Battery temperature reconstruction-based thermal runaway early warning method, device, equipment and medium

Abstract

The application discloses a battery temperature reconstruction-based thermal runaway early warning method and device, equipment and medium, and relates to the technical field of artificial intelligence and battery thermal safety management. The method comprises the following steps: determining a sparse temperature measurement point set according to battery pack simulation temperature field data; collecting sparse temperature measurement data, terminal voltage data and operating condition data, and combining spatial coordinates to generate a node feature sequence; constructing a battery pack graph structure according to a preset number of neighbors, a thermal resistance network and node distances; inputting the node feature sequence and the battery pack graph structure into a physically enhanced graph convolution long short-term memory model to obtain node temperature prediction data; performing residual correction autoregressive processing to obtain corrected temperature prediction data; generating full-field temperature reconstruction data based on the corrected temperature prediction data and the battery pack graph structure, and outputting a thermal runaway early warning result. The application can obtain full-field temperature reconstruction data and output a thermal runaway early warning result based on a small amount of temperature measurement data under the condition that the temperature measurement points of the power battery pack are sparse.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and battery thermal safety management technology, and in particular to a method, device, equipment and medium for early warning of thermal runaway based on battery temperature reconstruction. Background Technology

[0002] With the large-scale application of new energy vehicles, the thermal safety status of power battery packs has a significant impact on vehicle operation safety. Temperature is an important parameter reflecting battery performance degradation, lifespan changes, and the risk of thermal runaway. However, in actual deployment, due to limitations in cost, layout space, and wiring harness complexity, only a small number of temperature sensors can usually be placed inside the power battery pack, making it difficult to directly obtain the overall temperature distribution inside the battery pack.

[0003] Currently, temperature monitoring and reconstruction of power battery packs typically employ physical simulation models, thermal resistance network models, data-driven models, or a combination of physical models and neural networks. Physical simulation models can calculate the temperature field based on heat conduction mechanisms, thermal resistance network models can describe the heat transfer relationships between individual battery cells, data-driven models can learn temperature change patterns based on historical temperature data, and some solutions also introduce graph neural networks to model the relationships between different temperature measurement nodes within the battery pack.

[0004] However, physical simulation models are computationally intensive, making it difficult to meet the real-time warning requirements of vehicles; thermal resistance network models have limited ability to characterize the complex thermal coupling relationships of multiple cells; pure data-driven models are easily affected by insufficient observation information when temperature measurement points are sparse, and the prediction results lack physical constraints on thermal conduction; existing fusion methods also tend to have problems such as insufficient embedding of physical constraints and overly complex model structures. Therefore, how to obtain full-field temperature reconstruction data based on a small amount of temperature measurement data and output thermal runaway warning results under the condition of sparse temperature measurement points in the power battery pack has become an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this application is to provide a method, device, equipment and medium for thermal runaway early warning based on battery temperature reconstruction, aiming to solve the technical problem of how to obtain full-field temperature reconstruction data based on a small amount of temperature measurement data and output thermal runaway early warning results under the condition of sparse temperature measurement points in power battery pack.

[0006] To achieve the above objectives, this application proposes a thermal runaway early warning method based on battery temperature reconfiguration, the method comprising: Based on the simulated temperature field data of the power battery pack, a sparse temperature measurement point set is determined; Based on the sparse temperature measurement point set, sparse temperature measurement data, terminal voltage data, and operating condition data are collected, and a node feature sequence is generated according to the spatial coordinates of the sparse temperature measurement point set. A node set is established based on the spatial coordinates of the sparse temperature measurement point set. The connection relationship between the graph nodes in the node set is established based on the preset number of nearest neighbors. The thermal coupling edge weight of the connection relationship is determined based on the thermal resistance network and the node distance between the graph nodes, thus obtaining the battery pack graph structure. The node feature sequence and the battery pack graph structure are input into the trained Physical Augmented Graph Convolutional Long Short-Term Memory Model. Spatial encoding features are obtained through graph convolution with graph Laplacian constraints. The spatial encoding features are then subjected to long short-term memory temporal processing to obtain node temperature prediction data. The node temperature prediction data is subjected to residual correction autoregression processing to obtain corrected temperature prediction data; Full-field temperature reconstruction data is generated based on the corrected temperature prediction data and the battery pack diagram structure, and thermal runaway early warning results are output based on the full-field temperature reconstruction data.

[0007] Furthermore, to achieve the above objectives, this application also proposes a thermal runaway early warning device based on battery temperature reconstruction, the device comprising: The point selection module is used to determine the set of sparse temperature measurement points based on the simulated temperature field data of the power battery pack. The feature generation module is used to collect sparse temperature measurement data, terminal voltage data and operating condition data based on the sparse temperature measurement point set, and generate a node feature sequence according to the spatial coordinates of the sparse temperature measurement point set. The graph construction module is used to establish a node set based on the spatial coordinates of the sparse temperature measurement point set, establish the connection relationship between the graph nodes in the node set according to the preset number of nearest neighbors, and determine the thermal coupling edge weight of the connection relationship based on the thermal resistance network and the node distance between the graph nodes to obtain the battery pack graph structure. The model processing module is used to input the node feature sequence and the battery pack graph structure into the trained physical augmented graph convolutional long short-term memory model, obtain spatial encoding features through graph convolution processing with graph Laplacian constraints, and perform long short-term memory temporal processing on the spatial encoding features to obtain node temperature prediction data. The temperature correction module is used to perform residual correction autoregressive processing on the node temperature prediction data to obtain corrected temperature prediction data. The early warning module is used to generate full-field temperature reconstruction data based on the corrected temperature prediction data and the battery pack diagram structure, and output thermal runaway early warning results based on the full-field temperature reconstruction data.

[0008] In addition, to achieve the above objectives, this application also proposes a thermal runaway early warning device based on battery temperature reconstruction. The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the thermal runaway early warning method based on battery temperature reconstruction as described above.

[0009] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the thermal runaway early warning method based on battery temperature reconstruction as described above.

[0010] One or more technical solutions proposed in this application have at least the following technical effects: First, based on the simulated temperature field data, the internal temperature distribution characteristics of the battery pack are analyzed to determine sparse temperature measurement points that can characterize the temperature change pattern. Sparse temperature measurement data, terminal voltage data, and operating condition data are collected in chronological order of sampling time. Combined with the spatial coordinates of each temperature measurement point, the multi-source information is organized into a node feature sequence, ensuring continuous input of temperature, electrical status, and operating environment information over time. Subsequently, based on the spatial relationships of the sparse temperature measurement points, combined with the heat transfer paths and contact relationships in the thermal resistance network, and according to the distance between nodes, a battery pack graph structure is constructed, thereby representing the spatial proximity and thermal coupling between the temperature measurement points. The relationships are expressed in graph form. The node feature sequences and the battery pack graph structure are input into a physically enhanced graph convolutional long short-term memory model. Spatial thermal coupling features are extracted through graph convolution, and the dynamic characteristics of temperature change over time are characterized by the long short-term memory structure, yielding node temperature prediction results. Residual correction is applied to the prediction results to reduce the impact of error accumulation during continuous prediction, resulting in more stable corrected temperature prediction data. Finally, spatial interpolation and expansion are performed based on the corrected temperature prediction data and the battery pack graph structure to generate full-field temperature reconstruction data. Based on the reconstructed temperature distribution, thermal runaway judgment features are extracted, and thermal runaway early warning results are output. Therefore, this application can, under the condition of sparse temperature measurement points in the power battery pack, construct spatial relationships, fuse multi-source information, and perform spatiotemporal joint modeling to derive the full-field temperature distribution from local temperature measurement data, thereby obtaining full-field temperature reconstruction data and outputting thermal runaway early warning results based on a small amount of temperature measurement data. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating an embodiment of the thermal runaway early warning method based on battery temperature reconstruction provided in this application. Figure 2 This is a schematic diagram of the sparse temperature measurement point distribution provided in Embodiment 1 of the thermal runaway early warning method based on battery temperature reconstruction in this application. Figure 3 This is a schematic diagram of the battery pack structure provided in Embodiment 1 of the thermal runaway early warning method based on battery temperature reconstruction in this application; Figure 4 This is a schematic diagram of the structure of the physical augmentation graph convolutional long short-term memory model provided in Embodiment 1 of the thermal runaway early warning method based on battery temperature reconstruction in this application; Figure 5 This is a schematic diagram of the residual correction autoregressive processing provided in Embodiment 1 of the thermal runaway early warning method based on battery temperature reconstruction in this application; Figure 6 This is a flowchart illustrating Embodiment 2 of the thermal runaway early warning method based on battery temperature reconstruction in this application. Figure 7 This is a schematic diagram of the module structure of the thermal runaway early warning device based on battery temperature reconstruction according to an embodiment of this application; Figure 8 This is a schematic diagram of the device structure of the hardware operating environment involved in the thermal runaway early warning method based on battery temperature reconstruction in the embodiments of this application.

[0014] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0015] The embodiments of this application are described below with reference to the accompanying drawings. These embodiments are used to explain the technical solutions of this application and are not intended to limit the scope of protection of this application. The following description uses a temperature warning system as the implementing entity.

