Industrial motherboard computer heat dissipation optimization method and system based on AI intelligent control

Abstract

The present invention relates to the field of intelligent temperature control technology, and in particular to a method and system for optimizing heat dissipation for industrial motherboard computers based on AI intelligent control. The method comprises: obtaining real-time motherboard temperature data, performing time series preprocessing, and generating distributed time series data; extracting thermal features and classifying and identifying them to obtain the coordinate distribution of high-energy consumption areas; inputting a classification model to output classification results, screening areas exceeding a preset threshold to determine priority control labels; performing spatial interpolation based on the labels to generate a heat map and extract thermal distribution boundary data; analyzing thermal resistance based on the boundary data and identifying bottleneck areas to obtain a thermal resistance distribution; performing fan parameter optimization based on the thermal resistance to obtain an optimized heat dissipation strategy; inputting a load model to output dynamic load adaptation control parameters; performing real-time control and energy efficiency evaluation based on the control parameters, and screening and solidifying the final heat dissipation operating parameters. The method can solve the problem of lagging heat dissipation control.

Description

Technical Field

[0001] The present invention relates to the field of intelligent temperature control technology, and in particular to a method and system for optimizing heat dissipation of industrial motherboard computers based on AI intelligent regulation. Background Art

[0002] With the continuous advancement of industrial intelligence, industrial motherboard computers are widely used in scenarios such as automated production, smart manufacturing, edge computing, and real-time control. Compared to traditional equipment, industrial motherboards must combine high computing power with long-term stability, and must be able to operate reliably in environments with high temperatures, high humidity, and high dust levels. Because they are often deployed in enclosed spaces and perform multiple tasks concurrently, internal motherboard areas such as the CPU, GPU, and power management area are prone to localized overheating, creating significant temperature gradients and increasing the risk of downtime and aging. Therefore, achieving real-time sensing and dynamic control of motherboard temperature has become a key technical issue to ensure stable system operation.

[0003] Existing industrial motherboard cooling systems often employ fixed fan speeds, linear temperature control strategies, or passive cooling structures. These methods rely on preset temperature thresholds or simplified sensor response mechanisms, which can maintain basic temperature control effectiveness under light loads or uniform heat sources. However, in real industrial environments, the tasks performed by motherboards are highly uncertain and dynamic. The heat generated by different modules (such as the CPU, GPU, memory, and power supply) fluctuates rapidly due to varying loads, resulting in extremely uneven heat distribution. When the load suddenly changes or a local module operates at high frequency for extended periods, fixed-rate cooling mechanisms cannot respond promptly to sudden temperature increases, leading to heat accumulation in specific areas.

[0004] In summary, the existing technology has the problem of delayed heat dissipation control. Summary of the Invention

[0005] The present invention provides an industrial motherboard computer heat dissipation optimization method and system based on AI intelligent regulation to solve the problem of heat dissipation control lag.

[0006] In the first aspect, in order to solve the above technical problems, the present invention provides an industrial motherboard computer heat dissipation optimization method based on AI intelligent control, comprising:

[0007] Get real-time temperature data of the motherboard;

[0008] Performing time series preprocessing on the real-time temperature data of the mainboard to obtain distribution time series data;

[0009] Perform feature extraction and classification identification based on the distribution time series data to obtain the coordinate distribution of high energy consumption areas;

[0010] Input the coordinate distribution of the high energy consumption area into a pre-trained classification model and output the classification result;

[0011] Determine the area exceeding the preset temperature threshold in the classification result as a high energy consumption area, thereby obtaining a priority regulation label for the high energy consumption area;

[0012] Based on the priority control labels of the high-energy consumption areas, heat map generation and boundary extraction are performed to obtain boundary data of heat distribution perception;

[0013] Performing thermal resistance analysis and bottleneck identification based on the boundary data of the thermal distribution perception to obtain the thermal resistance distribution of the bottleneck area;

[0014] Optimizing and controlling fan parameters according to the thermal resistance distribution in the bottleneck area to obtain an optimized heat dissipation control strategy;

[0015] Input the optimized heat dissipation control strategy into a pre-trained load model and output control parameters adapted to the dynamic load;

[0016] According to the control parameters adapted to the dynamic load, real-time control and energy efficiency judgment are performed to obtain the final heat dissipation operating parameters.

[0017] Preferably, performing time series preprocessing based on the real-time temperature data of the mainboard to obtain distributed time series data includes:

[0018] Perform smoothing and filtering processing on the real-time temperature data of the mainboard to obtain a stable temperature change sequence;

[0019] According to the stable temperature change sequence, abnormal points are identified and structured to obtain distribution time series data.

[0020] Preferably, the feature extraction and classification identification are performed based on the distribution time series data to obtain the coordinate distribution of the high energy consumption area, including:

[0021] Extracting and analyzing the temperature slope according to the distribution time series data to obtain a set of thermal characteristic values;

[0022] The thermal feature value set is used to perform threshold judgment and area identification to obtain a high energy consumption area identification result;

[0023] Based on the high energy consumption area identification result, spatial mapping and coordinate positioning are performed to obtain the coordinate distribution of the high energy consumption area.

[0024] Preferably, inputting the coordinate distribution of the high energy consumption area into a pre-trained classification model and outputting the classification result includes:

[0025] The classification model is obtained by training the MLP neural network model;

[0026] Through the input layer of the classification model, the coordinate distribution of the high energy consumption area is converted into a unified format and dimensional encoded to obtain a standardized input vector;

[0027] Performing weighted transformation and nonlinear activation on the standardized input vector through the hidden layer of the classification model to obtain an intermediate expression vector;

[0028] The intermediate expression vector is classified and judged through the output layer of the classification model to obtain a classification result.

[0029] Preferably, the step of generating a heat map and extracting boundaries based on the priority control labels of the high energy consumption areas to obtain boundary data for heat distribution perception includes:

[0030] Constructing a spatial interpolation heat map based on the priority control labels of the high energy consumption areas to obtain two-dimensional heat map data;

[0031] Performing image clustering and segmentation processing based on the two-dimensional thermal map data to obtain high-temperature area segmentation data;

[0032] Based on the high-temperature area segmentation data, edge detection and boundary extraction analysis are performed to obtain boundary data for thermal distribution perception.

[0033] Preferably, performing thermal resistance analysis and bottleneck identification based on the boundary data of the thermal distribution perception to obtain the thermal resistance distribution of the bottleneck area includes:

[0034] Performing finite element meshing processing based on the boundary data of the thermal distribution perception to obtain a set of heat conduction units;

[0035] Perform node thermal resistance calculation and analysis based on the heat conduction unit set to obtain a thermal resistance distribution data set;

[0036] Based on a cluster analysis algorithm, bottleneck area clustering identification processing is performed on the thermal resistance distribution data set to obtain the thermal resistance distribution of the bottleneck area.

[0037] Preferably, the fan parameters are optimized and regulated according to the thermal resistance distribution of the bottleneck area to obtain an optimized heat dissipation control strategy, including:

[0038] According to the thermal resistance distribution of the bottleneck area, thermal resistance threshold screening and boundary fitting processing are performed to obtain high energy consumption area positioning data;

[0039] Based on the high energy consumption area positioning data, fan status collection and parameter iterative optimization are performed to obtain optimized fan control parameter data;

[0040] According to the optimized fan control parameter data, a regional control strategy is formulated to obtain an optimized heat dissipation control strategy.

[0041] Preferably, the optimized heat dissipation control strategy is input into a pre-trained load model to output control parameters adapted to the dynamic load, including:

[0042] The load model is obtained by training the LSTM neural network model;

[0043] Performing numerical normalization and vector encoding on the optimized heat dissipation control strategy through the input layer of the load model to obtain a time-structured input tensor;

[0044] Performing recursive gating processing on the time structured input tensor through the time recursive layer of the load model to obtain a state memory representation sequence;

[0045] Performing state aggregation processing on the state memory representation sequence through the state fusion layer of the load model to obtain a compact control feature vector;

[0046] Through the output layer of the load model, the compacted control feature vector is subjected to full connection calculation and nonlinear mapping to output multiple control variables, thereby obtaining control parameters adapted to the dynamic load.

[0047] Preferably, the real-time regulation and energy efficiency judgment are performed based on the regulation parameters adapted to the dynamic load to obtain the final heat dissipation operating parameters, including:

[0048] Perform execution control and collect response feedback according to the control parameters adapted to the dynamic load to obtain execution status data;

[0049] Performing evaluation, regulation, and energy efficiency judgment based on the execution status data to obtain an energy efficiency judgment label;

[0050] According to the energy efficiency judgment label and the dynamic load adaptation control parameters, final parameter screening and solidification are performed to obtain final heat dissipation operation parameters.

[0051] In a second aspect, the present invention provides an industrial motherboard computer heat dissipation optimization system based on AI intelligent control, comprising:

[0052] Data acquisition module, used to obtain real-time temperature data of the mainboard;

[0053] A time series module is used to perform time series preprocessing based on the real-time temperature data of the mainboard to obtain distribution time series data;

[0054] A coordinate distribution module is used to extract features and classify the distribution time series data to obtain the coordinate distribution of high energy consumption areas;

[0055] A classification module, configured to input the coordinate distribution of the high energy consumption area into a pre-trained classification model and output a classification result;

[0056] a priority regulation label module, configured to determine the areas exceeding a preset temperature threshold in the classification results as high energy consumption areas, thereby obtaining priority regulation labels for the high energy consumption areas;

[0057] A heat distribution perception module is used to generate a heat map and extract boundaries based on the priority control labels of the high-energy consumption areas to obtain boundary data for heat distribution perception;

[0058] A thermal resistance distribution module is used to perform thermal resistance analysis and bottleneck identification based on the boundary data of the thermal distribution perception to obtain the thermal resistance distribution of the bottleneck area;

[0059] a heat dissipation control module, configured to optimize and control fan parameters according to the thermal resistance distribution of the bottleneck area to obtain an optimized heat dissipation control strategy;

[0060] A control parameter module is used to input the optimized heat dissipation control strategy into a pre-trained load model and output control parameters adapted to the dynamic load;

[0061] The heat dissipation operation parameter module is used to perform real-time control and energy efficiency judgment according to the control parameters adapted to the dynamic load to obtain the final heat dissipation operation parameters.

