Based on the voltage response data and AC response current components of each cell, the independent AC internal resistance value of each cell is calculated as internal resistance timing data.

[0007] Furthermore, using the timestamps of the internal resistance time series data as a reference, interpolation and alignment operations are performed on the temperature time series data and the gas concentration time series data, followed by standardization, to generate a multimodal alignment feature matrix, including: Extract the timestamps from the internal resistance timing data and use them as the time alignment reference; Based on the time alignment benchmark, interpolation alignment is performed on the temperature time series data and the gas concentration time series data respectively to obtain the corresponding aligned temperature time series data and aligned gas concentration time series data. According to the preset sliding window, data for the corresponding time period are extracted from the aligned temperature time series data, aligned gas concentration time series data, and internal resistance time series data respectively, to obtain temperature window data sequence, gas concentration window data sequence, and internal resistance window data sequence. The mean and standard deviation of the temperature window data sequence, gas concentration window data sequence, and internal resistance window data sequence are calculated respectively to obtain the corresponding window mean and window standard deviation. Based on their respective window mean and window standard deviation, the temperature window data sequence, gas concentration window data sequence, and internal resistance window data sequence are standardized respectively to obtain temperature standardized feature sequence, gas concentration standardized feature sequence, and internal resistance standardized feature sequence. The temperature-normalized feature sequence, gas concentration-normalized feature sequence, and internal resistance-normalized feature sequence are spliced ​​together to generate a multimodal aligned feature matrix.

[0008] Furthermore, the thermal state evolution prediction model is trained in the following manner: Obtain the battery historical operation dataset; the battery historical operation dataset includes several historical training samples, each of which includes a historical multimodal aligned feature matrix and the corresponding real thermal runaway label; The historical multimodal aligned feature matrices from each historical training sample are sequentially input into the thermal state evolution prediction model for iterative training until the preset stopping iteration condition is reached, thus completing the training of the thermal state evolution prediction model. In each iteration, the thermal state evolution prediction model outputs prediction results based on the historical multimodal alignment feature matrix in the current historical training samples; the prediction results include the predicted probability of thermal runaway risk and the predicted temperature evolution parameters. Calculate the generated data distribution loss value based on the predicted probability of thermal runaway risk and the actual thermal runaway labels in the current historical training samples. The predicted temperature change gradient is extracted from the predicted temperature evolution parameters; Obtain the hidden state change parameters output by the thermal state evolution prediction model; Based on the predicted temperature change gradient and the hidden state change parameters, a physical constraint penalty loss value is generated. The data distribution loss value and the physical constraint penalty loss value are combined to generate a joint loss function value; The optimizer updates the network parameters in the thermal state evolution prediction model based on the joint loss function value.

[0009] Furthermore, the thermal conduction paths and charging / discharging circuits between cells in the battery module are identified. Using cells as nodes, the thermal conduction paths and charging / discharging circuits between cells as edges, and parameters representing the priority of thermal isolation between adjacent cells as edge weights, a weighted undirected graph is constructed, including: Obtain the device number for each battery cell; Based on the spatial adjacency of each battery cell, identify the heat conduction path between each battery cell; Based on the electrical connections between the cells, identify the charging and discharging circuits between the cells. Each battery cell is treated as a graph node, the corresponding device number is used as the node index of the graph node, the heat conduction path and the charging and discharging circuit are used as the edges connecting the corresponding graph nodes, and the parameters representing the thermal isolation priority calculated based on the preset maximum heat conduction power parameters between adjacent battery cells are used as the edge weights of the corresponding edges to generate a weighted undirected graph.

[0010] Furthermore, using cells with a thermal runaway risk probability exceeding a preset threshold as trigger nodes, a graph cut minimum cut partitioning calculation is performed in the weighted undirected graph to output the optimal isolation edge set, including: Cells with a thermal runaway risk probability exceeding a preset threshold are merged to generate a set of trigger nodes. Graph nodes other than the set of trigger nodes in the weighted undirected graph are extracted and merged to generate a set of non-trigger nodes. Using the set of triggering nodes as the source nodes and the set of non-triggering nodes as the sink nodes, a residual graph is generated based on the weighted undirected graph, the source nodes, and the sink nodes. Calculate the node height labels and edge preflow values ​​of the residual graph, and calculate the excess flow of the graph nodes in the residual graph based on the edge preflow values; Based on the node height label, excess traffic is pushed from the source node to the sink node, and the maximum flow value is calculated. Extract the set of minimum cut edges corresponding to the maximum flow value, and use the set of minimum cut edges as the optimal isolation edge set.

[0011] Furthermore, cutting off the heat conduction path and charging / discharging circuit corresponding to the optimal isolation edge set includes: Obtain the liquid cooling channel and charging / discharging circuit corresponding to the optimal isolation edge set; Identify the bidirectional solenoid valve associated with the liquid cooling channel and the electrical contactor associated with the charging and discharging circuit; Generate isolation control commands for driving the bidirectional solenoid valve to close and the electrical contactor to open; The bidirectional solenoid valve is driven to close to cut off the liquid cooling channel according to the isolation control command, and the electrical contactor is driven to open to cut off the charging and discharging circuit according to the isolation control command.

[0012] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0013] An embodiment of the present invention provides a battery thermal runaway isolation device based on risk prediction, comprising: a data acquisition module, a feature alignment module, a risk prediction module, a graph cut decision module, and an isolation execution module; The data acquisition module is used to acquire the time-series data of temperature, internal resistance, and gas concentration of the cells in the battery module. The feature alignment module is used to perform interpolation alignment operations and standardization processing on temperature time series data and gas concentration time series data based on the timestamp of internal resistance time series data, and generate a multimodal alignment feature matrix. The risk prediction module is used to input the multimodal aligned feature matrix into the pre-trained thermal state evolution prediction model and output the probability of thermal runaway risk of the battery cell. The graph cut decision module is used to identify the heat conduction paths and charging / discharging circuits between cells in the battery module. It constructs a weighted undirected graph using cells as nodes, the heat conduction paths and charging / discharging circuits between cells as edges, and parameters representing the thermal isolation priority between adjacent cells as edge weights. The thermal isolation priority is positively correlated with the heat conduction risk between cells, and the edge weights are negatively correlated with the heat conduction risk. Cells with a thermal runaway risk probability exceeding a preset threshold are used as trigger nodes. The module performs a minimum cut partition calculation in the weighted undirected graph and outputs the optimal isolation edge set. The isolation execution module is used to cut off the heat conduction path and charging / discharging circuit corresponding to the optimal isolation edge set.

[0014] Based on the above method embodiments, the present invention provides corresponding electronic device embodiments.

[0015] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the risk prediction-based battery thermal runaway isolation method described in any of the above-described method embodiments.

[0016] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments.

