Method and system for ctc-based early warning and control of thermal runaway of a battery pack based on machine learning

Abstract

The application relates to the technical field of machine learning, and discloses a CTG chassis thermal runaway early warning and control method and system based on machine learning. The method comprises the following steps: collecting data at the junction of the battery unit and the metal structure in the CTG chassis to obtain a temperature time sequence matrix and a structure sensitivity matrix; extracting the thermal characteristics of the battery and the structure coupling through a multi-scale thermal feature transformation neural network to obtain a space-time thermal feature matrix; analyzing the thermal flow interaction between the battery unit and the metal structure on the space-time thermal feature matrix to obtain a thermal runaway risk assessment probability value and generate a thermal runaway early warning signal; dividing the CTG chassis into multiple thermal management regions, and constructing a partitioned thermal management scheme for the multiple thermal management regions based on the thermal runaway early warning signal. The application can quickly respond to thermal runaway events, realize the combination of local rapid intervention and global stable control, and improve the safety and stability of the entire thermal management system.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a machine learning-based method and system for early warning and control of thermal runaway in CTC chassis. Background Technology

[0002] With the rapid development of the new energy vehicle industry, CTC (Cell-to-Chassis) technology, as a next-generation electric vehicle design concept, directly integrates the power battery into the chassis structure, achieving deep integration of the battery pack and the frame. While this integrated design significantly improves space utilization and vehicle rigidity, it also brings unprecedented thermal management challenges: a complex heat conduction network is formed between the battery cells and the metal structure, making the monitoring and control of thermal runaway risk extremely complex.

[0003] Traditional battery thermal management technologies primarily rely on temperature threshold monitoring and simple regression models. These methods are ineffective in addressing the unique battery-structure coupled thermal conduction characteristics of CTC chassis. Existing technologies lack a deep understanding of spatiotemporal correlations when processing multi-dimensional thermal sensor data, resulting in significant lag in early thermal runaway identification. Warnings are often only issued when significant temperature anomalies occur, missing crucial intervention opportunities. Furthermore, traditional global thermal control strategies cannot accurately handle the complex inter-regional heat flow interactions within CTC chassis, leading to slow control response times and failing to meet the demands for rapid local intervention. Summary of the Invention

[0004] This invention provides a machine learning-based method and system for early warning and control of thermal runaway in CTC chassis. This invention can quickly respond to thermal runaway events, and combine rapid local intervention with global stable control, thereby improving the safety and stability of the entire thermal management system.

[0005] In a first aspect, the present invention provides a machine learning-based method for early warning and control of CTC chassis thermal runaway, the machine learning-based method for early warning and control of CTC chassis thermal runaway includes:

[0006] Data was collected at the junction of the battery cells and the metal structure in the CTC chassis to obtain the temperature time series matrix and the structural sensitivity matrix.

[0007] The temperature time series matrix and the structure sensitivity matrix are input into a multi-scale thermal feature transformation neural network to extract the coupled thermal characteristics of the battery and structure, thereby obtaining a spatiotemporal thermal feature matrix.

[0008] The spatiotemporal thermal feature matrix is ​​analyzed for thermal flow interaction between battery cells and metal structures to obtain thermal runaway risk assessment probability values ​​and generate thermal runaway early warning signals.

[0009] The CTC chassis is divided into multiple thermal management zones, and a zoned thermal management scheme for the multiple thermal management zones is constructed based on the thermal runaway early warning signal.

[0010] Secondly, the present invention provides a machine learning-based CTC chassis thermal runaway early warning and control system, the machine learning-based CTC chassis thermal runaway early warning and control system comprising:

[0011] The data acquisition module is used to collect data at the junction of the battery cells and the metal structure in the CTC chassis, and obtain the temperature time series matrix and the structural sensitivity matrix.

[0012] The feature extraction module is used to input the temperature time series matrix and the structure sensitivity matrix into a multi-scale thermal feature transformation neural network to extract the coupled thermal characteristics of the battery and structure, and obtain a spatiotemporal thermal feature matrix.

[0013] The interactive analysis module is used to perform interactive analysis of the heat flow between the battery cell and the metal structure on the spatiotemporal thermal feature matrix, obtain the probability value of thermal runaway risk assessment and generate a thermal runaway early warning signal.

[0014] The module is used to divide the CTC chassis into multiple thermal management zones and to construct a zoned thermal management scheme for the multiple thermal management zones based on the thermal runaway early warning signal.

[0015] The technical solution provided by this invention addresses the battery-structure coupling thermal characteristics of CTC chassis by employing a spatial topology attention layer to specifically handle the interactive heat flow between battery cells and the metal structure. This accurately captures the unique structured heat conduction patterns of the CTC chassis, significantly improving the early identification capability of thermal runaway. A multi-scale thermal feature transformation neural network is used to simultaneously extract features from both fast and slow thermal diffusion phases, fully considering the heat conduction processes at different time scales within the CTC chassis, thus improving the comprehensiveness and accuracy of thermal feature representation. An adaptive hierarchical selection algorithm based on thermal feature correlation evaluation dynamically optimizes the network structure according to the actual thermal conduction characteristics of the CTC chassis, avoiding model redundancy while ensuring no loss of key features, thereby improving the model's accuracy and generalization ability. The global thermal control problem is decomposed into multiple local sub-problems, each region can independently formulate thermal control strategies, and a global thermal balance is maintained through inter-regional coordination mechanisms, achieving a balance between accuracy and real-time performance. This invention can rapidly respond to thermal runaway events, combining rapid local intervention with global stable control, improving the safety and stability of the entire thermal management system. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of an embodiment of the CTC chassis thermal runaway early warning and control method based on machine learning in this invention.

[0018] Figure 2 This is a schematic diagram of an embodiment of the CTC chassis thermal runaway early warning and control system based on machine learning in this invention. Detailed Implementation

[0019] This invention provides a machine learning-based method and system for early warning and control of CTC chassis thermal runaway. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0020] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the CTC chassis thermal runaway early warning and control method based on machine learning in this invention includes:

[0021] Step S101: Collect data at the junction of the battery cell and the metal structure in the CTC chassis to obtain the temperature time series matrix and the structural sensitivity matrix.

[0022] It is understood that the executing entity of this invention can be a machine learning-based CTC chassis thermal runaway early warning and control system, or it can be a terminal or a server; the specific implementation is not limited here. This embodiment of the invention will be described using a server as an example.

[0023] Specifically, a high-density temperature sensor array is deployed within the chassis structure. The sensors are preferentially positioned at the interface between the battery cells and the metal structure to achieve comprehensive coverage of key thermal coupling areas. The deployment principle is based on heat flux density analysis, increasing sensor density in contact interface areas with frequent heat flux abrupt changes, and ensuring at least 30 high-precision thermistors are deployed within the coverage area to establish a spatially sensing temperature monitoring network. With the support of this sensor array, the temperature at each sensor location is sampled in real time at a fixed frequency (e.g., 10Hz), forming a multi-channel raw temperature data matrix. Each row corresponds to one sensor, and each column corresponds to one time step, enabling continuous recording of the chassis thermal field evolution. A sliding window averaging filter algorithm is used to preprocess the raw data matrix. The filtering operation suppresses high-frequency noise fluctuations by calculating the temperature average of several time points before and after each time point. The filter window size is set to 5 sampling periods. After filtering, Z-score normalization is performed on the temperature data of each sensor channel. The data is normalized based on the mean and standard deviation of each channel to eliminate measurement biases caused by differences in thermal inertia at different sensor locations, resulting in normalized temperature data with comparable and balanced scale. The derivative in the time dimension and the gradient in the spatial dimension of the normalized data matrix are calculated separately. The former reflects the rate of temperature change over time, generating a temperature change rate matrix, while the latter constructs a spatially distributed gradient vector field based on the temperature difference between adjacent sensors, forming a temperature spatial distribution matrix. These two matrices are combined and fused through matrix multiplication to obtain a temperature time series matrix with rich thermodynamic characteristics. A heat flow path identification experiment is conducted. Based on the temperature disturbance responses of different battery cells under different power inputs during the experiment, the degree of their projected influence on the temperature at structural points is measured, thereby constructing a structural sensitivity matrix. Each element in this matrix represents the quantitative influence coefficient of a specific battery thermal change on the temperature response at a certain structural location.