[0062] In a third aspect, the present invention also provides an electronic device comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements any one of the above-mentioned methods for optimizing the heat dissipation of industrial motherboard computers based on AI intelligent control.

[0063] In a fourth aspect, the present invention also provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned industrial motherboard computer heat dissipation optimization methods based on AI intelligent control.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] (1) By acquiring real-time temperature data from an industrial motherboard and analyzing it based on the continuous change trend of the temperature data, the present invention can identify the heat distribution characteristics of various areas of the motherboard over different time periods. Compared to traditional methods that rely solely on a single-point temperature threshold trigger, the present invention can effectively identify areas with rising high temperatures, thereby preemptively identifying thermal risk areas and improving the targetedness and response speed of thermal control strategies.

[0066] (2) This invention achieves precise coordinate positioning of high-energy consumption areas by constructing spatial coordinate mapping relationships and combining them with temperature characteristic changes in time series. This coordinate data is further input into a pre-trained MLP neural network classification model. Combined with the neural network's ability to recognize nonlinear boundaries, it can efficiently complete thermal risk zoning and classification judgment, significantly improving the structured capability and robustness of thermal control strategy generation.

[0067] (3) This method constructs thermal distribution boundary data to further extract regional thermal resistance characteristics and identifies bottleneck areas through a clustering algorithm. This method can locate key areas of heat accumulation based on the thermal resistance distribution dataset, thereby providing guidance for fan parameter optimization and local control, effectively improving heat dissipation efficiency and extending device life.

[0068] (4) This invention processes the optimized heat dissipation control strategy by introducing an LSTM neural network load model, and outputs control parameters that can adapt to different load states. Combining the time recursive structure with the state aggregation mechanism, the system can dynamically match the energy consumption requirements of the motherboard during operation, generate stable and reliable control outputs, and improve the adaptability of energy efficiency control and the robustness of system operation.

[0069] In summary, this paper builds an AI-powered intelligent heat dissipation control method for multi-zone thermal risk identification and dynamic response for industrial motherboards through key steps including temperature sequence filtering preprocessing, high-energy consumption area coordinate identification, neural network classification and judgment, thermal resistance clustering bottleneck extraction, and load adaptive parameter reasoning. This method can address the issue of lagging heat dissipation control. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 This is a flow chart of an industrial motherboard computer heat dissipation optimization method based on AI intelligent control provided by the first embodiment of the present invention;

[0071] Figure 2 This is a structural diagram of an industrial motherboard computer heat dissipation optimization system based on AI intelligent control provided by the second embodiment of the present invention. DETAILED DESCRIPTION

[0072] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0073] Reference Figure 1 The first embodiment of the present invention provides an industrial motherboard computer heat dissipation optimization method based on AI intelligent control, comprising the following steps:

[0074] S11, obtain the real-time temperature data of the mainboard;

[0075] S12, performing time series preprocessing based on the real-time temperature data of the mainboard to obtain distribution time series data;

[0076] S13, performing feature extraction and classification recognition based on the distribution time series data to obtain the coordinate distribution of high energy consumption areas;

[0077] S14, inputting the coordinate distribution of the high energy consumption area into a pre-trained classification model and outputting the classification result;

[0078] S15, determining the area exceeding the preset temperature threshold in the classification result as a high energy consumption area, thereby obtaining a priority regulation label for the high energy consumption area;

[0079] S16, generating a heat map and extracting boundaries based on the priority control labels of the high-energy consumption areas to obtain boundary data for thermal distribution perception;

[0080] S17, performing thermal resistance analysis and bottleneck identification based on the boundary data of the thermal distribution perception to obtain a thermal resistance distribution of the bottleneck area;

[0081] S18, optimizing and controlling fan parameters based on the thermal resistance distribution in the bottleneck area to obtain an optimized heat dissipation control strategy;

[0082] S19, inputting the optimized heat dissipation control strategy into a pre-trained load model, and outputting control parameters adapted to the dynamic load;

[0083] S20 , performing real-time control and energy efficiency judgment according to the control parameters adapted to the dynamic load to obtain final heat dissipation operating parameters.

[0084] In step S11, real-time temperature data of the mainboard is obtained;

[0085] It is worth noting that in step S11, the process of obtaining the real-time temperature data of the motherboard is the basic operation for constructing the thermal management control process of the present invention. This process realizes continuous collection of the thermal status of each local position of the motherboard by deploying multiple high-precision temperature sensors in key areas of the motherboard, thereby establishing a real-time, structured temperature data system. The temperature change of the motherboard is dynamic and localized during operation, especially in areas such as the processor, graphics card, and power module, which are prone to heat accumulation. Therefore, data collection must be completed with high sampling frequency and spatial coverage accuracy.

[0086] During implementation, at least six digital temperature acquisition units were deployed across the motherboard, including those located in the CPU core area, GPU core area, power management unit, storage interface area, south bridge chip area, and the center of the motherboard bottom. Each acquisition unit records data at 0.5-second intervals and transmits the sampling results to a central control node via a communication interface. The transmission channel utilizes I2C or SMBus protocols, ensuring high timeliness and interference resistance.

[0087] The collected data includes each sensor node's timestamp, spatial coordinate number, and temperature value. During data structuring, the system binds the temperature value of each sampling point to its spatial coordinates, forming a triplet structure in the format of "node number - acquisition time - temperature value." For example, the CPU zone node data at a certain moment is "Node_1, 12:00:02, 65.2°C," indicating that the temperature at sampling point 1 at the specified time was 65.2 degrees Celsius. The system continuously collects data to form a time series.

[0088] In step S12, time series preprocessing is performed based on the real-time temperature data of the mainboard to obtain distributed time series data, including:

[0089] Perform smoothing and filtering processing on the real-time temperature data of the mainboard to obtain a stable temperature change sequence;

[0090] According to the stable temperature change sequence, abnormal points are identified and structured to obtain distribution time series data.

[0091] It's worth noting that when managing motherboard thermal management, the raw, real-time temperature data must first be smoothed and filtered to ensure the accuracy of subsequent temperature trend analysis, anomaly detection, and control optimization. The core goal of this process is to eliminate transient high-frequency noise caused by electromagnetic interference, data transmission jitter, or sudden operational fluctuations. This improves data stability and analyzability without disrupting the original temperature trend, thereby generating a stable temperature change series and providing a solid foundation for the subsequent construction of distributed time series data.

[0092] In actual operation, multiple sensors deployed on the motherboard collect the operating temperature of each functional module at a fixed sampling interval (e.g., once per second). Raw data is continuously generated over time and represented as a two-dimensional time series, where each row represents a timestamp and corresponds to temperature readings at multiple measurement points. Due to the complex sampling environment, abnormal fluctuations often occur, such as a sudden rise in GPU temperature from 46.8°C to 51.0°C and then a rapid drop back to 47.2°C. If unprocessed, these jumps can directly interfere with thermal trend analysis and even mislead thermal risk identification. Therefore, temperature smoothing is required on the raw data before constructing the time series.

[0093] This step uses the weighted sliding average method as the core processing algorithm. Its principle is to take each sampling point as the center, take multiple sampling points within a certain range before and after to form a sliding window, assign different weights to the temperature values ​​in the window, and calculate the weighted average to replace the original temperature value of the center point. Set the window length to 5, and the sampling values ​​in the window are to , the corresponding weights are 0.1, 0.15, 0.5, 0.15, 0.1, then the center point smooth temperature The calculation method is: This calculation is performed sequentially over the data sequence, generating a new smoothed value each time, ultimately forming a smoothed temperature sequence corresponding to the original timestamp. The weight setting can be adjusted based on the system's thermal inertia to meet the response requirements under different load conditions.

[0094] For example, in a system test, the temperature data of the CPU area at five consecutive time points were: 47.1°C, 47.6°C, 51.0°C, 48.2°C, and 47.8°C.

[0095] After the entire sliding process is complete, the temperature data from all measurement points is organized into a stable time series in chronological order and stored and managed uniformly. The system then categorizes the temperature series in different spatial regions using the measurement point number as a label and organizes them into a data structure based on the triplet of "sampling time, measurement point location, and smoothed temperature value" as the basic unit.

[0096] After smoothing and filtering the temperature series, the resulting stable temperature change series cannot be directly used for subsequent spatial thermal feature analysis. Therefore, outlier identification and structured organization must be performed to generate distributed time series data with engineering semantics and statistical continuity. The main goal of this step is to clean and reconstruct the temperature series to ensure temporal continuity, spatial correspondence, and data consistency, thereby providing a high-quality input data foundation for subsequent feature extraction and control strategies.

[0097] In the stage of outlier identification, in order to improve the stability of temperature data and the effectiveness of analysis, the numerical deviation judgment algorithm in the sliding window is used to perform point-by-point detection on the stable temperature change sequence. This method calculates the difference between the statistical characteristics of the temperature value at the current time point and several sampling points before and after it to determine whether it deviates from the normal variation range, and then identifies and eliminates interfering outliers. The specific operation process is as follows: Take the current time point as the starting point As the center, 5 time points before and after are selected to form a sliding window containing 11 sampling values, and the mean and standard deviation of the temperature value sequence in the window are calculated. The sum of the temperatures at the 11 points is divided by the number of samples, 11, and reflects the average temperature level during the period. The standard deviation Then calculate the square of the deviation of each temperature from the mean, take the mean and then square root it to express the temperature fluctuation range of the section. Execute deviation judgment, if satisfied The condition that a point deviates significantly from its neighboring samples constitutes an outlier. This threshold is based on normal distribution statistics: the probability of a sample point outside three standard deviations is less than 0.3%, providing a clear standard for identifying anomalies. For example, in a graphics card area, if the window mean is 53.99°C, the standard deviation is 2.33°C, and the current temperature is 61.5°C, the deviation is 7.51°C, exceeding the 6.99°C threshold. The system identifies this data as an outlier and deletes or interpolates it to prevent misleading subsequent trend analysis and parameter generation. After identification, the remaining data sequence is reorganized into a structured format, constructing a ternary record table based on the measurement point name, sampling timestamp, and temperature value fields to ensure temporal continuity and data consistency. This process provides a highly reliable input source free of anomaly interference for subsequent temperature spatial mapping and thermal feature extraction. This process can be efficiently implemented in practical embedded systems using a multi-threaded sliding window algorithm, ultimately outputting distributed time series data with engineering usability.

