Battery thermal conductive adhesive coating position optimization method and system based on machine learning

Abstract

The application relates to the technical field of battery heat management, and discloses a battery heat-conducting glue coating position optimization method and system based on machine learning. The method acquires a target battery pack structure and operating condition parameters, and divides a heat conduction sub-region; then, historical temperature distribution data sequences are collected, heat flow fluctuation extreme values and corresponding heat conduction path characteristics are extracted; subsequently, a topological network is constructed, taking the heat conduction sub-region as a node, the heat flow fluctuation extreme value as a node attribute, and the path characteristic as a connection weight; network features are learned by using a graph convolutional neural network, and a heat conduction influence factor set of each sub-region is output; according to the heat conduction influence factor set, a thermal resistance coefficient is calculated, and a candidate optimization set is formed by screening a sub-region exceeding the threshold value; finally, the path characteristics of the candidate regions are input into a pre-trained coating decision model, and the priority score is sorted according to the output; a predetermined number of high-score regions are selected as target coating positions, and the precision optimization of the heat-conducting glue coating is realized.

Description

Technical Field

[0001] This invention relates to the field of battery thermal management technology, specifically to a method and system for optimizing the coating position of thermally conductive adhesive for batteries based on machine learning. Background Technology

[0002] Against the backdrop of the rapid development of the new energy industry, battery packs, as the core energy source of new energy equipment, directly affect the overall performance of the equipment through their operational performance and safety. Battery packs generate a significant amount of heat during charging and discharging. If this heat cannot be dissipated in a timely manner, it will lead to increased battery temperature, decreased consistency, and even safety hazards such as thermal runaway. Therefore, effective thermal management is a crucial foundation for ensuring the long-term stable operation of battery packs.

[0003] Thermally conductive adhesive, a key material in battery pack thermal management systems, fills the gaps between individual battery cells and heat dissipation structures, reducing contact thermal resistance and accelerating heat conduction to achieve uniform battery temperature distribution. However, the application location of the thermally conductive adhesive plays a decisive role in the thermal management effect. Improper application not only wastes the adhesive but may also lead to localized heat buildup, preventing it from fully utilizing its thermal conductivity.

[0004] Traditional methods for determining the application location of thermally conductive adhesive often rely on engineers' experience or trial-and-error adjustments based on simple simulations. Experience-based judgment is limited by individual knowledge and struggles to fully consider the complex structural layout and diverse operating conditions of battery packs, easily leading to inaccurate optimization. While physical simulation-based methods can simulate heat conduction processes to some extent, they require establishing accurate physical models, involving numerous parameter settings and iterative calculations, which is time-consuming. Furthermore, for battery packs with complex structures and variable operating conditions, the accuracy of simulation results is often difficult to guarantee.

[0005] With the continuous advancement of battery technology, battery pack structures are becoming increasingly complex and integrated, while operating conditions are becoming more diverse, including varying charge / discharge rates, ambient temperatures, and load fluctuations. These changes result in a more nonlinear and dynamic heat transfer process within the battery pack, making traditional methods insufficient for the high-precision, high-efficiency optimization of coating placement. In practical applications, situations frequently arise where some areas experience insufficient heat dissipation and excessively high temperatures, while other areas suffer from overuse of thermal conductive adhesive, leading to resource waste and severely impacting battery pack performance and economics. Therefore, there is an urgent need for a method to optimize the placement of thermal conductive adhesive coating that can adapt to complex structures and diverse operating conditions, thereby improving thermal management effectiveness and resource utilization efficiency. Summary of the Invention

[0006] The purpose of this invention is to provide a method for optimizing the coating position of thermally conductive adhesive for batteries based on machine learning, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, this invention provides a machine learning-based method for optimizing the coating position of thermally conductive adhesive for batteries, the method comprising:

[0008] Obtain the structural parameters and operating condition parameters of the target battery pack, divide the heat conduction region of the battery pack based on the structural parameters, and generate multiple heat conduction sub-regions;

[0009] Collect temperature distribution data sequences of historical battery packs under multiple operating conditions, extract heat flow fluctuation features from the temperature distribution data sequences, and obtain multiple extreme values ​​of heat flow fluctuation and their corresponding heat conduction path features;

[0010] Construct a heat conduction path topology network, with multiple heat conduction sub-regions as network nodes, extreme values ​​of heat flow fluctuations as node attributes, and heat conduction path characteristics as connection weights between nodes.

[0011] A graph convolutional neural network is used to learn features from the heat conduction path topology network, and the set of heat conduction influencing factors for each heat conduction sub-region is output.

[0012] The thermal resistance coefficient of each thermal conduction sub-region is calculated based on the set of thermal conduction influencing factors, and a set of candidate optimization regions is generated by screening sub-regions whose thermal resistance coefficients exceed a preset threshold.

[0013] Extract the heat conduction path features of the candidate optimization region set, input them into the pre-trained coating decision model, and output the thermal adhesive coating priority score for each candidate optimization region;

[0014] The candidate optimization region set is sorted based on priority scores, and a predetermined number of regions with the highest scores are selected as the target coating locations.

[0015] Preferably, the operation of extracting heat flow fluctuation features from the temperature distribution data sequence includes:

[0016] Identify the moment points in a temperature distribution data sequence where the rate of temperature change exceeds a critical value, and extract the spatial temperature gradient distribution corresponding to that moment point.

[0017] Calculate the difference in heat flux intensity between adjacent monitoring points in the spatial temperature gradient distribution, and mark the pair of adjacent points with the largest difference as the abrupt heat conduction path.

[0018] Record the spatial coordinates and heat flow intensity difference of the heat conduction abrupt change path to generate heat conduction path characteristics.