[0016] Based on this, embodiments of this application provide a thermal runaway early warning method based on battery temperature reconstruction, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the thermal runaway early warning method based on battery temperature reconstruction in this application.

[0017] In this embodiment, the thermal runaway early warning method based on battery temperature reconstruction includes steps S10~S60: Step S10: Determine the sparse temperature measurement point set based on the simulated temperature field data of the power battery pack; It should be noted that the simulated temperature field data refers to the temperature data corresponding to each spatial location obtained through the battery pack thermal simulation model under different operating conditions. The sparse temperature measurement point set refers to the set of temperature measurement locations selected from multiple measurable locations within the power battery pack, and this set of temperature measurement locations is used to represent the temperature change information within the battery pack.

[0018] Understandably, in one example, the simulated temperature field data is generated from a 6S2P cylindrical battery pack 3D thermal conductivity model. 6S2P refers to a battery pack comprising 6 series-connected unit groups, each containing 2 parallel battery cells. This model considers the anisotropic thermal conductivity corresponding to the cylindrical battery winding structure, contact heat transfer between battery cells and connectors, and convective heat transfer between the battery pack's outer boundary and the environment. The transient heat transfer relationship inside the power battery pack can be expressed as: In the formula, For material density, For isobaric specific heat capacity, For temperature, For continuous time, For thermal conductivity tensor, This represents the volumetric heat generation rate. The thermal conductivity tensor is used to represent the thermal conductivity differences of individual cells in the radial, circumferential, and axial directions.

[0019] Multiple conventional operating conditions are formed by combining ambient temperature, discharge rate, and convective heat transfer coefficient. Thermal runaway conditions are created by increasing the volumetric heat generation rate in a localized area corresponding to at least one battery cell, or by applying a time-varying additional heat source to that localized area. A sparse set of temperature measurement points is determined based on the temperature changes and spatial distribution at various locations under different operating conditions. These measurement locations can cover the circumferential edges of battery cells, gaps between adjacent cells, ends, and areas near connectors.

[0020] Please refer to Figure 2 , Figure 2 This diagram illustrates the sparse temperature measurement point distribution provided in Embodiment 1 of the thermal runaway early warning method based on battery temperature reconstruction in this application. The light gray dots represent the full grid points in the simulated temperature field, while the purple, light green, and red dots represent the sparse temperature measurement points corresponding to three representative battery cells. The sparse temperature measurement points are mainly distributed near the circumferential edges of the battery cells, the gaps between adjacent cells, and the end boundaries. The temperature measurement points are relatively denser near the upper and lower boundaries. This distribution corresponds to the results of selecting temperature measurement points based on temperature gradient and spatial location.

[0021] The number of preset temperature measurement points can be determined based on the battery pack structure size, sensor placement space, and preset coverage threshold. The preset coverage threshold is used to limit the proportion of sparse temperature measurement points relative to the total number of nodes in the field, and can be in the range of less than 10%, for example, 3% to 8%. This step uses simulated temperature field data to help determine the temperature measurement locations, which can retain more representative temperature observation information under the condition of a limited number of sensors, reducing the dependence of subsequent temperature reconstruction on dense temperature measurement points.

[0022] Step S20: Based on the sparse temperature measurement point set, collect sparse temperature measurement data, terminal voltage data and operating condition data, and generate a node feature sequence according to the spatial coordinates of the sparse temperature measurement point set. It should be noted that sparse temperature measurement data refers to the temperature data collected at each temperature measurement location in the sparse temperature measurement point set within the sampling period. Terminal voltage data refers to the voltage data corresponding to the power battery pack or individual battery cells within the sampling period. Operating condition data can be one or more of the following: ambient temperature, discharge rate, heat transfer parameters, and state of charge. Node feature sequence refers to multiple sets of node features arranged in chronological order of sampling time. Each set of node features includes the initial node features corresponding to all graph nodes within the same sampling period. The time dimension of the node feature sequence corresponds to multiple consecutive sampling periods, the node dimension corresponds to multiple graph nodes in the sparse temperature measurement point set, and the feature dimension corresponds to the sparse temperature measurement temperature value, terminal voltage, operating condition data, and spatial coordinates.

[0023] Understandably, temperature data from each sparse temperature measurement point is collected according to a preset sampling period, while simultaneously acquiring terminal voltage data and operating condition data corresponding to the same sampling period. For each sparse temperature measurement point, the temperature data, terminal voltage data, operating condition data, and spatial coordinates of the measurement location are combined to obtain the node features of that measurement location under that sampling period. Then, the node features corresponding to multiple sampling periods are arranged in chronological order of sampling time to form a node feature sequence. The preset sampling period can be determined based on the data refresh frequency of the battery management system, for example, 0.5s to 5s. This step organizes temperature observation information, electrical status information, operating condition information, and spatial location information into time-series node features, enabling subsequent models to simultaneously utilize time-varying information and spatial location information for temperature prediction.

[0024] Step S30: Establish a node set based on the spatial coordinates of the sparse temperature measurement point set, establish the connection relationship between the graph nodes in the node set based on the preset number of nearest neighbors, and determine the thermal coupling edge weight of the connection relationship based on the thermal resistance network and the node distance between the graph nodes to obtain the battery pack graph structure. It should be noted that the node set refers to the set of graph nodes corresponding to each temperature measurement location in the sparse temperature measurement point set. The connectivity relationship refers to the edges between graph nodes used to represent heat transfer relationships. The thermal resistance network refers to the heat transfer relationship network formed based on the contact interface, heat transfer path, and heat dissipation boundary relationships between battery cells. The thermal coupling edge weight refers to the weight value corresponding to the connectivity relationship, which is used to represent the thermal coupling strength between two graph nodes. The battery pack graph structure refers to the graph data structure composed of the node set, connectivity relationships, and thermal coupling edge weights.

[0025] Understandably, each sparse temperature measurement point is treated as a graph node, and the distance between graph nodes is calculated based on their spatial coordinates. For each graph node, other graph nodes with nearest neighbor relationships are determined based on a preset number of nearest neighbors. Then, combined with the heat transfer correlation in the thermal resistance network, the connection relationships between graph nodes are established. Subsequently, the thermal coupling edge weights corresponding to the connection relationships are determined based on the node distances and heat transfer correlations, resulting in the battery pack graph structure.

[0026] Please refer to Figure 3 , Figure 3 This is a schematic diagram of the battery pack graph structure provided in Embodiment 1 of the thermal runaway early warning method based on battery temperature reconstruction of this application. In the diagram, red dots represent graph nodes corresponding to sparse temperature measurement points, blue line segments represent the connection relationships between graph nodes, and the color intensity of the line segments represents the weight of thermal coupling edges. The nodes are relatively densely connected near the edges of individual battery cells, the gaps between adjacent cells, and the ends, reflecting the battery pack graph structure formed by spatial distance and heat transfer correlation.

[0027] The preset number of nearest neighbors can be determined based on the battery pack structure density, the number of temperature measurement points, and computing resources, for example, 4 to 12, specifically 8. This step transforms the spatial proximity and heat transfer relationships between temperature measurement locations within the battery pack into a graph structure, providing an input basis for graph convolution processing and helping to express the thermal coupling relationships between different temperature measurement locations within the battery pack.

[0028] Step S40: Input the node feature sequence and the battery pack graph structure into the trained physical augmented graph convolutional long short-term memory model, obtain spatial encoding features through graph convolution with graph Laplacian constraints, and perform long short-term memory temporal processing on the spatial encoding features to obtain node temperature prediction data. It should be noted that the physically enhanced graph convolutional long short-term memory model refers to a neural network model that includes graph convolution processing and long short-term memory temporal processing, and introduces graph Laplacian constraints in the graph convolution processing. Graph Laplacian constraints are graph smoothing constraints based on the battery pack graph structure, used to constrain the temperature representation relationships between adjacent graph nodes. Spatial encoding features refer to the spatial association features of nodes obtained after graph convolution processing. Node temperature prediction data refers to the temperature data corresponding to each graph node at the prediction time output by the model.

[0029] Understandably, the node feature sequences and the battery pack graph structure are input into the trained physically augmented graph convolutional long short-term memory model. The model first performs graph convolution processing on the node feature sequences based on the battery pack graph structure, incorporating graph Laplacian constraints during the convolution process to obtain spatially encoded features. Then, these spatially encoded features are input into the long short-term memory units in chronological order to obtain temporal features containing historical temperature change information. Finally, the output layer generates the node temperature prediction data. In one example, the number of graph convolutional layers can be two, the hidden layer dimension can be 64, the preset physical constraint coefficient can be 0.1, and the output layer outputs a predicted temperature value for each graph node, with the number of output nodes matching the number of sparse temperature measurement points.

[0030] During continuous prediction, the Physically Augmented Graph Convolutional Long Short-Term Memory (LSTM) model outputs the predicted temperature values ​​for each graph node in the next prediction step. The first prediction step uses the initial node feature sequence as the model input; subsequent prediction steps update the historical temperature features based on the temperature data obtained from the previous prediction step and input them into the model again. The model is repeatedly called according to the preset prediction step size to obtain node temperature prediction data for multiple prediction steps arranged in chronological order.

