Circuit board thermal management method based on multi-level cooperation

Abstract

The application discloses a circuit board thermal management method based on multi-level cooperation, comprising the following steps: constructing a thermal field real-time acquisition system based on a multi-modal sensor array, fusing spatial layout, thermal state data and functional attributes, and forming a structured thermal field sampling data set; normalizing, filtering and feature dimension reduction processing the thermal field data to improve modeling accuracy and sensitivity; establishing a circuit board thermal field topology atlas based on the processed data, extracting structural features such as node centrality, clustering coefficient and connectivity, and realizing thermal distribution pattern recognition; according to a topology feature matching priority strategy library, dynamically generating multi-objective optimization thermal regulation conflict resolution instructions, and continuously optimizing strategy mapping through feedback and online learning mechanism, and the application significantly improves the intelligence, self-adaptation and response efficiency of the circuit board thermal management.

Description

Technical Field

[0001] This invention relates to the field of thermal management technology for circuit boards, and more particularly to a thermal management method for circuit boards based on multi-level collaboration. Background Technology

[0002] With the rapid development of electronic products, the integration and power consumption of circuit boards are constantly increasing, making thermal management an increasingly critical factor affecting system performance and reliability. Mainstream thermal management systems typically employ multi-objective optimization methods, using strategies such as real-time adjustment of fan speed, heat flow distribution, and power consumption distribution to achieve circuit board temperature control, energy management, and reliability assurance. Meanwhile, multi-level collaborative thermal regulation has become an important trend in high-end circuit board thermal management, achieving refined heat distribution management and resource sharing through intelligent scheduling and decomposition at the board, module, and even system levels.

[0003] Existing publicly available solutions are mostly based on threshold-based rule scheduling, mixed-integer programming algorithms, or static policy library mapping. Some high-end solutions combine finite element thermal simulation or machine learning-based thermal prediction models to achieve higher accuracy and better responsiveness. In recent years, the industry has focused on improving the thermal management system's ability to handle complex operating conditions (such as transient high power consumption, non-uniform hotspot distribution, and heat dissipation imbalance) through stronger data fusion, adaptive scheduling, and feedback mechanisms.

[0004] Despite some progress in existing multi-level thermal management and conflict resolution mechanisms for circuit boards, the following prominent technical shortcomings and unmet needs remain:

[0005] (1) The strategy is rigid. Traditional rule-driven or threshold-based conflict resolution methods are difficult to adjust autonomously for different thermal field distribution patterns. Their control logic relies on expert experience or fixed thresholds. When faced with the variable non-uniform heat flow and sudden hot spot transfer of the circuit board, the response is slow or misjudgment occurs.

[0006] (2) Insufficient topology awareness: Most existing control schemes are unable to automatically identify core topological features such as global structure, hot spot aggregation path, and heat flow migration trend in the heat flow network, making it difficult to discover potential systemic conflicts (such as multi-heat source coupling, local resource depletion, etc.). Therefore, the mitigation measures adopted are limited to local optima.

[0007] (3) The strategy matching is singular. Currently, many strategies can only match static templates or preset working conditions. They lack the ability to perceive and evolve in real time changes in the thermal management environment, making it difficult to support online dynamic learning and behavior transfer. This results in limited accuracy, robustness and energy consumption performance of thermal control strategies.

[0008] (4) Weak feedback and adaptive capabilities: Although some solutions introduce feedback control, most fail to achieve global online self-optimization based on the fusion of multi-dimensional thermal field deviation and topology state. This makes it impossible for the system to quickly correct strategies and parameters in situations such as sudden loads and obstructed heat dissipation, leading to problems such as local overheating and increased energy consumption. Summary of the Invention

[0009] In order to solve the above-mentioned technical problems, the present invention provides a circuit board thermal management method based on multi-level collaboration.

[0010] The technical solution of this invention is implemented as follows: a circuit board thermal management method based on multi-level collaboration, comprising:

[0011] S1: Collect temperature, power density and heat flux density data of multiple heat source areas and heat dissipation units on the circuit board, and record the spatial location and functional attributes of each sampling point in combination with the physical layout information of the thermal management system.

[0012] S2: Normalize and filter outliers on the collected thermal field data to eliminate the impact of transient disturbances and sensor drift on the accuracy of subsequent modeling.

[0013] S3: Based on the processed thermal field data, construct a thermal field topology map of the circuit board, where the graph nodes represent heat sources or heat dissipation units, and the graph edges represent the thermal flow coupling strength and direction.

[0014] S4: Extract structural features from the thermal field topology map, including node centrality, clustering coefficient and heat flow path connectivity, to characterize the current heat distribution pattern type;

[0015] S5: Input the extracted topological features into the preset conflict resolution strategy matching module, and match the corresponding priority scheduling rules and feedback adjustment parameter set in the feature similarity matching strategy library;

[0016] S6: Generate preliminary conflict resolution instructions based on the matching results, including multi-objective optimization weight allocation, thermal regulation action sequence and execution priority sorting;

[0017] S7: After executing the conflict resolution command, continuously collect the thermal field status feedback data of the circuit board and calculate the deviation index between the actual thermal field change and the prediction model.

[0018] S8: Based on the aforementioned deviation index, the mapping relationship between topological features and conflict resolution strategies is updated through a lightweight online learning mechanism to achieve dynamic optimization and adaptive evolution of the strategy library.

[0019] The circuit board thermal management method based on multi-level collaboration provided by this invention has the following beneficial effects:

[0020] (1) This invention differs from conventional processing methods that rely on fixed thresholds, preset rules, or complex mixed integer programming. It innovatively uses dynamic modeling of thermal field topology to identify different thermal distribution patterns (such as uniform, hotspot clustering, gradient mutation, etc.) in real time and adjusts the conflict resolution logic accordingly. Through topology feature-driven strategy matching and dynamic feedback mechanism, the policy library achieves environment-aware self-evolution. This method still has excellent intelligent decision-making capabilities in complex scenarios such as unsteady thermal environment and sudden power fluctuations, effectively improving the limitations of existing rigid policy execution and lagging pattern transfer.

[0021] (2) This invention achieves higher information compression efficiency and pattern recognition accuracy through multidimensional thermal field data normalization, graph structure feature extraction, and principal component analysis dimensionality reduction. The deep integration of topological structure features (such as node centrality, clustering coefficient, and connectivity) with multi-objective performance weights enables the weight allocation and optimization action sequence to ensure temperature control accuracy while taking into account multiple objectives such as energy consumption efficiency and control response speed. Compared with traditional hierarchical optimization or expert rule methods, this method significantly improves the overall thermal control performance.

[0022] (3) This invention, by constructing a weighted directed graph of the thermal field and applying graph theory algorithms such as shortest path and minimum spanning tree, greatly reduces the interference of redundant heat flow paths on multi-level collaborative control, resulting in a clearer logical link and a data structure that is more conducive to high-concurrency scheduling. The introduction of precise feature similarity matching and dynamic online learning in the strategy generation process enables rapid convergence to the optimal or suboptimal strategy based on real-time thermal field changes, avoiding the high computational burden caused by retraining and large-scale global search, and significantly enhancing the real-time performance and large-scale deployment feasibility of the industrial-grade thermal management system;

[0023] (4) This invention uses a lightweight online learning algorithm (such as incremental support vector machine) to correct the mapping relationship between topological features and policies in real time, forming a closed-loop feedback-driven adaptive update. This method can automatically reconstruct the optimal match for situations such as long-term operating condition drift, sensor drift, and structural changes caused by manufacturing tolerances, maintaining stable and efficient control performance and overcoming the problem of degradation over time in traditional static optimization schemes. Attached Figure Description

[0024] Figure 1 The flowchart shows the circuit board thermal management method based on multi-level collaboration of the present invention.

[0025] Figure 2 This is a sub-flowchart of the circuit board thermal management method based on multi-level collaboration of the present invention;

[0026] Figure 3 This is another sub-flowchart of the circuit board thermal management method based on multi-level collaboration of the present invention. Detailed Implementation

[0027] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0028] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0029] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising / including” or “having,” etc., specify the presence of the stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0030] Please see Figures 1-3 As shown, the circuit board thermal management method based on multi-level collaboration includes:

[0031] S1: Collect temperature, power density and heat flux density data of multiple heat source areas and heat dissipation units on the circuit board, and record the spatial location and functional attributes of each sampling point in combination with the physical layout information of the thermal management system.