[0098] After removing abnormal data, the remaining valid temperature samples need to be structured. This process not only requires the data to have a consistent time format, but also ensures that each record can correctly identify its spatial source. The system uses a hash map-based organization method to achieve rapid structural reconstruction: first, all valid records are bucketed using the "sensor number" as the hash key; then, the data entries in each bucket are sorted in ascending order according to the "timestamp" field to form a strictly ordered time series. The specific operations of the hash map include: setting the hash function H, inputting the sensor number (such as GPU, CPU, DIMM1, etc.) into the function, and mapping it to a unique bucket address, which stores all temperature records from the sensor. With the help of the hash table structure, The rapid positioning of sensor numbers to data sets is completed within the time complexity, significantly improving processing efficiency.

[0099] After completing the mapping and sorting, the system outputs the final time series structure in the form of triples. Each data item contains a clear measurement point number, acquisition time, and temperature value. Taking the GPU area of ​​a motherboard as an example, the preprocessed sequence includes the following data:<GPU, 2025-04-30 09:12:01, 52.3°C> ,<GPU, 2025-04-30 09:12:02, 52.6°C> …<GPU, 2025-04-30 09:13:00, 53.0°C>

[0100] All structured results are written uniformly into a distributed time series dataset, stored by region ID. This sequence not only provides a clear and stable data source for subsequent temperature slope calculations, classification identification, and spatial interpolation, but also serves as historical feedback for thermal trend visualization or fan control.

[0101] In step S13, feature extraction and classification recognition are performed based on the distribution time series data to obtain the coordinate distribution of high energy consumption areas, including:

[0102] Extracting and analyzing the temperature slope according to the distribution time series data to obtain a set of thermal characteristic values;

[0103] The thermal feature value set is used to perform threshold judgment and area identification to obtain a high energy consumption area identification result;

[0104] Based on the high energy consumption area identification result, spatial mapping and coordinate positioning are performed to obtain the coordinate distribution of the high energy consumption area.

[0105] It's worth noting that after smoothing the temperature data and removing anomalies, the system further extracts and analyzes the resulting distributed time series data to quantify the temperature rise rate at each measurement point over time. This process aims to establish a numerical evaluation index for thermal response intensity, providing a clear basis for identifying high-energy consumption areas.

[0106] First, at the start of operation, the system constructs a standard time series temperature data set for each sensor point. This data structure is represented as T = {T1, T2, ..., T}, and the corresponding sampling time series is t = {t1, t2, ..., t}. To describe the intensity of temperature changes, the system processes the temperature difference between consecutive time points and uses the first-order difference method to extract the temperature slope. The calculation formula is:

[0107]

[0108] in is the temperature change rate during the kth period, in °C / s. Since the sampling time interval is uniformly set to 1 second, the formula is simplified to the difference between the current temperature and the previous temperature, that is, . The whole sequence is calculated from the second point to , forming the temperature change rate sequence of the measuring point , as the first layer of thermal feature data for subsequent analysis.

[0109] To enhance data robustness and mitigate local slope jitter caused by power supply disturbances and heat dissipation fluctuations, the system performs a sliding average after the slope sequence calculation is completed. The sliding window length is set to 5 sampling points, and at each time position, the two slope values ​​before and after are taken together with the value at the current position to calculate the average value. The calculation process is as follows: This operation can balance single-point strong fluctuations, effectively retain the real warming trend, and suppress the impact of isolated abnormal growth segments.

[0110] After completing the sliding average processing, the system will obtain a smoothed slope sequence , used to more stably depict the temperature change trend over time. On this basis, the system further compresses and filters the numerical features for unified storage and call.

[0111] Next, the system uses the smooth slope sequence of each measuring point as the thermal response feature of the measuring point, summarizes it by the measuring point dimension, and forms a multi-channel thermal response data structure. The thermal feature set is ,in For the sensors, such as graphics card sensors, motherboard sensors, etc. This is the smoothed slope sequence of the measurement point within the current time period. For example, a motherboard may be deployed with multiple sensor nodes, such as the CPU, GPU, memory, and power supply module. The system generates an independent thermal rate of change vector for each node and combines them into a unified set to support subsequent regional clustering, threshold judgment, and spatial projection analysis.

[0112] After extracting and smoothing the temperature slope, the system generates a set of thermal signature values ​​with both temporal and spatial location information. To identify high-energy consumption areas with significant thermal responses and a sustained warming trend, threshold determination and region recognition are performed on this set. The core goal of this process is to identify the areas with the most dramatic thermal changes during the observation period by setting a clear temperature rise rate threshold, providing input for subsequent spatial coordinate positioning and control parameter generation.

[0113] First, the system extracts statistical indicators such as the maximum value, mean value, and fluctuation intensity of the thermal characteristic value of each measuring point, that is, the temperature change slope sequence. On this basis, a fixed temperature rise rate judgment threshold is introduced. The setting of this threshold is based on the design limit of the motherboard heat sink and the thermal capacity response of key components. After multiple rounds of experimental measurements, it was found that when the temperature rise rate exceeds 1.0°C / s, the temperature of local areas of most motherboards will continue to exceed the safe operating upper limit in a short period of time, which can easily lead to performance fluctuations or thermal failure. Therefore, the system sets the critical slope threshold for identifying high-energy consumption areas to 1.0°C / s. That is, if a certain measuring point has a slope data point exceeding this value for multiple consecutive time periods, it is determined that the measuring point has abnormal heat accumulation.

[0114] During the calculation process, the system first calculates the thermal characteristic slope sequence of each measuring point. A filter is performed to determine whether there are m consecutive time points (m ≥ 3) with slope values ​​greater than or equal to 1.0°C / s. If this condition is met, the measurement point is marked as a "high energy consumption point." For example, if the slope sequence corresponding to the inductor region of a graphics card is: 1.2, 1.3, 1.4, 0.9, 0.8, 1.1°C / s, then if the first three time periods meet the threshold condition, the region is marked as a high energy consumption point.

[0115] After completing single-point identification, the system performs spatial region aggregation. Specifically, the system maps the numbers of all measurement points marked as high-energy consumption points to a physical coordinate space and uses an adjacency list structure to determine whether there are high-heat blocks in two-dimensional space adjacent to these measurement points. If the number of high-energy consumption points within a contiguous area exceeds a set threshold (e.g., five points), the area is identified as a high-energy consumption area, resulting in a structured identification result.

[0116] For example, in the motherboard power MOS area, if a continuous area containing seven measurement points is identified, and its slope is greater than 1.1°C / s and lasts for more than 12 seconds, the area will be identified as a high-energy consumption block, and its boundary range X: 55–75 mm, Y: 30–45 mm will be recorded as the input basis for the next step of coordinate mapping and heat map construction.

[0117] After numerically identifying high-energy consumption areas, the system must further map the identification results to actual physical space, performing coordinate positioning operations to ultimately determine the coordinate distribution of high-energy consumption areas. The core task of this step is to combine thermal signatures with the motherboard's two-dimensional spatial structure to achieve visual representation and spatial location of hot areas, providing foundational support for subsequent thermal map construction, boundary extraction, and heat dissipation path analysis.

[0118] First, the system matches the number of each measurement point in the high energy consumption identification result with its hardware installation location, and extracts its physical coordinate information on the motherboard structure. Each measurement point corresponds to a unique layout number in the motherboard design and is given a clear 、 Coordinate values. For example, the measured point < The corresponding position may be : 56.3 mm, :38.7 mm. The system will heat the feature point < The identification label and its spatial coordinates are combined to form a triplet< , 56.3, 38.7>, and all the measurement point information that meets the high energy consumption conditions is stored in the coordinate mapping table.

[0119] To improve the regional continuity and spatial representation accuracy of the recognition results, the system further performs spatial aggregation calculations on high-energy consumption point sets. This process includes the following steps: First, all high-energy consumption points are projected onto a two-dimensional coordinate grid to construct a hotspot coordinate set. Second, a Euclidean distance-based adjacency analysis is performed on these coordinate sets to determine whether there are point clusters with a distance less than a preset threshold (e.g., 5 mm). If so, these point clusters are grouped into the same region, and a bounding box is constructed around them. The minimum and maximum coordinate values ​​of the region are extracted as the boundary of the high-energy consumption area.

[0120] For example, in the graphics card area of ​​the motherboard, if six high energy consumption measurement points are identified, their coordinates are as follows: , 56.3,38.7>,< 57.0, 39.2>,< , 55.8, 38.9>,< , 56.9, 37.8>,< , 57.1, 38.5>,< , 55.6, 39.0>, the system classifies these six measurement points into a continuous area and constructs a bounding box using a rectangular envelope method. The coordinate distribution range is: = 55.6 mm, = 57.1 mm, = 37.8 mm, = 39.2 mm. The final output coordinate distribution of the high energy consumption area is: area number R1, covering the range [55.6–57.1 mm] × [37.8–39.2 mm].

[0121] During the entire operation, the system executes the above mapping process for all identified high-energy consumption areas to ensure that each thermal anomaly area has a unique, continuous and computable coordinate expression in the motherboard space, and finally generates a set of coordinate distributions of high-energy consumption areas.

[0122] In step S14, the coordinate distribution of the high energy consumption area is input into a pre-trained classification model, and the classification results are output, including:

[0123] The classification model is obtained by training the MLP neural network model;

[0124] Through the input layer of the classification model, the coordinate distribution of the high energy consumption area is converted into a unified format and dimensional encoded to obtain a standardized input vector;

[0125] Performing weighted transformation and nonlinear activation on the standardized input vector through the hidden layer of the classification model to obtain an intermediate expression vector;

[0126] The intermediate expression vector is classified and judged through the output layer of the classification model to obtain a classification result.

[0127] It's worth noting that in practical applications, in order for the system to automatically identify the categories or cooling priorities of different high-energy-consuming areas, it's necessary to build an artificial neural network with classification capabilities and train it in advance using historical data. The neural network mentioned in this step is a multi-layer perception structure, which transforms and maps the input feature vectors step by step through connections between the input layer, hidden layer, and output layer, ultimately outputting the classification results. During the training phase, the goal of the neural network is to minimize the error between the predicted results and the actual category labels, thereby enabling it to correctly classify unknown data.