[0019] Preferably, the operations for constructing the heat conduction path topology network include:

[0020] The center coordinates of multiple heat conduction sub-regions are used as the network node locations;

[0021] Locate the node of the corresponding heat conduction sub-region based on the spatial coordinates in the heat conduction path characteristics;

[0022] The two sub-region nodes connected by the heat conduction abrupt change path are used as edge endpoints, and the difference in heat flow intensity is used as the edge weight;

[0023] Integrate all nodes and edges to form a weighted undirected topology network.

[0024] Preferably, the specific operations for feature learning of the heat conduction path topology network using a graph convolutional neural network include:

[0025] Neighborhood aggregation is performed on the node attributes of the heat conduction path topology network, and the aggregation range is determined by a preset adjacency matrix.

[0026] The node feature vector is updated iteratively through multiple learnable convolutional kernels;

[0027] The final layer outputs node features that include heat conduction efficiency attributes, which are defined as a set of heat conduction influencing factors.

[0028] Preferably, the operation of calculating the thermal resistance coefficient of each thermal conduction sub-region based on the set of thermal conduction influencing factors includes:

[0029] Extract the maximum heat loss rate from the set of factors affecting heat conduction;

[0030] Obtain the thermal conductivity and thickness parameters of the material in the corresponding heat-conducting sub-region;

[0031] The ratio of the maximum heat loss rate to the material's thermal conductivity is multiplied by the thickness parameter, and the result is used as the thermal resistance coefficient.

[0032] Preferably, the training operations for the coating decision model include:

[0033] Collect data on the thermal conductive adhesive coating effect in different areas of the historical battery pack. The effect data includes the change in the standard deviation of temperature before and after coating.

[0034] The characteristics of the heat conduction path are correlated and labeled with the change in the standard deviation of temperature;

[0035] The gradient boosting decision tree algorithm is used to learn the mapping relationship between features and performance data, and to generate a coating decision model.

[0036] Preferably, the operation of outputting the thermal conductive adhesive coating priority score for each candidate optimization region includes:

[0037] Input the heat conduction path characteristics of the candidate optimization region into the coating decision model;

[0038] Obtain the predicted change in temperature standard deviation from the model output;

[0039] The predicted change in temperature standard deviation is scaled proportionally to the scoring interval to generate a priority score.

[0040] Preferably, the method further includes:

[0041] Monitor the real-time temperature distribution data of the current battery pack;

[0042] When the deviation between the real-time temperature distribution data and the historical data sequence exceeds the allowable range, the dynamic update of the heat conduction path topology network is triggered.

[0043] The target coating location is regenerated using the updated topology network.

[0044] Preferably, the method further includes a coating location verification operation:

[0045] After the thermally conductive adhesive is applied to the target coating location, the battery pack operating temperature data is collected.

[0046] Analyze the temperature uniformity variation within a preset range around the coating location;

[0047] When the change in temperature uniformity does not reach the expected value, a correction signal is fed back to the coating decision model.

[0048] Preferably, the present invention further includes a battery thermal conductive adhesive coating position optimization system for implementing the above method, comprising:

[0049] The data acquisition and region division module is used to acquire the structural parameters and operating condition parameters of the target battery pack, and divide the heat conduction region of the battery pack based on the structural parameters to generate multiple heat conduction sub-regions.

[0050] The heat flow feature extraction module is used to collect temperature distribution data sequences of historical battery packs under multiple operating conditions, extract heat flow fluctuation features from the temperature distribution data sequences, and obtain multiple heat flow fluctuation extreme values ​​and their corresponding heat conduction path features.

[0051] The topology network construction module is used to construct a heat conduction path topology network, taking the multiple heat conduction sub-regions as network nodes, the extreme values ​​of heat flow fluctuations as node attributes, and the heat conduction path features as connection weights between nodes.

[0052] The feature learning module is used to learn features from the heat conduction path topology network using a graph convolutional neural network and output a set of heat conduction influencing factors for each heat conduction sub-region.

[0053] The candidate region filtering module is used to calculate the thermal resistance coefficient of each thermal conduction sub-region based on the set of thermal conduction influencing factors, and to filter sub-regions whose thermal resistance coefficient exceeds a preset threshold to generate a set of candidate optimization regions.

[0054] The coating priority decision module is used to extract the heat conduction path features of the candidate optimization region set, input them into the pre-trained coating decision model, and output the thermally conductive adhesive coating priority score for each candidate optimization region.

[0055] The target location determination module is used to sort the candidate optimization region set based on the priority score and select a predetermined number of regions with the highest score as the target coating location.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] This machine learning-based method for optimizing the coating location of thermally conductive adhesive in batteries achieves a refined breakdown of the battery pack's thermal conduction space by acquiring the structural and operational parameters of the target battery pack and dividing it into thermal conduction sub-regions. This allows subsequent thermal conduction analysis to focus on specific areas, avoiding potential omissions of details in the overall analysis. This regionalized approach better reflects the actual structural characteristics of the battery pack, providing a more precise analysis target for subsequent optimization efforts.

[0058] In the data processing stage, historical temperature distribution data sequences were collected, and extreme values ​​of heat flow fluctuations and corresponding heat conduction path characteristics were extracted, making full use of the heat conduction patterns contained in historical operating data. Extreme values ​​of heat flow fluctuations can reflect key nodes of heat accumulation in the battery pack under different operating conditions, while heat conduction path characteristics reveal the specific paths of heat transfer. The extraction of this information provides rich basic data for the subsequent construction of heat conduction models, enabling the models to more realistically reflect the heat conduction characteristics of the battery pack.

[0059] When constructing a heat conduction path topology network, heat conduction sub-regions are used as nodes, extreme values ​​of heat flow fluctuations are used as node attributes, and heat conduction path characteristics are used as connection weights. This network structure can intuitively and comprehensively characterize the heat conduction relationships between regions. Compared with traditional linear analysis models, topology networks can better capture the nonlinear heat conduction relationships between regions, allowing the transmission process of heat conduction effects to be more accurately reflected in the model.