[0031] Please refer to Figure 4 , Figure 4 This diagram illustrates the structure of the physically enhanced graph convolutional long short-term memory model provided in Embodiment 1 of the thermal runaway early warning method based on battery temperature reconstruction in this application. The diagram shows the transmission relationship between node feature sequences, battery pack graph structure, network layers, and physical constraints during node temperature prediction. Node feature sequences are input into the input layer to form initial hidden features. These initial hidden features, combined with the normalized adjacency matrix and normalized graph Laplacian matrix from the battery pack graph structure, are then input into the graph convolutional layer. After graph convolution processing with graph Laplacian constraints, spatial encoded features are obtained. These spatial encoded features are input into the long short-term memory layer according to the sampling time order to obtain temporal encoded features. The output layer maps the temporal encoded features to the predicted temperature values ​​corresponding to each graph node. The physical constraints in the diagram indicate that graph Laplacian constraints participate in the graph convolution processing and are not treated as independent network layers.

[0032] This step extracts the spatial relationships between temperature measurement nodes through graph convolution processing and extracts the relationships between temperature changes over time through long short-term memory temporal processing, so that the node temperature prediction data contains both spatial coupling information and temporal change information.

[0033] Step S50: Perform residual correction autoregressive processing on the node temperature prediction data to obtain corrected temperature prediction data; It should be noted that residual correction autoregressive processing refers to a method of adjusting subsequent prediction inputs using the deviation between predicted and measured temperatures during continuous prediction. Corrected temperature prediction data refers to the nodal temperature prediction results obtained after residual correction.

[0034] Understandably, the residual is calculated based on the predicted node temperature data and the measured node temperature data at the same sampling time, and the determination of whether to trigger correction is based on the residual. When the residual meets the preset correction conditions, the temperature features in the node feature sequence are updated using the measured historical temperature data, while retaining the terminal voltage data, operating condition data, and spatial coordinates of the corresponding sampling period to obtain the corrected node feature sequence. The corrected node feature sequence and the battery pack diagram structure are input into the trained physical augmentation graph convolutional long short-term memory model to obtain the single-step predicted temperature data corresponding to the current prediction step, and this single-step predicted temperature data is used as the corrected temperature data. When the residual does not meet the preset correction conditions, the predicted node temperature data corresponding to the current prediction step is used as the corrected temperature data.

[0035] Please refer to Figure 5 , Figure 5 This diagram illustrates the residual correction autoregressive processing provided in Embodiment 1 of the thermal runaway early warning method based on battery temperature reconstruction of this application. The horizontal axis represents the time step, and the vertical axis represents temperature. The solid blue line represents the actual temperature, and the dashed red line represents the residual correction autoregressive predicted temperature. The residual correction autoregressive predicted temperature is continuously updated with each prediction step, illustrating the process of triggering single-step correction based on the residual between the predicted and measured temperatures, and using the corrected temperature data for subsequent prediction steps.

[0036] The preset correction conditions can be determined based on the battery pack temperature sampling error, model validation error, and early warning response requirements. In one example, the preset correction conditions may include: the residual amplitude is greater than a preset first residual threshold, or the residuals have the same sign and the average residual value is greater than a preset second residual threshold within a consecutive preset number of steps; the preset first residual threshold can be 1.20℃, the preset second residual threshold can be 1.00℃, and the consecutive preset number of steps can be 5. This step constrains the continuous prediction process through residual information, which can reduce the impact of accumulated deviations in continuous prediction on subsequent temperature reconstruction, making the corrected temperature prediction data more suitable for subsequent full-field temperature reconstruction.

[0037] Step S60: Generate full-field temperature reconstruction data based on the corrected temperature prediction data and the battery pack diagram structure, and output thermal runaway early warning results based on the full-field temperature reconstruction data.

[0038] It should be noted that the full-field temperature reconstruction data refers to the reconstruction temperature data corresponding to multiple spatial locations within the power battery pack. The thermal runaway warning result refers to the output result obtained from the full-field temperature reconstruction data to characterize the thermal risk state, which can be one of the following: normal state, attention state, warning state, or the corresponding warning indicator.

[0039] Understandably, the corrected temperature prediction data is mapped to corresponding graph nodes in the battery pack graph structure, resulting in reconstructed node temperatures for multiple graph nodes. Then, for all-field nodes without temperature sensors, corresponding neighboring graph nodes are selected, and neighborhood weights are determined based on the node distance between the unsensed all-field nodes and their neighboring graph nodes. The reconstructed node temperatures corresponding to these neighboring graph nodes are then weighted based on these neighborhood weights to obtain the temperatures of the unsensed nodes. All-field temperature reconstruction data is generated based on the reconstructed node temperatures and the temperatures of the unsensed nodes. Thermal runaway judgment features are then extracted from this data, and each feature is compared with its corresponding first and second warning thresholds to obtain a normal state identifier, a concern state identifier, or a warning state identifier. A thermal runaway warning result is then output based on the thermal runaway state identifier. The first and second warning thresholds can be determined based on battery type, thermal management strategy, safety test data, and vehicle operating requirements. This step expands the corrected temperature prediction data corresponding to sparse temperature measurement points to the overall temperature distribution of the battery pack and performs thermal risk assessment based on the overall temperature reconstruction data, ensuring that the warning result no longer relies solely on the local temperature value of a single sensor.

[0040] This embodiment determines sparse temperature measurement points by simulating temperature field data, organizes temperature measurement data, terminal voltage data, operating condition data and spatial coordinates into a node feature sequence, and constructs a battery pack graph structure by combining thermal resistance network; predicts node temperature using physical augmented graph convolutional long short-term memory model, reconstructs the overall temperature of the battery pack after residual correction, and outputs thermal runaway early warning results, thereby obtaining the spatial temperature distribution of the battery pack when the number of temperature measurement points is limited.

[0041] As an example, the step of determining the sparse temperature measurement point set based on the simulated temperature field data of the power battery pack includes: extracting temperature data and spatial coordinates of multiple full-field nodes from the simulated temperature field data of the power battery pack; calculating the temperature gradient magnitude corresponding to the multiple full-field nodes based on the temperature data and spatial coordinates of the multiple full-field nodes; sorting the multiple full-field nodes according to the temperature gradient magnitude, and selecting candidate temperature measurement nodes from the sorted multiple full-field nodes according to a preset coverage threshold; clustering the candidate temperature measurement nodes according to the spatial coordinates and temperature gradient magnitude of the candidate temperature measurement nodes to obtain multiple temperature measurement node clusters; and selecting the cluster center node from each of the multiple temperature measurement node clusters to obtain the sparse temperature measurement point set.

[0042] It should be noted that a full-field node refers to a discrete location point in the simulated temperature field data of a power battery pack that possesses both temperature data and spatial coordinates. The temperature gradient modulus refers to a numerical value used to characterize the intensity of temperature changes around the full-field node. The preset coverage threshold refers to the proportion of candidate temperature measurement nodes selected from multiple full-field nodes, which can be determined based on the cost of temperature sensors, layout space, and temperature reconstruction accuracy requirements; for example, it can be 3% to 8%. Candidate temperature measurement nodes refer to nodes initially selected after sorting according to the temperature gradient modulus. A temperature measurement node cluster refers to a group of nodes formed after clustering the candidate temperature measurement nodes. The cluster center node refers to the node in the temperature measurement node cluster that represents the spatial location and temperature change characteristics of that temperature measurement node cluster.

[0043] Understandably, temperature data and three-dimensional spatial coordinates of each global node are extracted from the simulated temperature field data under different operating conditions. For any global node at spatial location r, nodes in the simulation mesh with direct mesh connections to that global node are considered as spatial neighbors. The temperature gradient can be directly extracted from the temperature gradient field output by the thermal simulation model; when the thermal simulation model does not directly output the temperature gradient field, the rate of temperature change in the three spatial coordinate directions can be determined based on the temperature difference and coordinate difference between the global node and its spatial neighbors, and then the temperature gradient magnitude can be calculated. In the formula, For the entire event node The corresponding temperature gradient modulus, For the entire event node The corresponding temperature , , This represents the spatial coordinate direction.

[0044] For the same full-field node, the average or peak value can be taken as the temperature gradient magnitude of the full-field node under multiple operating conditions or at multiple sampling times.

[0045] The nodes across the field are sorted according to their temperature gradient modulus, and candidate temperature measurement nodes are selected based on a preset coverage threshold. The 3D spatial coordinates and temperature gradient modulus of the candidate temperature measurement nodes are standardized, and the standardized 3D spatial coordinates and temperature gradient modulus are combined as clustering features. K-means clustering is then performed on these features. The number of clusters is the same as the preset number of temperature measurement points. From each temperature measurement node cluster, candidate temperature measurement nodes with small cluster feature distances to their respective cluster centers and meeting the sensor installation conditions are selected to form a sparse temperature measurement point set. Similarly, from each temperature measurement node cluster, candidate temperature measurement nodes with close proximity to their cluster centers and meeting the sensor installation conditions are selected to form a sparse temperature measurement point set. This ensures that the temperature measurement points cover areas with large temperature gradients, such as the edges of individual units, contact interfaces, connectors, and areas near heat dissipation boundaries.