[0024] Step S102: Input the temperature time series matrix and the structure sensitivity matrix into the multi-scale thermal feature transformation neural network to extract the coupled thermal characteristics of the battery and structure, and obtain the spatiotemporal thermal feature matrix;

[0025] Specifically, the temperature time series matrix and the structural sensitivity matrix are concatenated along the column direction to generate an input feature matrix with a unified structure. This input feature matrix is ​​then fed into three one-dimensional convolutional sub-networks in a multi-scale thermal feature transformation neural network for multi-scale time feature extraction. Each sub-network is configured with different convolutional kernel sizes, typically 3, 5, and 7. This differentiated receptive field design effectively captures the rapid and slow changes in the propagation of thermal runaway in the CTC chassis. Short convolutional kernels are suitable for extracting rapid heating and transient thermal diffusion signals, while long convolutional kernels are better suited for capturing chronic heat accumulation and low-frequency thermal evolution trends. Together, they construct a multi-scale time feature set covering multiple time scales. This set encodes the dynamic changes in the thermal diffusion rate and pattern under complex operating conditions of the CTC chassis in the feature space. Considering the complex thermal coupling channels between the battery module and structural connection unit in the CTC chassis, a spatial topology graph reflecting the connection relationships between sensors is constructed based on the pre-modeled physical structure of the CTC chassis. In this graph, nodes represent the locations of each sensor, and edges represent thermal connection paths between two nodes. Based on this topology, a spatial topology attention weight matrix is ​​generated by calculating the attention score between each node. Each element in this matrix is ​​normalized using a softmax function to the inner product of the query vector and the key vector, thus reflecting the response strength of a node to the thermal behavior of its neighboring nodes under the current thermal state. This attention weight matrix is ​​applied to a multi-scale temporal feature set, and a weighted calculation of the spatial topology attention mechanism is performed to obtain a spatial feature vector that simultaneously integrates the sensor's temporal variation trend and its spatial thermal coupling information. The spatial feature vector is input into the fully connected layer module of the multi-scale thermal feature transformation neural network. This module consists of two consecutive fully connected hidden layers, containing 128 and 64 neurons respectively. Batch normalization and dropout operations are added after each layer to improve the generalization ability and training stability of the model. The fully connected layer further compresses and transforms the spatial feature vector through nonlinear mapping to form the final spatiotemporal thermal feature matrix with spatiotemporal consistency, thermodynamic correlation and structural specificity.

[0026] Step S103: Perform thermal flow interaction analysis between battery cells and metal structure on the spatiotemporal thermal feature matrix to obtain the probability value of thermal runaway risk assessment and generate thermal runaway early warning signal;

[0027] Specifically, a structured analysis of the spatiotemporal thermal feature matrix is ​​performed. By calculating the correlation between features of different sensor nodes in the matrix, such as using Pearson correlation coefficient or mutual information, a high-dimensional correlation matrix is ​​constructed. Each element of this matrix represents the degree of coordinated change of thermal features of two sensors within a given time window, thereby revealing the topological diffusion characteristics of heat conduction within the battery structure complex. Combining this correlation matrix with the structural sensitivity matrix obtained through previous experimental measurements, the directional distribution information of heat flow is calculated. By mapping the coupling relationship between spatial temperature gradient and structural response into a vector field, this vector field reflects the dominant direction and intensity of heat flow between different sensors, thus characterizing the physical path and dynamic trend of heat source propagation to the metal frame. Based on this heat flow directional distribution information, the influence of each sensor location on the overall thermal state judgment is quantified. A feature importance scoring function is used to integrate heat flow intensity, structural connectivity, and local feature change rate into a comprehensive scoring index. Importance scores are assigned to each feature dimension, and a sensor feature weight distribution vector is constructed accordingly. This weight distribution can highlight the most sensitive measurement points and feature dimensions in the early stages of thermal runaway. All features are filtered based on weight values, eliminating weakly correlated features with weights below a set threshold. Only key features highly sensitive to thermal diffusion trends, energy accumulation phenomena, and abnormal structural temperature rise responses are retained, constructing a target feature subset for early warning discrimination. This target feature subset is input into the terminal classification module of the thermal runaway early warning neural network. The Sigmoid activation function outputs the probability value that the current thermal state belongs to the thermal runaway early warning category. This probability value is the thermal runaway risk assessment probability value, with a value between 0 and 1 reflecting the likelihood of thermal runaway occurring. This probability value is compared with a preset risk threshold. If the thermal runaway probability value exceeds the set threshold level, the system automatically generates a thermal runaway early warning signal. This signal will trigger a series of response mechanisms in the chassis thermal control module, including active cooling, load adjustment, and thermal channel reconfiguration, thereby achieving early intervention and control of thermal runaway risk.

[0028] Step S104: Divide the CTC chassis into multiple thermal management zones, and construct a zoned thermal management scheme for multiple thermal management zones based on thermal runaway early warning signals.

[0029] Specifically, based on the structural topology of the CTC chassis itself, a thermal connectivity map is constructed by analyzing the coupling relationships between battery cells, structural frames, and heat conduction paths. Building upon this, using heat flux intensity as the dividing criterion, a minimum heat flux path cutting operation is performed to divide the chassis into multiple relatively independent thermal management regions. These regions have strong internal thermal connectivity, while the heat flux interaction between regions is weak, thus achieving spatial decoupling in thermal management. After division, a state space is defined for each thermal management region. This state space includes temperature data from multiple temperature sampling points within the region, as well as information on the inflow and outflow intensity of heat within the region and thermal stress data of structural components due to temperature changes, reflecting the current thermodynamic state of the region. Simultaneously, a corresponding action space is designed for each region to describe the applicable control measures, including power adjustment parameters for the active cooling system to directly reduce the region temperature; load power limiting parameters for the battery system to reduce heat source intensity; and structural thermal conductivity adjustment parameters to change the heat flow propagation mode by adjusting thermally conductive materials or structural paths. Based on this, the difference between the current heat flow input and output in each thermal management zone is calculated to obtain the thermal energy balance state between zones. A global summary analysis yields the vehicle's overall thermal distribution equilibrium index, used to assist in assessing whether the overall system is in a stable thermal state or faces the risk of heat accumulation. When the system detects a thermal runaway risk signal based on the early warning model, the control module enters a response state. At this time, using the thermal state of each zone as input and combining the early warning level, dynamic control planning is carried out at both the strategy and execution layers. The strategy layer uses reinforcement learning algorithms as its core, generating corresponding control strategy directions based on the current thermal state and risk signal of the zone, such as prioritizing cooling, limiting heat generation, or dispersing heat flow. The execution layer calculates the specific adjustment action magnitude through model prediction and real-time feedback to ensure effective containment of potential thermal runaway trends within a limited response time. By integrating the status information, control strategy parameters, and global thermal balance indicators of each region, a thermal management execution plan with regional collaborative characteristics is constructed. The control commands that each region should take in the current cycle are clearly defined and sent to each region controller via a communication bus to achieve real-time response. This forms a highly efficient thermal management mechanism with regional autonomy and collaborative linkage in the entire chassis system, improving the thermal runaway response speed and regional thermal safety level.