[0032] S2: Normalize and filter outliers on the collected thermal field data to eliminate the impact of transient disturbances and sensor drift on the accuracy of subsequent modeling.

[0033] S3: Based on the processed thermal field data, construct a thermal field topology map of the circuit board, where the graph nodes represent heat sources or heat dissipation units, and the graph edges represent the thermal flow coupling strength and direction.

[0034] S4: Extract structural features from the thermal field topology map, including node centrality, clustering coefficient and heat flow path connectivity, to characterize the current heat distribution pattern type;

[0035] S5: Input the extracted topological features into the preset conflict resolution strategy matching module, and match the corresponding priority scheduling rules and feedback adjustment parameter set in the feature similarity matching strategy library;

[0036] S6: Generate preliminary conflict resolution instructions based on the matching results, including multi-objective optimization weight allocation, thermal regulation action sequence and execution priority sorting;

[0037] S7: After executing the conflict resolution command, continuously collect the thermal field status feedback data of the circuit board and calculate the deviation index between the actual thermal field change and the prediction model.

[0038] S8: Based on the aforementioned deviation index, the mapping relationship between topological features and conflict resolution strategies is updated through a lightweight online learning mechanism to achieve dynamic optimization and adaptive evolution of the strategy library.

[0039] Step S1: Collect temperature, power density, and heat flux density data of multiple heat source areas and heat dissipation units on the circuit board, and record the spatial location and functional attributes of each sampling point in conjunction with the physical layout information of the thermal management system. Specifically, this includes:

[0040] S1.1: Based on the physical layout design of the thermal management system, obtain the topological position coordinates of each heat source area and heat dissipation unit on the circuit board to form a spatial positioning information dataset to support subsequent thermal field topology mapping.

[0041] Based on the physical layout design drawing of the thermal management system, a computer-aided design (CAD) image analysis algorithm (parameters: resolution 0.01mm / pixel, coordinate system origin located at the lower left corner of the circuit board) is used to achieve accurate conversion of the physical layout drawing into a computable spatial model.

[0042] By using a hybrid localization method based on Hough transform and edge detection (parameters: Canny threshold 1.5, Hough cumulative threshold 100), the automatic identification of the heat source region and the boundary of the heat dissipation unit is achieved, and the coordinate set of the node boundary polygon is obtained.

[0043] Furthermore, the spatial center point of each heat source region and heat dissipation unit is calculated using a geometric centroid calculation algorithm (formula below):

[0044]

[0045] Where, x i With y i Let be the coordinates of the i-th sampling point on the node boundary, n be the number of sampling points, and C be the coordinates of the geometric center.

[0046] Based on a unified spatial reference system, the calculated center point coordinates are indexed and matched with the functional module database in electronic design automation (EDA) to realize the functional attribute binding of node topology positions, ensuring that the subsequent thermal field topology mapping has an accurate physical-functional mapping relationship.

[0047] The location dataset is numerically quantized by setting a coordinate accuracy threshold (e.g., 0.01 mm) to generate a formatted spatial positioning information dataset, which serves as the basic input for thermal field topology modeling.

[0048] Through the above algorithm processing method, the CAD physical layout information of the previous step is transformed into computable topological position coordinate data with spatial accuracy and functional labels, so as to realize the precise spatial positioning of heat source and heat dissipation unit nodes in the thermal field data structure.

[0049] For example, on a high-density circuit board measuring 120mm × 90mm, a CAD layout drawing with a resolution of 0.01mm / pixel is obtained. Twenty polygonal heat source boundaries and five heat dissipation unit boundaries are extracted using an edge detection threshold of 1.5 and a Hough transform cumulative threshold of 100. For each boundary, 200 points are sampled and input into the geometric centroid formula to calculate the center point; for example, the centroid of a heat source is calculated as (35.27, 42.63)mm. All node centroid coordinates are quantized to a precision of 0.01mm and associated with the EDA functional module ID, ultimately forming a spatial positioning information dataset containing 25 functional nodes (including spatial coordinates and functional labels). This dataset is input into the multimodal sensor array deployment algorithm in subsequent step S1.2, which can significantly improve sensor layout accuracy and reduce installation redundancy.

[0050] S1.2: Deploy multimodal sensor arrays in each heat source area and heat dissipation unit to synchronously collect temperature, power density and heat flux density to obtain multidimensional thermal field state data streams as raw inputs for thermal behavior feature extraction;

[0051] Based on the spatial positioning information dataset generated in step S1.1, the center point coordinates and functional labels of each heat source region and heat dissipation unit are used as spatial and functional constraint input conditions for sensor deployment.

[0052] A density-based multimodal sensor location algorithm (parameters: minimum number of points 3, neighborhood radius of 1-10mm, preferably 5mm) is adopted to achieve clustering and division of the local area where the heat source and heat dissipation unit are located, ensuring that the sensor array is densely arranged in the high heat flux gradient region and sparsely arranged in the low gradient region.

[0053] Furthermore, by using an improved greedy coverage optimization algorithm (parameters: coverage radius of 1-5mm, preferably 2.5mm, fault tolerance ratio of 1%-10%, preferably 5%), the minimum redundant installation of temperature sensor, power density sensor and heat flux density sensor is achieved, and the redundancy of each node is reduced by no less than 30%, thereby reducing layout duplication and improving the uniformity of signal coverage.

[0054] Furthermore, a multi-channel synchronous sampling control logic (clock synchronization accuracy 1μs) is adopted to achieve time-synchronous triggering sampling of three types of sensors: temperature (thermocouple), power density (Hall effect-based current sensing module), and heat flux density (heat flux meter), and to obtain the temperature sequence T. i Power density sequence P iWith heat flux density sequence Q i ;

[0055] Furthermore, by employing low-noise amplification and bandpass filtering (filter bandwidth 0.1Hz~10Hz) in the analog front-end signal conditioning circuit, electromagnetic interference and high-frequency jitter noise are suppressed, and a stable multidimensional analog signal stream is generated.

[0056] Furthermore, a high-precision analog-to-digital converter (16-bit resolution, 500Hz sampling rate) is used to achieve distortion-free conversion of analog signals to digital signals, resulting in a digital temperature matrix T. d Power density matrix P d With the heat flux density matrix Q d ;

[0057] Through the above multi-level acquisition and synchronous processing, the spatial positioning information and functional attributes of step S1.1 are combined and transformed into a multi-dimensional thermal field state data stream that covers the whole world and is time-aligned, so as to realize the complete original input construction of thermal behavior characteristics.

[0058] For example, on a 120mm × 90mm circuit board, the coordinates of the 25 functional nodes generated in step S1.1 are input. Based on a density clustering algorithm, the neighborhood radius is set to 5mm, dividing the area into 8 cluster regions. Regions with power density variations exceeding 10% are identified as high-gradient regions and assigned 3 sensors (temperature, power density, and heat flux density). The remaining low-gradient regions are assigned 1-2 sensors. The optimal sensor placement is calculated using a greedy coverage algorithm, reducing the total number of sensors from the initial 50 to 35, reducing redundancy by 30%. The sensor output signals are filtered by an analog front-end with a bandwidth of 0.1Hz–10Hz and acquired by a 16-bit ADC. The temperature matrix T... d The sampling accuracy reaches ±0.05℃, and the power density matrix P d The measurement resolution is 0.01 W / cm. 2 Heat flux density matrix Q d The measurement resolution is 0.5 W / m 2 The multidimensional thermal field state data stream meets the accuracy requirements of subsequent sensor calibration and thermal flux coupling modeling in S1.3;

[0059] S1.3: Zero-point drift compensation and nonlinear error correction are performed on the raw thermal field data collected by the sensor calibration algorithm to improve the data acquisition accuracy and provide reliable data support for subsequent modeling of thermal-fluid coupling relationship;

[0060] S1.4: Based on the functional module division information of the heat source area, the functional attributes of the collection points are labeled to generate a function-thermal mapping table, which is used for the construction of priority scheduling rules in the subsequent strategy matching module;

[0061] S1.5: Perform multi-source data fusion processing on the spatial positioning information dataset, multi-dimensional thermal field state data stream, and function-thermal state mapping table to generate a structured thermal field sampling dataset, which serves as the input basis for constructing the thermal field topology map.