[0017] One embodiment of the present invention provides a storage medium storing a computer program thereon, wherein, when the computer program is running, it controls the device where the storage medium is located to execute any of the risk prediction-based battery thermal runaway isolation methods described in the above-described method embodiments.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a battery thermal runaway isolation method, device, electronic device, and storage medium based on risk prediction. The method acquires time-series data of temperature, internal resistance, and gas concentration from cells in a battery module; aligns and standardizes the temperature and gas concentration time-series data based on the internal resistance time-series data to generate a multimodal alignment feature matrix; inputs the multimodal alignment feature matrix into a thermal state evolution prediction model to obtain the thermal runaway risk probability of the cells; constructs a weighted undirected graph based on the heat conduction paths and charge / discharge circuits between cells, and uses cells with thermal runaway risk probabilities exceeding a preset threshold as trigger nodes to perform graph cut minimum partitioning calculations to determine the optimal isolation edge set; and cuts the heat conduction paths and charge / discharge circuits corresponding to the optimal isolation edge set.

[0019] The technical solution of this invention uses the timestamp of internal resistance timing data as a benchmark to perform interpolation and alignment on multimodal data, solving the technical problem of deviations in pre-risk prediction caused by timing misalignment of sensing signals in existing technologies. Simultaneously, this invention constructs a weighted undirected graph using the battery cell, its heat conduction path, and the charging / discharging circuit, and performs graph cut minimum cut partitioning calculation to obtain the optimal isolation edge set when a risk is triggered, thereby implementing physical severance. This approach overcomes the deficiency of traditional empirical partitioning in quantifying the true physical topological connections of the battery, and solves the technical problem of existing systems being unable to accurately calculate the segmentation boundary for physical severance. Attached Figure Description

[0020] Figure 1 This is a schematic flowchart of a battery thermal runaway isolation method based on risk prediction provided in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of a battery thermal runaway isolation device based on risk prediction, provided in an embodiment of the present invention. Detailed Implementation

[0022] 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.

[0023] like Figure 1 As shown, to address the problems in existing technologies where timing misalignment of sensor signals leads to deviations in pre-emptive risk prediction, and the difficulty in quantifying the actual physical topology connections makes it impossible to accurately calculate suitable isolation boundaries for physical disconnection, an embodiment of the present invention provides a battery thermal runaway isolation method based on risk prediction, comprising at least the following steps: Step S1: Obtain the time-series data of temperature, internal resistance, and gas concentration of the cells in the battery module; In a preferred embodiment, each cell in the battery module is provided with a distributed temperature sensor on its surface, a microelectromechanical system gas sensor array is provided inside the battery module, a high-frequency Hall current sensor is independently deployed on the series branch of each cell, and a voltage sampling channel is deployed at both ends of each cell. Acquire time-series data of temperature, internal resistance, and gas concentration of the cells in the battery module, including: Receive temperature time-series data collected by the distributed temperature sensors; Receive gas concentration time-series data collected by the microelectromechanical system gas sensor array; With an AC excitation signal superimposed on the main bus of the battery module, the current data of each series branch collected by the high-frequency Hall current sensor and the voltage response data of each cell collected by the voltage sampling channel are received respectively. According to the preset frequency domain signal separation algorithm, the AC response current component of each cell branch is extracted from the current data. Based on the voltage response data and AC response current components of each cell, the independent AC internal resistance value of each cell is calculated as internal resistance timing data.

[0024] Specifically, constructing a multimodal sensing network covering the entire battery module requires the deployment of various physical monitoring nodes. Distributed temperature sensors are arranged on the surface of each cell within the battery module, forming a three-dimensional grid that closely adheres to the cell surface and the internal gaps of the module. Each standard battery module has at least nine temperature measurement points to accurately capture the spatial location of hotspots and the direction of heat diffusion. A microelectromechanical system (MEMS) gas sensor array is also deployed inside the battery module, fixedly positioned at the top. A high-frequency Hall current sensor is independently deployed on the series branch of each cell, and voltage sampling channels are deployed at both ends of each cell, establishing an independent electrical monitoring topology.

[0025] Based on the deployed multimodal sensor network, temperature time-series data collected by distributed temperature sensors is received. The sampling frequency of the distributed temperature sensors is set to 1Hz, and the collected temperature time-series data is aggregated to generate a temperature field distribution matrix. The mathematical expression of the temperature field distribution matrix is: In the formula, The temperature field distribution matrix is... The number of cells in the battery module is represented by the three columns of the three-dimensional matrix, which correspond to the temperature values ​​on the X-axis, Y-axis, and Z-axis in three-dimensional space, respectively.

[0026] The system synchronously receives time-series gas concentration data acquired by a microelectromechanical system (MEMS) gas sensor array. The MEMS gas sensor array monitors the concentrations of carbon monoxide, hydrogen, and carbon dioxide in real time at a sampling frequency of 0.2 Hz. The monitoring range for carbon monoxide covers 0 to 1000 ppm, for hydrogen 0 to 2000 ppm, and for carbon dioxide 0 to 5000 ppm. The acquired gas concentration time-series data is aggregated to generate a gas concentration matrix. The mathematical expression for the gas concentration matrix is: In the formula, This is a gas concentration matrix. The number of cells in the battery module is represented by the three columns of the three-dimensional matrix, which correspond to the carbon monoxide concentration, hydrogen concentration, and carbon dioxide concentration values, respectively.

[0027] To monitor the evolution of electrochemical interface impedance, an AC excitation signal is superimposed on the main bus of the battery module. The AC excitation signal is a small-amplitude AC signal with a frequency of 1 kHz and an amplitude less than 5 mA. With the AC excitation signal superimposed on the main bus of the battery module, current data from each series branch collected by a high-frequency Hall current sensor and voltage response data from each cell collected by a voltage sampling channel are received. The sampling frequency of both the high-frequency Hall current sensor and the voltage sampling channel is set to 0.2 Hz. According to a preset frequency domain signal separation algorithm, the AC response current component of each cell's branch is extracted from the current data. The frequency domain signal separation algorithm relies on the branch current distribution model of parallel circuits to filter out low-frequency charging and discharging DC characteristics, polarization impedance interference, and environmental electromagnetic noise, accurately locking the true AC response current component at a specific excitation frequency. In a specific embodiment, the preset frequency domain signal separation algorithm is implemented using a fast Fourier transform algorithm combined with a bandpass filter. The center frequency of the bandpass filter is set to 1 kHz, the same as the AC excitation signal, and the bandwidth is set to ±50 Hz. The current data in the time domain is converted to the frequency domain using the Fast Fourier Transform algorithm, and the current amplitude and phase information at the 1kHz frequency point are extracted by using a bandpass filter, thereby reconstructing a high-purity AC response current component without polarization interference.

[0028] Based on the voltage response data and AC response current components of each cell, the independent AC internal resistance value of each cell is calculated as the internal resistance timing data. The independent AC internal resistance value is calculated by combining the three-dimensional grid voltage drop of the corresponding cell node and the AC response current components, and then aggregating them to generate an internal resistance vector. The mathematical expression for the internal resistance vector is: In the formula, Let the internal resistance vector be... Let represent the number of cells in the battery module, and let each element of the internal resistance vector be the independent AC internal resistance value of the corresponding cell.