[0030] The system continuously monitors and records the heat inflow and outflow of each thermal management zone. Heat inflow originates from heat conducted from adjacent zones, battery heating within the zone, or the influence of ambient temperature. Heat outflow includes active cooling, heat diffusion, and heat conduction to external structures. Based on this data, heat flow difference calculations are performed on a zone-by-zone basis to calculate the current heat flow balance deviation value for each thermal management zone. This value accurately reflects whether the zone is in a state of heat accumulation, heat dissipation, or thermal stability. The heat flow balance deviation values ​​of all zones are globally aggregated to form a total balance deviation value used to evaluate the overall thermal distribution of the CTC chassis. The smaller this value, the closer the overall thermal state of the system is to equilibrium. To more precisely characterize the heat conduction relationships between zones, neighborhood information exchange calculations are performed based on real-time temperature and heat flow differences at zone boundary nodes. This assesses the thermal coupling strength between any two adjacent zones, yielding a zone heat flow interaction strength index. This index helps identify zones with high thermal dependence or significant heat exchange under current operating conditions. Based on the control capabilities of each region within the current control cycle, such as cooling regulation, battery load limiting, or structural thermal conductivity regulation, the set of executable actions for each region is extracted from the action space. These actions are then matched with the intensity index of inter-regional heat flow interaction to identify regions with close heat flow connections and joint control capabilities. This process constructs a target region network topology, reflecting the network structure where there is a possibility of coordinated control among regions in a thermodynamic sense. Based on this target region network topology, the response consistency of each region's nodes during heat conduction is evaluated. By analyzing the differences in heat flow conduction balance between nodes, their dispersion is calculated, determining whether the thermal behavior between region nodes is coordinated and whether there is uneven heat migration. This yields a quantitative index of global heat distribution consistency; a lower index indicates a more uniform heat energy distribution between regions and a more ideal system thermal control effect. The three core evaluation values ​​mentioned above—namely, the total balance deviation value representing the degree of difference in the overall heat input and output of the chassis, the heat flow interaction intensity index reflecting the tightness of thermal coupling between adjacent areas, and the global heat distribution consistency evaluation index measuring the synchronicity of thermal behavior in multiple areas—are fused together according to preset weights to output an integrated global heat distribution equilibrium index.

[0031] A multi-level response mechanism is established in the thermal runaway early warning module. This mechanism classifies the current thermal state based on the probability value of thermal runaway risk assessment and the level judgment rules of the corresponding thermal runaway early warning signal, generating a thermal control level identifier. This identifier is used to express the current thermal risk level faced by the CTC chassis, such as general warning, moderate risk, or emergency intervention. The classification logic of the thermal control level identifier not only refers to the probability value range of the neural network output, but also combines dynamic characteristics such as heat diffusion rate, the range of the heated area of ​​the structure, and the temperature rise trend to ensure that the early warning level judgment is timely and forward-looking. Based on the current thermal control level identifier of the CTC chassis, a matching set of control command candidates for each area is selected from the established zonal thermal management scheme. These candidate sets are pre-configured sets of executable actions for different early warning levels, covering a series of control measures such as cooling power adjustment, battery power current limiting, thermal channel change, and structural heat conduction strategy. The selection strategy of candidate commands will fully consider the intensity of thermal control response and energy consumption balance, ensuring that unnecessary power waste or resource overload is reduced while meeting control requirements. To achieve parallel optimization and distributed execution of thermal control tasks, the global thermal control problem is divided into several local subproblems, each corresponding to an independent thermal management region. This division not only improves response speed and computational efficiency but also facilitates differentiated processing of each region based on its own state. During the establishment of local subproblems, a set of control optimization equations is constructed for each thermal management region. The design of these equations comprehensively considers internal temperature changes, heat flow distribution, structural stress state, and heat exchange behavior between adjacent regions. The goal is to achieve precise regulation of local thermal dynamics by control commands while ensuring a safe temperature range and structural stability. In the solution phase of control optimization, an iterative solution process is executed using the current candidate instruction sets for each region as input. Constraint analysis and variable adjustments are performed on each set of control optimization equations. In each iteration, the control commands are continuously corrected to approach the optimal control solution. The entire process gradually converges through multiple iterations, ultimately outputting a real-time control command sequence covering all thermal management regions.

[0032] In this embodiment of the invention, targeting the battery-structure coupled thermal characteristics of the CTC chassis, a spatial topology attention layer is used to specifically handle the interactive heat flow between the battery cells and the metal structure. This accurately captures the unique structured heat conduction patterns of the CTC chassis, significantly improving the early identification capability of thermal runaway. A multi-scale thermal feature transformation neural network is employed to simultaneously extract features from both fast and slow thermal diffusion phases, fully considering the heat conduction processes at different time scales within the CTC chassis, thus improving the comprehensiveness and accuracy of thermal feature representation. An adaptive hierarchical selection algorithm based on thermal feature correlation evaluation dynamically optimizes the network structure according to the actual thermal conduction characteristics of the CTC chassis, avoiding model redundancy while ensuring that key features are not lost, thereby improving the model's accuracy and generalization ability. The global thermal control problem is decomposed into multiple local sub-problems, each region can independently formulate a thermal control strategy, while maintaining global thermal balance through inter-regional coordination mechanisms, achieving a balance between accuracy and real-time performance. This invention can quickly respond to thermal runaway events, combining rapid local intervention with global stable control, improving the safety and stability of the entire thermal management system.

[0033] In one specific embodiment, the process of performing step S101 may specifically include the following steps:

[0034] A temperature sensor array is deployed at the junction of the battery cell and the metal structure within the CTC chassis, and temperature data sampling is performed based on the temperature sensor array to obtain the original temperature data matrix.

[0035] The original temperature data matrix is ​​subjected to sliding window averaging and Z-Score normalization to obtain normalized temperature data.

[0036] The time-dimension derivative and spatial-dimension gradient are calculated based on the normalized temperature data to obtain the temperature change rate matrix and the temperature spatial distribution matrix.

[0037] The temperature change rate matrix and the temperature spatial distribution matrix are multiplied to obtain the temperature time series matrix. The influence coefficient of the thermal change of the battery cell on the temperature of the structural point is determined by heat flow experiment to obtain the structural sensitivity matrix.