[0060] By utilizing graph convolutional neural networks (GNNs) for feature learning in topological networks, the advantages of machine learning in processing complex network data are fully leveraged. GNNs can automatically uncover potential relationships and feature transfer patterns between nodes, and the output set of heat conduction influencing factors can quantify the impact of each sub-region on overall heat conduction. This quantification provides an objective basis for subsequent region selection, overcoming the limitations of traditional methods that rely on subjective experience.

[0061] By calculating the thermal resistivity based on the thermal conductivity influencing factor and screening candidate optimization regions, it is possible to accurately locate key areas that significantly impede thermal conduction. Screening using preset thresholds ensures the relevance of candidate regions, avoids interference from irrelevant areas, and allows subsequent optimization work to focus on areas that truly require improvement in thermal conductivity, thus improving optimization efficiency.

[0062] The heat conduction path characteristics of candidate regions are input into a pre-trained coating decision model to obtain priority scores. Leveraging the decision-making experience learned from a large amount of historical data by the pre-trained model, the necessity and effectiveness of coating each candidate region can be scientifically evaluated. The generation of priority scores is based on data-driven analysis, comprehensively considering the influence of multiple factors on the coating effect, resulting in a more rational selection of the final coating location.

[0063] Priority scoring is used to select target coating locations, ensuring that resources are allocated to areas with the most significant improvement in thermal conductivity. This optimization method maximizes the improvement of battery pack thermal conductivity with limited thermal conductive adhesive usage, reduces localized overheating, and avoids indiscriminate use of thermal conductive adhesive, thus improving thermal management effectiveness while maintaining economic efficiency. The entire methodology is tightly integrated, forming a complete optimization loop from data acquisition to model learning and decision output. It can adapt to battery packs with different structures and operating conditions, providing a systematic and efficient solution for optimizing thermal conductive adhesive coating locations. Attached Figure Description

[0064] Figure 1 This is a timing diagram of the battery thermal conductive adhesive coating position optimization method based on machine learning described in this invention.

[0065] Figure 2 A flowchart of the operation for constructing the heat conduction path topology network;

[0066] Figure 3 This is a flowchart of the heat conduction path topology network feature learning operation based on graph convolutional neural network;

[0067] Figure 4 This is a flowchart of the thermal resistance coefficient calculation process. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Please see Figure 1This invention provides a method and system for optimizing the coating position of thermally conductive adhesive for batteries based on machine learning. The method includes:

[0070] By analyzing the thermal conductivity characteristics of the battery pack and combining a graph convolutional neural network (GNN) with a coating decision model, intelligent selection of the thermal conductive adhesive coating location is achieved. Specifically, the steps include: obtaining the structural and operating parameters of the target battery pack; dividing the battery pack into thermal conduction regions based on the structural parameters to generate multiple thermal conduction sub-regions; collecting historical temperature distribution data sequences of the battery pack under multiple operating conditions, extracting extreme values ​​of heat flow fluctuations and their corresponding thermal conduction path features; constructing a thermal conduction path topology network, using the thermal conduction sub-regions as network nodes, extreme values ​​of heat flow fluctuations as node attributes, and thermal conduction path features as connection weights between nodes; using a graph convolutional neural network to learn features of the topology network, outputting a set of thermal conduction influence factors for each thermal conduction sub-region; calculating the thermal resistance coefficient of each sub-region based on the set of thermal conduction influence factors, and selecting sub-regions with thermal resistance coefficients exceeding a preset threshold to generate a set of candidate optimization regions; extracting the thermal conduction path features of the candidate optimization regions, inputting them into a pre-trained coating decision model, and outputting a thermal conductive adhesive coating priority score for each candidate optimization region; ranking the candidate optimization regions based on the priority scores, and selecting a predetermined number of regions with the highest scores as the target coating locations.

[0071] Example 1: See Figure 2 The system identifies time points in a temperature distribution data series where the rate of temperature change exceeds a critical value. This data series comprises temperature readings from multiple monitoring points on or inside a battery pack over a time period. The critical value is a preset parameter, such as 0.5 degrees Celsius per second, used to distinguish between normal fluctuations and significant changes. In practice, the system iterates through each time point, calculating the temperature change between adjacent time points divided by the time interval to obtain the rate of change. If this value exceeds the critical value, the system marks that time point as a hotspot. Subsequently, the spatial temperature gradient distribution corresponding to that time point is extracted. This involves spatial analysis of the temperature data from all monitoring points at the current time, generating a gradient vector for each point, indicating the magnitude and direction of the temperature change. The gradient distribution map can be visualized in a geospatial coordinate system, for example, using x, y, and z coordinates to represent location, with the gradient value as an additional attribute.

[0072] Based on the spatial temperature gradient distribution, the difference in heat flux intensity between adjacent monitoring points is calculated. Heat flux intensity is a derived parameter representing the heat transfer rate per unit area, which is positively correlated with the temperature gradient. The system retrieves adjacent point pairs in the gradient distribution; these pairs may be connected in a grid format, such as a triangular grid or a regular array. For each adjacent point pair, the absolute difference in their heat flux intensity is calculated; this difference represents the numerical difference in heat flux intensity between the two points. Then, the magnitudes of the differences between all adjacent point pairs are compared, and the pair with the largest difference is marked as a heat conduction abrupt change path. This represents the most significant discontinuity in the heat conduction process, such as the junction between high and low thermal conductivity materials, or bottlenecks caused by structural constraints. Subsequently, the spatial coordinates of the heat conduction abrupt change path and the heat flux intensity difference are recorded. The spatial coordinates include the location identifiers of the start and end points, and the heat flux intensity difference serves as a quantitative indicator. All this information is integrated to generate a complete heat conduction path feature, stored as a structured dataset containing path identifiers, coordinate sequences, and difference values ​​for subsequent processing.