[0128] In the training data preparation phase, we first need to build a structured training sample set. Each piece of training data consists of two parts: the input feature vector and the corresponding category label. The input feature vector comes from the coordinate distribution information of the high energy consumption area, including the coordinates of the center point of the area. , minimum boundary value , maximum boundary value A set of unified numerical values ​​across dimensions such as area size, shape ratio, and thermal feature mean and peak values. Category labels are manually annotated classification levels, such as high-risk, medium-risk, or low-risk areas, or more fine-grained device type labels such as CPU area, graphics card area, and power supply area. The total number of samples must exceed several thousand and cover diverse scenarios and structures to improve generalization capabilities.

[0129] Next, we enter the network structure definition and initialization phase. Taking a typical three-layer structure as an example, the number of input layer nodes matches the dimensionality of the feature vector, assuming 8 dimensions. The hidden layer consists of two fully connected neural layers, each with 64 and 32 neurons, respectively, using ReLU as the activation function. The number of output layer nodes equals the number of classification categories. If the classification is three-category, three output nodes are used, and the Softmax function is used for probability normalization. Initial values ​​for network weights and bias parameters are randomly generated using a uniform or normal distribution to ensure convergence during the initial training phase.

[0130] During the training process, the system uses a supervised learning strategy to iteratively optimize weight parameters through a gradient descent algorithm. The specific steps are as follows: First, a forward propagation calculation is performed on each set of training samples, and the input feature vector is sequentially input into each layer of the neural network to obtain the current prediction result; then, the prediction result is compared with the true label and the loss function value is calculated. The commonly used loss function is cross entropy loss, which is calculated as follows:

[0131]

[0132] in, is the number of classification categories, is the one-hot encoded value of the actual label, is the predicted probability output by the neural network. After calculating the loss, the system performs gradient calculations on the weights and biases of each layer in the network according to the backpropagation algorithm. The update direction is automatically expanded by the chain rule. The update formula is:

[0133]

[0134] in is the learning rate, which ranges from 0.001 to 0.01 and reflects the step size of each iterative parameter update. is the updated weight parameter, is the weight parameter before updating, is the loss function Weight The partial derivative of .

[0135] The entire training process is performed in batch mode. In each iteration, a batch of samples is selected and fed into the network for forward and backward propagation. A complete training cycle is called an epoch, with a range of 50 to 200 epochs. The system performs an accuracy evaluation on the validation set after each epoch to monitor overfitting. If the validation accuracy decreases for three consecutive rounds, training is terminated early.

[0136] After training is complete, the network weights are solidified into a set of stable parameters for subsequent classification and identification of high-energy-consuming areas. In actual deployment, simply normalize the new coordinate distribution data and input it into the network. After step-by-step calculations through the input layer, hidden layer, and output layer, the classification result for the area is directly obtained. This result will serve as the basis for subsequent regulatory priority decisions.

[0137] For example, if a set of high-energy consumption area coordinate data is the center point (54,72)(54, 72)(54,72), the boundary points are (50,70)(50, 70)(50,70) to (58,75)(58, 75)(58,75), the thermal characteristic mean is 1.6°C / s, and the area is 120 square millimeters, the network outputs three categories of probabilities after input: [0.82,0.12,0.06][0.82, 0.12, 0.06][0.82,0.12,0.06], the system classifies it as a high-heat risk area and sets its control level to the highest priority.

[0138] After normalizing the coordinate distribution of high-energy-consumption areas, the system inputs this standardized vector into the first layer of the structure used for classification and judgment, the input layer. The primary function of the input layer is to uniformly arrange and encode all input dimensions, ensuring that the positional relationship of each coordinate value in the spatial coordinate system can be accurately expressed as a vector. The system scales each dimension of each data point to fall within the standard range of zero to one. This unified scale ensures numerical comparability and stability in subsequent calculations.

[0139] Next, we enter the hidden layer. The function of the hidden layer is to redistribute the importance of information in different dimensions by assigning specific weight coefficients to each dimension of the input vector. For example, in a six-dimensional coordinate distribution, the system may believe that changes in the center point of the region are more critical to the classification result. Therefore, it will set higher weight coefficients for the first and second dimensions (such as the center position of the horizontal and vertical coordinates), such as 0.8 or 0.9, while the dimensions representing the boundary range of the region (such as the boundary length or boundary span) are set to medium or low weight values, such as 0.3 or 0.2. These weights are not set randomly, but are automatically optimized through the system pre-training process. Their role is to enhance the information that plays a key role in judging high-energy consumption areas, while reducing the interference of secondary dimensions on the results.

[0140] In practice, the system multiplies the value of each dimension in the input vector by its corresponding weight, then combines this with an internal bias term to produce an intermediate composite score. The weighted results of all dimensions are aggregated into a new set of feature data, essentially mapping the input regions to pre-defined classification criteria. This weighted integration allows the system to extract core features with discernible value from complex spatial structures.

[0141] To enhance the ability to identify nonlinear classification boundaries, the system introduces a nonlinear processing step at the output stage of the hidden layer, using a ReLU activation mechanism. The ReLU logic is simple: only values ​​greater than zero are retained, and if a weighted result is negative or invalid, it is set to zero. This processing has the advantage of filtering out spurious signals and retaining only those numerical features that contribute positively to the current judgment task, thereby improving classification accuracy and stability.

[0142] All activated data forms an intermediate expression vector, a compressed representation of the current input region in a multidimensional thermal response space. It no longer contains raw coordinate data, but rather the extracted response intensities of key features. For example, a region with a high response value due to a dramatic change in its center coordinates indicates a significant abrupt change in its temperature distribution, consistent with the typical behavior of high-energy-consuming areas.

[0143] Ultimately, these intermediate features are passed to the output layer, where the system classifies them based on the classification boundaries learned during training. The output layer is designed as a binary classification system. Based on the distance between the intermediate expression vector and the internal reference value, the system outputs a continuous value close to zero or one, indicating the confidence level in classifying the area as high-energy or non-high-energy. For example, an output of 0.91 indicates that the system believes there is a 91% probability that the current area is in a high-energy consumption state.

[0144] In step S15, the areas exceeding a preset threshold in the classification results are determined as high energy consumption areas, thereby obtaining priority regulation labels for the high energy consumption areas;

[0145] It is worth noting that in step S15, the system performs classification output based on the multi-layer perceptron neural network used in step S14, and further performs the operation of determining the high energy consumption area and generating the priority control label. Specifically, when the coordinate distribution of the high energy consumption area is input in step S14, the system simultaneously saves the corresponding original spatial attribute information, including the center point coordinates ( , ), boundary start and end coordinates ( ), area, thermal characteristic slope mean, and other numerical attributes. After calculation by the classification model of S14, the system obtains the classification confidence value corresponding to each input area, ranging from zero to one, which is used to indicate the possibility of the area being judged as a high energy consumption state.

[0146] In this step, the system performs a threshold determination for each classification confidence value. The preset threshold is 0.65, which balances recognition sensitivity and false positives. When a classification confidence value is greater than or equal to 0.65, the system identifies the corresponding area as a high-energy consumption area. At this point, the system extracts the area's spatial coordinate information, area data, and temperature variation characteristics from previously recorded data, combining these elements into a set of labeled data in a fixed format. The specific combination process is as follows: first, the center point coordinates are used as the primary positioning field. Next, the boundary range description is added to clarify spatial scalability. The corresponding area value is then added to represent the heat load scale. Finally, the thermal characteristic slope value is added to reflect the local temperature rise rate. The system then combines these fields into a standard record entry with a unified structure, such as [center point (54,72), boundary (50,70)-(58,75), area 120 mm², average slope 1.6°C / s], and assigns it the priority regulation label for high-energy consumption areas.

[0147] Through the above processing, each area identified as high-energy consumption has complete physical location, spatial boundaries, area magnitude and thermal activity characteristics. All information is generated by directly linking the previous input data and classification reasoning, ensuring the continuity of the data source and the engineering availability of the structure.

[0148] In step S16, heat map generation and boundary extraction are performed based on the priority control labels of the high energy consumption areas to obtain boundary data of heat distribution perception, including:

[0149] Constructing a spatial interpolation heat map based on the priority control labels of the high energy consumption areas to obtain two-dimensional heat map data;

[0150] Performing image clustering and segmentation processing based on the two-dimensional thermal map data to obtain high-temperature area segmentation data;

[0151] Based on the high-temperature area segmentation data, edge detection and boundary extraction analysis are performed to obtain boundary data for thermal distribution perception.

[0152] It is worth noting that after completing the extraction of the high energy consumption area priority control labels, in order to present the heat distribution status of the motherboard surface, the system performs spatial interpolation heat map construction based on the priority control labels, and finally generates a complete two-dimensional heat map data. The input object is a set of structured high energy consumption area parameters, including the center point coordinates , ( ), area, and thermal characteristic mean. Each high-energy consumption area is considered a spatial sampling point and assigned a specific thermal intensity value, which serves as the basis for subsequent interpolation calculations.

[0153] In the specific interpolation process, the system uses Kriging interpolation to deduce the heat distribution of each point in the target space. The core of Kriging interpolation is to construct a spatial autocovariance matrix to describe the degree of spatial correlation between different sampling points. First, the system calculates the Euclidean distance between any two sampling points based on the spatial positions of all sampling points. and The distance calculation formula is:

[0154]

[0155] Then, based on the distance , To indicate the The horizontal coordinate position of the sampling point in the high energy consumption area, To indicate the The horizontal coordinate position of the sampling point in the high energy consumption area, To indicate the The vertical coordinate position of the sampling point in the high energy consumption area, To indicate the The vertical coordinate position of the sampling point in the high energy consumption area.

[0156] Define the covariance function , to reflect the relationship between distance and spatial correlation. The covariance function uses an exponential function, defined as:

[0157]

[0158] in, is the total variance, which is set to 1.0 for normalization. It is a related dimensional parameter, which is determined according to the size of the motherboard. For example, if the short side of the motherboard is 150 mm, then 30 mm is used.

[0159] Through the above method, the system calculates the specific covariance value for each pair of sampling points, and all covariance combinations form a complete spatial autocovariance matrix , the matrix dimension is , is the number of sampling points.