[0038] Specifically, a high-sensitivity, high-sampling-rate temperature sensor array is deployed in key areas of the chassis structure, namely the junction between the battery cells and the metal structure. The sensor deployment principle is to cover all battery modules and structural connection nodes, with localized densification based on heat flux density distribution patterns. Particular emphasis is placed on areas with prominent thermal coupling, such as battery casing edges, module seams, and frame connection plates, to increase sensor density and enhance spatial resolution of data acquisition. For example, by deploying one temperature sensor every 25 square centimeters, a high-density network of dozens or even hundreds of measurement points is formed throughout the chassis, establishing a widely covered and rapidly responding thermal monitoring system. This temperature sensor array samples the temperature at each deployment point at fixed time intervals. Each sensor generates a temperature value in each sampling cycle, and the measurements from all measurement points on the same time axis sequentially constitute a temperature sampling sequence, forming the original temperature data matrix. The rows of this matrix represent sensor numbers, and the columns represent time series, recording the continuous evolution of the internal heat distribution of the CTC chassis in both time and space. The original temperature data matrix is ​​then denoised and normalized to enhance the stability and comparability of subsequent analyses. A sliding window averaging filter algorithm is introduced to preprocess the temperature data. The filter window size is set to five sampling points. By weighted averaging of multiple measurements before and after each time point, high-frequency noise caused by environmental disturbances, electromagnetic interference, or hardware inaccuracies is effectively suppressed, improving the continuity and smoothness of the data. After filtering, Z-Score normalization is performed on the temperature sequence of each sensor channel. By subtracting the mean of each data set and dividing by its standard deviation, a unified scale transformation is achieved for temperature measurements at different sensor locations. This allows for lateral comparison of thermal changes between measurement points and eliminates measurement biases caused by differences in installation location, background temperature, or material heat capacity. After this stage of processing, normalized temperature data is obtained. The time-dimensional derivative and spatial-dimensional gradient are calculated based on the normalized temperature data. In the time dimension, the first derivative operation is performed on the normalized temperature sequence of each sensor, i.e., the rate of temperature change over time is calculated, resulting in a temperature rate of change matrix. This matrix reflects the trend, rate, and acceleration of thermal state changes, and is used to identify the speed of the heating process, abrupt changes in local overheating, and nonlinear stages of thermal diffusion. In the spatial dimension, by analyzing the temperature difference between adjacent sensors, a spatial temperature gradient field is constructed. That is, the temperature difference between each measuring point and its adjacent points is divided by the physical distance to form a temperature spatial distribution matrix, thereby revealing the direction, speed and intensity of heat diffusion along the structural path in the chassis. This matrix can not only reflect the direction of heat flow, but also reflect whether there are abnormal phenomena such as non-uniform accumulation, thermal bridge or thermal conduction barrier in the thermal field.To enhance the modeling capability of chassis heat conduction paths, the temperature change rate matrix and the temperature spatial distribution matrix are multiplied together, and the temporal variation and spatial diffusion are jointly encoded into a composite feature, resulting in a temperature time series matrix. Through heat flow experiments, the battery cells are heated and controlled under multiple typical thermal conditions, and the response characteristics of structural nodes to changes in the heat source are recorded by the temperature response of these nodes. Through multiple experimental comparisons and analyses, the quantitative impact of the heating of different battery cells under different conditions on the temperature changes of various structural measuring points is analyzed, establishing a structural sensitivity matrix. Each element of this matrix reflects the contribution of a specific battery thermal disturbance to the temperature of a particular structural point, serving as an important physical parameter for measuring the intensity of the battery heat source's thermal response to the structure.

[0039] In one specific embodiment, the process of performing step S102 may specifically include the following steps:

[0040] The temperature time series matrix and the structure sensitivity matrix are concatenated to obtain the input feature matrix;

[0041] The input feature matrix is ​​input into three one-dimensional convolutional sub-networks in the multi-scale thermal feature transformation neural network to extract multi-scale time features, resulting in a multi-scale time feature set containing fast thermal diffusion phase and slow thermal diffusion phase.

[0042] Based on the CTC chassis topology, a spatial topology attention weight matrix is ​​constructed, and spatial topology attention is calculated on the multi-scale time feature set based on the spatial topology attention weight matrix to obtain a spatial feature vector that integrates battery and structural interaction heat flow information.

[0043] The spatial feature vector is input into a fully connected neural network of a multi-scale thermal feature transformation neural network to perform feature transformation, thereby obtaining the spatiotemporal thermal feature matrix.

[0044] Specifically, the temperature time series matrix and the structural sensitivity matrix are concatenated to form a high-dimensional input feature matrix composed of multiple channels and variables. Each row of this feature matrix represents the location of a sensor or structural node, and each column represents the temperature change pattern of that node within a certain time period, its thermal diffusion trend, and its sensitivity to the thermal response of other nodes. The input feature matrix is ​​then fed into the temporal convolution module of the multi-scale thermal feature transformation neural network. This module consists of three one-dimensional convolutional sub-networks, each with a convolution kernel size set to a different scale. For example, smaller convolution kernels capture short-term temperature rise fluctuations during rapid thermal diffusion, while larger kernels capture temperature change trends during slow thermal conduction, thereby extracting high-frequency and low-frequency temporal features respectively. These convolutional sub-networks operate independently in parallel, performing sliding convolution operations on the same input feature matrix. Each convolution operation can sense temperature change patterns, transient disturbance responses, or slow heat accumulation effects within a local time window, forming a multi-scale temporal feature set. This feature set is expressed as three sets of temporal feature tensors with the same dimension but different sensing ranges, representing the thermal diffusion behavior within the CTC chassis at different time scales, and reflecting the non-uniform and nonlinear thermal dynamic evolution process caused by physical coupling between the battery and the structural system. In the highly integrated environment of the CTC chassis, the thermal diffusion process not only exhibits temporal continuity and multi-scale characteristics, but also spatial structural dependence and asymmetric interactions between nodes. Therefore, to extract the distribution pattern of thermal diffusion in the spatial structure, a spatial topological attention mechanism is introduced after temporal convolution. This mechanism constructs a structural topology graph based on the actual physical layout of the chassis, treating each sensor or structural node as a vertex in the graph, and the thermal connections between nodes as edges, thus generating a chassis thermal topology graph. Based on this graph structure, a spatial topological attention weight matrix is ​​calculated, where each element measures the importance of one node to the thermal response of another node. The calculation process of this attention weight not only considers the physical distance and thermal connection strength between nodes, but also introduces parameters such as the current temperature gradient, response speed, and historical heat conduction path of the node, achieving spatial focusing of attention by dynamically adjusting the weight values. The attention weight matrix is ​​applied to the multi-scale temporal feature set obtained in the previous step, and the thermal features of each spatial node are weighted and summed to generate a new spatial feature vector. This vector integrates the temperature change patterns on the time scale with the thermal interaction relationships in the spatial dimension, effectively capturing how thermal anomalies in the CTC chassis propagate, converge, and feed back between nodes, especially demonstrating higher expression accuracy and risk identification capabilities in multi-node coordinated heating or asynchronous response scenarios. The spatial feature vector is then input into a fully connected neural network of a multi-scale thermal feature transformation neural network for feature transformation.This part consists of two or more fully connected layers. Through continuous linear transformations and nonlinear activation operations, it reduces the dimensionality, compresses, and semantically reconstructs the input features, mapping the original high-dimensional redundant features into more compact embedding vectors with stronger classification capabilities. Batch normalization and Dropout mechanisms are introduced into each fully connected layer to improve training stability and model generalization ability. In this process, the neural network continuously enhances its response to key features while suppressing invalid or redundant local changes, ultimately outputting a spatiotemporal hot feature matrix with high abstraction capabilities.

[0045] In one specific embodiment, the process of executing step S103 may specifically include the following steps:

[0046] The correlation between the location features of different sensors is calculated based on the spatiotemporal thermal feature matrix to obtain correlation data reflecting the thermal conduction topology of the CTC chassis;

[0047] The heat flow direction distribution information is calculated based on correlation data and structural sensitivity matrix. The heat flow direction distribution information is used to characterize the directional flow characteristics of heat conduction at the interface between the battery and the structure.