[0073] The construction of the heat conduction path topology network is integrated with feature extraction. Construction begins with multiple heat conduction sub-regions. These sub-regions are pre-divided based on the structural parameters of the battery pack, for example, by decomposing the battery pack structure into physically separated or meshed regions, with each sub-region representing an independent thermal unit. The system determines the network node location based on the center coordinates of each sub-region. These center coordinates are obtained through geometric calculations, such as the coordinates of the center point of a 3D geometric object. Then, the system locates the heat conduction sub-region nodes to which these coordinates belong based on the spatial coordinates in the generated heat conduction path features. The localization process includes matching coordinate points with sub-region boundaries: if the start or end coordinates of a heat conduction path fall within the geometric range of a certain sub-region, the path is associated with the corresponding sub-region node. This operation ensures an accurate correspondence between nodes and heat conduction features.

[0074] The system uses nodes in two sub-regions connected by a heat conduction mutation path as the endpoints of an edge. Each mutation path corresponds to one edge, with its endpoints being nodes in the sub-regions where the path's start and end points are located. Edge weights are set based on the difference in heat flow intensity in the heat conduction path features; the weight values ​​are directly taken from the recorded difference values ​​to reflect the intensity of the thermal mutation along the path. For example, a high difference corresponds to a high weight, representing a strong connection requirement. The construction process also considers network type: forming a weighted undirected topology network, meaning that edge weights are effective in both directions, and there are no directional restrictions on connections between network nodes. Finally, the system integrates all nodes and edges to form a complete topology network. Integration includes storing a list of nodes (containing center coordinates), a list of edges (containing endpoints and weights), and additional attributes such as extreme values ​​of heat flow fluctuations. Extreme values ​​of heat flow fluctuations are added as node attributes; these extreme values ​​originate from key fluctuation indicators in the heat conduction path features, representing the maximum or minimum fluctuation amplitude of heat flow at the node. In this way, the topology network becomes a structured representation, with nodes representing heat conduction sub-regions and edge weights quantifying the intensity of the conduction path, providing basic input for subsequent feature learning. After the network is generated, it is saved in a graphical format, such as an adjacency list or matrix, using standard data processing methods.

[0075] The implementation of the heat conduction path topology network involves iteratively processing multiple heat conduction sub-regions. For each sub-region, the system first loads its physical description, including volume and boundary points. Center coordinates are calculated using an average location method, such as averaging the coordinates of the region's boundary points. When locating heat conduction path coordinates, the system performs spatial queries, such as using bounding box cross-checking or point-polygon containment algorithms. Once all paths are associated with nodes, edge definitions and weight assignments are automatically completed. After weight settings, network integration uses a graphics processing library to progressively add nodes and edges, ensuring no duplicate or isolated nodes. The entire topology network construction process can be repeatedly applied to data under different operating conditions, ensuring consistency. Details of network attributes, such as the allocation of extreme heat flux fluctuations, are associated with corresponding nodes based on the maximum or minimum heat flux values ​​extracted from the original temperature data. The final output topology network file is used by downstream modules to support further analysis of the battery thermal management system.

[0076] Example 2: See Figure 3 The feature learning process of the graph convolutional neural network on the heat conduction path topology network is described below. The heat conduction path topology network, as the initial input data structure, contains three core types of information: the node set represents the center coordinates of the heat conduction sub-regions, node attributes record the extreme values ​​of heat flow fluctuations in the corresponding regions, edge connections define the conduction paths between regions, and edge weights store the numerical values ​​of heat flow intensity differences. This topology network is stored in the form of an adjacency list, with a mapping table established between node indices and edge relationships.

[0077] At the start of feature learning, the system loads the adjacency matrix as a preset parameter. The adjacency matrix exists as a two-dimensional array, with dimensions equal to the total number of network nodes. Matrix elements represent the connection state between nodes: if two nodes are connected by an edge, the element value is taken from the edge weight; otherwise, it is marked as zero. This matrix controls the range of information propagation during graph convolution. Neighborhood aggregation is performed based on the adjacency matrix, specifically by traversing each node and obtaining its set of direct neighbors. Subsequently, the system performs weighted aggregation calculations on the attribute vectors of neighboring nodes based on the weight values ​​of the adjacency matrix. The attribute vector of each neighboring node is multiplied by its corresponding edge weight and then summed to form the aggregated feature value. After weighted aggregation, the system concatenates the current node's own attribute vector with the aggregated feature vector to form an enhanced feature representation.

[0078] A multi-layer convolutional kernel structure is responsible for iteratively updating node features. The input dimension of the first convolutional kernel is set to the sum of the original dimension of the node attributes and the dimension of the aggregated features. For example, if the initial extreme values ​​of heat flow fluctuations contain five statistics (maximum, minimum, mean, etc.), and the dimension is expanded to ten after aggregation, then the input size of the first convolutional kernel is set to ten. Each convolutional layer adopts a fully connected neural network structure, containing a learnable weight parameter matrix and a bias vector. The convolution operation is defined as a linear transformation of the feature vector and the weight matrix followed by a non-linear activation function. The ReLU function is used to handle negative values ​​and maintain effective gradient propagation. During forward propagation, the output vector of each layer is passed layer by layer as the input of the next layer. The system presets the number of convolutional layers, typically set to three to five layers, with the output dimension of each layer decreasing sequentially to achieve feature compression. Batch standardization is performed between layers to automatically adjust the feature distribution to a zero-mean, unit-variance state.