[0046] This example can select representative temperature measurement locations from simulated temperature field data, making the distribution of temperature measurement points take into account both the intensity of temperature changes and the spatial coverage. This reduces the loss of local information caused by relying solely on experience to arrange temperature measurement points, which is beneficial for preserving the more critical temperature change information within the battery pack under the condition of a limited number of temperature measurement points, and provides a data foundation for subsequent node feature sequence generation and full-field temperature reconstruction.

[0047] As an example, the step of collecting sparse temperature measurement data, terminal voltage data, and operating condition data based on the sparse temperature measurement point set, and generating a node feature sequence according to the spatial coordinates of the sparse temperature measurement point set includes: collecting sparse temperature measurement data for the current sampling period and multiple historical sampling periods according to the temperature measurement positions corresponding to the sparse temperature measurement point set; collecting terminal voltage data and operating condition data corresponding to the sparse temperature measurement data; combining the sparse temperature measurement data, terminal voltage data, operating condition data, and spatial coordinates under the same sampling period to obtain initial node features; standardizing the initial node features according to preset standardization parameters to obtain standardized node features; arranging the standardized node features of multiple sampling periods in the order of sampling time to generate a node feature sequence; the initial node features are represented as follows: In the formula, For graph nodes During the sampling period Initial characteristics of nodes, For graph nodes During the sampling period The sparse thermometric temperature values, Sampling period The terminal voltage, Sampling period Operating condition vector, , , Graph nodes The three-dimensional coordinates; the upper right corner This represents the transpose operator, which has a different meaning from the temperature variable T.

[0048] It should be noted that sparse temperature measurement data refers to the temperature data collected at each temperature measurement location in the sparse temperature measurement point set within the sampling period. The current sampling period refers to the target sampling period corresponding to the generation of the node feature sequence. The historical sampling period refers to the sampling period before the current sampling period. The terminal voltage data can be the total terminal voltage of the power battery pack within the sampling period, or the terminal voltage of each individual battery cell within the sampling period. When using the total terminal voltage of the battery pack, each graph node within the same sampling period corresponds to the same total terminal voltage; when using the terminal voltage of each individual battery cell, each graph node corresponds to the terminal voltage of the individual battery cell to which the graph node is located. The terminal voltage data in the same embodiment adopts the same data level. The operating condition data can be one or more of the following: ambient temperature, discharge rate, heat transfer parameters, and state of charge.

[0049] In this example, the preset standardization parameters include the mean and standard deviation of each continuous feature obtained statistically from the training samples. For any continuous feature, the node-standardized feature is obtained by dividing the difference between the continuous feature and its corresponding mean by the corresponding standard deviation. The same set of means and standard deviations obtained statistically from the training samples are used in the model training, validation, testing, and application phases.

[0050] Understandably, based on the temperature measurement locations corresponding to the sparse temperature measurement point set, sparse temperature measurement data are collected for the current sampling period and multiple historical sampling periods, and terminal voltage data and operating condition data corresponding to the same sampling period are obtained. When the sampling times of various types of data are inconsistent, the data are matched to the corresponding sampling period based on the timestamp. The sparse temperature measurement data, terminal voltage data, operating condition data, and spatial coordinates under the same sampling period are combined into the initial features of the nodes, and standardized using the mean and standard deviation corresponding to the training samples.

[0051] The node feature sequence is generated by arranging the node normalized features from multiple sampling periods in chronological order. In one example, the sampling period is 1 second, and the node feature sequence includes node normalized features from 150 consecutive sampling periods.

[0052] This example organizes sparse temperature measurement data, terminal voltage data, operating condition data, and spatial coordinates into a unified node feature sequence. This reduces format and dimensional differences between different data sources and preserves temperature change information at each measurement location over multiple sampling periods, providing a data foundation for subsequent temperature prediction based on the node feature sequence.

[0053] As an example, the steps of establishing a node set based on the spatial coordinates of the sparse temperature measurement point set, establishing connection relationships between graph nodes in the node set based on a preset number of nearest neighbors, and determining the thermal coupling edge weights of the connection relationships based on the thermal resistance network and the node distances between graph nodes to obtain the battery pack graph structure include: taking each sparse temperature measurement point in the sparse temperature measurement point set as a graph node to obtain a node set; calculating the node distance between any two graph nodes in the node set based on their spatial coordinates; selecting nearest neighbor graph nodes for each graph node based on the preset number of nearest neighbors and the node distances; determining heat transfer associated node pairs between each graph node and its corresponding nearest neighbor graph nodes based on the thermal resistance network, and establishing connection relationships based on the heat transfer associated node pairs; calculating the thermal coupling edge weights of the connection relationships; generating an adjacency matrix based on the connection relationships and the thermal coupling edge weights, and obtaining the battery pack graph structure based on the adjacency matrix and the node set.

[0054] It should be noted that the preset number of nearest neighbors refers to the number of nearest neighbor graph nodes selected for each graph node. This number can be determined based on the number of sparse temperature measurement points, the spatial structure of the battery pack, and computational resources, for example, 4 to 12. A heat transfer associated node pair refers to two graph nodes in the thermal resistance network that have a heat transfer association and satisfy the nearest neighbor relationship. The preset kernel width parameter is a parameter used to control the degree of influence of node distance on the weight of thermally coupled edges. This parameter can be determined based on the node distance distribution, for example, taking 0.5 to 2 times the median node distance.

[0055] Understandably, each sparse temperature measurement point in the sparse temperature measurement point set is treated as a graph node, and the corresponding three-dimensional spatial coordinates of each graph node are retained. The node distance is calculated based on the three-dimensional spatial coordinates of any two graph nodes, and a preset number of nearest neighbor graph nodes are selected for each graph node in order of increasing distance. When a graph node... graph nodes Selected as a nearest neighbor graph node, or graph node graph nodes When a node is selected as a nearest neighbor in the graph, and the two nodes form a heat transfer associated node pair, in the graph node... With graph nodes Establish undirected connections between them, and ensure that corresponding elements in the adjacency matrix satisfy... In one example, the default number of nearest neighbors is 8.

[0056] Each graph node is mapped to its corresponding cell region, connector region, or heat dissipation boundary region in the thermal resistance network. For a graph node and its corresponding nearest neighbor nodes, if there is a direct thermal resistance branch between the regions corresponding to the two graph nodes in the thermal resistance network, the two graph nodes are identified as a heat transfer associated node pair; if there is no direct thermal resistance branch between the two regions, no connection is established between the two graph nodes. For any heat transfer associated node pair, the thermal coupling edge weight is calculated according to the following formula: In the formula, For graph nodes With graph nodes The thermal coupling edge weights between them For graph nodes With graph nodes The distance between nodes The preset core width parameter can be determined based on the node distance distribution of the heat transfer associated node pairs, for example, by taking 0.5 to 2 times the median node distance.

[0057] An adjacency matrix is ​​generated based on connectivity relationships and thermally coupled edge weights. In the adjacency matrix, the matrix elements corresponding to two graph nodes with connectivity relationships are set to their respective thermally coupled edge weights, while the matrix elements corresponding to two graph nodes without connectivity relationships are set to 0. The battery pack graph structure is obtained based on the adjacency matrix and the node set.

[0058] This example transforms the spatial proximity and heat transfer relationships between sparse temperature measurement points into a battery pack graph structure, enabling subsequent models to use the thermal coupling relationships between temperature measurement points for temperature prediction. At the same time, by controlling the connection scale by presetting the number of nearest neighbors, the impact of irrelevant node connections on the calculation process can be reduced, thus lowering the redundancy of the graph structure.

[0059] As an example, the step of performing residual correction autoregressive processing on the node temperature prediction data to obtain corrected temperature prediction data includes: determining a prediction step sequence based on the sampling time order and preset prediction step size corresponding to the node temperature prediction data; for the m-th prediction step in the prediction step sequence, obtaining the node temperature prediction data corresponding to the m-th prediction step, and obtaining the measured node temperature data of the sparse temperature measurement point set at the sampling time corresponding to the m-th prediction step; calculating the residual data corresponding to the m-th prediction step based on the node temperature prediction data and the measured node temperature data; and based on... The residual data corresponding to each prediction step within the preset residual window before the m-th prediction step are used to determine the residual amplitude, residual mean, and residual sign sequence corresponding to the m-th prediction step. Based on the residual amplitude, residual mean, and residual sign sequence corresponding to the m-th prediction step, the corrected temperature data corresponding to the m-th prediction step is determined. Based on the corrected temperature data corresponding to the m-th prediction step, the historical temperature window in the node feature sequence is updated. The prediction step sequence is traversed to obtain the corrected temperature data corresponding to multiple prediction steps, and the corrected temperature prediction data is obtained based on the corrected temperature data corresponding to multiple prediction steps.

[0060] It should be noted that the prediction step sequence refers to multiple temperature prediction moments arranged in chronological order of sampling time. The preset prediction step size refers to the time interval between two adjacent prediction steps, which can be determined based on the temperature sampling period and early warning response requirements, for example, 0.5s to 5s. The m-th prediction step refers to the m-th temperature prediction moment in the prediction step sequence, where m is a positive integer. The measured node temperature data refers to the node temperature data collected at the sampling moment corresponding to the m-th prediction step. The node residual refers to the difference between the predicted node temperature value and the measured temperature value for the same graph node at the same prediction step. The preset residual window refers to the range of consecutive prediction steps ending at the current prediction step, used to statistically analyze the changes in residuals for nodes in the same graph; it can be determined based on the temperature change rate and model validation error, for example, 3 to 20 prediction steps. The historical temperature window refers to the range of historical temperature data in the node feature sequence that will be used in the next prediction.