[0048] The feature importance score of each sensor location is calculated based on the heat flow direction distribution information, and the sensor feature weight distribution is generated based on the feature importance score.

[0049] Feature filtering is performed based on sensor feature weight distribution to obtain a target feature subset for thermal runaway early warning. The probability value of the current thermal state belonging to the thermal runaway early warning state is calculated based on the target feature subset to obtain the thermal runaway risk assessment probability value.

[0050] The probability value of thermal runaway risk assessment is compared with a preset threshold. When the probability value exceeds the preset threshold, a thermal runaway warning signal is generated.

[0051] Specifically, the correlation between sensor locations is calculated based on the spatiotemporal thermal feature matrix. The correlation assessment employs statistical correlation quantification methods, such as calculating the covariance or correlation coefficient of different sensor nodes along the thermal feature dimension, to measure the degree of coordination in temperature evolution trends among multiple sensors within the same time window. When the thermal feature changes of two sensors are highly synchronized, it is inferred that they have a close coupling relationship along the heat conduction path; conversely, if their thermal response modes differ significantly, it indicates that they are located in thermal isolation zones or areas of strong heat dissipation. By calculating the correlation between all pairs of sensors, a high-dimensional correlation matrix reflecting the interconnected structure of the CTC chassis's internal thermal behavior is constructed. Based on this, the aforementioned correlation data is coupled with the structural sensitivity matrix obtained through previous thermal flow experiments to calculate the heat flow direction distribution information. The correlation between the thermal features of sensor nodes is combined with the intensity of the thermal response influence between the battery and the structure, and the dominant flow direction of heat flow between each node is quantified by constructing a weighted vector field. For example, in a certain area, if a structural node exhibits high thermal response sensitivity to multiple battery nodes, and these battery nodes show a high correlation in temperature changes, it is inferred that the structural node is a concentrated response point on these heat flow paths, thus forming a directional heat flow propagation trend. Based on this reasoning logic, the system constructs a heat flow direction map for the entire chassis area. Each connection path not only reflects whether heat is transferred but also specifically indicates its directionality and influence weight, thereby obtaining a dynamic, adjustable, and physically interpretable set of heat flow direction distribution information, depicting the conduction law of heat energy between the battery module and the chassis structure interface. Based on the obtained heat flow direction distribution information, the system evaluates the role strength and early warning value of each sensor node in the overall thermal network. For each node, the system calculates parameters such as the number of critical paths it undertakes in the entire heat flow conduction network, the cumulative value of its thermal response weight, and the connection strength with surrounding nodes. These data are then combined to generate a feature importance score. This score reflects the dominance of a node in heat conduction; a higher score indicates a more critical position in the heat anomaly propagation chain. Based on feature importance scores, a feature weight distribution map of sensor nodes is constructed. This map highlights the most representative measurement point locations for thermal runaway detection under the current operating conditions, while eliminating irrelevant nodes that are stable or have a weak impact on the current thermal field distribution. According to the sensor feature weight distribution, all thermal features are filtered, and the feature vectors corresponding to several nodes with the highest weight ranking and the most sensitive performance are selected to form a target feature subset. This target feature subset is input into the output layer of the thermal runaway classification module or neural network. The trained classifier is then used to probabilistically classify the current thermal state, calculating the probability that the thermal state belongs to the thermal runaway warning category, thus obtaining a thermal runaway risk assessment probability value. This probability value is numerically between 0 and 1; the closer it is to 1, the higher the risk of thermal runaway; the closer it is to 0, the more stable the system is.The probability value of the thermal runaway risk assessment is compared in real time with a pre-set risk discrimination threshold. This threshold is optimized based on historical thermal runaway data and model accuracy indicators. Once the current probability value exceeds the threshold, it indicates that an abnormal state with a tendency for thermal runaway has occurred in the chassis, and the system immediately triggers a thermal runaway warning signal.

[0052] In one specific embodiment, the process of executing step S104 may specifically include the following steps:

[0053] Multiple thermal management zones are obtained by cutting the minimum heat flow path based on the CTC chassis topology.

[0054] Define the state space for each thermal management region. The state space includes the temperature vector, heat flow vector, and structural thermal stress vector within the region. At the same time, define the corresponding action space for the state space. The action space includes active cooling power regulation, battery load power limitation, and structural thermal conductivity enhancement control.

[0055] Calculate the difference between input and output heat flow in each heat management zone to obtain the global heat distribution balance index;

[0056] A two-layer control architecture, comprising a strategy layer and an execution layer, is constructed based on the state space and action space. Control strategy parameters for each region are generated based on the thermal runaway early warning signal and the two-layer control architecture.

[0057] Based on the control strategy parameters of each region and the global heat distribution balance index, create a zoned heat management scheme for multiple heat management zones.

[0058] Specifically, based on the thermal structural characteristics of the CTC chassis, a thermal conduction topology is constructed according to the physical connections between its internal battery modules, structural connectors, cooling pipes, heat-conducting plates, and other thermally coupled components. This topology uses temperature monitoring nodes or thermal control elements as nodes, and establishes edges based on the thermal conduction paths or thermal coupling relationships between nodes, forming a comprehensive graph structure model describing the heat flow distribution of the CTC chassis. Based on this thermal topology, the heat flow intensity represented by all edges in the graph is calculated, and physical properties such as thermal resistance, heat capacity, structural density, and material thermal conductivity are introduced to weight the thermal channels in the entire network. Applying the minimum cut method in graph theory, with the goal of minimizing the total cross-regional heat flow, a heat flow path cutting operation is performed, dividing the CTC chassis into several sub-regions with high thermal connection strength and minimal inter-regional heat exchange. This forms multiple independent thermal management regions that maintain boundary thermal controllability. Each region performs dynamic temperature regulation, energy distribution, and stress control independently. A state space is defined for each thermal management zone. This state space must comprehensively describe the current thermophysical state, dynamic behavior, and thermal stress effects of the zone, including three dimensions: a temperature vector, which is the set of temperature values ​​currently collected by all sensors within the zone, used to characterize the instantaneous distribution and trend changes of the local thermal state; a heat flow vector, representing the heat input and heat output intensity per unit time within the zone, reflecting the direction of heat exchange and the velocity of energy flow; and a structural component thermal stress vector, which, through analysis of the temperature gradient of heated components such as metal supports and thermally conductive connecting plates, infers the potential thermal expansion and contraction or structural deformation trends, thus extending thermal safety issues to the structural safety level. Simultaneously, to effectively regulate the regional thermal state, an action space corresponding to the state space is defined. This action space includes three core control methods: active cooling power adjustment, which changes the local heat dissipation rate by adjusting the output power of the air-cooling or liquid-cooling system; battery load power limitation, used to reduce the local battery operating power to reduce internal heat generation; and structural thermal conductivity enhancement control, which enhances the heat diffusion rate in the structure by adjusting the state of thermally conductive materials, switching thermal paths, or changing the state of thermal interfaces. After obtaining the status of each thermal management zone, the difference between the heat flow input and output of each zone is calculated. By monitoring the heat flow intensity entering and leaving the thermal channels at the boundary of each zone per unit time, the heat flow balance deviation value of each zone is obtained. The heat flow deviation values ​​of all zones are weighted and summed to form a global thermal distribution balance index. This index is used to assess whether there are problems such as severe heat accumulation, uneven distribution, or thermal resistance imbalance between zones in the entire CTC chassis.A two-layer control architecture, comprising a strategy layer and an execution layer, is constructed based on state space and action space. The strategy layer is responsible for establishing decision-making logic and formulating intelligent strategies. It employs reinforcement learning algorithms or neural policy networks, learning the mapping relationship between thermal diffusion trends and control behavior by inputting the current state space and thermal runaway early warning signals, and outputting the optimal control direction. The execution layer, based on the direction information provided by the strategy layer and combined with the actual physical model, uses model predictive control methods to perform real-time rolling predictions of the current temperature, heat flow, and stress state. Within the prediction time window, it selects a set of control command sequences that meet boundary conditions, resource constraints, and optimization objectives to precisely adjust the cooling power, power load, and heat conduction channel state of each region. Through the collaborative action of the strategy and execution layers, the system intelligently generates control strategy parameters for each thermal management region based on different thermal control levels and dynamic thermal states. These parameters include specific executable content such as cooling response level, battery power adjustment ratio, and structural heat conduction switching status. Based on the control strategy parameters of each region and the aforementioned global thermal distribution equilibrium index, a zoned thermal management scheme with dynamic adaptability, cross-regional collaboration, and risk perception capabilities is created by comprehensively considering the overall thermal network load and local response capabilities. The scheme clearly defines the control measures required for each region during the current control cycle, the tolerable temperature fluctuation range, the stress time window, and the thermal interconnection strategy of adjacent regions. At the same time, it maintains the dynamic balance of the overall temperature distribution and the controllability and predictability of structural stress at the system level, thereby achieving the goal of ensuring the thermal safety operation of the CTC chassis under complex working conditions and the coordinated control of multiple regions.