[0029] The branch current decoupling calculation method eliminates the interference of AC impedance shunting between parallel branches on the measurement of the internal resistance of a single cell, and avoids the shunting calculation error caused by using the total load current. A multimodal sensing system incorporating physical measurement points and electrical decoupling logic is established, effectively eliminating measurement interference caused by dimensional differences and parallel circuit topology, providing high-purity and physically meaningful underlying time-series data support for subsequent capture of multidimensional signal collaborative variation characteristics. In specific application scenarios, battery modules often face strong electromagnetic interference and drastic temperature fluctuations. The distributed temperature sensor uses a thermocouple thin-film patch structure for encapsulation, and the microelectromechanical system gas sensor array is equipped with a PTFE dustproof, moisture-proof, and breathable membrane. The high-frequency Hall current sensor is equipped with a permalloy magnetic shield at the physical level to isolate the alternating magnetic field interference generated by the large current charging and discharging of adjacent branches. A hardware and software co-design combining physical shielding with a preset frequency domain signal separation algorithm completely ensures that the extracted AC response current component is not contaminated by harsh external environments. Furthermore, the extraction process of the AC response current component not only focuses on the current amplitude information, but also extracts the phase difference characteristics between the voltage response data and the current data. The phase difference characteristics can more sensitively reflect the microstructural damage at the solid electrolyte phase interface inside the cell, expanding the single-dimensional amplitude monitoring dimension to a two-dimensional joint amplitude and phase monitoring dimension, which significantly improves the physical information density of the underlying raw data.

[0030] Step S2: Using the timestamp of the internal resistance time series data as a reference, perform interpolation and alignment operations on the temperature time series data and the gas concentration time series data and perform standardization processing to generate a multimodal alignment feature matrix; In a preferred embodiment, using the timestamp of the internal resistance time series data as a reference, interpolation and alignment operations are performed on the temperature time series data and the gas concentration time series data, and then standardized to generate a multimodal alignment feature matrix, including: Extract the timestamps from the internal resistance timing data and use them as the time alignment reference; Based on the time alignment benchmark, interpolation alignment is performed on the temperature time series data and the gas concentration time series data respectively to obtain the corresponding aligned temperature time series data and aligned gas concentration time series data. According to the preset sliding window, data for the corresponding time period are extracted from the aligned temperature time series data, aligned gas concentration time series data, and internal resistance time series data respectively, to obtain temperature window data sequence, gas concentration window data sequence, and internal resistance window data sequence. The mean and standard deviation of the temperature window data sequence, gas concentration window data sequence, and internal resistance window data sequence are calculated respectively to obtain the corresponding window mean and window standard deviation. Based on their respective window mean and window standard deviation, the temperature window data sequence, gas concentration window data sequence, and internal resistance window data sequence are standardized respectively to obtain temperature standardized feature sequence, gas concentration standardized feature sequence, and internal resistance standardized feature sequence. The temperature-normalized feature sequence, gas concentration-normalized feature sequence, and internal resistance-normalized feature sequence are spliced ​​together to generate a multimodal aligned feature matrix.

[0031] Specifically, eliminating timing misalignments and dimensional differences between multimodal signals is a prerequisite for outputting the feature matrix. The timestamps of the internal resistance timing data are extracted and used as a time alignment reference. Because the sampling frequencies of the temperature field distribution matrix and the gas concentration matrix often differ from the sampling frequency of the internal resistance timing data, resulting in timestamp discrepancies, establishing a unified time axis as the time alignment reference is crucial.

[0032] Based on a time alignment benchmark, interpolation alignment is performed on temperature time series data and gas concentration time series data respectively to obtain corresponding aligned temperature time series data and aligned gas concentration time series data. Cubic spline interpolation is used for time series alignment. Cubic spline interpolation constructs a third-order polynomial between every two known sampling points, requiring adjacent segments to simultaneously satisfy three constraints at the connection points: continuity of function values, continuity of the first derivative, and continuity of the second derivative. This achieves frequency uniformity while maintaining signal smoothness and avoids inflection point distortion introduced by linear interpolation. The mathematical expression of the interpolation polynomial is: In the formula, For the first Segment cubic spline polynomial, For the interpolation time, For the first At a known sampling time, For the first Zero-order coefficients of a segment polynomial For the first First-order coefficients of the segment polynomial, For the first Second-order coefficients of the segment polynomial For the first Third-order coefficients of a segmental polynomial.

[0033] After time-series alignment is completed, data for corresponding time periods are extracted from the aligned temperature time-series data, aligned gas concentration time-series data, and internal resistance time-series data according to a preset sliding window, resulting in temperature window data sequences, gas concentration window data sequences, and internal resistance window data sequences. The sliding window operates by using a preset time length as the window length and a fixed time span as the sliding step size to extract continuous data segments across the entire time axis. In a specific embodiment, the preset time length (i.e., window length) is set to 10 minutes, and the fixed time span (i.e., sliding step size) is set to 10 seconds. Extracting data using this time scale preserves sufficient historical information about the thermal state evolution while ensuring high-frequency real-time updates of the feature matrix, preventing the omission of crucial, minute abrupt changes.

[0034] For the extracted data segments, the mean and standard deviation of the temperature window data sequence, gas concentration window data sequence, and internal resistance window data sequence are calculated respectively, yielding their respective window mean and standard deviation. Based on these window mean and standard deviation, the temperature window data sequence, gas concentration window data sequence, and internal resistance window data sequence are standardized respectively, resulting in standardized feature sequences for temperature, gas concentration, and internal resistance. The standardization process specifically employs a standardization deviation calculation method to eliminate dimensional differences between different physical quantities and local baseline drift. The mathematical expression of the standardization formula is: In the formula, For the first The first window Each feature at time The standardized value, For the first Each feature at time The original value, For the first The first window The mean of each feature, For the first The first window The standard deviation of each feature. (Subscript) It represents a specific feature or sensor device index, rather than a general mathematical summation index, thus ensuring that each standardized operation strictly corresponds to a unique physical monitoring channel.

[0035] During the standardization process, data points with deviations exceeding three standard deviations within the window are simultaneously removed as outliers to prevent impulse noise introduced by sensor malfunctions or electromagnetic interference from contaminating the subsequent output data matrix. Finally, the temperature-standardized feature sequence, gas concentration-standardized feature sequence, and internal resistance-standardized feature sequence are concatenated to generate a multimodal aligned feature matrix. Performing interpolation alignment operations and standardization processing can completely eliminate the barriers in time and dimension of multimodal heterogeneous sensing signals, laying a high-quality data foundation for the subsequent network model to accurately extract local physical features and cross-modal global temporal correlations. The generated temperature-standardized feature sequence, gas concentration-standardized feature sequence, and internal resistance-standardized feature sequence, after concatenation, constitute the multimodal aligned feature matrix. The multimodal aligned feature matrix is ​​presented in the form of a high-order tensor in data structure, with each dimension of the high-order tensor mapping the physical time sliding step, sensor spatial node index, and multimodal sensing physical quantity channels, respectively. By performing rigorous standardization, temperature (Celsius), gas concentration (parts per million), and internal resistance (milliohms), which originally had different physical dimensions, are uniformly mapped to a dimensionless continuous mapping space with zero mean and a variance of one. The construction of this dimensionless continuous mapping space effectively prevents feature channels with large absolute values ​​from being overwhelmed by feature channels with smaller absolute values ​​but containing crucial abrupt changes during backpropagation gradient calculation. Simultaneously, removing data points with deviations exceeding three standard deviations within the window effectively constructs a low-level data cleaning barrier, physically intercepting fly-spot data caused by transient sensor contact problems due to external mechanical vibrations.