[0061] Understandably, the prediction step sequence is determined according to the sampling time order corresponding to the node temperature prediction data and the preset prediction step size. When the sampling time corresponding to the m-th prediction step arrives, the measured temperature value of each graph node at that sampling time is obtained. For any graph node, the predicted node temperature value corresponding to the m-th prediction step is subtracted from the measured temperature value to obtain the node residual. The absolute value of the node residual is used as the residual amplitude corresponding to the m-th prediction step. The residual mean is calculated based on the absolute values ​​of the residuals of multiple nodes within the preset residual window of the m-th prediction step, and the positive and negative signs of the multiple signed node residuals are arranged according to the prediction step order to obtain the residual sign sequence.

[0062] Based on the residual magnitude, residual mean, and residual sign sequence corresponding to each graph node, it is determined whether each graph node triggers correction. For graph nodes that trigger correction, the historical temperature features of the graph node in the node feature sequence are updated using measured historical temperature data to obtain the correction node feature sequence. The correction node feature sequence and the battery pack graph structure are then input into the trained Physical Augmentation Graph Convolutional Long Short-Term Memory Model. The single-step predicted temperature value corresponding to the graph node output by the model is used as the correction temperature value of the graph node. For graph nodes that do not trigger correction, the node temperature prediction value corresponding to the m-th prediction step is used as the correction temperature value of the graph node.

[0063] The correction temperature vector for the m-th prediction step is obtained by combining the correction temperature values ​​of each graph node in the order of the graph nodes. The number of elements in the correction node temperature vector is the same as the number of graph nodes, and each element corresponds to the correction temperature value of one graph node.

[0064] The temperature vector of the corrected node corresponding to the m-th prediction step is added to the historical temperature window in the node feature sequence, and the oldest node temperature vector in the historical temperature window is removed, so that the length of the updated historical temperature window remains unchanged. Each prediction step in the prediction step sequence is processed in turn, and the temperature vectors of the corrected node corresponding to each prediction step are arranged in the order of prediction steps to obtain the corrected temperature prediction data.

[0065] This example incorporates the residuals generated during continuous prediction into the temperature prediction data correction process, reducing the impact of the deviation of the previous prediction step on the subsequent prediction step. At the same time, the historical temperature window is updated using the corrected temperature data, enabling subsequent predictions to utilize more stable temperature inputs and improving the reliability of continuous temperature prediction data when used for full-field temperature reconstruction.

[0066] As an example, the step of determining the corrected temperature data corresponding to the m-th prediction step based on the residual amplitude, residual mean, and residual sign sequence corresponding to the m-th prediction step includes: when the residual amplitude corresponding to the m-th prediction step is greater than a preset first residual threshold, or when the residual sign sequence corresponding to the m-th prediction step has the same sign within a preset consecutive number of steps and the residual mean corresponding to the m-th prediction step is greater than a preset second residual threshold, the single-step predicted temperature data obtained based on the measured historical temperature window before the m-th prediction step is used as the corrected temperature data corresponding to the m-th prediction step; when the residual amplitude corresponding to the m-th prediction step is less than or equal to the preset first residual threshold, and the residual sign sequence corresponding to the m-th prediction step does not satisfy the condition that the signs are the same within a preset consecutive number of steps and the residual mean corresponding to the m-th prediction step is greater than the preset second residual threshold, the node temperature predicted data corresponding to the m-th prediction step is used as the corrected temperature data corresponding to the m-th prediction step.

[0067] It should be noted that the residual magnitude refers to the absolute value of the node residual corresponding to the m-th prediction step of the graph node. The residual mean refers to the average of the absolute values ​​of the node residuals corresponding to multiple prediction steps of the same graph node within the preset residual window. The residual sign sequence refers to the sequence of positive and negative signs of multiple signed node residuals of the same graph node within the preset residual window, arranged in the order of prediction steps. The preset first residual threshold is a threshold used to determine whether the deviation of a single prediction step exceeds the allowable range, which can be determined based on the temperature sensor measurement error and the residual of the model validation set, for example, 0.5℃ to 3℃.

[0068] The preset second residual threshold is a threshold used to determine whether the continuous deviation level exceeds the allowable range, and can be less than or equal to the preset first residual threshold. The preset number of consecutive steps refers to the number of prediction steps used to determine the continuity of the residual direction, which can be determined based on the sampling period and the rate of temperature change, for example, 2 to 10 prediction steps. Single-step predicted temperature data refers to the node temperature data corresponding to the m-th prediction step, output after updating the temperature features in the node feature sequence using measured historical temperature data up to the m-th prediction step, and then inputting the updated node feature sequence and battery pack graph structure into the trained physical augmentation graph convolutional long short-term memory model.

[0069] Understandably, for any graph node, if the residual magnitude corresponding to the node in the m-th prediction step is greater than a preset first residual threshold, or if the residual sign sequence corresponding to the node has the same sign within a preset number of consecutive steps and the average residual value is greater than a preset second residual threshold, then the node is determined to trigger single-step correction, and the single-step predicted temperature value of the node obtained based on the correction node feature sequence is used as the correction temperature value; otherwise, the node temperature prediction value corresponding to the node in the m-th prediction step is used as the correction temperature value. The correction temperature values ​​of each graph node are combined in the order of the graph nodes to obtain the correction node temperature vector corresponding to the m-th prediction step.

[0070] In one example, the preset first residual threshold is 1.20℃, the preset second residual threshold is 1.00℃, and the preset number of consecutive steps is 5.

[0071] This example can select corrected temperature data based on the magnitude of single-step deviation and the changes in continuous deviation, reducing the impact of error accumulation on temperature prediction results during continuous prediction. At the same time, it retains node temperature prediction data when the residual is within the allowable range, avoiding input fluctuations caused by frequent corrections, and making the corrected temperature prediction data more suitable for subsequent full-field temperature reconstruction.

[0072] As an example, the step of generating full-field temperature reconstruction data based on the corrected temperature prediction data and the battery pack graph structure, and outputting a thermal runaway early warning result based on the full-field temperature reconstruction data includes: determining the reconstruction node temperatures corresponding to multiple graph nodes based on the corrected temperature prediction data and the battery pack graph structure; selecting graph nodes with heat transfer associations with full-field nodes without temperature sensors from the multiple graph nodes in the battery pack graph structure as neighboring graph nodes; determining the corresponding neighbor weights according to the calculation rules of the thermal coupling edge weights based on the node distance between the full-field nodes without temperature sensors and the neighboring graph nodes, and performing weighted calculations on the reconstruction node temperatures corresponding to the neighboring graph nodes based on the neighbor weights to obtain multiple unmeasured node temperatures; generating full-field temperature reconstruction data based on the multiple reconstruction node temperatures and the multiple unmeasured node temperatures; determining thermal runaway judgment features based on the full-field temperature reconstruction data; comparing each feature in the thermal runaway judgment features with the corresponding preset early warning threshold to obtain a thermal runaway state identifier, and outputting a thermal runaway early warning result based on the thermal runaway state identifier.

[0073] It should be noted that reconstructed node temperature refers to the corrected temperature data corresponding to the nodes in the battery pack diagram structure. Unmeasured node temperature refers to the reconstructed temperature data corresponding to the spatial nodes within the power battery pack that do not have temperature sensors. Full-field temperature reconstruction data refers to the battery pack spatial temperature distribution data composed of reconstructed node temperatures and unmeasured node temperatures. Thermal runaway determination features refer to temperature-related characteristics used to determine the thermal risk state of the battery pack, which may include temperature value, temperature rise rate, and spatial temperature difference. Thermal runaway state identifiers refer to identifiers used to indicate the thermal risk level of the power battery pack, including normal state identifiers, concern state identifiers, and warning state identifiers. In the example using three thermal risk levels, each thermal runaway determination feature corresponds to a first warning threshold and a second warning threshold, respectively, with the second warning threshold being higher than the corresponding first warning threshold. The first and second warning thresholds can be determined based on battery type, safety test data, thermal management strategies, and vehicle operating requirements.

[0074] Understandably, the corrected temperature prediction data is mapped to the corresponding graph nodes in the battery pack graph structure according to the graph node number or spatial coordinates to obtain the reconstructed node temperature. For any full-field node without a temperature sensor, graph nodes with a heat transfer path to that full-field node are selected as candidate neighbor graph nodes based on the thermal resistance network. From the candidate neighbor graph nodes, a preset number of reconstructed nearest neighbors are selected as the neighborhood graph nodes in order of node distance from closest to farthest. The preset number of reconstructed nearest neighbors can be determined based on the full-field temperature reconstruction error corresponding to the validation sample, for example, 4 to 12.

[0075] Based on the node distances between the global node and its neighboring graph nodes, the initial neighborhood weights for each neighboring graph node are determined according to the calculation rules for thermal coupling edge weights. Each initial neighborhood weight is divided by the sum of all initial neighborhood weights to obtain the corresponding normalized neighborhood weight. Then, the reconstructed node temperatures of each neighboring graph node are weighted and summed based on these normalized neighborhood weights to obtain the temperatures of the unmeasured nodes. Based on the reconstructed node temperatures and the temperatures of the unmeasured nodes, global temperature reconstruction data corresponding to the battery pack's spatial coordinates is generated.