[0059] In one specific embodiment, the process of calculating the difference between input and output heat flow in each thermal management zone to obtain the global heat distribution balance index can specifically include the following steps:

[0060] The difference between the heat inflow and heat outflow in each heat management zone is calculated to obtain the heat flow balance deviation value of each zone.

[0061] The heat flow balance deviation values ​​of each region are summed to obtain the total balance deviation value that reflects the degree of heat distribution uniformity of the entire CTC chassis.

[0062] Based on the temperature and heat flow differences between adjacent regions, neighborhood information exchange calculations are performed to obtain the intensity index of heat flow interaction between regions.

[0063] By matching the intensity index of heat flow interaction between regions with the set of executable actions in each region of the action space, strong interaction pairs of adjacent regions are identified, and the network topology of the target region is obtained.

[0064] The dispersion of heat flow conduction balance index between nodes in each region of the target region network topology is calculated to obtain the global heat distribution consistency evaluation index.

[0065] The total balance deviation value, the intensity index of inter-regional heat flow interaction and the global heat distribution consistency evaluation index are weighted and fused to obtain the global heat distribution balance index.

[0066] Specifically, the difference between the heat inflow and heat outflow in each thermal management zone is calculated. Heat inflow includes heat energy conducted from adjacent zones through boundary nodes, heat generated by battery cells within the zone during operation, and heat backflow from structural components. Heat outflow mainly consists of heat carried away by the active cooling system, heat energy diffused to surrounding structures, and energy released through thermal channels. By calculating the difference between heat inflow and heat outflow, the thermal balance deviation value for each thermal management zone in the current cycle is obtained. The closer this deviation value is to zero, the more thermally balanced the zone is; conversely, a larger absolute value indicates more severe heat accumulation or overcooling, requiring adjustment and control. After obtaining the thermal balance deviation values ​​for all zones, the deviation values ​​are summed to calculate a total global thermal distribution balance deviation value. This value is considered a coarse-grained indicator of the consistency of thermal distribution across the entire CTC chassis system at the current moment. If this indicator is significantly higher than the safety setting range, it indicates that the current thermal management system has failed to achieve coordinated control between multiple regions, resulting in some regions experiencing excessively high temperatures while other regions waste heat dissipation capacity. Overall optimization and resource reallocation in the control strategy are required. Based on the temperature differences and actual heat flow exchange differences at the boundaries of each thermal management region, neighborhood information exchange calculations are performed to obtain the inter-regional heat flow interaction intensity index. This calculation process considers the physical characteristics of heat conduction, i.e., the temperature gradient is the direct driving factor for the direction and intensity of heat flow. It also combines parameters such as the thermal conductivity of the heat transfer path, structural coupling relationships, and thermal conductivity area to comprehensively determine whether there is a significant heat flow interdependence between two adjacent regions. If a persistent temperature difference exists between two regions and the heat flow intensity is significant, it indicates that they exhibit strong coupling characteristics in thermodynamic behavior, requiring coordinated adjustment in subsequent control strategies. Matching analysis is performed based on the inter-regional heat flow interaction intensity index and the set of executable actions for each region in the action space. The action space encompasses the control capabilities currently available to each region, such as adjustable cooling capacity, battery power limiting capabilities, and heat conduction enhancement methods. By matching the executable actions with the thermal coupling strength between regions, it identifies which adjacent region pairs currently possess both strong thermodynamic coupling and the feasibility of synchronous linkage in terms of control methods. These region pairs, once extracted, constitute the target region network topology graph, where nodes represent regions and edges represent co-controllable coupling relationships. Based on this target region network topology graph, a dispersion analysis is performed on the heat flow conduction balance index between nodes in the graph. By quantitatively calculating the differences in heat conduction direction, intensity, and control response along each edge, the degree of consistency in heat flow conduction among the regions in the graph is determined.If significant differences in heat flow exist between certain nodes and control actions fail to synchronize, it indicates a heat conduction imbalance in that region, causing heat energy to concentrate at a particular node, forming a hotspot. Conversely, if the differences in heat flow distribution between nodes are small and the control behaviors are highly coordinated, it indicates that the thermal network structure is operating stably. By accumulating and standardizing these inconsistencies in local thermal behavior, a global heat distribution consistency evaluation index is generated. The lower the index value, the higher the internal coordination of the thermal management system and the more balanced the temperature control; conversely, a higher value suggests the need to strengthen the control coupling strategy between certain regions. The three core data indicators—total balance deviation, inter-regional heat flow interaction intensity, and global heat distribution consistency evaluation index—are weighted and integrated to construct a unified global heat distribution equilibrium index.

[0067] In one specific embodiment, the method for implementing CTC chassis thermal runaway early warning and control based on machine learning further includes the following steps:

[0068] The warning level is determined by classifying the thermal runaway warning signal and the probability value of thermal runaway risk assessment, thus obtaining the current CTC chassis thermal control level identifier.

[0069] Based on the current CTC chassis thermal control level identifier, select the candidate set of control commands for each zone from the zone thermal management scheme;

[0070] The global thermal control problem is decomposed into N local subproblems, each of which corresponds to a thermal management region, and a set of control optimization equations for the N local subproblems is created.

[0071] The control optimization equations are iteratively solved based on the candidate sets of control commands for each region, and a real-time control command sequence is output.