[0160] On this basis, for each grid point to be interpolated, the system calculates the covariance column vector from this point to all sampling points , and then determine the interpolation weight vector by solving the following linear equations :

[0161]

[0162] Solved It is the weighted coefficient of the thermal characteristic value of each sampling point at the grid point. The calculation formula is:

[0163]

[0164] in, For sampling points The mean thermal characteristic value of is the corresponding weight, is the total number of sampling points involved in the interpolation calculation. For example, if the covariances between an interpolation grid point and three sampling points are 0.75, 0.42, and 0.12, respectively, and the weights obtained after solving the linear equations are 0.65, 0.30, and 0.05, respectively, then the estimated calorific value of the point is: , the entire motherboard space is divided into a uniform and detailed grid, and the heat estimation of all grid points is completed through the above method, thereby constructing a complete, continuous and physical distribution-compliant two-dimensional thermal map data.

[0165] After constructing the two-dimensional heat map data based on spatial interpolation, the system performs image clustering and segmentation on the data to further identify high-temperature areas with dense heat distribution. The core goal of this operation is to divide the entire heat map into several internally consistent, well-defined sub-regions based on the temperature distribution characteristics of each region, thereby accurately extracting potential areas of heat accumulation.

[0166] In its implementation, the system first treats the thermal map data matrix as a grayscale image, where the grayscale value of each pixel directly corresponds to the temperature at that location. The system then uses a K-means clustering algorithm to cluster and classify all pixels in the image. The basic calculation process of K-means clustering includes the following steps:

[0167] First, we set the number of cluster categories, K. Based on the system's analysis of the motherboard's thermal distribution characteristics, we set K between 3 and 5 to distinguish between different temperature ranges: low, medium, and high. In this step, we chose K=4, corresponding to the background, normal, medium, and high temperature zones.

[0168] The system then uses the grayscale value of each pixel as input and randomly initializes four center values ​​as the initial cluster centers. The system then calculates the Euclidean distance between each pixel's grayscale value and all the center values, selects the center with the closest distance, and assigns the pixel to the corresponding category.

[0169] After completing the initial assignment of all pixels, the system recalculates the average grayscale value of each pixel type and uses it as the new center value. The assignment process repeats based on the new center value. This process continues iteratively, stopping when the center value changes by less than a preset convergence threshold (e.g., less than 0.5°C) or when the number of iterations reaches a set upper limit (e.g., 50).

[0170] Through the above calculations, the two-dimensional heat map is divided into four sub-regions. While the internal temperature changes within each sub-region are relatively consistent, there are significant temperature gradient differences between different sub-regions. Based on the final clustering results, the system extracts the regions belonging to the high-temperature category as the high-temperature region segmentation data. The judgment criterion is that the center temperature of the category is at least 20% higher than the average temperature of the overall heat map.

[0171] For example, in an actual motherboard test, the interpolated heat map for a specific area ranges from 0 to 80°C. K-means clustering yields four clusters with center values ​​of 15°C, 35°C, 55°C, and 72°C. The system identifies the clusters corresponding to 55°C and 72°C as high-temperature regions and extracts all corresponding pixels, generating high-temperature region segmentation data.

[0172] After clustering and extracting the high-temperature region segmentation data, the system further performs edge detection and boundary extraction analysis to accurately characterize the geometric form of the heat distribution spatially. The goal of this step is to extract continuous and clear contours from the identified high-temperature regions, forming thermally sensitive boundary data and providing a basic spatial framework for subsequent heat conduction path modeling and optimization.

[0173] In the specific processing process, the system first inputs the high-temperature area segmentation data in the form of a binary image, that is, assigning a value of 1 to pixels belonging to the high-temperature category and a value of 0 to pixels belonging to the non-high-temperature category. Based on this input, the system uses the Sobel operator to perform edge detection operations. The Sobel operator performs convolution calculations on the image along the horizontal (X-axis) and vertical (Y-axis) directions respectively, extracts the position where the grayscale change is most significant, and then locates the edge. Specifically, the system constructs a small 3×3 window around each pixel, applies the convolution kernel in the horizontal and vertical directions respectively, and calculates and Two gradient components. Then, the system follows the formula:

[0174]

[0175] Calculate the total gradient magnitude of each pixel, where and Represents the grayscale change rate in the horizontal and vertical directions, respectively. The system determines whether to mark the current pixel as an edge point based on the gradient magnitude. To avoid misjudgment due to weak noise, the system sets a fixed gradient threshold, such as 20. Only when the total gradient magnitude exceeds this threshold is the pixel considered a valid edge.

[0176] After completing preliminary edge detection, the system must perform boundary extraction analysis. To this end, the system uses a connectivity analysis-based approach to traverse the set of pixels marked as edge points. During this traversal, the system gradually connects adjacent edge points according to the 8-neighborhood rule to form a continuous boundary curve. Furthermore, to eliminate false boundaries and isolated noise points, the system sets a minimum closed length threshold, for example, 50 pixels. If the total length of a closed curve is less than this threshold, it is considered a noise structure and is removed.

[0177] Taking actual motherboard thermal distribution data as an example, in a heat map of a graphics card area, segmentation and extraction revealed several high-temperature regions. The largest hot zone had a boundary contour of approximately 120 pixels long and a regular, continuous shape. Through the aforementioned edge detection and boundary extraction process, the system successfully obtained the precise boundary point set for this high-temperature region, including a complete sequence of spatial coordinates, such as {(50,70),(51,70),(52,71),…,(58,75)}.

[0178] Ultimately, the system encodes each extracted high-temperature region boundary into a separate data structure, each containing a continuous sequence of coordinate points and the parameters of the bounding rectangle. All of this boundary data forms the thermal distribution perception boundary data relied upon for subsequent heat conduction modeling and optimization analysis, providing a rigorous spatial basis for the system's subsequent path planning and thermal control strategies.

[0179] In step S17, thermal resistance analysis and bottleneck identification are performed based on the boundary data of the thermal distribution perception to obtain the thermal resistance distribution of the bottleneck area, including:

[0180] Performing finite element meshing processing based on the boundary data of the thermal distribution perception to obtain a set of heat conduction units;

[0181] Perform node thermal resistance calculation and analysis based on the heat conduction unit set to obtain a thermal resistance distribution data set;

[0182] Based on a cluster analysis algorithm, bottleneck area clustering identification processing is performed on the thermal resistance distribution data set to obtain the thermal resistance distribution of the bottleneck area.

[0183] It is worth noting that after extracting the boundary data for thermal distribution perception, the system must further perform finite element meshing on the boundary range to establish a spatial discretization basis for subsequent thermal resistance analysis. Specifically, the core goal of finite element meshing is to subdivide the continuous and complex thermal distribution area into a number of structured local cells. Each cell can be treated as a locally uniform body in terms of geometry and physical properties, thereby facilitating subsequent numerical calculations and thermal resistance solutions.

[0184] In the actual execution process, the system will first read the thermal distribution boundary contour data obtained in the previous stage. This data contains the coordinate point set of each hot zone boundary, defining a closed high-temperature area. In order to discretize the continuous area, the system adopts a rule division method. The specific process is: the entire area is divided into a fixed grid size. and Segmentation is performed. For example, the space within the entire boundary range is divided into rectangular cells of size 1mm×1mm. The vertex coordinates of each small cell are aligned with the original boundary points through linear interpolation to ensure that the grid can fully cover the entire high-temperature area without missing or crossing the boundary.

[0185] In the process of dividing the grid, in order to improve the calculation efficiency and take into account the accuracy, the system adopts a density adaptive optimization strategy, that is, in areas where the heat distribution changes drastically or the boundary is complex, the grid is appropriately refined. and The optimization strategy is to reduce the grid size to 0.5mm to capture more local details; in areas with flat heat distribution and simple boundaries, the basic grid size is maintained at 1mm to avoid unnecessary increase in calculation amount. Make a judgment, if If the angle is greater than the preset threshold of 10 degrees, the grid will be refined.

[0186] After the aforementioned division and optimization process, the system ultimately generates a set of small cells with a complete geometric description. Each cell is represented by four vertex coordinates plus its area, forming a standardized heat conduction unit. Each cell not only identifies its precise location within the overall area but also contains the essential physical parameters for subsequent temperature difference calculations, heat flow derivation, and thermal resistance estimation.

[0187] For example, if the overall area of ​​a high-temperature region is 200 mm² and a 1 mm × 1 mm grid is used, approximately 200 heat conduction units will theoretically be generated. If the grid is locally refined to 0.5 mm × 0.5 mm in areas with complex boundaries, the number of generated units will increase further, ensuring that the overall thermal distribution characteristics can be meticulously restored and captured.

[0188] After generating a collection of heat transfer units, the system further calculates and analyzes nodal thermal resistance based on the resulting small unit set to obtain a complete thermal resistance distribution dataset. Calculating nodal thermal resistance is a key step in assessing the ease of heat transfer in various regions and plays a decisive role in subsequent identification of bottleneck areas.

[0189] Specifically, in the initial stage of the node thermal resistance calculation, the system first organizes the internal geometric parameters of each heat conduction unit, extracting the unit's area A, thickness d, and the thermal conductivity k of the unit material. These basic physical parameters are essential input conditions for subsequent thermal resistance calculations. Area A is determined during meshing, thickness d is set based on the actual physical dimensions of the motherboard stacked structure, and ranges from 1 mm to 2 mm. The thermal conductivity k depends on the type of material, such as 398 W / (m·K) for copper and 149 W / (m·K) for silicon substrates.

[0190] After the parameter extraction is completed, the system executes the standard one-dimensional steady-state heat conduction formula to calculate the node thermal resistance element by element. The formula is:

[0191] in, is the thermal resistance value, unit is , is the unit thickness on the heat flow path, is the thermal conductivity of the material, is the cross-sectional area of ​​the unit. The physical meaning of this formula is: the larger the thermal resistance value, the greater the resistance to heat passing through the unit area, and the worse the heat dissipation performance.

[0192] The system independently executes the above calculation process on each heat transfer unit, generating a complete set of nodal thermal resistance data upon batch completion. To facilitate spatial distribution analysis, the system binds each node's thermal resistance value to its spatial coordinate position in the grid, forming a thermal resistance spatial distribution table. Each record is stored in a standardized structure of {unit number, node coordinate X, node coordinate Y, node thermal resistance value}, ensuring direct access during the subsequent bottleneck area identification phase.