[0079] The feature vector undergoes multiple rounds of evolution during convolution. The initial input vector contains raw statistical information about heat flow fluctuations. The first convolution output captures heat conduction correlation patterns within a local neighborhood, such as the synchronous fluctuation characteristics of areas affected by the same heat source. The second convolution continues to aggregate information within an expanded neighborhood, covering nodes within two hops of the current node, identifying indirect influence relationships in heat conduction paths. The final convolution gives each node a global receptive field of the entire graph, and the output vector fuses heat conduction path features across regions. Intermediate feature vectors are cached during each iteration for use in backpropagation. Backpropagation updates the weight parameters using a gradient descent algorithm: first, the error between the predicted and expected features is calculated, using a mean squared error function; then, the partial derivatives of the error with respect to the weight parameters of each layer are calculated; finally, the parameter values ​​are adjusted according to a preset learning rate to make the feature representation more consistent with the isomorphic graph characteristics. This process is repeated for a preset number of iterations or until the error converges and stabilizes.

[0080] After multiple iterations, the system extracts feature vectors from the final output layer of the network. These output vectors are compressed to a preset size, typically a four- to six-dimensional numerical vector. Each dimension corresponds to a different heat conduction pattern encoding, containing implicit heat flow efficiency characteristics. For example, a higher-value dimension in the vector might reflect thermal resistance anomalies, while a lower-value dimension might represent an efficient conduction path. The output features are normalized and converted into a probability distribution, ultimately defined as a set of heat conduction influencing factors and persistently stored. Each factor in the set corresponds to a unique identifier for a heat conduction sub-region, and the factor values ​​are normalized to between zero and one. This numerical set is organized in key-value pairs and used as a quantification indicator for subsequent thermal resistance coefficient calculations. All preset parameters involved in feature learning, including the adjacency matrix, number of convolutional layers, layer dimension settings, and activation function types, are recorded in the configuration file. The entire graph convolutional neural network is trained offline, and the model weights are fixed as embedded parameters, ensuring real-time computation and inference during deployment. When the network topology is updated, the system reloads the new adjacency matrix while maintaining the existing network weights, executing the same process to output the updated set of influencing factors.

[0081] Example 3: See Figure 4 When calculating the thermal resistance coefficient of each thermal conduction sub-region using the set of heat conduction influence factors, the system first extracts the feature vector corresponding to each sub-region from the set. This vector contains values ​​in multiple dimensions, with a specific dimension identified as the maximum heat loss rate parameter. The maximum heat loss rate parameter is determined by analyzing the numerical distribution of each dimension of the feature vector, typically selecting the dimension with the largest value as representative. The system records this parameter value for each sub-region and associates it with the physical property data of that region. The physical property data includes the material thermal conductivity and thickness parameters, which are obtained from the battery pack structure database. The material thermal conductivity reflects the thermal conductivity of the main constituent materials in that region, and the thickness parameter represents the geometric thickness of the heat conduction path in that region.

[0082] The thermal resistivity is calculated using the following formula:

[0083] ;

[0084] in, Indicates the thermal resistance coefficient. Indicates the maximum heat loss rate. Indicates the thermal conductivity of a material. This represents the thickness parameter. The calculation process is performed independently for each candidate sub-region. The system first retrieves the feature vector of the current sub-region to locate the maximum heat flow loss rate. Then, it queries the material database to obtain the thermal conductivity value of the region. This value may adjust with temperature changes, and the system selects the corresponding coefficient value based on the current operating temperature. The thickness parameter is extracted from the structural parameter table and is typically the average thickness of the region along the heat flow direction. These three parameters are input into the calculation module, which performs division and multiplication operations in the order of the formula. The output result is the thermal resistance coefficient of the sub-region. The calculation result is stored in floating-point form, retaining four decimal places of precision.

[0085] The training process of the coating decision model involves multiple data preparation and processing stages. Historical data on the thermal conductive adhesive coating effects in different areas of the battery pack are collected from maintenance records. Each record includes the coating area identifier, pre-coating temperature distribution data, post-coating temperature distribution data, and operating parameters. The change in temperature standard deviation is calculated as a key indicator, obtained by comparing the standard deviations of temperature data from the same set of monitoring points before and after coating. The system performs the following processing for each historical case: extracting the heat conduction path features of the coating area, which are consistent with the feature extraction process described in Example 1; calculating the difference in temperature standard deviation before and after coating in that area; and associating and labeling the two to form training samples. The sample data structure includes an input feature vector and an output change value.

[0086] The gradient boosting decision tree algorithm is used to learn the mapping relationship between features and performance data. The algorithm employs an iterative boosting approach. In the initial stage, a basic decision tree model is built, which is divided into split nodes based on input features. Each leaf node predicts the initial change in temperature standard deviation. Subsequently, the algorithm enters multiple boosting rounds, adding a new decision tree in each round. This new tree learns from the prediction residuals of the previous model, gradually correcting the overall prediction accuracy. The growth process of the decision tree is controlled by hyperparameters, including maximum tree depth, minimum number of samples per leaf node, and learning rate. The splitting criterion for each tree is the combination of features and thresholds that maximize the reduction in mean squared error. After a predetermined number of iterations, multiple decision trees are combined to form a complete model. This model can receive new heat conduction path feature inputs and output the predicted change in temperature standard deviation.

[0087] Cross-validation is implemented during model training. The original dataset is randomly divided into multiple subsets, with some subsets selected alternately as validation data and the remainder as training data. Each validation iteration calculates the model's prediction error on unseen data, and the average error is used as the model's performance evaluation. Training terminates when the validation error converges or the maximum number of iterations is reached. The trained model parameters are serialized and saved, including the structural parameters of each decision tree and node splitting rules. During deployment, the model file is loaded, and the prediction inference engine is initialized. The model performs a multi-tree traversal on the new input feature vector, with each tree outputting a predicted value. The final prediction result is a weighted sum of the outputs of all trees, with the weights controlled by the learning rate parameter. This predicted value represents the expected change in the standard deviation of temperature after thermal adhesive coating and is used for subsequent priority scoring calculations.