[0072] Specifically, based on the thermal runaway warning signal and the thermal runaway risk assessment probability value, a comprehensive risk level judgment is made, and the current chassis thermal control status is classified. The thermal runaway warning signal is the activation output generated by a multi-scale thermal feature neural network after detecting that the combination of multiple thermophysical features such as temperature, thermal gradient, and structural sensitivity has reached a critical mode. It is expressed as a Boolean value or a trigger flag indicating whether a high-risk zone has been entered. The thermal runaway risk assessment probability value is a quantitative result of the confidence that the current state belongs to the thermal runaway risk state. Its value is between 0 and 1 and can be divided into multiple intervals for further classification of risk levels. On this basis, a set of warning level classification standards is set. For example, a probability value between 0.0 and 0.6 is defined as a normal state; between 0.6 and 0.75 is classified as a Level 1 warning state, indicating that a slight thermal anomaly has not yet caused structural impact; between 0.75 and 0.9 is classified as a Level 2 warning state, indicating that there is a trend of heat accumulation and the possibility of local propagation; and a state above 0.9 is classified as a Level 3 warning state, representing that a structural coupling effect has been triggered and may quickly lead to global thermal runaway. By combining the above evaluation results with the thermal runaway early warning signal, a thermal control level identifier for the current CTC chassis is generated. Once the thermal control level identifier is clear, it is used as an index to select a candidate set of control commands matching that level from a predefined zoned thermal management strategy library. This candidate set is organized by region, with each region receiving different levels of control commands based on its risk level. For example, in a Level 1 early warning state, only low-power cooling and light load limiting measures are activated; in a Level 2 early warning state, local heat flow channel reconstruction and structural heat conduction enhancement functions are activated; and in a Level 3 early warning state, all emergency control measures are activated, including forced current limiting, full-power cooling, and fault isolation commands. The selected candidate set will serve as the parameter boundary or feasible control action set for each region controller during optimization scheduling and execution calculations, providing a constraint basis for subsequent control strategy solutions. The thermal control problem of the entire CTC chassis is decomposed from a global perspective, constructing a multi-sub-problem joint optimization model. The entire chassis thermal management system is divided into N independent thermal management regions, each region being modeled as a subsystem, forming N corresponding local control sub-problems. The core objective of each sub-problem is to ensure that the internal temperature of the region remains within a safe range, the heat flux is stably distributed, and the structural stress is controlled within allowable thresholds, while achieving the control objective while minimizing energy consumption and control intensity. Based on this objective, a corresponding set of control optimization equations is created for each sub-problem. This set of equations includes the current temperature state, heat flux boundary conditions, thermal stress constraints, the influence model of control actions, and thermal coupling feedback from adjacent regions, constraining a multivariate optimization problem jointly influenced by dynamic state, spatial structural relationships, and control response limitations. Each set of control equations, combined with the aforementioned candidate control instruction set, is used to constrain the solution space, ensuring that all control strategies are engineering-feasible, logically sound, and responsive.After modeling is completed, the current state of each thermal management region is used as the initial input, along with the corresponding candidate set of control commands, to initiate a multi-region parallel control solution process. During the solution process, each local control equation set evaluates the cost function value based on the current region's temperature, heat flux, and structural state. The system performs iterative optimization calculations based on these cost functions, updating control command variables, adjusting action combination strategies, and evaluating the expected temperature response and heat flux change trends in real time during each iteration. Simultaneously, to ensure that the control behaviors of different regions do not conflict, consistency correction is performed on the boundary heat flux and structural response between adjacent regions in each iteration. This involves data synchronization and feedback constraint adjustment at the inter-regional conduction channels, ensuring that the local control strategy maintains thermal power continuity and structural coordination within the global heat flux network. The entire iterative process is scheduled using a convergent control mechanism. When the control solutions of all regions simultaneously satisfy the current temperature response target, power constraint, and regional boundary thermal consistency criterion, the iteration is considered to have reached the optimal solution, and the system outputs the real-time control command sequence for the current cycle. The control command sequence will be sent to each area controller or execution unit in real time. The command content includes the output power of the cooling system, the adjustment range of the battery load, the switching status of the heat conduction path, etc. In each cycle, it will receive thermal status feedback for closed-loop correction, forming a multi-area thermal control strategy framework with hierarchical response, autonomous adaptation and collaborative linkage.

[0073] The above describes the CTC chassis thermal runaway early warning and control method based on machine learning in the embodiments of the present invention. The following describes the CTC chassis thermal runaway early warning and control system based on machine learning in the embodiments of the present invention. Please refer to [link / reference]. Figure 2 One embodiment of the CTC chassis thermal runaway early warning and control system based on machine learning in this invention includes:

[0074] The data acquisition module 201 is used to acquire data at the junction of the battery cell and the metal structure in the CTC chassis to obtain a temperature time series matrix and a structural sensitivity matrix.

[0075] The feature extraction module 202 is used to input the temperature time series matrix and the structure sensitivity matrix into the multi-scale thermal feature transformation neural network to extract the coupled thermal characteristics of the battery and structure, and obtain the spatiotemporal thermal feature matrix.

[0076] The interactive analysis module 203 is used to perform interactive analysis of the heat flow between the battery cell and the metal structure on the spatiotemporal thermal feature matrix, obtain the probability value of thermal runaway risk assessment and generate a thermal runaway early warning signal.

[0077] Module 204 is used to divide the CTC chassis into multiple thermal management zones and to construct a zoned thermal management scheme for multiple thermal management zones based on thermal runaway early warning signals.

[0078] Through the collaborative efforts of the aforementioned components, this invention addresses the battery-structure coupling thermal characteristics of the CTC chassis. A spatial topology attention layer specifically handles the interactive heat flow between battery cells and the metal structure, accurately capturing the unique structured heat conduction patterns of the CTC chassis and significantly improving early thermal runaway identification capabilities. A multi-scale thermal feature transformation neural network is employed to simultaneously extract features from both fast and slow thermal diffusion phases, fully considering the heat conduction processes at different time scales within the CTC chassis, thus improving the comprehensiveness and accuracy of thermal feature representation. An adaptive hierarchical selection algorithm based on thermal feature correlation evaluation dynamically optimizes the network structure according to the actual thermal conduction characteristics of the CTC chassis, avoiding model redundancy while ensuring no loss of key features, thereby improving model accuracy and generalization ability. The global thermal control problem is decomposed into multiple local sub-problems, each region can independently formulate thermal control strategies, while maintaining global thermal balance through inter-regional coordination mechanisms, achieving a balance between accuracy and real-time performance. This invention can rapidly respond to thermal runaway events, combining rapid local intervention with global stable control, improving the safety and stability of the entire thermal management system.

[0079] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0080] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a machine learning-based CTC chassis thermal runaway early warning and control device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0081] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A machine learning-based method for early warning and control of CTC chassis thermal runaway, characterized in that, include: Data was collected at the junction of the battery cells and the metal structure in the CTC chassis to obtain the temperature time series matrix and the structural sensitivity matrix. The temperature time series matrix and the structure sensitivity matrix are input into a multi-scale thermal feature transformation neural network to extract the coupled thermal characteristics of the battery and structure, thereby obtaining a spatiotemporal thermal feature matrix. The process involves performing thermal flow interaction analysis between the battery cells and the metal structure on the spatiotemporal thermal feature matrix to obtain a thermal runaway risk assessment probability value and generate a thermal runaway early warning signal. This includes: calculating the correlation between features at different sensor locations based on the spatiotemporal thermal feature matrix to obtain correlation data reflecting the thermal conduction topology of the CTC chassis; calculating heat flow direction distribution information based on the correlation data and the structural sensitivity matrix, whereby the heat flow direction distribution information characterizes the directional flow characteristics of heat conduction at the battery and structural interface; calculating the feature importance score of each sensor location based on the heat flow direction distribution information, and generating a sensor feature weight distribution based on the feature importance score; performing feature filtering based on the sensor feature weight distribution to obtain a target feature subset for thermal runaway early warning, and calculating the probability value that the current thermal state belongs to a thermal runaway early warning state based on the target feature subset, thus obtaining a thermal runaway risk assessment probability value; comparing the thermal runaway risk assessment probability value with a preset threshold, and generating a thermal runaway early warning signal when the probability value exceeds the preset threshold. The CTC chassis is divided into multiple thermal management zones, and a partitioned thermal management scheme for these zones is constructed based on the thermal runaway early warning signal. This includes: cutting the minimum heat flow path based on the CTC chassis topology to obtain multiple thermal management zones; defining a state space for each thermal management zone, which includes a temperature vector, a heat flow vector, and a structural component thermal stress vector within the zone; defining an action space corresponding to the state space, which includes active cooling power adjustment, battery load power limitation, and structural thermal conductivity enhancement control; calculating the input-output heat flow difference for each thermal management zone to obtain a global thermal distribution balance index; constructing a two-layer control architecture including a strategy layer and an execution layer based on the state space and the action space; generating control strategy parameters for each zone based on the thermal runaway early warning signal and the two-layer control architecture; and creating a partitioned thermal management scheme for the multiple thermal management zones based on the control strategy parameters for each zone and the global thermal distribution balance index.