[0193] After batch nodal thermal resistance calculations are completed, the system also performs local smoothing optimization on the nodal thermal resistance data to further enhance the continuity and visibility of the analysis. This operation uses a 3×3 local sliding window as the unit, taking the average of the nine nodal thermal resistances within the window and updating the thermal resistance value of the central node. This reduces occasional extreme numerical fluctuations and enhances the physical rationality of the overall data distribution.

[0194] For example, in the hot zone of a motherboard processor, the results obtained after calculating the node thermal resistance show that the node thermal resistance in the central area is generally as high as 3.0 to 3.5K / W, while the node thermal resistance in the peripheral area is only around 1.2 to 1.5K / W, indicating that heat accumulates severely in the central area and the heat dissipation capacity is relatively poor.

[0195] In summary, this step accurately and efficiently generates a structured thermal resistance distribution dataset based on a set of heat conduction units through a standardized nodal thermal resistance calculation formula and a local optimization strategy.

[0196] After the node thermal resistance distribution dataset is constructed, bottleneck area clustering is performed to identify areas of the system with insufficient heat dissipation and concentrated thermal resistance. This step uses a density-based clustering analysis algorithm to aggregate the thermal resistance data to accurately extract high-resistance bottleneck areas on the motherboard. Bottleneck area identification essentially identifies the contiguous areas with the most severe heat transfer obstruction among the numerous thermal resistance nodes, providing a basis for subsequent heat dissipation optimization.

[0197] At the beginning of the processing, the system first performs standardization preprocessing on the properties of each node. Each node contains spatial coordinates (X, Y) and the corresponding thermal resistance value. ,The system scales the X, Y and thermal resistance values ​​of all nodes to a unified numerical range of 0 to 1 according to the maximum and minimum normalization method, ensuring that different dimensional features have equal influence in subsequent distance calculations.

[0198] Next, the system performs a preliminary cluster analysis based on the Density Peaks Clustering algorithm. The algorithm operation process is as follows: First, calculate the Euclidean distance between all nodes and define the node To Node The distance is:

[0199]

[0200] In order to perform local density estimation, the system sets the cutoff distance dc as the neighborhood radius. The value of dc is dynamically set according to the density of the node distribution, specifically the 5% quantile value of the distance between all nodes in the sample set. In this application example, the system selects dc as 2% of the long side size of the motherboard mapping area. That is, if the long side of the motherboard thermal management analysis area is 100 mm, then dc is set to 2 mm. This value can ensure that each high thermal resistance cluster has at least 5 to 10 valid neighbor nodes, avoiding both isolated nodes caused by local sparseness and excessive aliasing within the cluster caused by overcrowding.

[0201] Based on this DC value, the system counts the number of neighbors of each node within its DC neighborhood radius, which is recorded as the local density. The greater the local density, the denser the thermal resistance nodes are around the node. Then, for each node, the minimum distance to the node with a higher local density than itself is calculated. , which is used to measure the degree of isolation in a dense group.

[0202] System and In the constructed two-dimensional feature plane, nodes with both high local density and large minimum distance are selected as initial cluster centers. The specific criteria are: the local density must be greater than twice the average local density of all nodes, and the minimum distance must be greater than 1.5 times the average minimum distance of all nodes. This selection method effectively ensures that the cluster center node is located at the core of the area with abnormal thermal resistance accumulation.

[0203] After cluster center selection, the system assigns all other nodes to the bottleneck cluster represented by the nearest cluster center based on the shortest distance principle. Each cluster corresponds to a local bottleneck block. To ensure physical plausibility, the system performs a posteriori cleaning on the initial clustering results: clusters with fewer than five nodes or a spatial span less than twice the DC are removed to prevent isolated small clusters from interfering with overall recognition performance.

[0204] For example, in the thermal analysis of the processor power supply area of ​​a certain test motherboard, the system identified three thermal resistance bottleneck clusters through the above processing. The largest bottleneck area covered 20 nodes, and the local average thermal resistance reached 3.8K / W, which was significantly higher than the global average thermal resistance of 1.5K / W, confirming that there was a serious heat dissipation bottleneck in this area.

[0205] In step S18, fan parameters are optimized and regulated according to the thermal resistance distribution of the bottleneck area to obtain an optimized heat dissipation control strategy, including:

[0206] According to the thermal resistance distribution of the bottleneck area, thermal resistance threshold screening and boundary fitting processing are performed to obtain high energy consumption area positioning data;

[0207] Based on the high energy consumption area positioning data, fan status collection and parameter iterative optimization are performed to obtain optimized fan control parameter data;

[0208] According to the optimized fan control parameter data, a regional control strategy is formulated to obtain an optimized heat dissipation control strategy.

[0209] It is worth noting that after completing the extraction of the thermal resistance distribution in the bottleneck area, the system needs to screen key areas and construct spatial contours based on the thermal resistance values ​​to achieve accurate positioning of high-energy consumption hotspot areas. Specifically, the system first performs numerical threshold screening on the entire thermal resistance distribution data set. The system's preset thermal resistance screening threshold is set to 1.3 times the global average thermal resistance value. This value standard is determined based on a large number of sample tests and can effectively balance high sensitivity and low misjudgment rate. In this solution, if the overall average thermal resistance of the bottleneck area is 2.5K / W, the screening threshold is set to 3.25K / W. During the screening process, the system traverses the thermal resistance distribution data point by point, retains all nodes with thermal resistance values ​​greater than or equal to 3.25K / W, and eliminates the remaining low thermal resistance nodes. This calculation process is performed in the data structure in a hash index manner to ensure that the time complexity of the screening operation remains at O(N).

[0210] After selecting high-thermal-resistance nodes, the system performs boundary fitting based on the results. First, the minimum bounding rectangle algorithm is used to perform a preliminary enveloping of the high-thermal-resistance node set. The α-Shape spatial reconstruction algorithm is then used to further fit the boundary contours to accurately reflect the actual distribution of high-energy-consuming areas. The α-Shape algorithm controls boundary curvature by setting a scale parameter α (here set to 1.5 times the unit grid step size). This effectively smooths out the interference of isolated nodes and provides a reasonable approximation of complex boundary structures.

[0211] After boundary fitting is complete, the system extracts the core positioning parameters of each high-energy consumption area, including the center point coordinates (calculated as the arithmetic mean of all node coordinates), the maximum boundary range (i.e., the set of vertices in the outermost fitted boundary), the area (calculated by the number of nodes and unit area within the fitted outline), and the overall mean thermal resistance (the mean of the thermal resistance values ​​of the selected nodes). These parameters are organized into a structured data format to provide accurate input for subsequent fan status collection and control strategy formulation.

[0212] For example, during the analysis of the thermal resistance distribution of the motherboard's power supply area, a specific region was found to contain 38 nodes with thermal resistance exceeding 3.25K / W. After fitting, the region's area was approximately 220 square millimeters, with a center coordinate of (68, 45) and a boundary span of 12mm x 18mm. The average thermal resistance reached 3.9K / W. This high-energy consumption area was officially designated as a priority target for control and will serve as a key input for fan optimization strategy development in subsequent steps.

[0213] After obtaining location data for high-energy consumption areas, the system further integrates real-time fan operating status to conduct targeted collection and iterative optimization of fan operating parameters to improve cooling and reduce system energy consumption. First, during the fan status collection phase, the system continuously obtains key operating parameters for each fan through the motherboard interface, including speed (RPM), current consumption (mA), blade angle, wind direction control angle, and temperature control response delay. These parameters are recorded through a timed sampling process with a sampling period of 5 seconds. The data is then organized into a time series structure through an internal cache for subsequent analysis.

[0214] After completing basic state acquisition, the system analyzes the existing fan layout and coverage of high-heat zones based on the center location and boundaries of the high-energy consumption areas obtained in the previous step. If insufficient coverage, wind force offset, or speed response lag are detected, the system initiates an iterative parameter optimization process. This parameter optimization utilizes a micro-step adjustment strategy based on gradient approximation. Using the current fan operating parameters as the initial values, small adjustments are made to the speed, angle, response delay, and other parameters, observing the temperature response changes in the high-energy consumption areas after these adjustments. The system uses the heat reduction caused by each adjustment as an evaluation metric, prioritizing parameter adjustments that maximize cooling benefits.

[0215] During the calculation process, the speed adjustment range is set to within 20% of the current speed, with a step size of 50 RPM; the wind direction angle adjustment range is set to ±30 degrees, with a step size of 5 degrees; and the response delay adjustment range is set to between a decrease of 0.2 seconds and an increase of 0.5 seconds, with a step size of 0.1 seconds. After each fine-tuning, the system measures the temperature rise and decrease in the target area during the corresponding time period and calculates the objective function value ΔT (temperature drop). If the ΔT value exceeds a preset minimum threshold (e.g., 0.8°C), the current adjustment is fixed; otherwise, exploration continues in other directions.

[0216] During the optimization process, the system employs a greedy search strategy. Once a particular adjustment direction is found to yield positive optimization results, it continues fine-tuning in that direction until ΔT stops increasing or reaches the hardware parameter limit. To prevent side effects such as vibration and noise from excessive adjustments, the system incorporates multi-objective constraints. For example, the speed increase must not exceed 1.2 times the original design limit, and wind direction adjustments must not cause airflow to interfere with other critical areas.

[0217] For example, the cooling fan in the power supply area of ​​a motherboard was initially set to 1800 RPM. After data collection and analysis, the system discovered severe heat accumulation in the core power supply area. After performing parameter optimization, the fan speed was increased to 2200 RPM, the air direction was adjusted by 12 degrees to point toward the center of the hot zone, and the response delay was shortened by 0.3 seconds. Testing showed that the average temperature in the target high-energy consumption area dropped by 2.1°C, meeting the preset optimization criteria. Ultimately, this set of optimized speed, angle, and delay parameters was solidified as the fan's new operating strategy and recorded as optimized fan control parameter data.

[0218] After optimizing and adjusting fan control parameters, the system further develops specific cooling control strategies for each region based on the actual heat load characteristics of different high-energy consumption areas, thereby simultaneously improving overall cooling efficiency and local response speed. First, based on the obtained optimized fan control parameter data, the system associates each fan with its corresponding high-energy consumption area according to its spatial location. Each fan is clearly assigned to a specific thermal zone number through a regional mapping table, ensuring spatial accuracy and logical consistency in subsequent control.