[0088] The thermal resistance coefficient calculation module and the coating decision model form a serial processing flow. The system first filters sub-regions where the thermal resistance coefficient exceeds a preset threshold, which is set according to the battery pack's thermal management requirements and typically selects a specific percentile value of the thermal resistance coefficient from historical data. The selected candidate optimization regions enter the decision model evaluation stage, where the heat conduction path characteristics of each region are extracted and formatted for input into the model. After the model outputs the predicted change, the system records the correspondence between this value and the region identifier. After all candidate regions are processed, a complete list of prediction results is generated, which serves as the basis for priority scoring calculations. Throughout the process, the thermal resistance coefficient calculation focuses on identifying heat conduction bottleneck regions, while the decision model evaluation quantifies the expected effect of coating measures in these regions. The two work together to complete the transformation from thermal analysis to optimization decision-making. The parameters and thresholds involved in the calculation process can be adjusted in the system configuration to adapt to the requirements of different battery pack models or operating environments.

[0089] Example 4: The process of generating priority scores for thermally conductive adhesive coating in candidate optimization regions is illustrated by the following specific example. Assume a battery pack system contains six thermally conductive sub-regions. After calculating the thermal resistance coefficient, four candidate optimization regions are selected, numbered A2, B1, C3, and D4. The thermal conduction path characteristics of these regions have been extracted in previous steps, and the feature vector of each region contains a five-dimensional numerical description. The coating decision model receives these feature vectors as input and performs predictive inference operations.

[0090] Taking region A2 as an example, its feature vector values ​​are [0.82, 1.15, 0.63, 1.07, 0.91]. Multiple decision trees within the model process this input sequentially. The first decision tree assigns the feature vector to the left child node based on the comparison between the third dimension value of 0.63 and the preset threshold of 0.7. The second tree examines the relationship between the first dimension value of 0.82 and the threshold of 0.9, selecting the right branch path. Subsequent trees refine the prediction step by step according to similar rules. Finally, the weighted sum of the output values ​​of each tree yields a predicted change in the temperature standard deviation of region A2 of 0.38 degrees Celsius. The system performs the same processing on all four candidate regions, obtaining the prediction results shown in Table 1.

[0091] Table 1: Prediction results obtained by performing the same processing on all four candidate regions.

[0092]

[0093] Priority scores are generated using a linear transformation method. The system sets the score range to 50 to 100 points, corresponding to the full range mapping of temperature standard deviation variation. The table shows the predicted variation range is 0.35 to 0.45 degrees Celsius. The minimum value of 0.35 is mapped to 50 points, the maximum value of 0.45 to 100 points, and intermediate values ​​are converted proportionally. Region D4, with a temperature of 0.45 degrees Celsius, receives a full score of 100 points; region C3, with a temperature of 0.35 degrees Celsius, receives 50 points; and region B1, with a temperature of 0.42 degrees Celsius, is converted to (0.42-0.35) / (0.45-0.35)×50+50=85 points. The final priority score ranking for the four regions is: D4 (100 points), B1 (85 points), A2 (70 points), and C3 (50 points). The system selects the two regions with the highest scores, D4 and B1, as the target coating locations for this round of optimization.

[0094] The real-time monitoring system continuously collects temperature data during battery pack operation. An array of temperature sensors deployed on the battery pack surface records readings from each monitoring point every minute, forming real-time temperature distribution data. The system maintains a sliding time window, storing historical data sequences for the most recent two hours. Each time new data arrives, the system performs the following check: calculating the difference between the current temperature distribution and the historical baseline distribution; the difference is defined as the sum of the squares of the temperature deviations at each monitoring point. When the difference exceeds a preset tolerance value, a topology network update process is triggered. For example, if a monitoring detects that the difference jumps to three times the normal value, the system immediately initiates the update procedure.

[0095] The topology network update process begins by re-extracting heat flux fluctuation characteristics. The system analyzes the latest temperature data sequence to identify anomalous fluctuation points. Assuming anomalies in temperature change rates are detected at three monitoring points at time T, the system extracts the spatial temperature gradient distribution at these points. New abrupt heat conduction paths are marked; for example, a significant difference in heat flux intensity is found between the original region B1 and the newly added region E5. The network nodes are then adjusted accordingly, with a new node E5 added and the connectivity of node B1 updated. A new row and column are added to the adjacency matrix to represent node E5, and the edge weight between node E5 and node B1 is set to the latest calculated difference in heat flux intensity, 1.23.

[0096] The heat conduction influence factor is recalculated using the updated topology network. The graph convolutional neural network loads a new adjacency matrix, keeping the original weight parameters unchanged, and performs forward propagation calculations. The initial feature vector of the newly added node E5 is generated through feature interpolation of neighboring nodes, and after three layers of convolution operations, its heat conduction influence factor is obtained as [0.88, 0.76, 0.92, 0.85]. The system reruns the entire optimization process: when calculating the thermal resistance coefficient of each region, the thermal resistance coefficient of region E5 is 0.68; when screening candidate regions, E5 is added to the candidate set because it exceeds the threshold; the coating decision model outputs a predicted change of 0.40 degrees Celsius for the feature vector [0.88, 0.76, 0.92, 0.85] of the new region E5, corresponding to a priority score of 75. The new candidate region set is changed to D4, B1, E5, A2, and C3, and its score ranking is updated to D4 (100 points), B1 (85 points), E5 (75 points), A2 (70 points), and C3 (50 points). If two target coating locations are still selected, D4 and B1 remain unchanged; if three locations are selected, region E5 is added. This dynamic update mechanism allows the system to adjust its optimization strategy based on the actual state of the battery pack, adapting to changes in thermal conductivity.