[0062] Step S2 involves normalizing and filtering the collected thermal field data to eliminate the impact of transient disturbances and sensor drift on the accuracy of subsequent modeling. Specifically, this includes:

[0063] S2.1: Normalize the temperature, power density and heat flux density data of each area of ​​the circuit board, and perform linear transformation on the multidimensional heterogeneous data based on the min-max normalization algorithm to eliminate the influence of different physical dimensions on the thermal field topology modeling and obtain a standardized thermal field data matrix under a unified dimension.

[0064] For the three physical quantities of temperature, power density and heat flux density in the structured thermal field sampling dataset after S1.5 multi-source fusion, the minimum-maximum normalization algorithm (parameter: normalization interval [0,1]) is used to realize the linear interval mapping of data under different physical units and eliminate the influence of physical unit differences on subsequent thermal field topology modeling.

[0065] Furthermore, by employing a grouping normalization strategy based on the type of physical quantity, the minimum and maximum values ​​of temperature, power density, and heat flux density within the sampling period are used as mapping benchmarks to calculate the normalization conversion coefficient k. g With translation coefficient b g And execute the formula:

[0066]

[0067] Where X is the original observation value, X min With X max These are the minimum and maximum values ​​for the corresponding physical quantity groups, X. norm This is the normalized output value;

[0068] Furthermore, a range protection mechanism (parameters: lower limit 0, upper limit 1) is introduced to perform truncation compensation on abnormal values ​​that exceed the boundary of the normalization result, so as to prevent extreme noise values ​​from amplifying the perturbation effect on the feature space during the normalization process.

[0069] Furthermore, batch matrix operations are employed to accelerate the process. Based on column vectorization operations, the temperature, power density, and heat flux density matrices are simultaneously input into the normalization operator to generate a three-dimensional normalized thermal field data matrix M under a unified dimension. norm ;

[0070] By employing a min-max normalization and range protection strategy, the original thermal field observations with varying physical units are transformed into a standardized numerical domain matrix, thereby achieving consistent control of characteristic dimensions and unified data input for subsequent thermal field topology modeling.

[0071] For example, on a circuit board measuring 120mm × 90mm, three types of physical quantities from an array of 35 sensors are sampled 500 times per day, with a temperature range of [35.32, 87.46]℃ and a power density range of [0.015, 0.128]W / cm². 2 Heat flux density range [215.7, 645.2] W / m 2 For the temperature channel, execute the following formula:

[0072]

[0073] The power density and heat flux density are subjected to the same normalization mapping, and normalization values ​​exceeding [0,1] are truncated. The final generated M norm The mean volatility of the matrix across all channels is reduced to 35% of its original value, providing a stable input for the S2.2 outlier filter;

[0074] S2.2: Based on the sliding window mechanism, perform time-series filtering on the standardized thermal field data matrix, and use an improved median filtering algorithm to detect and correct outliers in the temperature and power density sequences, so as to eliminate false thermal signals caused by sensor transient disturbances or sudden changes in local hot spots, and obtain a time-stable filtered thermal field dataset.

[0075] S2.3: Perform spatial consistency verification on each sampling point in the filtered thermal field dataset. Determine whether there is local sensor drift based on the gradient change threshold of heat flux density in adjacent regions. If drift is detected, use a Kalman filter to dynamically compensate the current sampling point data to improve the physical rationality of spatial distribution data and the accuracy of topology modeling.

[0076] S2.4: Construct a multidimensional feature vector space based on the corrected spatially consistent thermal field data, and use principal component analysis (PCA) to compress the data dimensions and extract the main thermal field change features to reduce the computational complexity of subsequent thermal field topology modeling and improve the model's sensitivity to heat distribution patterns.

[0077] S2.5: Perform dynamic window sliding mean calculation on the thermal field feature vector after dimensionality reduction by principal component analysis to generate a thermal field state sequence with time sliding characteristics. This is used to support the dynamic updating of the subsequent thermal field topology map and the real-time identification of thermal distribution patterns, ensuring that the multi-level collaborative conflict resolution mechanism has sufficient response sensitivity and predictive ability.

[0078] Step S3: Based on the processed thermal field data, construct a thermal field topology map of the circuit board, where graph nodes represent heat sources or heat dissipation units, and graph edges represent the thermal flux coupling strength and direction. For example... Figure 2 As shown, it specifically includes:

[0079] S3.1: Classify the heat sources and heat dissipation units in the processed thermal field data by functional attributes, and divide the topology node set based on the spatial location information of their physical layout in order to establish a graph structure basic representation of the thermal behavior of the circuit board.

[0080] S3.2: Based on the Fourier thermal conductivity relationship between heat flux density and temperature gradient, calculate the thermal coupling strength between adjacent heat sources and heat dissipation units to obtain a quantitative characterization of the heat transfer path;

[0081] Based on the set of heat source and heat dissipation unit nodes and their spatial layout information obtained in step S3.1, adjacent node pairs are selected as the objects of heat flow coupling calculation.

[0082] The steady-state thermal flux density between node pairs is calculated using the Fourier thermal conductivity relation (parameter: thermal conductivity k is taken from material measurements or thermal databases). The formula is:

[0083]

[0084] Where q is the heat flux density, T is the temperature, and x is the spatial distance;

[0085] Furthermore, by substituting the temperature difference between adjacent nodes and the spatial center distance into the discretized Fourier formula, the heat flux density is calculated:

[0086]

[0087] Among them, T i and T j The temperatures of node i and node j are respectively, d ij The spatial distance between the two;

[0088] Furthermore, the calculated heat flux density is compared with the interaction area A between nodes. ij Multiply to obtain the total heat flow Q ij :

[0089] Q ij =q ij A ij

[0090] Furthermore, the obtained total heat flux value is subjected to symmetric normalization (parameter: the normalization factor is the maximum heat flux value in the figure), and the dimensionless heat flux coupling strength is calculated:

[0091]

[0092] Among them, W ij The normalized intensity takes values ​​in the range [0,1].

[0093] Through the above calculation chain, the temperature gradient, spatial distance and material thermal conductivity are combined and transformed into quantitative coupling strength data of the heat flow transfer path between node pairs, so as to realize the physical accurate assignment of the thermal field topology edge weights.

[0094] For example, on a circuit board with dimensions of 120mm × 90mm, the center distance between node A and node B is 15mm, the thermal conductivity of the material is k = 205W / (m·K), and the temperature T of node A is... A =80.2℃, node B temperature T B =72.6℃, contact area between nodes A AB = 2.5 cm². Substituting into the formula, we get the heat flux density q = -205 × ((80.2 - 72.6) / 0.015) ≈ -105,067 W / m², and the total heat flux Q. AB ≈-262.67W. The absolute value of the maximum heat flux across the entire graph is 300W, then the normalized coupling strength W AB ≈0.875. This data will be used as the edge weight input in the subsequent S3.3 directional modeling to improve the accuracy and stability of heat flow migration trend identification;

[0095] S3.3: The heat flow transfer path is modeled directionally using the directional pattern modeling method to construct a heat flow transfer direction matrix, which is used to represent the migration trend of heat energy from high-density areas to low-density areas;

[0096] Based on the normalized heat flux coupling strength data between nodes calculated in step S3.2, node pairs whose absolute heat flux value exceeds the direction determination threshold are selected as directional modeling objects. A directional pattern modeling method (parameter: threshold 0.05) is used to determine and encode the heat flux migration direction.