[0036] Step S3: Input the multimodal aligned feature matrix into the pre-trained thermal state evolution prediction model and output the probability of thermal runaway risk of the battery cell. In a preferred embodiment, the thermal state evolution prediction model is trained in the following manner: Obtain the battery historical operation dataset; the battery historical operation dataset includes several historical training samples, each of which includes a historical multimodal aligned feature matrix and the corresponding real thermal runaway label; The historical multimodal alignment feature matrices from each historical training sample are sequentially input into the thermal state evolution prediction model for iterative training until a preset stopping condition is reached, thus completing the training of the thermal state evolution prediction model. In a specific embodiment, the base network of the thermal state evolution prediction model is constructed using a one-dimensional residual network (1D-ResNet). The preset stopping condition is set as follows: the joint loss function value on the independent validation set no longer decreases for 20 consecutive iterations, or the maximum number of iterations reaches 200. This explicit stopping condition effectively avoids overfitting during model training.

[0037] In each iteration, the thermal state evolution prediction model outputs prediction results based on the historical multimodal alignment feature matrix in the current historical training samples; the prediction results include the predicted probability of thermal runaway risk and the predicted temperature evolution parameters. Calculate the generated data distribution loss value based on the predicted probability of thermal runaway risk and the actual thermal runaway labels in the current historical training samples. The predicted temperature change gradient is extracted from the predicted temperature evolution parameters; Obtain the hidden state change parameters output by the thermal state evolution prediction model; Based on the predicted temperature change gradient and the hidden state change parameters, a physical constraint penalty loss value is generated. The data distribution loss value and the physical constraint penalty loss value are combined to generate a joint loss function value; The optimizer updates the network parameters in the thermal state evolution prediction model based on the joint loss function value.

[0038] Specifically, the multimodal alignment feature matrix is ​​input into a pre-trained thermal state evolution prediction model, which outputs the probability of thermal runaway risk of the battery cell. This requires pre-training the network parameters of the thermal state evolution prediction model. A historical battery operation dataset is acquired, comprising several historical training samples, each including a historical multimodal alignment feature matrix and a corresponding real thermal runaway label. To augment the scarce positive thermal runaway samples, an electrochemical thermal coupling simulation model is used to generate synthetic time-series samples covering runaway modes such as overcharging, internal short circuits, and external heating. This data is then augmented using a conditional generative adversarial network (GAN) to ensure that the ratio of positive to negative samples in the historical training samples reaches a preset equilibrium standard. In one specific embodiment, the preset equilibrium standard is set to a ratio of 1:3 to 1:1 between the number of positive thermal runaway samples and the number of negative samples under normal operating conditions in the historical training samples. In a preferred embodiment, the output stopping condition of the GAN is controlled to ensure that the ratio of positive to negative samples is strictly 1:1, thereby effectively eliminating the gradient bias problem in the backpropagation process of the thermal state evolution prediction model.

[0039] The historical multimodal alignment feature matrices from each historical training sample are sequentially input into the thermal state evolution prediction model for iterative training until a preset stopping condition is met, thus completing the training of the thermal state evolution prediction model. In each network iteration, the thermal state evolution prediction model outputs a prediction result based on the historical multimodal alignment feature matrices in the current historical training samples. The prediction result fully includes the predicted thermal runaway risk probability and the predicted temperature evolution parameters. A data distribution loss value is calculated based on the predicted thermal runaway risk probability and the actual thermal runaway labels in the current historical training samples. The calculation of the data distribution loss value relies on the focus loss mechanism, which is mainly used to apply larger network optimization weights to a minority of positive thermal runaway samples. The mathematical expression for the data distribution loss value is: In the formula, This represents the data distribution loss value. The total number of historical training samples input in a single iteration; Pre-defined minority class sample weights; To focus on adjusting parameters; For the first Real thermal runaway labels for historical training samples; For the first The predicted thermal runaway risk probability is calculated from historical training samples. In a specific embodiment, to effectively address the extreme imbalance between positive thermal runaway samples and negative samples under normal operating conditions, a pre-set minority class sample weight is used. The value range is set to 5 to 10 (for example, preferably 7), focusing on the adjustment parameter. Set to 2.

[0040] At the level of physical constraints, the predicted temperature change gradient is analyzed from the parameters of predicted temperature evolution. The differential form of the Arrhenius formula reveals the physical law that the reaction rate monotonically accelerates with increasing temperature, and its mathematical expression is: In the formula, The reaction rate constant; Absolute temperature; It is the activation energy; This is the universal gas constant.

[0041] To embed the prior knowledge that temperature evolution and reaction rate evolution must move in the same direction, as required by the aforementioned rules, into the model, a corresponding physical constraint penalty loss value was constructed. Specifically, the rate of change of the latent state norm in the internal calculation output of the thermal state evolution prediction model is extracted as a latent state change parameter. The physical constraint penalty loss value requires that the rate of change of the latent state norm output by the thermal state evolution prediction model maintain a high degree of consistency with the evolution sign of the predicted temperature change gradient, and applies a differentiable penalty constraint for abnormal deviations in sign. The mathematical expression for the physical constraint penalty loss value is: In the formula, This represents the physical constraint penalty loss value. Standard symbolic functions; It is a linear rectification activation function; To predict temperature gradient; This is the rate of change of the implicit state norm calculated internally by the model.

[0042] Finally, the data distribution loss value and the physical constraint penalty loss value are fused to generate a joint loss function value. The optimizer then updates the network parameters in the thermal state evolution prediction model based on the joint loss function value. The network parameter update stage employs an adaptive moment estimation optimization algorithm combined with a cosine annealing strategy to dynamically adjust the learning rate, guiding the network parameters to converge stably towards the optimal solution. By deeply integrating the multimodal data feature distribution patterns with the underlying chemical reaction kinetics physical boundary, the erroneous extrapolation evolution direction of the thermal state evolution prediction model is effectively limited when positive samples are extremely scarce, significantly improving the algorithm's risk prediction accuracy and generalization ability in various extreme boundary conditions. For the iterative optimization process of the thermal state evolution prediction model, the introduction of a physical constraint penalty loss value fundamentally changes the optimization convergence trajectory of a purely data-driven black-box model. Traditional deep learning networks, when faced with extremely imbalanced historical training samples, are prone to falling into the illusion of local optima where all battery cells are predicted as safe. By fusing the data distribution loss value and the physical constraint penalty loss value to generate a joint loss function, the optimizer is forced to simultaneously satisfy the external data statistical distribution fitting requirement and the internal chemical reaction kinetics monotonically increasing requirement when updating network parameters. Specifically, when the hidden state output of the thermal state evolution prediction model shows an abnormal downward trend, but the extracted predicted temperature change gradient shows a continuous upward trend, the physical constraint penalty loss value will be exponentially amplified, transmitting a very large penalty gradient to the optimizer. After receiving the penalty gradient, the optimizer forcibly pulls the shifted network weights back into a reasonable parameter space that conforms to the Arrhenius chemical kinetic formula. Through repeated joint iterative training, the final output thermal state evolution prediction model not only memorizes the thermal runaway feature classification boundaries in historical data, but also internalizes the underlying physical mechanism of the thermal runaway chain reaction of battery materials, ensuring that it can still output a conservative thermal runaway risk probability prediction value that does not violate physical common sense when facing unknown abnormal operating conditions.