[0097] All data flow and processing steps involved in the entire implementation process are recorded in the system log. Each time a priority score is generated, the system saves a complete list of candidate regions and their score details; each time the topology network is updated, it records the changed node and edge information; each time the final target location is determined, it stores the selection result and decision timestamp. This data forms the basis of a closed-loop feedback loop, used for subsequent analysis of the effectiveness of optimization strategies. Parameters used in data processing, such as the score interval range, difference tolerance, and the number of candidate regions selected, are all managed through configuration files, supporting flexible adjustments based on actual application needs. The system does not rely on preset fixed rules during operation, but rather dynamically generates decisions based on real-time data analysis, achieving adaptive optimization of battery pack thermal management.

[0098] Example 5: Coating Location Verification Operation is initiated after the thermally conductive adhesive is applied to the target location. A temperature sensor system is deployed around the coating area in a specific pattern designed based on the area's geometry. For rectangular coating areas, sensors are arranged in a grid array; for circular areas, they are distributed in concentric circles. Each sensor collects temperature data at fixed time intervals, with a sampling frequency of five minutes per data point. The data includes the absolute temperature value and timestamp for each monitoring point. The monitoring period covers the entire battery pack charging and discharging process, typically lasting seventy-two hours. The collected raw data is preprocessed and then input into the analysis module. The main processing steps include outlier filtering and time zone alignment to ensure data quality meets analysis requirements.

[0099] The calculation of temperature uniformity change is performed within a preset spatial range. The system defines the center point of the coating location as the center of a circle, with a radius of 30 centimeters as the analysis range. All monitoring points within this range participate in the uniformity assessment. The uniformity index is represented by the temperature standard deviation, and the calculation formula is implemented in two stages: first, the temperature standard deviation of all monitoring points at each sampling time is calculated; second, the average of the standard deviations at all times is taken to obtain the temperature uniformity baseline value after coating. The system simultaneously retrieves historical data of the same area before coating and uses the same algorithm to calculate the historical baseline value. The change is defined as the difference between the baseline value after coating and the historical baseline value before coating; a positive value indicates improved uniformity, and a negative value indicates deterioration. This calculation process is completed automatically, and the results are stored in the validation database.

[0100] The effect evaluation process includes quantitative judgment rules. The system presets an expected value range, for example, the temperature standard deviation change should be between 0.35 and 0.55 degrees Celsius. The actual calculated change is compared with the expected value range: when the change falls within this range, the coating effect is considered to meet expectations; if the change exceeds the upper limit, it indicates that excessive local heat dissipation may cause a temperature gradient reversal; if it is below the lower limit, it means that the coating has not achieved the expected heat conduction optimization effect. When the change does not meet the expected value, the system automatically generates a correction signal. The signal type is graded according to the degree of deviation: a slight deviation triggers a yellow correction signal, and a severe deviation triggers a red correction signal. Each signal corresponds to a different parameter adjustment strategy.

[0101] The correction signal is transmitted to the coating decision model server, triggering online updates of the model parameters. The update process is based on an incremental learning mechanism. The system extracts relevant data from the candidate region that caused the correction: the feature vector of the original heat conduction path in the region, the actual coating effect value (i.e., the change in temperature standard deviation), and the original predicted value of the model. The difference between the actual value and the predicted value is used as the loss, and the gradient direction optimization algorithm is used to adjust the decision tree parameters. Specific adjustments include fine-tuning the split point threshold and redistributing the weights of the leaf nodes. The correction magnitude of each decision tree is positively correlated with the correction signal level: a yellow signal only adjusts the terminal leaf nodes, while a red signal corrects the branch decision points on the entire prediction path. Model version management is performed synchronously; a new model version number is generated after each update, and historical versions are retained for comparative analysis.

[0102] The calibration and verification process forms a closed-loop workflow. After the model update is complete, the system re-evaluates the priority scores of all candidate regions in the current batch. The updated decision model is used to recalculate the original input features, outputting the corrected predicted changes. The new scores guide subsequent coating tasks: for uncoated candidate regions, the coating order is re-planned according to the new scores; for coated regions, the new scores determine whether additional coating is needed. The entire process is automated: when a calibration signal is generated, the system records the trigger time, specific region number, and deviation value; during model updates, the adjusted parameter positions and modification amounts are recorded; during re-evaluation, the score change differences for each region are saved. These data constitute the key records of the verification closed loop, used to track the long-term evolution trend of the system's optimization capabilities.

[0103] In addition to the regular operating cycle, the system also features periodic overall verification. A full verification program is initiated at fixed intervals or after a set number of coating tasks are completed. This program summarizes all historical verification records and statistically analyzes the model's prediction accuracy. Accuracy is calculated using a segmented matching method: the proportion of predicted values ​​falling within ±10% of the actual value. If the overall accuracy falls below a set threshold, a systematic retraining is triggered: the complete training set from the historical database is retrieved, and the gradient boosting decision tree model training process is re-executed. After retraining, the current online model is completely replaced, ensuring that long-term predictive capability remains at a high confidence level. The full verification results generate periodic evaluation reports, supporting dynamic optimization and adjustments for operational decisions.

[0104] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0105] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing the coating position of thermally conductive adhesive for batteries based on machine learning, characterized in that, include: Obtain the structural parameters and operating condition parameters of the target battery pack, divide the heat conduction region of the battery pack based on the structural parameters, and generate multiple heat conduction sub-regions; Collect temperature distribution data sequences of historical battery packs under multiple operating conditions, extract heat flow fluctuation features from the temperature distribution data sequences, and obtain multiple extreme values ​​of heat flow fluctuation and their corresponding heat conduction path features; Construct a heat conduction path topology network, with multiple heat conduction sub-regions as network nodes, extreme values ​​of heat flow fluctuations as node attributes, and heat conduction path characteristics as connection weights between nodes. A graph convolutional neural network is used to learn features from the heat conduction path topology network, and the set of heat conduction influencing factors for each heat conduction sub-region is output. The thermal resistance coefficient of each thermal conduction sub-region is calculated based on the set of thermal conduction influencing factors, and a set of candidate optimization regions is generated by screening sub-regions whose thermal resistance coefficients exceed a preset threshold. Extract the heat conduction path features of the candidate optimization region set, input them into the pre-trained coating decision model, and output the thermal adhesive coating priority score for each candidate optimization region; The candidate optimization region set is sorted based on priority scores, and a predetermined number of regions with the highest scores are selected as the target coating locations.