2. The machine learning-based CTC chassis thermal runaway early warning and control method according to claim 1, characterized in that, The process of collecting data at the interface between the battery cells and the metal structure within the CTC chassis to obtain a temperature time series matrix and a structural sensitivity matrix includes: A temperature sensor array is deployed at the junction of the battery cell and the metal structure within the CTC chassis, and temperature data sampling is performed based on the temperature sensor array to obtain the original temperature data matrix. Perform sliding window averaging filtering and Z-Score normalization on the original temperature data matrix to obtain normalized temperature data; Based on the normalized temperature data, the time dimension derivative and spatial dimension gradient are calculated to obtain the temperature change rate matrix and the temperature spatial distribution matrix. The temperature change rate matrix and the temperature spatial distribution matrix are multiplied to obtain a temperature time series matrix. The influence coefficient of the thermal change of the battery cell on the temperature of the structural point is determined by heat flow experiment to obtain the structural sensitivity matrix.

3. The machine learning-based CTC chassis thermal runaway early warning and control method according to claim 1, characterized in that, The step of inputting the temperature time series matrix and the structure sensitivity matrix into a multi-scale thermal feature transformation neural network to extract the coupled thermal characteristics of the battery and structure, thereby obtaining a spatiotemporal thermal feature matrix, includes: The temperature time series matrix and the structure sensitivity matrix are concatenated to obtain the input feature matrix. The input feature matrix is ​​input into three one-dimensional convolutional sub-networks in the multi-scale thermal feature transformation neural network to extract multi-scale time features, thereby obtaining a multi-scale time feature set containing fast thermal diffusion phase and slow thermal diffusion phase. A spatial topological attention weight matrix is ​​constructed based on the CTC chassis topology, and spatial topological attention is calculated on the multi-scale time feature set based on the spatial topological attention weight matrix to obtain a spatial feature vector that integrates battery and structural interactive heat flow information. The spatial feature vector is input into the fully connected neural network of the multi-scale thermal feature transformation neural network for feature transformation to obtain the spatiotemporal thermal feature matrix.

4. The machine learning-based CTC chassis thermal runaway early warning and control method according to claim 1, characterized in that, The calculation of the input-output heat flow difference in each heat management zone to obtain the global heat distribution balance index includes: The difference between the heat inflow and heat outflow in each heat management zone is calculated to obtain the heat flow balance deviation value of each zone. The heat flow balance deviation values ​​of each region are summed to obtain the total balance deviation value that reflects the degree of heat distribution uniformity of the entire CTC chassis. Based on the temperature and heat flow differences between adjacent regions, neighborhood information exchange calculations are performed to obtain the intensity index of heat flow interaction between regions. Based on the matching analysis of the inter-regional heat flow interaction intensity index and the set of executable actions in each region of the action space, the strong interaction of adjacent regions is identified, and the target region network topology is obtained. The dispersion of heat flow conduction balance index between nodes in each region of the target region network topology is calculated to obtain the global heat distribution consistency evaluation index. The total balance deviation value, the intensity index of inter-regional heat flow interaction, and the global heat distribution consistency evaluation index are weighted and fused to obtain the global heat distribution balance index.

5. The machine learning-based CTC chassis thermal runaway early warning and control method according to claim 1, characterized in that, The machine learning-based CTC chassis thermal runaway early warning and control method also includes: Based on the thermal runaway warning signal and the thermal runaway risk assessment probability value, the warning level is classified to obtain the current CTC chassis thermal control level identifier; Based on the current CTC chassis thermal control level identifier, select a set of control command candidates for each region from the zonal thermal management scheme; The global thermal control problem is decomposed into N local subproblems, each of which corresponds to a thermal management region, and a set of control optimization equations is created for the N local subproblems. The control optimization equations are iteratively solved based on the candidate sets of control commands for each region, and a real-time control command sequence is output.

6. A machine learning-based CTC chassis thermal runaway early warning and control system, characterized in that, For implementing the machine learning-based CTC chassis thermal runaway early warning and control method as described in any one of claims 1-5, the machine learning-based CTC chassis thermal runaway early warning and control system comprises: The data acquisition module is used to collect data at the junction of the battery cells and the metal structure in the CTC chassis, and obtain the temperature time series matrix and the structural sensitivity matrix. The feature extraction module is used to input the temperature time series matrix and the structure sensitivity matrix into a multi-scale thermal feature transformation neural network to extract the coupled thermal characteristics of the battery and structure, and obtain a spatiotemporal thermal feature matrix. An interactive analysis module is used to perform thermal flow interaction analysis between the battery cells and the metal structure on the spatiotemporal thermal feature matrix, obtain a thermal runaway risk assessment probability value, and generate a thermal runaway early warning signal. This includes: calculating the correlation between the characteristics of different sensor locations based on the spatiotemporal thermal feature matrix to obtain correlation data reflecting the thermal conduction topology of the CTC chassis; calculating heat flow direction distribution information based on the correlation data and the structural sensitivity matrix, whereby the heat flow direction distribution information characterizes the directional flow characteristics of heat conduction at the battery and structural interface; calculating the feature importance score of each sensor location based on the heat flow direction distribution information, and generating a sensor feature weight distribution based on the feature importance score; performing feature filtering based on the sensor feature weight distribution to obtain a target feature subset for thermal runaway early warning, and calculating the probability value that the current thermal state belongs to a thermal runaway early warning state based on the target feature subset, thus obtaining a thermal runaway risk assessment probability value; comparing the thermal runaway risk assessment probability value with a preset threshold, and generating a thermal runaway early warning signal when the probability value exceeds the preset threshold. A construction module is used to divide the CTC chassis into multiple thermal management zones and construct a partitioned thermal management scheme for the multiple thermal management zones based on the thermal runaway early warning signal. This includes: cutting the minimum heat flow path based on the CTC chassis topology to obtain multiple thermal management zones; defining a state space for each thermal management zone, which includes a temperature vector, a heat flow vector, and a structural component thermal stress vector within the zone; defining an action space corresponding to the state space, which includes active cooling power adjustment, battery load power limitation, and structural thermal conductivity enhancement control; calculating the input and output heat flow difference for each thermal management zone to obtain a global thermal distribution balance index; constructing a two-layer control architecture including a strategy layer and an execution layer based on the state space and the action space; generating control strategy parameters for each zone based on the thermal runaway early warning signal and the two-layer control architecture; and creating a partitioned thermal management scheme for the multiple thermal management zones based on the control strategy parameters for each zone and the global thermal distribution balance index.

Images