[0097] Furthermore, by calculating the temperature difference ΔT at the nodes... ij The sign of the heat flux is used to determine the direction of heat flow from high-temperature nodes to low-temperature nodes. The formula is:

[0098] D ij =sign(T) i -T j )

[0099] Among them, D ij As a directional indicator, 1 indicates that it points from i to j, and -1 indicates that it points from j to i;

[0100] Furthermore, the directional identifier is coupled with the normalized heat flux coupling strength W. ij Combine them to construct the elements of the heat flow transfer direction matrix:

[0101] F ij =D ij ×W ij

[0102] To achieve data encoding that simultaneously describes the intensity and direction characteristics of heat flux;

[0103] Furthermore, the above calculations are performed sequentially for all node pairs, and the results are filled into the corresponding positions of the heat flow direction matrix F. Non-adjacent node pairs are filled with 0 to maintain matrix sparsity and computational efficiency.

[0104] Furthermore, by performing matrix normalization (parameter: normalization by row vector magnitude), consistent control of the total outflow from different nodes is achieved, preventing directional imbalance caused by excessively large absolute values ​​at local hotspot nodes.

[0105] By using the directional graph modeling method, the thermal flux coupling intensity data between nodes is transformed into a quantized matrix containing intensity and direction, realizing a formal expression of the migration trend of thermal energy from high-density areas to low-density areas, and providing accurate directional input for the subsequent construction of a weighted directed graph model;

[0106] For example, on a circuit board with dimensions of 120mm × 90mm, the temperature of node A is 80.2℃, the temperature of node B is 72.6℃, and the normalized thermal flux coupling strength W... AB =0.875. Calculating the temperature difference using the formula, 80.2 - 72.6 = 7.6℃, with a positive sign, we obtain the directional indicator D. AB =1. Combining the strength, we obtain the matrix element F. AB =0.875, assigned a value at the corresponding position in the direction matrix, realizing the directional expression of heat flow from high-temperature node A to low-temperature node B. Performing the same process on all node pairs on the entire plate, the non-zero elements of the sparse direction matrix are concentrated in the hotspot neighborhood, verifying the effectiveness of heat migration trend identification;

[0107] S3.4: Based on the thermal flux coupling strength and direction information between nodes, a weighted directed graph model is constructed, where nodes represent heat sources or heat dissipation units, and edges represent thermal flux coupling paths and their directionality, in order to form the preliminary structure of the thermal field topology map;

[0108] Based on the normalized heat flux coupling intensity matrix and heat flux transfer direction matrix between nodes calculated by steps S3.2 and S3.3, a weighted directed graph modeling method is adopted (parameter: the set of graph nodes is derived from the spatial function partitioning result of step S3.1) to realize the structured representation of node topological relationships;

[0109] Furthermore, by combining the direction matrix F and the intensity matrix W, a weighted directed edge weight is constructed, and the calculation formula is as follows:

[0110] E ij =F ij ×W ij

[0111] Among them, E ij The weighted directed edge weights from node i to node j comprehensively reflect the coupling characteristics of heat flow direction and intensity.

[0112] Furthermore, the aforementioned edge weights are filled into the weighted directed adjacency matrix A, and the corresponding positions of non-adjacent nodes are assigned a value of 0 to form a sparse structure storage form, while retaining node identifiers and function labels to support subsequent feature extraction and policy mapping.

[0113] Furthermore, the constructed weighted adjacency matrix is ​​normalized (parameter: calculated by the out-degree normalization factor), as follows:

[0114]

[0115] The denominator is the sum of all outgoing edge weights of node i, which is used to eliminate the global skew caused by the large absolute heat flow value of local hotspot nodes.

[0116] Furthermore, through the serialization interface of the graph data structure, the weighted directed topological graph is encoded into a set of node entities and a set of edge entities. The node entities contain position vectors, functional attributes and topological indexes, and the edge entities contain start nodes, end nodes and normalized weights, which are used for efficient access and algorithm interface calls.

[0117] By using a weighted directed graph model construction method, the directional and intensity data obtained in the previous step are transformed into a preliminary thermal field topology structure with both quantitative and structural characteristics, thereby realizing the formal expression of the heat transfer network and its availability for subsequent computation.

[0118] For example, on a circuit board with dimensions of 120mm × 90mm, assuming nodes A, B, and C represent a high-power chip, a medium-power chip, and a heatsink, respectively, the normalized heat flux intensity matrix W contains W AB 0.875, W BC 0.652, F in the direction matrix F AB 1. F BC 1. Substitute into the formula to calculate the weighted boundary value E. AB 0.875, E BC 0.652. Let the sum of the outgoing edges of node A be 0.875 and the sum of the outgoing edges of node B be 0.652. After normalization, A... AB 1.0, A BCVersion 1.0 shows that the non-zero elements of the weighted directed adjacency matrix are concentrated in the paths A→B and B→C, reflecting the main links of hotspot sequence transmission and providing structured input for subsequent minimum spanning tree optimization and pattern recognition.

[0119] S3.5: Perform graph structure optimization processing on the constructed thermal field topology map, and use the minimum spanning tree algorithm to remove redundant heat flow paths to improve the logical clarity and physical interpretability of the topology map.

[0120] Step S4: Extract structural features from the thermal field topology map, including node centrality, clustering coefficient, and heat flow path connectivity, to characterize the current heat distribution pattern type. For example... Figure 3 As shown, it specifically includes:

[0121] S4.1: Calculate the node centrality of the graph nodes in the completed circuit board thermal topology map. Based on the node degree centrality and proximity centrality index, quantify the relative importance of each heat source or heat dissipation unit in the heat flow network to obtain the node influence feature vector.

[0122] For each node in the completed circuit board thermal topology map, the node degree centrality calculation method is adopted (parameter: node degree is defined as the number of edges connecting adjacent nodes, including edge weighting) to realize the function of quantifying the direct connection capability of nodes in the local network structure.

[0123] Furthermore, by using a proximity centrality calculation method (parameter: the reciprocal of the average shortest path distance from a node to all other nodes), the function of characterizing the information transmission efficiency of a node in the global heat flux network is realized, and the quantitative value of the influence range of each node in the topology relative to the global network is obtained.

[0124] Furthermore, for degree centrality calculation, based on the weighted directed adjacency matrix A, the following formula is performed:

[0125]

[0126] Where d is the total number of weighted edges directly connecting nodes, N is the total number of nodes in the graph, and C is obtained. Di The value range is [0,1], which is used to represent the local connectivity density of a node;

[0127] Furthermore, for the calculation of near-centrality, the formula is used:

[0128]

[0129] Where d ij The shortest path length calculated by Dijkstra's algorithm on the weighted directed graph is represented by the inverse ratio, which reflects the efficiency of node information transmission.

[0130] Furthermore, normalization is performed on degree centrality and proximity centrality (parameter: normalize by global maximum value) to unify the dimensions of different centralities and eliminate numerical skew caused by differences in network size;

[0131] Furthermore, by connecting the degree centrality and proximity centrality vectors, a two-dimensional node influence feature vector C is constructed. D C C The vector is then stored in the topological feature space data structure as the basic input data for subsequent pattern classification and policy matching.

[0132] By using node centrality calculation and feature fusion processing, the weighted directed graph structure of the previous step is transformed into a two-dimensional centrality feature that can reflect local connectivity and global reachability, thereby realizing a quantitative characterization of the relative importance and potential key role of each heat source or heat dissipation unit in the heat flow network.

[0133] For example, in the thermal topology diagram of a 100mm × 80mm circuit board, node A connects to nodes B and C, with edge weights of 0.85 and 0.65 respectively, and the total number of nodes in the network is 6. Calculate the weighted degree d = 2 for node A, and substitute it into the formula to obtain... Using Dijkstra's algorithm, the sum of the shortest path distances from A to all other nodes is found to be 7.5. Substituting this into the proximity centrality formula, we get... Normalizing both centralities to their global maximum values ​​yields a normalized centrality vector (0.8, 0.95), which is then stored in the feature space. This result indicates that node A exhibits efficient heat transfer paths and dense local connections within the global network, making it a crucial node for priority regulation in hotspot clustering patterns.

[0134] S4.2: Based on the local adjacency relationship in the graph, calculate the clustering coefficient of each node to reflect the tightness of heat flow coupling in its neighborhood, so as to extract the local heat field aggregation characteristics and form a local heat interaction intensity characteristic index.