[0076] The highest temperature, highest temperature rise rate, and highest spatial temperature difference are extracted from the full-field temperature reconstruction data within a preset warning time window. Each thermal runaway determination feature corresponds to a first warning threshold and a second warning threshold, respectively, with the second warning threshold being higher than the corresponding first warning threshold. The highest temperature, highest temperature rise rate, and highest spatial temperature difference are determined using the first and second warning thresholds with their corresponding dimensions.

[0077] A normal state identifier is generated when the highest temperature, the highest temperature rise rate, and the highest spatial temperature difference are all below their respective first warning thresholds. A concern state identifier is generated when at least one thermal runaway determination feature reaches its corresponding first warning threshold, and all thermal runaway determination features are below their respective second warning thresholds. A warning state identifier is generated when at least one thermal runaway determination feature reaches its corresponding second warning threshold. The location of all nodes corresponding to the thermal runaway determination feature that reaches the highest state level is designated as an anomaly location, and at least one of the following is output based on the thermal runaway state identifier: anomaly location, sampling time, and corresponding temperature data.

[0078] This example expands the corrected temperature prediction data to the overall temperature distribution data of the battery pack, so that thermal runaway early warning does not only rely on the local temperature values ​​of a few sensors. At the same time, by combining the overall temperature reconstruction data to extract thermal runaway judgment features, it can provide a more complete temperature basis for battery pack thermal risk assessment and improve the correlation between thermal runaway early warning results and battery pack space temperature state.

[0079] As an example, the step of determining the thermal runaway judgment characteristics based on the full-field temperature reconstruction data includes: extracting the node reconstruction temperatures of multiple full-field nodes within the preset warning time window from the full-field temperature reconstruction data according to the target warning sampling period and a preset warning time window before the target warning sampling period; selecting the highest temperature value from the node reconstruction temperatures of multiple full-field nodes within the preset warning time window; calculating the temperature rise rate of multiple full-field nodes within the preset warning time window based on the node reconstruction temperature difference of the same full-field node in adjacent sampling periods and the sampling time interval, and selecting the highest temperature rise rate from the multiple temperature rise rates; calculating the spatial temperature difference corresponding to multiple sampling periods based on the highest and lowest temperature values ​​among the node reconstruction temperatures of multiple full-field nodes within the same sampling period, and selecting the highest spatial temperature difference from the multiple spatial temperature differences; and obtaining the thermal runaway judgment characteristics based on the highest temperature value, the highest temperature rise rate, and the highest spatial temperature difference.

[0080] It should be noted that the target warning sampling period refers to the sampling period for which thermal runaway judgment results need to be output. The preset warning time window refers to the time range used to extract temperature change features, ending at the target warning sampling period. This can be determined based on the sampling period, thermal management response time, and battery type, for example, 5s to 300s, or 5 to 100 consecutive sampling periods. Node reconstruction temperature refers to the temperature value of each full-field node in the full-field temperature reconstruction data within the corresponding sampling period. Temperature rise rate refers to the rate of temperature change of the same full-field node between adjacent sampling periods. Spatial temperature difference refers to the temperature difference between multiple full-field nodes within the same sampling period. In this example, the thermal runaway judgment feature can be a feature group consisting of the highest temperature value, the highest temperature rise rate value, and the highest spatial temperature difference value.

[0081] Understandably, taking the target early warning sampling period as the end point, the node reconstruction temperature of each full-field node within the preset early warning time window is extracted from the full-field temperature reconstruction data, and the highest temperature value is selected. The temperature rise rate is calculated based on the difference in node reconstruction temperature of the same full-field node in adjacent sampling periods and the sampling time interval, and the highest temperature rise rate is selected. The spatial temperature difference is calculated based on the highest and lowest node reconstruction temperatures within the same sampling period, and the highest spatial temperature difference within the preset early warning time window is selected. The highest temperature value, the highest temperature rise rate, and the highest spatial temperature difference constitute the thermal runaway determination feature.

[0082] The thermal runaway determination features in this example simultaneously reflect the battery pack's temperature level, temperature change rate, and spatial temperature imbalance within the preset warning time window. This allows the thermal runaway determination to go beyond the local temperature value of a single temperature measurement point, improving the utilization of full-field temperature reconstruction data in thermal risk identification.

[0083] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 6 , Figure 6 This is a flowchart illustrating the second embodiment of the thermal runaway early warning method based on battery temperature reconstruction of this application. The physically enhanced graph convolutional long short-term memory model includes an input layer, a graph convolutional layer, a long short-term memory layer, and an output layer. Step S40 of the thermal runaway early warning method based on battery temperature reconstruction includes steps S41 to S45: Step S41: Input the node feature sequence into the input layer to obtain the initial hidden features; Step S42: Generate a normalized adjacency matrix and a normalized graph Laplacian matrix based on the adjacency matrix in the battery pack graph structure. Step S43: Input the initial hidden features, the normalized adjacency matrix, and the normalized graph Laplacian matrix into the graph convolutional layer to perform graph convolutional processing with graph Laplacian constraints, and use the hidden features of the last layer as spatial encoding features; Step S44: Input the spatial coding features into the long short-term memory layer in the order of sampling time to obtain the temporal coding features; Step S45: The temporal coding features are mapped to the predicted temperature values ​​corresponding to each graph node through the output layer to obtain node temperature prediction data.

[0084] It should be noted that the input layer refers to the network layer used to receive the node feature sequence and convert it into the model's hidden representation. The graph convolutional layer refers to the network layer that performs spatial correlation processing on the graph node features based on the battery pack graph structure. The normalized adjacency matrix is ​​the matrix obtained after scaling the adjacency matrix, used to represent the connection strength between graph nodes. The normalized graph Laplacian matrix is ​​the graph constraint matrix obtained based on the battery pack graph structure, used to introduce physical correlations between adjacent graph nodes in graph convolution processing. The temporal encoded feature refers to the temperature change feature obtained by the Long Short-Term Memory layer according to the sampling time sequence. The preset physical constraint coefficient can be a parameter used to control the degree of participation of the graph Laplacian constraint, and can be selected according to the prediction error of the validation samples, for example, 0.01 to 1.

[0085] Understandably, the node feature sequence is input into the input layer to obtain the initial hidden features. The input layer can be a linear mapping layer to convert node features of different dimensions into hidden features of a unified dimension. The dimension of the hidden features can be determined based on the number of nodes, the length of the input sequence, and computational resources, for example, from 32 to 128.

[0086] Based on the adjacency matrix A in the battery pack graph structure, generate an adjacency matrix containing self-connections according to the following formula: In the formula, It is an adjacency matrix that includes self-connections. This is the identity matrix corresponding to the number of nodes in the graph. Based on the adjacency matrix including self-connections... Generate a degree matrix D, and the degree matrix D's degree is... The diagonal elements are an adjacency matrix containing self-connections. No. The sum of all elements in each row. The normalized adjacency matrix and the normalized graph Laplacian matrix are generated according to the following formulas: In the formula, For the normalized adjacency matrix, This is the normalized graph Laplace matrix.

[0087] The elements in the adjacency matrix represent the weights of hot-coupled edges between graph nodes. Normalization can reduce the impact of differences in the number of connections between different graph nodes on the graph convolution result.

[0088] Next, the initial hidden features, the normalized adjacency matrix, and the normalized graph Laplacian matrix are input into the graph convolutional layer for graph Laplacian-constrained graph convolution. The graph convolutional layer can obtain the (k+1)th layer hidden features according to the following formula: In the formula, For the first Hidden features of layers For the first Hidden features of layers For the normalized adjacency matrix, For the first Layer weight matrix, For activation function, To preset physical constraint coefficients, This is the normalized graph Laplacian matrix. After at least one layer of graph convolution, the hidden features of the last layer are used as spatial encoding features.

[0089] Spatial coding features are input into the Long Short-Term Memory (LSTM) layer in chronological order of sampling time to obtain temporal coding features. The LSM layer receives spatial coding features corresponding to multiple sampling periods, extracts the correlation information of node temperature changes over time, and outputs temporal coding features corresponding to the prediction time.

[0090] Finally, the temporal encoded features are mapped to the predicted temperature values ​​corresponding to each graph node through the output layer, resulting in node temperature prediction data. The output layer can be a linear mapping layer to convert the temporal encoded features into temperature-dimensional data, so that each graph node outputs a predicted temperature value at the corresponding prediction step.

[0091] This embodiment performs hidden representation transformation on the node feature sequence through the input layer, extracts the spatial correlation between graph nodes by combining the normalized adjacency matrix and the normalized graph Laplacian matrix through the graph convolutional layer, and extracts the change relationship of spatial coding features in the sampling time sequence through the long short-term memory layer. This makes the node temperature prediction data simultaneously contain thermal coupling information and temperature time sequence change information in the battery pack graph structure, providing a data foundation for subsequent residual correction and full-field temperature reconstruction.