[0043] Step S4: Identify the heat conduction paths and charging / discharging circuits between cells in the battery module, and construct a weighted undirected graph using cells as nodes, heat conduction paths and charging / discharging circuits between cells as edges, and parameters characterizing the thermal isolation priority between adjacent cells as edge weights; wherein the thermal isolation priority is positively correlated with the thermal conduction risk between cells, and the edge weights are negatively correlated with the thermal conduction risk; using cells whose thermal runaway risk probability exceeds a preset threshold as trigger nodes, perform graph cut minimum partitioning calculation in the weighted undirected graph, and output the optimal isolation edge set; In a preferred embodiment, the heat conduction paths and charging / discharging circuits between cells in the battery module are identified. A weighted undirected graph is constructed, using cells as nodes, heat conduction paths and charging / discharging circuits between cells as edges, and parameters characterizing the priority of thermal isolation between adjacent cells as edge weights. Obtain the device number for each battery cell; Based on the spatial adjacency of each battery cell, identify the heat conduction path between each battery cell; Based on the electrical connections between the cells, identify the charging and discharging circuits between the cells. Each battery cell is treated as a graph node, the corresponding device number is used as the node index of the graph node, the heat conduction path and the charging and discharging circuit are used as the edges connecting the corresponding graph nodes, and the parameters representing the thermal isolation priority calculated based on the preset maximum heat conduction power parameters between adjacent battery cells are used as the edge weights of the corresponding edges to generate a weighted undirected graph.

[0044] In a preferred embodiment, a cell with a thermal runaway risk probability exceeding a preset threshold is used as a trigger node. A minimum cut partitioning calculation is performed on the weighted undirected graph to output the optimal set of isolated edges, including: Cells with a thermal runaway risk probability exceeding a preset threshold are merged to generate a set of trigger nodes. Graph nodes other than the set of trigger nodes in the weighted undirected graph are extracted and merged to generate a set of non-trigger nodes. Using the set of triggering nodes as the source nodes and the set of non-triggering nodes as the sink nodes, a residual graph is generated based on the weighted undirected graph, the source nodes, and the sink nodes. Calculate the node height labels and edge preflow values ​​of the residual graph, and calculate the excess flow of the graph nodes in the residual graph based on the edge preflow values; Based on the node height label, excess traffic is pushed from the source node to the sink node, and the maximum flow value is calculated. Extract the set of minimum cut edges corresponding to the maximum flow value, and use the set of minimum cut edges as the optimal isolation edge set.

[0045] Specifically, identifying the heat conduction paths and charging / discharging circuits between the cells in the battery module and constructing the topology network requires first obtaining the device number of each cell. Based on the spatial adjacency between cells, the heat conduction paths between different cells are identified. Simultaneously, based on the electrical connections between cells, the charging / discharging circuits formed by all cells are identified. Each cell is treated as a graph node, and the corresponding device number is directly used as the node index. The heat conduction paths and charging / discharging circuits are used as edges connecting the corresponding graph nodes. To accurately quantify the heat transfer patterns and enable the graph cut algorithm to prioritize cutting high-risk paths, a parameter representing the priority of thermal isolation between adjacent cells is introduced as an edge weight. The thermal isolation priority is positively correlated with the heat conduction risk between cells, and the edge weight is negatively correlated with the heat conduction risk. In a specific embodiment, a preset maximum heat conduction power parameter between adjacent cells (which positively represents the heat conduction risk) is calibrated based on the physical parameters of the thermally conductive structural adhesive actually filled within the battery module. Specifically, the thermal conductivity of the thermally conductive structural material is set to 2.5 Kelvin per meter, the contact area between adjacent cells is set to 0.05 square meters, the physical distance between adjacent cells is set to 0.002 meters, and the maximum allowable temperature difference is set to 50 Kelvin. Based on the above-set physical constants, the maximum heat transfer power parameter between adjacent cells is calculated. To meet the algorithm requirement that edge weights are negatively correlated with heat transfer risk, and to ensure that the minimum cut partition can accurately cut off the path with the maximum heat transfer power, the reciprocal of this maximum heat transfer power parameter is taken as the edge weight of the final weighted undirected graph, ensuring that the edge weights of the weighted undirected graph have real physical meaning. The mathematical expression for the edge weight is: In the formula, For nodes With nodes The dimensionless edge weights between them are reciprocals of the maximum thermal conductivity parameter; This refers to the thermal conductivity of the heat-conducting material between the corresponding battery cells; This refers to the contact area between adjacent battery cells; This refers to the physical distance between adjacent battery cells. This is the maximum permissible temperature difference set. This is a normalized reference temperature, taken as the rated operating temperature of the battery module. (Subscript) and It strictly represents the specific cell device number index in three-dimensional physical space, and is by no means a general mathematical summation index, so as to ensure that the constructed network topology is completely mapped to the actual physical structure of the battery module.

[0046] Based on a given set of graph nodes, connecting edges, and edge weights, a weighted undirected graph is generated. The mathematical expression for a weighted undirected graph is: In the formula, Represents a weighted undirected graph; Represents the set of graph nodes; It represents the set of edges connecting the nodes in a graph; This represents the set of edge weights for connecting edges. In a weighted undirected graph, the core objective of minimum cut partitioning computation is to find the set of cut edges with the minimum total weight. The mathematical expression for the optimization objective function is: In the formula, To find the optimal set of isolated edges; Let be the set of candidate cut edges.

[0047] To find the optimal isolation edge set, cells with a thermal runaway risk probability exceeding a preset threshold are merged to generate a trigger node set. In a specific embodiment, the preset threshold is set between 0.90 and 0.98 based on the safety level requirements of the energy storage system, for example, 0.98. When the thermal runaway risk probability of any cell exceeds 0.98, the system strictly classifies it as an extremely high-risk state and uses it as a trigger node, immediately initiating the underlying graph cut isolation calculation. Graph nodes other than the trigger node set are extracted from the weighted undirected graph and merged to generate a non-trigger node set. The trigger node set is used as the source node, and the non-trigger node set is used as the sink node. The push-pull labeling algorithm is then used to find the optimal isolation edge set. The residual graph is generated based on the weighted undirected graph, source nodes, and sink nodes. Specifically, this includes: using all graph nodes in the weighted undirected graph as graph nodes in the residual graph; for each edge in the weighted undirected graph, constructing a forward edge in the residual graph corresponding to the same connection relationship, and assigning the capacity of the forward edge to the corresponding edge weight; and simultaneously constructing a reverse edge in the residual graph corresponding to the same connection relationship, and assigning the capacity of the reverse edge to zero.