[0219] During the strategy formulation process, to avoid calculation errors caused by inconsistent dimensions between different heat load indicators, the system first normalizes each raw indicator. The temperature rise is normalized to a range of 0 to 1 based on the maximum temperature rise value in the area; the thermal resistance level is normalized based on the maximum thermal resistance value in the system; and the cumulative energy consumption indicator is normalized based on the highest cumulative energy consumption in the entire area during the set sampling period. The normalization method uniformly adopts the linear normalization method, and the specific formula is:

[0220]

[0221] in is the original indicator value, and are the minimum and maximum values ​​of the indicator respectively. is the normalized index value after normalization.

[0222] The system then uses the three normalized indicators as the basis for a weighted scoring method to generate a comprehensive priority score for each high-energy consumption area. Weights are set at 0.4 for temperature rise, 0.4 for thermal resistance, and 0.2 for cumulative energy consumption, reflecting a focus on the rate of temperature rise and the severity of heat accumulation. Each hot zone's final priority score is calculated using a weighted summation formula, ranging from 0 to 1, with higher scores indicating a greater urgency for cooling the area.

[0223] Based on the comprehensive priority score, the system divides all high-energy-consuming areas into three levels: emergency response zone, key support zone, and conventional control zone. For the emergency response zone, the fan strategy is set to increase the fan speed to the maximum speed limit allowed by the optimization parameters, while adjusting the wind direction angle to point toward the center of the hot zone and shortening the fan response time to less than 200 milliseconds. For the key support zone, a moderate speed increase and flexible wind direction coverage are set, with the response time controlled within 400 milliseconds. For the conventional control zone, normal cooling is maintained according to the basic speed and wind direction set in the optimization parameters, and dynamic adjustments are triggered only when the overall ambient temperature is abnormal.

[0224] During the control strategy issuance phase, the system generates a separate instruction set for each fan, including target speed, wind direction adjustment angle, response trigger threshold, and protection strategy for maximum power generation load. Each instruction set is indexed by thermal zone number, ensuring rapid, on-demand switching of cooling solutions under varying load conditions and supporting coordinated strategy optimization across multiple fans.

[0225] In step S19, the optimized heat dissipation control strategy is input into a pre-trained load model, and control parameters for dynamic load adaptation are output, including:

[0226] The load model is obtained by training the LSTM neural network model;

[0227] Performing numerical normalization and vector encoding on the optimized heat dissipation control strategy through the input layer of the load model to obtain a time-structured input tensor;

[0228] Performing recursive gating processing on the time structured input tensor through the time recursive layer of the load model to obtain a state memory representation sequence;

[0229] Performing state aggregation processing on the state memory representation sequence through the state fusion layer of the load model to obtain a compact control feature vector;

[0230] Through the output layer of the load model, the compacted control feature vector is subjected to full connection calculation and nonlinear mapping to output multiple control variables, thereby obtaining control parameters adapted to the dynamic load.

[0231] It's worth noting that in this step, in order for the system to output the corresponding dynamic load adaptation control parameters based on the optimized thermal control strategy, the load model training process must first be completed in the initial stage. This load model uses a long short-term memory (LSTM) structure. Its training task is to learn the response patterns of each strategy to the dynamic changes in the motherboard's real-time load based on historical thermal control strategy change data. This allows the corresponding dynamic load response parameters to be directly inferred based on the new thermal control strategy in actual applications. The entire training process includes five stages: training dataset construction, input preprocessing, parameter initialization, iterative training, and final solidification and storage. Each stage has clear and specific processing methods and standards.

[0232] First, during the training data construction phase, the system collects a large number of cooling strategy adjustment records and corresponding load response parameters under actual working conditions to form training samples. Each sample data contains two parts: the input end records a series of continuous cooling control strategy changes, such as fan speed curves, fan angle deflection curves, and dynamic execution duration sequences; the output end records the actual measured dynamic load response parameters, such as load power consumption change curves, temperature rise rate change curves, and load fluctuation amplitude change curves. The sampling frequency of the data is fixed at 1Hz, and the time range covered by a single sample is 300 seconds, ensuring that sufficient short-term and medium-term dynamic change characteristics are covered. The final training set contains 50,000 complete samples, covering high, medium, and low thermal load scenarios to improve the wide applicability of the training process.

[0233] Next, during the input preprocessing phase, all input and output data must be normalized. This ensures that all input and output features are normalized to the range of 0 to 1, preventing gradient skew or convergence difficulties caused by data of different dimensions and value ranges during training.

[0234] During the parameter initialization phase, the system initializes the weights and bias parameters of the LSTM units and fully connected output layer according to the configured network structure. Weight parameters are initialized using a standard normal distribution with a mean of 0 and a standard deviation of 0.05; bias parameters are uniformly initialized to zero. This ensures that the initial network output remains linear, preventing exploding or vanishing gradients and improving convergence speed during the initial iterations.

[0235] Then it enters the core iterative training phase. The system adopts a supervised training mechanism and the loss function is the mean square error (MSE), which is used to measure the numerical difference between the dynamic load parameters output by the network and the actual sampled load parameters. The specific calculation formula is:

[0236]

[0237] in, is the number of output dimensions, is the actual load parameter, Output the predicted value for the network. In each training iteration, the system performs forward propagation on a batch of samples to calculate the output value, and then backpropagates the error information according to the above loss function, and uses the Adam optimization algorithm to update all learnable parameters. The Adam optimizer parameters are set to a learning rate of 0.001 and a momentum factor of is 0.9, the momentum factor It is 0.999, which takes both numerical stability and convergence speed into consideration.

[0238] After each epoch (i.e., a complete run through the entire training set), the system evaluates the validation set and calculates the mean squared error (MSE) and load forecast accuracy. If the validation set MSE does not decrease within 10 consecutive epochs, the system triggers an early stopping mechanism, terminating training prematurely to prevent overfitting. The maximum number of training rounds is set to 150. At the 92nd round, the validation set MSE converged to 0.0035, meeting the pre-defined convergence criteria.

[0239] After completing load model training, the system performs layer-by-layer processing on each set of optimized thermal control strategy data during the inference phase, ultimately outputting control parameters for dynamic load adaptation. The entire inference process, from input to output, consists of four sequential processing stages: input normalization and vector encoding, time recursion processing, state fusion processing, and fully connected output calculation. Each stage utilizes clear mathematical calculations and algorithmic methods, eliminating ambiguity or jumps.

[0240] First, in the input stage, the system performs normalization and vector encoding processing on the received heat dissipation control strategy data.

[0241] Normalization ensures that the numerical scale of each input field is uniformly distributed between zero and one, eliminating dimension and magnitude differences. The system then arranges the normalized feature values ​​into a fixed-dimensional one-dimensional vector in a preset order, such as a three-dimensional vector consisting of [fan speed percentage, fan angle percentage, execution duration percentage], to form a standardized input tensor.

[0242] Entering the time recursive processing stage, the system uses the long short-term memory (LSTM) unit to perform state recursion and gate control on the input time structured tensor. Specifically, the LSTM contains three basic operation units: input gate, forget gate and output gate. , first calculate the current input vector through the input gate The importance weight of is:

[0243]

[0244] in, is the hidden state at the previous moment, is the input gate weight matrix, is the bias vector, is the Sigmoid activation function. The input gate weight measures the contribution of the current input to the accumulated memory.

[0245] Then the degree of historical memory retention is controlled through the forget gate, and the formula is:

[0246]

[0247] is the output value of the forget gate at the current moment, ranging from 0 to 1, indicating the retention ratio of the memory state of the previous step. It is a Sigmoid activation function that compresses the input to the (0,1) interval. is the weight matrix of the forget gate, is the hidden state of the previous time step, is the input feature vector at the current moment, is the bias vector of the forget gate. The formula is:

[0248]

[0249] is the output gate value at the current moment, Sigmoid activation function compresses the input to between 0 and 1 for gate control. is the weight matrix of the output gate, is the hidden state at the previous moment, is the input feature vector at the current moment, is the bias term of the output gate.

[0250] Through these gating mechanisms, the system performs memory compression and recursive state propagation on the input sequence in the time dimension, gradually forming a state memory sequence that can reflect the trajectory of strategy changes.

[0251] Next, in the state fusion phase, the system uses a time series pooling mechanism to compress the hidden state set of the entire time series into an intermediate feature vector of fixed length. The specific processing method is to perform a maximum pooling operation on each hidden state vector to extract the maximum response value of each feature channel in the time dimension. The formula is:

[0252]

[0253] in, For the The hidden state of the moment, is the total length of the input sequence. Max pooling ensures that the most active response signal in the time series is retained, effectively enhancing the discriminability and compactness of feature expression.

[0254] In the final output stage, the system inputs the intermediate feature vector obtained by state fusion into the fully connected computing unit to perform multiple predictions of the control parameters. The calculation process of the fully connected unit is:

[0255]

[0256] in, is the output weight matrix, is the bias term. This linear transformation maps the high-dimensional feature vector to a specific control parameter space. In order to increase the nonlinear fitting capability, the system also applies a ReLU activation process before output, namely:

[0257]

[0258] Ensure that all output control parameters are non-negative values ​​to meet the actual working requirements of the cooling device. For example, the three output items represent the fan acceleration rate, the fan blade angle adjustment range, and the control trigger delay time.

[0259] For example, if the input optimized heat dissipation control strategy data is: [speed ratio 0.7, angle deflection 0.4, execution time 0.6], after normalization, the input tensor is formed, and the system performs recursive calculation and maximum pooling processing to output the intermediate vector [0.82, 0.64, 0.71]. Finally, after full connection and activation processing, the output parameters [1.20, 0.25, 2.0] are obtained, which respectively represent an additional 20% increase in the basic speed, a 25-degree increase in the fan blade angle, and a control delay of 2 seconds.

[0260] In step S20, real-time control and energy efficiency judgment are performed according to the control parameters adapted to the dynamic load to obtain the final heat dissipation operating parameters, including:

[0261] Perform execution control and collect response feedback according to the control parameters adapted to the dynamic load to obtain execution status data;

[0262] Performing evaluation, regulation, and energy efficiency judgment based on the execution status data to obtain an energy efficiency judgment label;

[0263] According to the energy efficiency judgment label and the dynamic load adaptation control parameters, final parameter screening and solidification are performed to obtain final heat dissipation operation parameters.