[0135] Based on the task requirements of the S4 main step, the weighted directed topology graph and node adjacency information input from the S4.1 step are used as initial conditions, and an appropriate clustering coefficient calculation model is selected to extract the local heat flow coupling tightness features.

[0136] A local clustering coefficient calculation method based on the adjacency matrix is ​​adopted (parameter: weighted undirection is used to eliminate the interference of directionality on the clustering coefficient calculation) to quantify the degree of heat flow closure in the neighborhood of each node;

[0137] Furthermore, an adjacency submatrix for node i is constructed using adjacency matrix A, which is used to count the actual number E edges between neighboring nodes. i ;

[0138] Furthermore, an improved weighted clustering coefficient formula is used to calculate node C. Li :

[0139]

[0140] Where, k i Let be the number of neighbors of node i, and correct it in the weighted graph by averaging the weights of the adjacent edges;

[0141] Furthermore, perform an undirected operation on the weighted adjacency matrix (parameter: The weights are normalized to reduce the bias effect of extreme heat flux edge weights on the clustering coefficients.

[0142] Furthermore, the clustering coefficients of each node are calculated to form a local thermal interaction intensity feature vector, which is then bound and stored with the node ID and functional label to ensure that it can be directly called in the subsequent S4.4 feature fusion step.

[0143] By calculating and normalizing clustering coefficients, the weighted directional topological information from the previous step is transformed into a quantitative index that can characterize the degree of local heat flow convergence, thereby achieving accurate extraction and quantitative description of the local heat field convergence characteristics.

[0144] For example, in the thermal topology of a 100mm × 80mm circuit board, suppose node P is connected to Q and R with edge weights of 0.8 and 0.6 respectively, and there is an edge weight of 0.7 between nodes Q and R. The number of neighbors of node P is k. P =2, the actual number of edges between neighbors E P =1. Substitute into the formula to calculate the weighted undirected clustering coefficients: The value remains 1.0 after normalization, indicating that the neighborhood heat flow structure of node P is completely closed, and the local thermal interaction intensity is extremely high. In this example test, using the clustering coefficient distribution of all nodes on the plate as input, it was found that high values ​​are concentrated in the hot spot area of ​​the triangular closed structure, which is consistent with the high temperature area measured by infrared thermal imaging, thus realizing the effective quantification of local heat flow accumulation characteristics.

[0145] S4.3: Perform topological connectivity analysis on the heat flow paths, and use the shortest path algorithm to calculate the heat flow reachability between any two nodes in the graph to obtain the global heat flow path connectivity matrix, which characterizes the overall structural connectivity of the thermal field.

[0146] Based on the task requirements of the S4 main step, the weighted undirected topology matrix and node set input from the S4.2 step are used as the initial conditions for topology connectivity analysis. A shortest path search algorithm suitable for weighted directed graphs is selected to calculate the heat flow reachability between any two nodes.

[0147] Dijkstra's shortest path algorithm is used (parameter: path weights are taken from the edge inverses of the weighted directed adjacency matrix to ensure that the path length is shorter as the heat flux intensity increases), to solve for the shortest path length between each pair of nodes and generate a global shortest path distance matrix D. ij ;

[0148] Furthermore, a path reachability determination method (parameter: if the shortest path length is less than an infinite threshold, the node pair connectivity is recorded as 1; otherwise, it is recorded as 0) is used to implement the binary encoding of heat flow reachability between any node pair, resulting in the heat flow reachability matrix R. ij ;

[0149] Furthermore, the global connectivity calculation formula is adopted:

[0150]

[0151] Where N is the total number of nodes, D ij To achieve the shortest path length, the global heat flow path connectivity is quantified;

[0152] Furthermore, the obtained global connectivity index and reachability matrix are normalized (parameter: normalized according to the maximum possible value of the entire graph) to obtain the dimensionless connectivity matrix C. ij This is to allow it to be integrated with other features such as node centrality and clustering coefficients in the same feature space;

[0153] By using the shortest path algorithm and reachability analysis, the global structural information of the topology graph is transformed into connectivity features that can directly reflect the overall heat flow transfer efficiency and connection integrity, thereby achieving an accurate characterization of the global structural connectivity of the thermal field.

[0154] For example, on a circuit board with dimensions of 100mm × 80mm, there are four nodes A, B, C, and D. In the weighted directed adjacency matrix, the weights for A→B are 0.8, B→C is 0.6, A→D is 0.4, and D→C is 0.7. Using the edge inverse weights as the path length input to Dijkstra's algorithm, the shortest path from A to C is found to be A→B→C, with a total length of 2.92 (obtained by adding the lengths of 1.25 + 1.67 meters). Based on reachability determination, the connectivity between A and C is set to 1. A reachability matrix is ​​constructed, and the global connectivity is calculated. Let N = 4. The reciprocals of the distances of all reachable paths are summed and substituted into the global connectivity formula to obtain L. c ≈0.82. After normalization, the connectivity value between A and C is 0.85. The final global connectivity matrix shows higher values ​​for A, B, and C, and relatively lower values ​​for node D, verifying that this node is located at the edge of the heat flow network and contributes little to the global heat transfer.

[0155] S4.4: Multi-dimensional feature fusion of node centrality, clustering coefficient and heat flow path connectivity matrix to construct thermal field topological feature space, so as to form a unified topological feature description vector;

[0156] S4.5: Based on the preset heat distribution pattern classification rules, perform pattern recognition classification processing on the topological feature description vector, and use the K-nearest neighbor algorithm or support vector machine model to determine whether the current heat field belongs to a uniform, hot spot clustering or gradient mutation heat distribution pattern, and output the pattern classification label as the basis for subsequent strategy matching.

[0157] Step S5: The extracted topological features are input into a preset conflict resolution strategy matching module, which matches the priority scheduling rules and feedback adjustment parameter set corresponding to the feature similarity matching strategy library. Specifically, this includes:

[0158] S5.1: Encode the structural features of the thermal field topology map, and generate feature vectors based on node centrality, clustering coefficient and heat flow path connectivity to construct a digital representation of the current heat distribution pattern;

[0159] S5.2: Execute a similarity measurement algorithm based on feature vectors to match and calculate the feature templates corresponding to various heat distribution patterns in the pre-stored strategy library in order to identify the degree of similarity between the current heat field topology and the existing patterns in the strategy library;

[0160] The current heat distribution pattern feature vector output from step S5.1 is used as input, and a feature similarity measurement algorithm based on Euclidean distance is adopted (parameter: summing the squared differences of each dimension component of the feature vector under a unified normalized dimension and taking the square root) to achieve a preliminary calculation of the similarity between the current heat field topology and each pattern template in the strategy library.

[0161] Furthermore, by using a weighted Euclidean distance algorithm (parameters: the weight coefficients of each feature dimension are set according to their discriminative contribution rate in historical matching results), the importance of different topological feature dimensions is adjusted, and a weighted similarity distance value d is obtained. wi .

[0162] Furthermore, the weighted Euclidean similarity is calculated using the following formula:

[0163]

[0164] Where α k Let x be the weight of the k-th dimension feature. k Let y be the current thermal field eigenvector component. k The template features are the feature values ​​of the strategy library, and n is the total dimension of the features;

[0165] Furthermore, a cosine similarity algorithm (parameter: performing a dot product on the feature vectors and dividing by the product of their magnitudes) is employed to measure the consistency of vector directions, reflecting the relative trend matching degree between the current pattern and the template pattern. Its calculation formula is as follows:

[0166]

[0167] Furthermore, the weighted Euclidean distance and cosine similarity are normalized and fused (parameter: fusion coefficient β determined empirically) to form a comprehensive similarity score S. f This is used for ranking the adaptability of strategy templates;

[0168] By using feature similarity calculation and multi-index fusion processing, the encoded feature vector from the previous step is transformed into a numerical index that can directly characterize the degree of fit between the current thermal field mode and the strategy library template, thereby achieving accuracy and adaptability in the conflict resolution strategy matching process.