[0092] As an example, the training steps of the Physical Augmented Graph Convolutional Long Short-Term Memory (PSM) model include: constructing a training sample set and a validation sample set based on the simulated temperature field data of the power battery pack; the training samples in the training sample set include training node feature sequences, training battery pack graph structures, and actual node temperature data; inputting the training node feature sequences and the training battery pack graph structures into the PSM model to be trained to obtain training node temperature prediction data; calculating the model training loss based on the training node temperature prediction data, the actual node temperature data, and model weight parameters; updating the model parameters of the PSM model to be trained based on the model training loss to obtain the updated PSM model; inputting the validation sample set into the updated PSM model to obtain the validation loss; stopping training when the validation loss corresponding to a consecutive preset number of training rounds is not less than the recorded minimum validation loss, or when the number of training rounds reaches a preset training round threshold; and selecting the model parameters with the lowest validation loss from the model parameters corresponding to multiple training rounds to obtain the trained PSM model.

[0093] It should be noted that the training sample set is used to update model parameters, the validation sample set is used to determine training stopping conditions and select model parameters, and the test sample set is used for model evaluation after training is completed. The simulated temperature field data can be hierarchically divided according to operating condition type, ambient temperature, discharge rate, and thermal runaway state, so that the training sample set, validation sample set, and test sample set do not overlap; in one example, the sample size ratio of the training sample set, validation sample set, and test sample set is approximately 4:1:1.

[0094] Understandably, a sliding time window is used to segment the continuous simulated temperature field data. For any training sample, sparse temperature measurement data, terminal voltage data, operating condition data, and spatial coordinates within a preset number of sampling periods are used to form a training node feature sequence. The actual temperature of each graph node corresponding to the prediction step after the end of the training node feature sequence is taken as the actual node temperature data. In one example, the training node feature sequence includes 150 consecutive sampling periods, and the actual node temperature data corresponds to the 151st sampling period. The training node feature sequence and the training battery pack graph structure are input into the Physical Augmentation Graph Convolutional Long Short-Term Memory model to be trained to obtain the training node temperature prediction data. The training node temperature prediction data and the actual node temperature data maintain a correspondence in graph nodes and prediction time steps.

[0095] Based on the predicted temperature data of the training nodes, the actual temperature data of the nodes, and the model weight parameters, the model training loss is calculated according to the following formula: In the formula, For model training loss, The number of nodes in the graph. The number of target prediction time points contained in a training batch, the th Each target prediction time corresponds to a training node feature sequence and the actual temperature data of the node in the next prediction step after that training node feature sequence. For graph nodes In the Temperature prediction values ​​of training nodes at each time step For graph nodes In the The actual temperature value of each node at each time step. To preset the regularization coefficient, For the first The model weight matrix of the layer, For the first The squared Frobenius norm of the model weight matrix of the layer.

[0096] The prediction error term in the model training loss only calculates the error between the predicted and actual node temperatures; terminal voltage data is used as a node input feature and not as the model's prediction target. Graph Laplacian constraints are added to the forward processing of the graph convolutional layer and are not treated as an independent loss term. In one example, the preset regularization coefficient is... .

[0097] The Adam optimizer is used to update the model parameters based on the model training loss. In one example, the initial learning rate is... The first-order moment attenuation coefficient is 0.9, the second-order moment attenuation coefficient is 0.999, and the numerical stability parameter is... During training, the overall norm of the gradients corresponding to all trainable parameters of the model is calculated. When the overall norm exceeds a preset gradient norm threshold, the gradients corresponding to each model parameter are scaled synchronously according to the ratio of the overall norm to the preset gradient norm threshold, so that the scaled overall norm equals the preset gradient norm threshold. The preset gradient norm threshold can be determined based on the gradient norm distribution of multiple training batches in the early stage of training.

[0098] After each training round, the validation sample set is input into the updated model to obtain the temperature prediction data for the validation nodes. The validation loss is then calculated based on the mean squared error between the predicted temperature data and the actual temperature data of the corresponding nodes. The calculation of the validation loss does not include the model weight regularization term, and the model parameters are not updated during the calculation of the validation loss. When the current validation loss is less than the recorded minimum validation loss, the current model parameters are saved and the minimum validation loss is updated. When the validation loss for a consecutive preset number of training rounds is not less than the recorded minimum validation loss, or when the number of training rounds reaches a preset training round threshold, training stops, and the model parameters saved when the validation loss is lowest are used as the trained model parameters. In one example, the preset number of consecutive training rounds is 10.

[0099] This application also provides a thermal runaway early warning device based on battery temperature reconstruction. Please refer to [link / reference]. Figure 7 The thermal runaway early warning device based on battery temperature reconstruction includes: The point selection module 10 is used to determine the sparse temperature measurement point set based on the simulated temperature field data of the power battery pack. The feature generation module 20 is used to collect sparse temperature measurement data, terminal voltage data and operating condition data based on the sparse temperature measurement point set, and generate a node feature sequence according to the spatial coordinates of the sparse temperature measurement point set. Graph construction module 30 is used to establish a node set based on the spatial coordinates of the sparse temperature measurement point set, establish the connection relationship between the graph nodes in the node set based on the preset number of nearest neighbors, and determine the thermal coupling edge weight of the connection relationship based on the thermal resistance network and the node distance between the graph nodes to obtain the battery pack graph structure. The model processing module 40 is used to input the node feature sequence and the battery pack graph structure into the trained physical augmented graph convolutional long short-term memory model, obtain spatial encoding features through graph convolution processing with graph Laplacian constraints, and perform long short-term memory temporal processing on the spatial encoding features to obtain node temperature prediction data. Temperature correction module 50 is used to perform residual correction autoregressive processing on the node temperature prediction data to obtain corrected temperature prediction data. The early warning module 60 is used to generate full-field temperature reconstruction data based on the corrected temperature prediction data and the battery pack diagram structure, and output thermal runaway early warning results based on the full-field temperature reconstruction data.

[0100] This application provides a thermal runaway early warning device based on battery temperature reconstruction. The thermal runaway early warning device based on battery temperature reconstruction includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the thermal runaway early warning method based on battery temperature reconstruction in the above embodiment 1.

[0101] The following is for reference. Figure 8 This diagram illustrates a suitable structure for implementing a thermal runaway early warning device based on battery temperature reconstruction, as described in the embodiments of this application. The thermal runaway early warning device includes a processing unit 1001, a ROM 1002, a storage unit 1003, a RAM 1004, a bus 1005, an I / O interface 1006, an input device 1007, an output device 1008, and a communication device 1009. The input device 1007 may include a data acquisition interface for receiving sparse temperature measurement data, terminal voltage data, and operating condition data. The output device 1008 outputs thermal runaway early warning results. The communication device 1009 exchanges data with a battery management system, a vehicle control system, or a display terminal. When the processing unit 1001 executes a stored computer program, it implements the aforementioned thermal runaway early warning method based on battery temperature reconstruction.

[0102] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the thermal runaway early warning method based on battery temperature reconstruction in the above embodiments.

[0103] The thermal runaway early warning device, equipment and storage medium based on battery temperature reconstruction provided in this application adopt the thermal runaway early warning method based on battery temperature reconstruction in the above embodiments. The relevant technical effects are described in the method embodiments and will not be repeated here.

[0104] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A thermal runaway early warning method based on battery temperature reconstruction, characterized in that, The method includes: Based on the simulated temperature field data of the power battery pack, a sparse temperature measurement point set is determined; Based on the sparse temperature measurement point set, sparse temperature measurement data, terminal voltage data, and operating condition data are collected, and a node feature sequence is generated according to the spatial coordinates of the sparse temperature measurement point set. A node set is established based on the spatial coordinates of the sparse temperature measurement point set. The connection relationship between the graph nodes in the node set is established based on the preset number of nearest neighbors. The thermal coupling edge weight of the connection relationship is determined based on the thermal resistance network and the node distance between the graph nodes, thus obtaining the battery pack graph structure. The node feature sequence and the battery pack graph structure are input into the trained Physical Augmented Graph Convolutional Long Short-Term Memory Model. Spatial encoding features are obtained through graph convolution with graph Laplacian constraints. The spatial encoding features are then subjected to long short-term memory temporal processing to obtain node temperature prediction data. The node temperature prediction data is subjected to residual correction autoregression processing to obtain corrected temperature prediction data; Full-field temperature reconstruction data is generated based on the corrected temperature prediction data and the battery pack diagram structure, and thermal runaway early warning results are output based on the full-field temperature reconstruction data. The physically enhanced graph convolutional long short-term memory model includes an input layer, a graph convolutional layer, a long short-term memory layer, and an output layer. The steps of inputting the node feature sequence and the battery pack graph structure into the trained Physical Augmented Graph Convolutional Long Short-Term Memory (LSTM) model, obtaining spatial encoded features through graph Laplacian-constrained graph convolution, and performing LSTM temporal processing on the spatial encoded features to obtain node temperature prediction data include: The node feature sequence is input into the input layer to obtain the initial hidden features; Based on the adjacency matrix in the battery pack graph structure, generate a normalized adjacency matrix and a normalized graph Laplacian matrix; The initial hidden features, the normalized adjacency matrix, and the normalized graph Laplacian matrix are input into the graph convolutional layer for graph Laplacian-constrained graph convolution processing. The hidden features of the last layer are used as spatial encoding features. The graph convolution processing is represented as follows: In the formula, For the first Hidden features of layers For the first Hidden features of layers For the normalized adjacency matrix, For the first Layer weight matrix, For activation function, To preset physical constraint coefficients, The normalized graph Laplacian matrix; The spatial coding features are input into the long short-term memory layer in the order of sampling time to obtain the temporal coding features; The output layer maps the temporal coding features to the predicted temperature values ​​corresponding to each graph node, thus obtaining node temperature prediction data.