[0048] After the residual graph is generated, the node height labels and edge preflow values ​​are calculated. The edge preflow value refers to the process flow allocation value on the edge when the flow into a certain intermediate node is allowed to be greater than the flow out of that node during the push-pull labeling algorithm. The specific calculation method for the edge preflow value is as follows: Obtain the initial edge preflow value, initial node height label, and initial excess flow of the residual graph. Specifically, assign the initial node height label of the source node to the total number of graph nodes in the residual graph, and assign the initial node height label of all other graph nodes to zero. For all forward edges directly emanating from the source node in the residual graph, assign their initial edge preflow value to the capacity of that forward edge. For all other edges in the residual graph, assign their initial edge preflow value to zero. Calculate the initial excess flow of each graph node based on the initial edge preflow value. At this point, the initial edge preflow value satisfies a strict capacity constraint condition. The mathematical expression for the capacity constraint condition is: In the formula, For nodes in the residual graph To the node The pre-flow value of the pushed edges is used to calculate the excess flow of the graph nodes in the residual graph. The mathematical expression for excess flow is: In the formula, For the current graph node in the residual graph Excess traffic; For graph nodes Connected adjacent nodes; the first item represents the node flowing into the graph. The sum of all preflow values; the second term represents the sum of values ​​from the graph nodes. The sum of all outflowing preflow values.

[0049] Repeat the pre-flow push and relabeling iteration operations until the residual graph has no graph nodes whose current excess flow is greater than zero, except for the source node and sink node. The target edge preflow value generated when the iteration stops is determined as the maximum flow allocation state of the network, and the total flow out of the source node is extracted as the maximum flow value of the network based on the maximum flow allocation state. The pre-stream push and relabeling iteration operations include: For the current graph nodes in the residual graph that, excluding the source and sink nodes, have a current excess flow greater than zero, determine and execute the calculation branch to generate the target edge preflow value, the target node height label, and the target excess flow; wherein, the initial edge preflow value is the initial edge preflow value; the initial node height label is the initial node height label; and the initial excess flow is the initial excess flow. If there exists an adjacent node connected to the current graph node, and the remaining available capacity of the forward edge from the current graph node to the adjacent node is greater than zero, and the current node height label of the current graph node is strictly equal to the current node height label of the adjacent node plus one, then the minimum value between the current excess flow and the remaining available capacity of the aforementioned forward edge is taken as the single flow update value. The current edge preflow value of the aforementioned forward edge is added to the single flow update value, and the current edge preflow value of the corresponding reverse edge is subtracted from the single flow update value to generate the target edge preflow value. The target excess flow is recalculated based on the target edge preflow value, and the current node height label is taken as the target node height label. If for the current graph node, the current node height label of all adjacent nodes pointed to by the positive edges with remaining available capacity is greater than or equal to the current node height label of the current graph node, then the minimum value among the current node height labels of the aforementioned adjacent nodes is selected, and the minimum value is added by one to generate the target node height label. The current edge preflow value and the current excess flow are respectively used as the target edge preflow value and the target excess flow. Update the target edge preflow value, target node height label, and target excess flow to the current edge preflow value, current node height label, and current excess flow, respectively.

[0050] The minimum cut edge set corresponding to the maximum flow value is extracted, and this minimum cut edge set is ultimately output as the optimal isolation edge set. The abstract physical network of heat propagation and electrical connection is transformed into an optimization mathematical model with capacity constraints to solve for the optimal cut set. This accurately quantifies the real topological connection relationship and rapidly calculates the optimal physical boundary for cutting off heat and energy propagation in polynomial time, completely avoiding the defects of insufficient isolation or over-response caused by artificially solidified partitions. Performing graph cut minimum cut partition calculation in a weighted undirected graph essentially finds the optimal physical topological breakpoint to block the mutual propagation of heat and electrical energy. The set of trigger nodes is used as the source node, representing the central source of the violent energy burst of thermal runaway; the set of non-trigger nodes is used as the sink node, representing the normal cell group that needs to be protected. The edge pre-current value physically represents the priority barrier of thermoelectric isolation between adjacent cells, and it is negatively correlated with the sum of the maximum heat and electricity actually transferred per unit time. The process of pushing excess flow from the source node to the sink node according to the node height label perfectly simulates the physical evolution process of heat conduction and diffusion from high-temperature nodes to low-temperature nodes in the three-dimensional physical space of the battery module. When the excess flow of a node in the residual graph reaches saturation and can no longer push pre-flow values ​​to the sink node, the weighted undirected graph reaches the maximum flow state. The physical meaning of the maximum flow value is the overall thermal isolation bottleneck resistance of the battery module under the current topology. Extracting the minimum cut edge set corresponding to the maximum flow value precisely identifies those physical connections that, once cut, can achieve the maximum energy blocking effect with minimal hardware action. Outputting the optimal isolation edge set based on real-time physical state parameters maximizes the continuous working capability of healthy cells and avoids the crude response mode of directly disconnecting the entire battery cluster due to the failure of a single cell in traditional strategies.

[0051] Step S5: Cut off the heat conduction path and charging / discharging circuit corresponding to the optimal isolation edge set.

[0052] In a preferred embodiment, cutting off the heat conduction path and charging / discharging circuit corresponding to the optimal isolation edge set includes: Obtain the liquid cooling channel and charging / discharging circuit corresponding to the optimal isolation edge set; Identify the bidirectional solenoid valve associated with the liquid cooling channel and the electrical contactor associated with the charging and discharging circuit; Generate isolation control commands for driving the bidirectional solenoid valve to close and the electrical contactor to open; The bidirectional solenoid valve is driven to close to cut off the liquid cooling channel according to the isolation control command, and the electrical contactor is driven to open to cut off the charging and discharging circuit according to the isolation control command.

[0053] Specifically, after obtaining the optimal isolation edge set, it is necessary to accurately map and disconnect the heat conduction path and charging / discharging circuit corresponding to the optimal isolation edge set. In terms of execution, the liquid cooling channel and charging / discharging circuit corresponding to the optimal isolation edge set are first obtained.

[0054] Based on the established physical connection network, the specific hardware actuators are located, including the bidirectional solenoid valves associated with the liquid cooling channels and the electrical contactors associated with the charging and discharging circuits. The bidirectional solenoid valves are deployed at the branch nodes of each liquid cooling microchannel within the battery module, responsible for controlling the physical circulation of the convective heat transfer medium; the electrical contactors are deployed on the series branches of each cell, responsible for controlling the continuous input of charging and discharging energy.

[0055] After accurately locating the specific hardware actuator, the system proceeds to the control command generation stage, generating isolation control commands to drive the bidirectional solenoid valve to close and the electrical contactor to open. Encapsulating the specific closing and opening actions directly into the command definition ensures that the isolation control commands are targeted clearly and that the actions are legitimate, avoiding conflicts between multiple control signals.