2. The method for optimizing the coating position of battery thermal conductive adhesive as described in claim 1, characterized in that, The operations for extracting heat flow fluctuation features from temperature distribution data sequences include: Identify the moment points in a temperature distribution data sequence where the rate of temperature change exceeds a critical value, and extract the spatial temperature gradient distribution corresponding to that moment point. Calculate the difference in heat flux intensity between adjacent monitoring points in the spatial temperature gradient distribution, and mark the pair of adjacent points with the largest difference as the abrupt heat conduction path. Record the spatial coordinates and heat flow intensity difference of the heat conduction abrupt change path to generate heat conduction path characteristics.

3. The method for optimizing the coating position of battery thermal conductive adhesive as described in claim 2, characterized in that, The operations for constructing a heat conduction path topology network include: The center coordinates of multiple heat conduction sub-regions are used as the network node locations; Locate the node of the corresponding heat conduction sub-region based on the spatial coordinates in the heat conduction path characteristics; The two sub-region nodes connected by the heat conduction abrupt change path are used as edge endpoints, and the difference in heat flow intensity is used as the edge weight; Integrate all nodes and edges to form a weighted undirected topology network.

4. The method for optimizing the coating position of battery thermal conductive adhesive as described in claim 3, characterized in that, The specific operations for feature learning of heat conduction path topology networks using graph convolutional neural networks include: Neighborhood aggregation is performed on the node attributes of the heat conduction path topology network, and the aggregation range is determined by a preset adjacency matrix. The node feature vector is updated iteratively through multiple learnable convolutional kernels; The final layer outputs node features that include heat conduction efficiency attributes, which are defined as a set of heat conduction influencing factors.

5. The method for optimizing the coating position of battery thermal conductive adhesive as described in claim 4, characterized in that, The operation of calculating the thermal resistance coefficient of each thermal conduction sub-region based on the set of thermal conduction influencing factors includes: Extract the maximum heat loss rate from the set of factors affecting heat conduction; Obtain the thermal conductivity and thickness parameters of the material in the corresponding heat-conducting sub-region; The ratio of the maximum heat loss rate to the material's thermal conductivity is multiplied by the thickness parameter, and the result is used as the thermal resistance coefficient.

6. The method for optimizing the coating position of battery thermal conductive adhesive as described in claim 1, characterized in that, The training operations for the coating decision model include: Collect data on the thermal conductive adhesive coating effect in different areas of the historical battery pack. The effect data includes the change in the standard deviation of temperature before and after coating. The characteristics of the heat conduction path are correlated and labeled with the change in the standard deviation of temperature; The gradient boosting decision tree algorithm is used to learn the mapping relationship between features and performance data, and to generate a coating decision model.

7. The method for optimizing the coating position of battery thermal conductive adhesive as described in claim 6, characterized in that, The operation of outputting the thermal conductive adhesive coating priority score for each candidate optimization region includes: Input the heat conduction path characteristics of the candidate optimization region into the coating decision model; Obtain the predicted change in temperature standard deviation from the model output; The predicted change in temperature standard deviation is scaled proportionally to the scoring interval to generate a priority score.

8. The method for optimizing the coating position of battery thermal conductive adhesive as described in claim 1, characterized in that, Also includes: Monitor the real-time temperature distribution data of the current battery pack; When the deviation between the real-time temperature distribution data and the historical data sequence exceeds the allowable range, the dynamic update of the heat conduction path topology network is triggered. The target coating location is regenerated using the updated topology network.

9. The method for optimizing the coating position of battery thermal conductive adhesive as described in claim 1, characterized in that, It also includes coating location verification operations: After the thermally conductive adhesive is applied to the target coating location, the battery pack operating temperature data is collected. Analyze the temperature uniformity variation within a preset range around the coating location; When the change in temperature uniformity does not reach the expected value, a correction signal is fed back to the coating decision model.

10. A battery thermal conductive adhesive coating position optimization system, characterized in that, For implementing the method according to any one of claims 1-9, comprising: The data acquisition and region division module is used to acquire the structural parameters and operating condition parameters of the target battery pack, and divide the heat conduction region of the battery pack based on the structural parameters to generate multiple heat conduction sub-regions. The heat flow feature extraction module is used to collect temperature distribution data sequences of historical battery packs under multiple operating conditions, extract heat flow fluctuation features from the temperature distribution data sequences, and obtain multiple heat flow fluctuation extreme values ​​and their corresponding heat conduction path features. The topology network construction module is used to construct a heat conduction path topology network, taking the multiple heat conduction sub-regions as network nodes, the extreme values ​​of heat flow fluctuations as node attributes, and the heat conduction path features as connection weights between nodes. The feature learning module is used to learn features from the heat conduction path topology network using a graph convolutional neural network and output a set of heat conduction influencing factors for each heat conduction sub-region. The candidate region filtering module is used to calculate the thermal resistance coefficient of each thermal conduction sub-region based on the set of thermal conduction influencing factors, and to filter sub-regions whose thermal resistance coefficient exceeds a preset threshold to generate a set of candidate optimization regions. The coating priority decision module is used to extract the heat conduction path features of the candidate optimization region set, input them into the pre-trained coating decision model, and output the thermally conductive adhesive coating priority score for each candidate optimization region. The target location determination module is used to sort the candidate optimization region set based on the priority score and select a predetermined number of regions with the highest score as the target coating location.

Images