[0169] For example, on a circuit board measuring 150mm × 120mm, the feature vector obtained from S5.1 is 0.82, 0.76, 0.65, and the hotspot clustering template feature in the strategy library is 0.80, 0.70, 0.60. Let the weight coefficients be 0.5, 0.3, and 0.2 respectively. Substituting into the weighted Euclidean formula, the weighted sum of squared differences is calculated as 0.0004 + 0.00108 + 0.00005 = 0.00153. Taking the square root, the weighted distance is approximately 0.0391. Using the cosine similarity formula, the dot product is 1.446, and the product of moduli is approximately 1.453, resulting in a cosine similarity of approximately 0.995. The fusion coefficient β is set to 0.6, and the overall similarity S... f =0.6×(1-0.0391)+0.4×0.995≈0.976. This value exceeds the matching threshold of 0.95, and is included in the candidate strategy template set, verifying the high-precision matching capability of this step under different feature dimension combinations;

[0170] S5.3: Based on the similarity measurement results, a set of candidate strategy templates with a matching degree higher than the set threshold are selected as the input candidate set for the conflict resolution strategy matching module to ensure the accuracy and adaptability of strategy matching;

[0171] The comprehensive similarity score S output from step S5.2 is used to calculate the similarity score. f The matching threshold parameter is input into the filtering algorithm module to realize the dynamic filtering function of candidate strategy templates;

[0172] Furthermore, a threshold-based determination method is adopted (parameter: the matching threshold T is set within the range of 0.8-1.0, preferably 0.95, based on the statistical results of historical strategy execution accuracy), to determine the S of all strategy templates in the strategy library. fThe values ​​are compared sequentially, and a binary mask matrix is ​​generated to mark the values ​​that satisfy S. f The set of template indices for the condition ≥T;

[0173] Furthermore, an index mapping extraction method (parameter: based on the row index with a label value of 1 in the mask matrix) is used to extract the corresponding policy template data structure from the policy library, including priority scheduling rules, feedback adjustment parameter groups, and historical execution performance labels, in order to construct an initial list of candidate policies;

[0174] Furthermore, a duplicate detection algorithm (parameter: based on template ID and rule fingerprint hash value for duplicate determination) is used to filter the templates in the initial list of candidate strategies for uniqueness, eliminating duplicate templates with completely identical instruction sets or whose parameter fine-tuning is insufficient to significantly improve performance, thereby reducing redundant computational overhead;

[0175] Furthermore, the quicksort algorithm is adopted (parameter: by S...). f (arranged in descending order of value), the candidate strategy set after deduplication is prioritized, and the top K are selected as the final candidate strategy set. K is determined by the system's parallel strategy evaluation capability and execution timing constraints.

[0176] Through the above screening and sorting process, the similarity evaluation results output by S5.2 are transformed into a set of candidate conflict resolution strategy templates that are highly adaptable and non-redundant under the current hot field mode. The size of the candidate strategy is controlled within 20% of the total template library, so as to achieve a balance between accuracy and execution efficiency in the strategy matching stage.

[0177] For example, in a high-density circuit board thermal management system with dimensions of 160mm × 120mm, S5.2 outputs a comprehensive similarity score vector for 10 strategy templates: 0.976, 0.943, 0.912, 0.955, 0.890, 0.861, 0.981, 0.915, 0.905, 0.799. A matching threshold T = 0.95 is set, and a mask matrix is ​​constructed to obtain the index set {1, 4, 7}. Among the templates extracted based on this index set, template 1 and template 7 are completely identical in priority scheduling rules and parameter sets. They are determined to be duplicate templates by hash detection and are removed. The remaining templates are sorted in descending order of similarity to obtain the final candidate template sequence [7, 4], and K = 2 is limited by the system's parallel execution capability, serving as the input for step S5.4. Verification showed that this process, while ensuring the accuracy of the screening results, kept the candidate set size within 20%, significantly reduced the computational load of subsequent multi-objective optimization, and maintained the optimal thermal control performance of the strategy execution.

[0178] S5.4: Based on the priority scheduling rules and feedback adjustment parameter set of the candidate strategy template, execute the multi-objective optimization weight allocation algorithm to generate a multi-level collaborative control strategy combination that is adapted to the current heat distribution pattern;

[0179] S5.5: Output the matched strategy combination to the conflict resolution instruction generation module as the strategy input basis for generating preliminary conflict resolution instructions, so as to realize the environmental perception and adaptive selection of thermal control strategy.

[0180] Step S6: Generate preliminary conflict resolution instructions based on the matching results, including multi-objective optimization weight allocation, thermal regulation action sequence, and execution priority ranking. Specifically, it includes:

[0181] S6.1: Based on the priority scheduling rules and feedback adjustment parameter set output by the strategy matching module, the weight factors of the multi-objective optimization are initialized and configured to construct a multi-objective optimization function for temperature control accuracy, energy efficiency and response speed.

[0182] S6.2: Apply dynamic weighting factors to each objective term in the multi-objective optimization function, and perform weighted summation based on the thermal field topological feature type (such as hot spot clustering or gradient mutation) to generate a comprehensive performance optimization objective function expression;

[0183] Based on the initialized multi-objective optimization function input, a dynamic weight allocation method is adopted (parameter: weight factor set {w i Based on the thermal field topology feature type, extract parameters from the candidate strategy template set to load initial weights for target items such as temperature control accuracy, energy efficiency, and response speed.

[0184] Furthermore, a feature-type driven weight adjustment algorithm is used (parameters: the heat distribution mode values ​​are uniform, hotspot clustering, and gradient mutation, corresponding to a preset set of weight adjustment coefficients {γ}). j This enables dynamic adjustment of the weight factor ratio of target items under different modes, ensuring priority matching of thermal control targets in specific scenarios;

[0185] Furthermore, by using a normalization operator (parameter: L1 norm normalization of the weight vector), the adjusted target weight factors are proportionally standardized to obtain the desired result. The final set of constraint weights.

[0186] Furthermore, through a weighted summation operator (parameter: the fusion coefficient depends on the historical optimization performance evaluation index), the optimization functions {f} of each objective term are combined. i Calculate the overall performance optimization objective function by superimposing the corresponding weights.

[0187] Furthermore, by using an objective function structure detection algorithm (parameters: detecting changes in gradient sign and the activity of constraint boundaries), the smoothness and differentiability of the comprehensive performance optimization objective function are confirmed, providing mathematical feasibility for subsequent gradient descent solutions.

[0188] By applying dynamic weights and performing weighted summation, the multi-objective optimization problem is transformed into a single scalar optimization objective under the adaptive condition of the current thermal field topology, thereby achieving global performance coordination and optimization path determination of the conflict resolution strategy during the execution phase.

[0189] For example, in a high-density circuit board thermal management system with dimensions of 200mm × 150mm, the initial weights (generated by S6.1) of the temperature accuracy objective function f1, energy efficiency objective function f2, and response speed objective function f3 are set to 0.4, 0.35, and 0.25, respectively. When the thermal field pattern is identified as hot spot clustering, the corresponding adjustment coefficients are 1.2, 0.9, and 0.9, respectively. The adjusted weight set is 0.48, 0.315, and 0.225. After L1 norm normalization, the final weights are 0.48 / 1.02, 0.315 / 1.02, 0.225 / 1.02 ≈ 0.4706, 0.3088, and 0.2206. Substituting this weight set into the comprehensive objective function F = 0.4706 × f1 + 0.3088 × f2 + 0.2206 × f3, the comprehensive optimization direction of the corresponding thermal control strategy is generated in real time. In a 15-minute high-power computing heat dissipation test, the solution achieved an 8°C reduction in hot spot temperature, kept energy consumption improvement within 4%, and a 12% reduction in response time, indicating that the comprehensive objective function generated by dynamic weights significantly improves the system's robustness and overall performance.

[0190] S6.3: Based on the comprehensive performance optimization objective function expression, the initial control parameters of each thermal regulation action are calculated using the gradient descent method to generate preliminary planning results for the thermal regulation action sequence;

[0191] S6.4: Prioritize the preliminary planning results of the thermal regulation action sequence, determine the execution order of each action based on the heat flow path connectivity index and the node centrality index, and generate an execution priority sequence.