[0056] Upon receiving the isolation control command, the hardware actuator immediately responds, driving the bidirectional solenoid valve to close to cut off the liquid cooling channel, and simultaneously driving the electrical contactor to open to cut off the charging and discharging circuit.

[0057] By directly mapping the optimal isolation edge set output by the algorithm to the valve and switching action commands of the front-end hardware, the convection mass transfer path and energy supply channel are completely cut off at the physical level, achieving a real and efficient thermoelectric dual physical isolation effect for high-risk fault areas. The generation and execution process of the isolation control command are configured with strict millisecond-level timing logic constraints. After generating the isolation control commands to drive the bidirectional solenoid valve to close and the electrical contactor to open, the underlying hardware execution module independently distributes the isolation control commands to each hardware execution mechanism through the controller area network bus. Among them, the opening action of the electrical contactor is set with the highest action priority, requiring the complete separation of the physical contacts within 50 milliseconds after receiving the isolation control command, quickly cutting off the continuous short-circuit current or overcharge current, and cutting off the continuous source of Joule internal heat generation. The closing action of the bidirectional solenoid valve is executed immediately afterward, requiring the valve to be mechanically locked within 200 milliseconds, confining the high-temperature heat-conducting medium within the physical range defined by the minimum cut edge set, and preventing the high-temperature liquid cooling medium from flowing back along the liquid cooling channel and scouring the healthy battery cell. The precise timing coordination between the physical power cut-off action and the physical thermal blockade action effectively limits potential catastrophic thermal runaway events to a very limited number of triggering nodes, thus constructing a battery module thermal runaway isolation and defense system with extremely high safety redundancy.

[0058] To further illustrate the working process of the method of this invention, the following is a complete application scenario example: Taking a lithium iron phosphate battery module of an energy storage power station as an example, the system continuously collects multimodal raw data of the cells. After time-series alignment processing, a multimodal feature matrix is ​​generated using a 10-minute sliding window and a 10-second sliding step size. After inputting a pre-trained one-dimensional residual network model, the system outputs the thermal runaway risk probability of each cell in real time. When the risk probability of cell No. 3 in the battery module reaches 0.99 and exceeds the preset threshold of 0.98, the system immediately uses cell No. 3 as the trigger node. Based on the actual physical topology of the battery module, a graph cut minimum cut calculation is performed, and the optimal isolation edge set is output within 100 milliseconds. Subsequently, control commands are directly issued to drive the corresponding bidirectional solenoid valve to close to cut off the liquid cooling channel in the area where cell No. 3 is located, and simultaneously drive the corresponding electrical contactor to open to cut off its charging and discharging circuit, successfully limiting the thermal runaway risk to the smallest physical partition and completely blocking the chain thermal propagation.

[0059] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0060] like Figure 2 As shown, an embodiment of the present invention provides a battery thermal runaway isolation device based on risk prediction, including: a data acquisition module 101, a feature alignment module 102, a risk prediction module 103, a graph cut decision module 104, and an isolation execution module 105; The data acquisition module 101 is used to acquire the temperature time-series data, internal resistance time-series data, and gas concentration time-series data of the cells in the battery module. The feature alignment module 102 is used to perform interpolation alignment operations and standardization processing on temperature time series data and gas concentration time series data based on the timestamp of internal resistance time series data to generate a multimodal alignment feature matrix. The risk prediction module 103 is used to input the multimodal aligned feature matrix into the pre-trained thermal state evolution prediction model and output the probability of thermal runaway risk of the battery cell. The graph cut decision module 104 is used to identify the heat conduction paths and charging / discharging circuits between cells in the battery module, and to construct a weighted undirected graph with cells as nodes, heat conduction paths and charging / discharging circuits between cells as edges, and parameters representing the thermal isolation priority between adjacent cells as edge weights. The thermal isolation priority is positively correlated with the thermal conduction risk between cells, and the edge weights are negatively correlated with the thermal conduction risk. Cells with a thermal runaway risk probability exceeding a preset threshold are used as trigger nodes. The graph cut minimum partition calculation is performed in the weighted undirected graph to output the optimal isolation edge set. The isolation execution module 105 is used to cut off the heat conduction path and charging / discharging circuit corresponding to the optimal isolation edge set.

[0061] It should be noted that the embodiments of the device described above correspond to the embodiments of the present invention described above, and can realize the battery thermal runaway isolation method based on risk prediction described in any one of the above embodiments of the present invention. Furthermore, the embodiments of the device described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without creative effort.

[0062] Based on the above-described method embodiments of the present invention, a corresponding embodiment of an electronic device is provided.

[0063] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the battery thermal runaway isolation method based on risk prediction as described in any one of the present invention, or the processor executes the computer program to implement the functions of each module in the above-described device embodiments.

[0064] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the terminal device.

[0065] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0066] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0067] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD card), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0068] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments; Another embodiment of the present invention provides a storage medium including a stored computer program, wherein, when the computer program is executed, the device where the storage medium is located executes any of the above-described risk prediction-based battery thermal runaway isolation methods of the present invention.

[0069] The aforementioned storage medium is a computer-readable storage medium, and the computer program includes computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0070] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0071] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A battery thermal runaway isolation method based on risk prediction, characterized in that, include: Acquire the time-series data of temperature, internal resistance, and gas concentration of the cells in the battery module; Using the timestamps of the internal resistance time series data as a reference, interpolation and alignment operations are performed on the temperature time series data and the gas concentration time series data, and then standardized to generate a multimodal alignment feature matrix. The multimodal aligned feature matrix is ​​input into a pre-trained thermal state evolution prediction model, which outputs the probability of thermal runaway risk of the battery cell. Identify the heat conduction paths and charging / discharging circuits between cells in a battery module, and construct a weighted undirected graph with cells as nodes, heat conduction paths and charging / discharging circuits between cells as edges, and parameters characterizing the thermal isolation priority between adjacent cells as edge weights; wherein the thermal isolation priority is positively correlated with the thermal conduction risk between cells, and the edge weights are negatively correlated with the thermal conduction risk; Using cells whose thermal runaway risk probability exceeds a preset threshold as trigger nodes, graph cut minimum cut partitioning calculation is performed in a weighted undirected graph to output the optimal isolation edge set; Cut off the heat conduction path and charging / discharging circuit corresponding to the optimal isolation edge set.

2. The battery thermal runaway isolation method based on risk prediction as described in claim 1, characterized in that, Each cell in the battery module is equipped with a distributed temperature sensor on its surface, a microelectromechanical system gas sensor array is arranged inside the battery module, a high-frequency Hall current sensor is independently deployed on the series branch of each cell, and a voltage sampling channel is deployed at both ends of each cell. Acquire time-series data of temperature, internal resistance, and gas concentration of the cells in the battery module, including: Receive temperature time-series data collected by the distributed temperature sensors; Receive gas concentration time-series data collected by the microelectromechanical system gas sensor array; With an AC excitation signal superimposed on the main bus of the battery module, the current data of each series branch collected by the high-frequency Hall current sensor and the voltage response data of each cell collected by the voltage sampling channel are received respectively. According to the preset frequency domain signal separation algorithm, the AC response current component of each cell branch is extracted from the current data. Based on the voltage response data and AC response current components of each cell, the independent AC internal resistance value of each cell is calculated as internal resistance timing data.