2. The method as described in claim 1, characterized in that, The step of collecting sparse temperature measurement data, terminal voltage data, and operating condition data based on the sparse temperature measurement point set, and generating a node feature sequence according to the spatial coordinates of the sparse temperature measurement point set includes: According to the temperature measurement location corresponding to the sparse temperature measurement point set, collect sparse temperature measurement data for the current sampling period and multiple historical sampling periods; Collect terminal voltage data and operating condition data corresponding to the sparse temperature measurement data; The initial features of the node are obtained by combining sparse temperature measurement data, terminal voltage data, operating condition data and spatial coordinates under the same sampling period. The initial features of the node are standardized according to preset standardization parameters to obtain the standardized features of the node. The standardized features of the nodes from multiple sampling periods are arranged in chronological order to generate a node feature sequence. The initial features of the node are represented as follows: In the formula, For graph nodes During the sampling period Initial characteristics of nodes, For graph nodes During the sampling period The sparse thermometric temperature values, where V(t) is the sampling period. The terminal voltage, C(t) is the sampling period. Operating condition vector, , , Graph nodes 3D coordinates; superscript This indicates transpose.

3. The method as described in claim 1, characterized in that, The steps of establishing a node set based on the spatial coordinates of the sparse temperature measurement point set, establishing the connection relationship between the graph nodes in the node set according to a preset number of nearest neighbors, and determining the thermal coupling edge weight of the connection relationship based on the node distance between the thermal resistance network and the graph nodes to obtain the battery pack graph structure include: Each sparse temperature measurement point in the sparse temperature measurement point set is used as a graph node to obtain a node set; Calculate the node distance between any two graph nodes in the node set based on their spatial coordinates. Based on the preset number of nearest neighbors and the node distance, select nearest neighbor graph nodes for each graph node; Based on the thermal resistance network, heat transfer associated node pairs are determined between each graph node and its corresponding nearest neighbor graph node, and connection relationships are established based on the heat transfer associated node pairs; The weights of the thermally coupled edges in the aforementioned connections are calculated using the following formula: In the formula, For graph nodes With graph nodes The thermal coupling edge weights between them For graph nodes With graph nodes The distance between nodes This is the preset kernel width parameter; An adjacency matrix is ​​generated based on the connection relationships and the thermally coupled edge weights, and a battery pack graph structure is obtained based on the adjacency matrix and the node set.

4. The method as described in claim 1, characterized in that, The step of performing residual correction autoregressive processing on the node temperature prediction data to obtain corrected temperature prediction data includes: The prediction step sequence is determined based on the sampling time sequence corresponding to the node temperature prediction data and the preset prediction step size; For the m-th prediction step in the prediction step sequence, obtain the node temperature prediction data corresponding to the m-th prediction step, and obtain the measured node temperature data of the sparse temperature measurement point set at the sampling time corresponding to the m-th prediction step. Based on the node temperature prediction data and measured node temperature data corresponding to the m-th prediction step, calculate the residual data corresponding to the m-th prediction step; Based on the m-th prediction step and the residual data corresponding to each prediction step within the preset residual window before the m-th prediction step, the residual magnitude, residual mean, and residual sign sequence corresponding to the m-th prediction step are determined. The residual magnitude refers to the absolute value of the node residual corresponding to the graph node in the m-th prediction step. The residual mean refers to the average value of the absolute values ​​of the node residuals corresponding to the same graph node in multiple prediction steps within the preset residual window. The residual sign sequence refers to the positive and negative sign sequence of multiple signed node residuals of the same graph node within the preset residual window, arranged in the order of prediction steps. Based on the residual amplitude, residual mean, and residual sign sequence corresponding to the m-th prediction step, determine the corrected temperature data corresponding to the m-th prediction step; Update the historical temperature window in the node feature sequence according to the corrected temperature data corresponding to the m-th prediction step, traverse the prediction step sequence to obtain the corrected temperature data corresponding to multiple prediction steps, and obtain the corrected temperature prediction data according to the corrected temperature data corresponding to multiple prediction steps. The step of determining the corrected temperature data corresponding to the m-th prediction step based on the residual amplitude, residual mean, and residual sign sequence corresponding to the m-th prediction step includes: when the residual amplitude corresponding to the m-th prediction step is greater than a preset first residual threshold, or when the residual sign sequence corresponding to the m-th prediction step has the same sign within a preset consecutive number of steps and the residual mean corresponding to the m-th prediction step is greater than a preset second residual threshold, the single-step predicted temperature data obtained based on the measured historical temperature window before the m-th prediction step is used as the corrected temperature data corresponding to the m-th prediction step; when the residual amplitude corresponding to the m-th prediction step is less than or equal to the preset first residual threshold, and the residual sign sequence corresponding to the m-th prediction step does not satisfy the condition that the signs are the same within a preset consecutive number of steps and the residual mean corresponding to the m-th prediction step is greater than the preset second residual threshold, the node temperature predicted data corresponding to the m-th prediction step is used as the corrected temperature data corresponding to the m-th prediction step.

5. The method as described in claim 3, characterized in that, The steps of generating full-field temperature reconstruction data based on the corrected temperature prediction data and the battery pack diagram structure, and outputting thermal runaway early warning results based on the full-field temperature reconstruction data, include: Based on the corrected temperature prediction data and the battery pack diagram structure, the reconstructed node temperatures corresponding to multiple diagram nodes are determined; From the multiple graph nodes in the battery pack graph structure, select the graph node that has a heat transfer relationship with the full field node without a temperature sensor as the neighborhood graph node. Based on the node distance between the full-field node without a temperature sensor and the neighboring graph node, the corresponding neighborhood weight is determined according to the calculation rule of the thermal coupling edge weight, and the reconstructed node temperature corresponding to the neighboring graph node is weighted and calculated based on the neighborhood weight to obtain the temperature of multiple unmeasured nodes. Based on the temperatures of multiple reconstructed nodes and the temperatures of multiple unmeasured nodes, full-field temperature reconstruction data is generated. The thermal runaway determination characteristics are determined based on the full-field temperature reconstruction data. Each feature in the thermal runaway determination feature is compared with its corresponding preset warning threshold to obtain a thermal runaway state identifier, and a thermal runaway warning result is output based on the thermal runaway state identifier.

6. The method according to any one of claims 1 to 5, characterized in that, The step of determining the sparse temperature measurement point set based on the simulated temperature field data of the power battery pack includes: Temperature data and spatial coordinates of multiple full-field nodes are extracted from the simulated temperature field data of the power battery pack; Based on the temperature data and spatial coordinates of multiple global nodes, calculate the temperature gradient magnitude corresponding to the multiple global nodes; The multiple full-field nodes are sorted according to the temperature gradient modulus, and candidate temperature measurement nodes are selected from the sorted multiple full-field nodes according to a preset coverage threshold. Based on the spatial coordinates and temperature gradient magnitude of the candidate temperature measurement nodes, the candidate temperature measurement nodes are clustered to obtain multiple temperature measurement node clusters. A sparse temperature measurement point set is obtained by selecting the cluster center node from each of the multiple temperature measurement node clusters.

7. A thermal runaway early warning device based on battery temperature reconstruction, characterized in that, The device employs the thermal runaway early warning method based on battery temperature reconstruction as described in any one of claims 1 to 6, and the device comprises: The point selection module is used to determine the set of sparse temperature measurement points based on the simulated temperature field data of the power battery pack. The feature generation module is used to collect sparse temperature measurement data, terminal voltage data and operating condition data based on the sparse temperature measurement point set, and generate a node feature sequence according to the spatial coordinates of the sparse temperature measurement point set. The graph construction module is used to establish a node set based on the spatial coordinates of the sparse temperature measurement point set, establish the connection relationship between the graph nodes in the node set according to the preset number of nearest neighbors, and determine the thermal coupling edge weight of the connection relationship based on the thermal resistance network and the node distance between the graph nodes to obtain the battery pack graph structure. The model processing module is used to input the node feature sequence and the battery pack graph structure into the trained physical augmented graph convolutional long short-term memory model, obtain spatial encoding features through graph convolution processing with graph Laplacian constraints, and perform long short-term memory temporal processing on the spatial encoding features to obtain node temperature prediction data. The temperature correction module is used to perform residual correction autoregressive processing on the node temperature prediction data to obtain corrected temperature prediction data. The early warning module is used to generate full-field temperature reconstruction data based on the corrected temperature prediction data and the battery pack diagram structure, and output thermal runaway early warning results based on the full-field temperature reconstruction data.

8. A thermal runaway early warning device based on battery temperature reconstruction, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the thermal runaway early warning method based on battery temperature reconfiguration as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the thermal runaway early warning method based on battery temperature reconstruction as described in any one of claims 1 to 6.

Images