[0192] S6.5: Based on the execution priority sequence and the thermal regulation action sequence, generate a complete conflict resolution instruction set containing multi-objective optimization weights, action sequences and execution order, as input instructions for the multi-level collaborative thermal regulation execution module.

[0193] Step S7: After executing the conflict resolution command, continuously collect the thermal field status feedback data of the circuit board and calculate the deviation index between the actual thermal field change and the prediction model. Specifically, this includes:

[0194] S7.1: Perform multi-dimensional real-time sampling of the thermal field state of the circuit board after executing the conflict resolution command, collect temperature, heat flux density and power density data of each heat source area and heat dissipation unit to obtain complete thermal field feedback information;

[0195] S7.2: Based on the collected thermal field feedback data, the data of each sampling point is processed for time synchronization and spatial alignment to eliminate the impact of sensor response delay and layout deviation on subsequent deviation calculation;

[0196] S7.3: Input the synchronized thermal field feedback data into the output interface of the prediction model, and generate the predicted thermal field distribution data based on the current thermal field topology map and the initial conflict resolution strategy, so as to establish a comparable relationship between the actual thermal field and the predicted thermal field.

[0197] Using time-synchronized and spatially aligned thermal field feedback data as input, a topology consistency parsing algorithm (parameters: node position matrix, heat flow direction matrix, node attribute mapping table) is employed to achieve structural mapping between the feedback data and the current thermal field topology map.

[0198] Furthermore, the mapped feedback data is transformed into a set of node state vectors {v} using a state vector generator (parameters: temperature component, heat flux density component, power density component). k} and maintain a one-to-one correspondence with the topology node index;

[0199] Furthermore, based on the initial conflict resolution strategy, the physical modeling-numerical solution hybrid prediction module (parameters: Fourier thermal conductivity matrix, convective heat transfer coefficient, initial boundary conditions) is invoked to perform iterative calculation of the heat conduction partial differential equation on the nodal state vector to obtain the predicted heat transfer path and steady-state distribution.

[0200] Furthermore, a prediction result reconstruction algorithm (parameters: topological adjacency matrix and heat flow transfer weights) is used to map the numerical solution back to the topological graph space, forming the predicted thermal field distribution matrix T. pred Matrix element T predij This represents the relationship between the temperature of node i and the heat flow between node j at the predicted time.

[0201] Furthermore, by using the data alignment operator (parameters: same timestamp, node index mapping rules), the predicted thermal field distribution matrix and the actual thermal field feedback data matrix are combined into a binary comparison set, providing a directly comparable input basis for subsequent deviation analysis;

[0202] Through the above prediction modeling and result alignment, the synchronized thermal field feedback data was successfully transformed into predicted thermal field distribution data corresponding to the actual topology, achieving a complete match between the actual and the prediction in spatial and parameter dimensions.

[0203] For example, in a high-density circuit board thermal management system with 48 nodes, the node location matrix is ​​a 48×3 spatial coordinate table, the heat flow direction matrix uses a weighted directed adjacency representation, the Fourier thermal conductivity matrix ranges from 150 to 250 W / (m·K), and the convective heat transfer coefficient is 35 W / (m·K). 2 The initial boundary condition is set at a temperature of 300K for the heat dissipation unit. The heat conduction equation is then solved using the finite difference method. The process involves iterations with a time step of 0.1 s, for 100 iterations to obtain the predicted thermal field distribution matrix T. pred T is formed by aligning the comparison operator with the actual measurement matrix. pred ,T real A control set was used. After multiple runs, the average difference between the predicted and actual temperatures was controlled within 1.5K, ensuring high accuracy and reliability for subsequent deviation analysis.

[0204] S7.4: The differential analysis method is used to compare the actual thermal field distribution with the predicted thermal field distribution point by point, and calculate the temperature deviation, heat flow direction offset and power density error of each sampling point to form a multidimensional deviation vector.

[0205] S7.5: Based on the multidimensional deviation vector, a global deviation index is generated through a weighted fusion algorithm. The global deviation index includes the mean square error of the thermal field, the hot spot offset distance, and the control response lag time, which are used to quantitatively evaluate the execution effect of the conflict resolution strategy.

[0206] Step S8: Based on the deviation index, the mapping relationship between topological features and conflict resolution strategies is updated through a lightweight online learning mechanism to achieve dynamic optimization and adaptive evolution of the strategy library. Specifically, this includes:

[0207] S8.1: Collect and time-align the feedback data of the circuit board thermal field status after executing the conflict resolution command to obtain the actual thermal field distribution data that is consistent with the output time of the prediction model.

[0208] S8.2: Based on the collected actual thermal field distribution data and the thermal field distribution data output by the prediction model, the mean square error algorithm is used to calculate the thermal field deviation index between the two to quantify the execution effect of the conflict resolution strategy.

[0209] S8.3: Input the calculated thermal field deviation index into the lightweight online learning module, and generate a strategy deviation feature vector by combining the current thermal field topology features to characterize the adaptation error of strategy matching;

[0210] The thermal field deviation index calculated by S8.2 and the current thermal field topological feature description vector extracted from S4 are used as joint inputs to the feature fusion unit of the lightweight online learning module to realize the correlation representation between error and structural features.

[0211] The feature normalization method (parameter: minimum-maximum normalization interval is [0,1]) is used to perform dimensionless processing on the temperature deviation, hot spot offset distance and control response lag time in the thermal field deviation index, so as to eliminate the influence of different physical units on subsequent fusion calculations.

[0212] Furthermore, a comprehensive deviation feature vector e is constructed using a feature-weighted concatenation method (parameters: temperature deviation weight 0.4, hotspot offset distance weight 0.35, lag time weight 0.25). bias And maintain alignment with the dimensions of the topological feature description vector;

[0213] Furthermore, a vector fusion mapping algorithm is employed (parameters: fusion method is the direct product of feature vectors and element-wise weighted summation) to combine the comprehensive deviation feature vector with the topological feature description vector f. topo Mapped to policy bias feature vector v policy The calculation formula is as follows:

[0214] v policy =W f f topo +W e e bias

[0215] Among them W f With W e These are the fusion weight matrices for topological features and deviation features, which are initialized offline in historical running data through an online learning module and continuously fine-tuned during operation.

[0216] Furthermore, a principal component preservation algorithm is adopted (parameter: retain the top few components with a cumulative contribution rate of 95%) to compress the dimension of the policy bias feature vector while retaining the components that are most explanatory to the policy fit error, so as to improve the computational efficiency of subsequent policy updates.

[0217] Through the above fusion and dimensionality reduction processing, the original deviation index and topological features are jointly transformed into a policy deviation feature vector with consistent structure, clear physical meaning and suitable for machine learning processing, so as to realize the quantitative characterization of the adaptability error of conflict resolution strategy.

[0218] For example, in a thermal management system of a circuit board with 48 nodes, the maximum calculated temperature deviation is 2.5K, and its normalized value after minimum-maximum normalization is 0.625; the hot spot offset distance is 8mm, with a normalized value of 0.53; and the control response lag time is 0.4s, with a normalized value of 0.4. The comprehensive deviation eigenvector e is obtained by weighting and summing these three values ​​in a ratio of 0.4:0.35:0.25. bias 0.53 (scalar example). Topological feature description vector ftopo It includes a mean node centrality of 0.72, a mean clustering coefficient of 0.64, and a path connectivity of 0.81. The policy bias eigenvector v is calculated using the fusion weight matrix. policy The cumulative contribution rate of the 0.615, 0.582, and 0.668 values, after being reduced to two dimensions using the principal component preservation algorithm, is 96.2%. This strategy deviation feature vector accurately reflects the matching deviation between the current thermal field topology features and the conflict resolution strategy, providing optimal input for the incremental support vector machine mapping update of S8.4.

[0219] S8.4: Based on the policy deviation feature vector, an incremental support vector machine algorithm is used to locally update the mapping relationship between topological features and conflict resolution strategies in order to optimize the decision accuracy of the policy matching module.