3. The battery thermal runaway isolation method based on risk prediction as described in claim 2, characterized in that, Using the timestamps of the internal resistance time series data as a reference, interpolation and alignment operations are performed on the temperature time series data and the gas concentration time series data, followed by standardization, to generate a multimodal alignment feature matrix, including: Extract the timestamps from the internal resistance timing data and use them as the time alignment reference; Based on the time alignment benchmark, interpolation alignment is performed on the temperature time series data and the gas concentration time series data respectively to obtain the corresponding aligned temperature time series data and aligned gas concentration time series data. According to the preset sliding window, data for the corresponding time period are extracted from the aligned temperature time series data, aligned gas concentration time series data, and internal resistance time series data respectively, to obtain temperature window data sequence, gas concentration window data sequence, and internal resistance window data sequence. The mean and standard deviation of the temperature window data sequence, gas concentration window data sequence, and internal resistance window data sequence are calculated respectively to obtain the corresponding window mean and window standard deviation. Based on their respective window mean and window standard deviation, the temperature window data sequence, gas concentration window data sequence, and internal resistance window data sequence are standardized respectively to obtain temperature standardized feature sequence, gas concentration standardized feature sequence, and internal resistance standardized feature sequence. The temperature-normalized feature sequence, gas concentration-normalized feature sequence, and internal resistance-normalized feature sequence are spliced ​​together to generate a multimodal aligned feature matrix.

4. The battery thermal runaway isolation method based on risk prediction as described in claim 3, characterized in that, The thermal state evolution prediction model is trained using the following method: Obtain the battery historical operation dataset; the battery historical operation dataset includes several historical training samples, each of which includes a historical multimodal aligned feature matrix and the corresponding real thermal runaway label; The historical multimodal aligned feature matrices from each historical training sample are sequentially input into the thermal state evolution prediction model for iterative training until the preset stopping iteration condition is reached, thus completing the training of the thermal state evolution prediction model. In each iteration, the thermal state evolution prediction model outputs prediction results based on the historical multimodal alignment feature matrix in the current historical training samples; the prediction results include the predicted probability of thermal runaway risk and the predicted temperature evolution parameters. Calculate the generated data distribution loss value based on the predicted probability of thermal runaway risk and the actual thermal runaway labels in the current historical training samples. The predicted temperature change gradient is extracted from the predicted temperature evolution parameters; Obtain the hidden state change parameters output by the thermal state evolution prediction model; Based on the predicted temperature change gradient and the hidden state change parameters, a physical constraint penalty loss value is generated. The data distribution loss value and the physical constraint penalty loss value are combined to generate a joint loss function value; The optimizer updates the network parameters in the thermal state evolution prediction model based on the joint loss function value.

5. The battery thermal runaway isolation method based on risk prediction as described in claim 4, characterized in that, Identify the heat conduction paths and charging / discharging circuits between cells in the battery module, and construct a weighted undirected graph using cells as nodes, heat conduction paths and charging / discharging circuits as edges, and parameters representing the priority of thermal isolation between adjacent cells as edge weights. This includes: Obtain the device number for each battery cell; Based on the spatial adjacency of each battery cell, identify the heat conduction path between each battery cell; Based on the electrical connections between the cells, identify the charging and discharging circuits between the cells. Each battery cell is treated as a graph node, the corresponding device number is used as the node index of the graph node, the heat conduction path and the charging and discharging circuit are used as the edges connecting the corresponding graph nodes, and the parameters representing the thermal isolation priority calculated based on the preset maximum heat conduction power parameters between adjacent battery cells are used as the edge weights of the corresponding edges to generate a weighted undirected graph.

6. The battery thermal runaway isolation method based on risk prediction as described in claim 5, characterized in that, Using cells with a thermal runaway risk probability exceeding a preset threshold as trigger nodes, graph cut minimum cut partitioning calculation is performed in a weighted undirected graph, outputting the optimal set of isolated edges, including: Cells with a thermal runaway risk probability exceeding a preset threshold are merged to generate a set of trigger nodes. Graph nodes other than the set of trigger nodes in the weighted undirected graph are extracted and merged to generate a set of non-trigger nodes. Using the set of triggering nodes as the source nodes and the set of non-triggering nodes as the sink nodes, a residual graph is generated based on the weighted undirected graph, the source nodes, and the sink nodes. Calculate the node height labels and edge preflow values ​​of the residual graph, and calculate the excess flow of the graph nodes in the residual graph based on the edge preflow values; Based on the node height label, excess traffic is pushed from the source node to the sink node, and the maximum flow value is calculated. Extract the set of minimum cut edges corresponding to the maximum flow value, and use the set of minimum cut edges as the optimal isolation edge set.

7. The battery thermal runaway isolation method based on risk prediction as described in claim 6, characterized in that, Cut off the heat conduction path and charging / discharging circuit corresponding to the optimal isolation edge set, including: Obtain the liquid cooling channel and charging / discharging circuit corresponding to the optimal isolation edge set; Identify the bidirectional solenoid valve associated with the liquid cooling channel and the electrical contactor associated with the charging and discharging circuit; Generate isolation control commands for driving the bidirectional solenoid valve to close and the electrical contactor to open; The bidirectional solenoid valve is driven to close to cut off the liquid cooling channel according to the isolation control command, and the electrical contactor is driven to open to cut off the charging and discharging circuit according to the isolation control command.

8. A battery thermal runaway isolation device based on risk prediction, characterized in that, include: The module includes a data acquisition module, a feature alignment module, a risk prediction module, a graph cut decision module, and an isolation execution module. The data acquisition module is used to acquire the time-series data of temperature, internal resistance, and gas concentration of the cells in the battery module. The feature alignment module is used to perform interpolation alignment operations and standardization processing on temperature time series data and gas concentration time series data based on the timestamp of internal resistance time series data, and generate a multimodal alignment feature matrix. The risk prediction module is used to input the multimodal aligned feature matrix into the pre-trained thermal state evolution prediction model and output the probability of thermal runaway risk of the battery cell. The graph cut decision module is used to identify the heat conduction paths and charging / discharging circuits between cells in the battery module. It constructs a weighted undirected graph using cells as nodes, the heat conduction paths and charging / discharging circuits between cells as edges, and parameters representing the thermal isolation priority between adjacent cells as edge weights. The thermal isolation priority is positively correlated with the heat conduction risk between cells, and the edge weights are negatively correlated with the heat conduction risk. Cells with a thermal runaway risk probability exceeding a preset threshold are used as trigger nodes. The module performs a minimum cut partition calculation in the weighted undirected graph and outputs the optimal isolation edge set. The isolation execution module is used to cut off the heat conduction path and charging / discharging circuit corresponding to the optimal isolation edge set.

9. An electronic device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the risk prediction-based battery thermal runaway isolation method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the storage medium to perform the battery thermal runaway isolation method based on risk prediction as described in any one of claims 1 to 7.

Images