[0220] S8.5: The updated topology-policy mapping relationship is fed back to the conflict resolution policy library to realize the dynamic evolution of priority scheduling rules and feedback adjustment parameter sets in the policy library, so as to enhance the system's adaptive decision-making ability in unsteady thermal environments.

[0221] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0222] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A circuit board thermal management method based on multi-level collaboration, characterized in that, Includes the following steps: S1: Collect temperature, power density and heat flux density data of multiple heat source areas and heat dissipation units on the circuit board, and record the spatial location and functional attributes of each sampling point in combination with the physical layout information of the thermal management system. S2: Normalize and filter outliers on the collected thermal field data to obtain preprocessed thermal field data; S3: Based on the preprocessed thermal field data, construct a thermal field topology map of the circuit board, where the graph nodes represent heat sources or heat dissipation units, and the graph edges represent the thermal flow coupling strength and direction. S4: Extract structural features from the thermal topology map of the circuit board, including node centrality, clustering coefficient and heat flow path connectivity, to characterize the current heat distribution pattern type; S5: Input the extracted topological features into the preset conflict resolution strategy matching module, and match the corresponding priority scheduling rules and feedback adjustment parameter set in the feature similarity matching strategy library; S6: Generate preliminary conflict resolution instructions based on the matching results, including multi-objective optimization weight allocation, thermal regulation action sequence and execution priority sorting; S7: After executing the initial conflict resolution command, continuously collect the thermal field status feedback data of the circuit board and calculate the deviation index between the actual thermal field change and the prediction model. S8: Based on the deviation index, update the mapping relationship between topological features and conflict resolution strategies through a lightweight online learning mechanism.

2. The circuit board thermal management method based on multi-level collaboration according to claim 1, characterized in that, Step S1 specifically includes: Based on the physical layout design diagram of the thermal management system, the topological position coordinates of each heat source area and heat dissipation unit on the circuit board are obtained to form a spatial positioning information dataset. Multimodal sensor arrays are deployed in each heat source area and heat dissipation unit to synchronously collect temperature, power density and heat flux density, and obtain multidimensional thermal field state data streams; Zero-point drift compensation and nonlinear error correction are performed on the collected raw thermal field data through sensor calibration algorithms; Based on the functional module division information of the heat source area, the functional attributes of the collection points are labeled to generate a function-thermal mapping table. The spatial positioning information dataset, the multidimensional thermal field state data stream, and the function-thermal state mapping table are subjected to multi-source data fusion processing to generate a structured thermal field sampling dataset.

3. The circuit board thermal management method based on multi-level collaboration according to claim 2, characterized in that, In step S1, the multimodal sensor array is arranged in partitions using density clustering and greedy coverage algorithms. The neighborhood radius is 1-10 mm, the coverage radius is 1-5 mm, the fault tolerance ratio is 1%-10%, and the redundancy of each node is reduced by no less than 30%.

4. The circuit board thermal management method based on multi-level collaboration according to claim 1, characterized in that, Step S2 specifically includes: The collected temperature, power density, and heat flux density data of each region of the circuit board were normalized, and the multidimensional heterogeneous data were linearly transformed to obtain a standardized thermal field data matrix. A time-series filtering operation is performed on the standardized thermal field data matrix based on a sliding window mechanism, and an improved median filtering algorithm is used to detect and correct outliers in the temperature and power density sequences to obtain a filtered thermal field dataset. Spatial consistency verification is performed on each sampling point in the filtered thermal field dataset. Based on the gradient change threshold of heat flux density in adjacent regions, it is determined whether there is a local sensor drift phenomenon. If drift is detected, dynamic compensation is performed on the current sampling point data. A multidimensional feature vector space is constructed based on the corrected spatially consistent thermal field data. Principal component analysis is used to compress the data dimensions and extract the main thermal field change features. Dynamic window sliding mean calculation is performed on the thermal field eigenvectors after dimensionality reduction by principal component analysis to generate a thermal field state sequence.

5. The circuit board thermal management method based on multi-level collaboration according to claim 4, characterized in that, In step S2, the linear transformation is specifically performed on the multidimensional heterogeneous data based on the min-max normalization algorithm. The normalization interval for each type of physical quantity is [0,1], and linear interval mapping of data under different physical dimensions is performed.

6. The circuit board thermal management method based on multi-level collaboration according to claim 1, characterized in that, Step S3 specifically includes: The heat sources and heat dissipation units in the processed thermal field data are classified by functional attributes, and the topology node set is divided based on the spatial location information of their physical layout to establish a graph structure basic representation of the thermal behavior of the circuit board. Based on the Fourier thermal conductivity relationship between heat flux density and temperature gradient, the thermal flux coupling strength between adjacent heat sources and heat dissipation units is calculated to obtain a quantitative characterization of the heat transfer path. The heat flow transfer path is modeled directionally using a directional pattern modeling method to construct a heat flow transfer direction matrix; Based on the thermal flux coupling strength and direction information between nodes, a weighted directed graph model is constructed, where nodes represent heat sources or heat dissipation units, and edges represent thermal flux coupling paths and their directionality, forming the preliminary structure of the thermal field topology map; The constructed thermal field topology map is subjected to graph structure optimization processing, and the minimum spanning tree algorithm is used to remove redundant heat flow paths.

7. The circuit board thermal management method based on multi-level collaboration according to claim 6, characterized in that, Step S3 further includes calculating the heat flow coupling strength between nodes using the Fourier thermal conductivity relation and the center distance of the node space, and limiting the strength to [0,1] through normalization operation. At the same time, the heat flow direction is determined by combining the temperature difference, and the outflow of nodes is made consistent by using a weighted directed adjacency matrix and row normalization method, and finally forming a weighted directed graph.

8. The circuit board thermal management method based on multi-level collaboration according to claim 1, characterized in that, Step S4 specifically includes: Node centrality is calculated for the nodes in the completed thermal topology map of the circuit board. Based on the node degree centrality and proximity centrality indices, the relative importance of each heat source or heat dissipation unit in the heat flow network is quantified, and the node influence feature vector is obtained. Based on the local adjacency relationships in the graph, the clustering coefficient of each node is calculated, the local clustering characteristics of the thermal field are extracted, and a local thermal interaction intensity characteristic index is formed. A topological connectivity analysis of the heat flow path is performed, and the shortest path algorithm is used to calculate the heat flow reachability between any two nodes in the graph to obtain the global heat flow path connectivity matrix. The node centrality, the clustering coefficient, and the heat flow path connectivity matrix are fused in multiple dimensions to construct a thermal field topological feature space and form a unified topological feature description vector. Based on the preset heat distribution pattern classification rules, the topological feature description vector is subjected to pattern recognition classification processing to determine whether the current heat field belongs to a uniform, hot spot clustering or gradient abrupt heat distribution pattern, and outputs the pattern classification label.

9. The circuit board thermal management method based on multi-level collaboration according to claim 1, characterized in that, Step S5 specifically includes: The structural features of the thermal field topology map are encoded, and feature vectors are generated based on node centrality, clustering coefficient and heat flow path connectivity to construct a digital representation of the current heat distribution pattern. Based on the feature vector, a similarity measurement algorithm is executed to match and calculate the feature templates corresponding to various heat distribution patterns in the pre-stored strategy library, and to identify the degree of similarity between the current heat field topology and the existing patterns in the strategy library. Based on the similarity measurement results, a set of candidate strategy templates with a matching degree higher than a set threshold are selected as the input candidate set for the conflict resolution strategy matching module; Based on the priority scheduling rules and feedback adjustment parameter set of the candidate strategy template, a multi-objective optimization weight allocation algorithm is executed to generate a multi-level collaborative control strategy combination that is adapted to the current heat distribution pattern. The multi-level collaborative control strategy combination is output to the conflict resolution instruction generation module.

10. The circuit board thermal management method based on multi-level collaboration according to claim 9, characterized in that, Step S5 further includes fusing the feature vector weighted Euclidean distance and cosine similarity as dual indicators and introducing empirical weight coefficients. The comprehensive similarity matching threshold is 0.8-1.

0. Duplicate templates are deduplicated using hash fingerprints, and the size of the candidate strategy is controlled within 20% of the total template library.

Images