[0264] It is worth noting that when executing this step, the system first reads the dynamic load adaptation control parameters generated in the previous stage. This parameter set contains specific instruction information in multiple dimensions, such as target fan speed, fan start and stop status, adjustment direction angle, and control duration. The system parses these values ​​one by one and completes the parameter distribution operation based on the instruction field. Specifically, the system sends a standardized instruction signal to the corresponding heat dissipation execution unit through the underlying interface of the motherboard, such as the I2C bus or the PWM control line. This process requires that the instruction encoding format fully matches the hardware control protocol to ensure that the numerical field will not be distorted due to transmission errors during the distribution process.

[0265] After a control command is issued, the system simultaneously opens a response feedback channel to monitor the actual response of each execution unit in real time. Response data collected includes, but is not limited to, the following: current fan speed (in rpm), operating status (on or off), actual wind direction offset angle (in degrees), and control command execution delay (in milliseconds). The collection frequency is set between 500ms and 1000ms to balance response speed and system load balance. Data collection uses active polling, recording the actual status once per poll, ensuring complete coverage of all key dynamic changes during the control period.

[0266] The collected raw response data undergoes preliminary verification. Verification rules include checking command correspondence (i.e., whether the actual execution action is consistent with the preset command), filtering value fluctuations (eliminating abnormal jumps caused by sensor jitter), and adjusting time synchronization (ensuring that all data points have accurate timestamps). If an anomaly is detected during the verification process, such as no fan response after the command is issued or the speed deviation exceeds the preset tolerance band by 10%, the system automatically records the anomaly flag and appends the anomaly information to the data entry for reference in subsequent energy efficiency assessments.

[0267] Ultimately, all verified execution response data is organized into a structured execution status dataset. Each data record contains the following fields: control object number, issued instruction content, actual response value, response deviation percentage, anomaly flag, and corresponding timestamp. The data storage format uses an efficient indexing structure, such as a key-value hash table or ordered array, to improve the efficiency of subsequent query and analysis.

[0268] To illustrate, let's assume the system issues a control command to cooling fan number 1: a target speed of 2800 rpm, continuous operation for 120 seconds, and a 10-degree wind direction deflection. The collected response data shows that the fan starts within the first second, stabilizes at 2785 rpm within 3 seconds, with a deviation of 0.54%, and achieves a wind direction deflection of 9.8 degrees, with a deviation of 2%. These collected results are all within the allowable error range, so the system records the command's execution status as "successful" and simultaneously generates the corresponding execution status data record.

[0269] Based on the status data collected after the control is executed, the system performs energy efficiency evaluation and control result judgment, and finally generates an energy efficiency judgment label to support the subsequent parameter screening and solidification process. First, after the control action is completed, the system collects the temperature response data of the target area, the equipment energy consumption change record and the fan operation characteristic curve in real time to form a complete execution status data set. This data includes the average temperature before and after the control, the total power consumption value and the fan speed change process, all with clear timestamps to ensure data consistency and time series comparability.

[0270] The system performs preliminary preprocessing on these collected data, including removing missing values, eliminating invalid sampling points, and unifying all physical quantities involving temperature, energy consumption, and time into standard units to ensure the accuracy and consistency of subsequent analysis. On this basis, the system conducts evaluations around three core indicators, namely the temperature reduction amplitude, the energy consumption change ratio, and the degree of response time delay. For each indicator, the system uses the set judgment criteria for analysis. The temperature reduction amplitude requires that the temperature after regulation is at least 5% lower than before regulation, the energy consumption change ratio requires that the new energy consumption does not exceed 10% of the original power consumption, and the response time is required to be completed within 1.2 times the standard response time.

[0271] The system processes each set of data in a fixed order: first, it compares the average temperature difference before and after the adjustment. If the temperature reduction ratio meets the set threshold, it is recorded as meeting the temperature control standard; then it compares the change in total energy consumption before and after the adjustment. If the energy consumption increase is less than 10%, it is recorded as meeting the energy consumption control standard; finally, it extracts the actual response time of the fan or other actuator and compares it with the standard response time. If it is within an acceptable range, it is recorded as timely response.

[0272] When all three core indicators meet the set standards, the system deems the control result valid, generates an Energy Efficiency Judgment Tag of 1, and records it. If any indicator fails to meet the requirements, the control is deemed invalid, and an Energy Efficiency Judgment Tag of 0 is generated. Along with the tag, the system also records a snapshot of the corresponding control parameters for subsequent traceability and optimization.

[0273] For example, in a typical execution, the CPU temperature dropped from 68.5°C to 64.1°C, energy consumption increased by 6.7%, and the fan response time was 1.15 times the standard time. After a step-by-step evaluation, the system determined that this regulation met all standards and output an energy efficiency label of 1. Conversely, if in another set of data, although the temperature dropped by 6%, energy consumption increased by 12.3%, it is determined that energy efficiency has deteriorated and the output energy efficiency label is 0.

[0274] Based on the generated energy efficiency judgment tags and the control parameters adapted to the dynamic load, the system performs final parameter screening and curing operations to determine the optimal cooling operating parameters. First, the system selects records with valid energy efficiency judgment tags from all executed control parameters. Specifically, it selects those control parameters that meet the preset requirements for core indicators such as temperature control range, energy consumption growth, and response time. Parameter records with invalid energy efficiency tags are eliminated, ensuring that the curing process is based only on data samples that have been verified to have reliable cooling effects.

[0275] During the specific screening process, the system sorts multiple indicators according to the temperature control range, energy efficiency change range, response delay and other values ​​corresponding to each valid parameter group, and sets priority weight coefficients. Priority weights are assigned based on actual heat dissipation optimization needs. For example, the temperature control range weight is set to 0.5, the energy efficiency increase weight is set to 0.3, and the response time weight is set to 0.2. The system calculates the comprehensive score through weighted scoring. The comprehensive score formula is:

[0276] Comprehensive score = Normalized value of temperature control range × 0.5 + Normalized value of energy efficiency change × 0.3 + Normalized value of response delay × 0.2

[0277] After all valid samples have achieved a uniform, comparable comprehensive score using the aforementioned method, the system selects the parameter group with the highest comprehensive score as the final thermal dissipation operating parameters to be solidified. To ensure the rationality of the scoring process, each raw indicator must be normalized before participating in the comprehensive score calculation. Using a minimum-maximum normalization formula, values ​​of different dimensions are compressed to a uniform range of zero to one, avoiding evaluation imbalances caused by unit differences.

[0278] For example, in a batch of valid records, parameter group A had a standardized value of 0.88 for temperature control amplitude, 0.93 for energy efficiency change, and 0.81 for response delay, resulting in a comprehensive score of 0.88 × 0.5 + 0.93 × 0.3 + 0.81 × 0.2 = 0.877. The comprehensive score for another parameter group, B, was 0.862, so the system prioritized parameter group A as the final cooling operating parameters.

[0279] After determining the final parameter set, the system solidifies it. This solidification process involves writing fan speed settings, temperature control trigger thresholds, dynamic speed regulation logic, and other parameters into the mainboard control system's parameter table. It also locks the corresponding parameter version numbers to prevent unauthorized modification. The system also generates a solidification parameter log file, documenting the parameter selection source, score calculation process, and solidification time, ensuring traceability of parameter management.

[0280] It is worth noting that the final cooling operating parameters are not just static settings, but have a certain dynamic adaptation mechanism. During execution, the system will monitor actual load changes and, based on the current temperature rise trend, make subtle dynamic adjustments to the fan speed or speed control logic to further reduce the cooling response delay and the probability of lag. Specifically, when the sampled temperature rise rate is higher than the set threshold, the system can activate the fan acceleration mechanism in advance, without having to wait for the temperature to reach a fixed set value before starting, thus achieving an early warning cooling response.

[0281] For example, with traditional static control, when the CPU temperature reaches 85°C, the fan starts running at full speed, with significant lag. However, with the finalized dynamic cooling operating parameters, when the system detects a temperature increase of 0.5°C per second, even if the temperature is only 80°C, it will increase the fan speed in advance, thus avoiding the heat accumulation effect caused by lag and greatly improving the immediacy and reliability of the overall system cooling control.

[0282] In summary, through rigorous energy efficiency judgment, comprehensive scoring screening and parameter solidification mechanism, combined with dynamic load monitoring and fine-tuning optimization, the system finally obtained a set of heat dissipation operating parameters that can significantly suppress hysteresis.

[0283] Reference Figure 2 The second embodiment of the present invention provides an industrial motherboard computer heat dissipation optimization system based on AI intelligent control, including:

[0284] Data acquisition module, used to obtain real-time temperature data of the mainboard;

[0285] A time series module is used to perform time series preprocessing based on the real-time temperature data of the mainboard to obtain distribution time series data;

[0286] A coordinate distribution module is used to extract features and classify the distribution time series data to obtain the coordinate distribution of high energy consumption areas;

[0287] A classification module, configured to input the coordinate distribution of the high energy consumption area into a pre-trained classification model and output a classification result;

[0288] a priority regulation label module, configured to determine the areas exceeding a preset temperature threshold in the classification results as high energy consumption areas, thereby obtaining priority regulation labels for the high energy consumption areas;

[0289] A heat distribution perception module is used to generate a heat map and extract boundaries based on the priority control labels of the high-energy consumption areas to obtain boundary data for heat distribution perception;

[0290] A thermal resistance distribution module is used to perform thermal resistance analysis and bottleneck identification based on the boundary data of the thermal distribution perception to obtain the thermal resistance distribution of the bottleneck area;

[0291] a heat dissipation control module, configured to optimize and control fan parameters according to the thermal resistance distribution of the bottleneck area to obtain an optimized heat dissipation control strategy;

[0292] A control parameter module is used to input the optimized heat dissipation control strategy into a pre-trained load model and output control parameters adapted to the dynamic load;

[0293] The heat dissipation operation parameter module is used to perform real-time control and energy efficiency judgment according to the control parameters adapted to the dynamic load to obtain the final heat dissipation operation parameters.

[0294] It should be noted that the industrial motherboard computer heat dissipation optimization system based on AI intelligent control provided in an embodiment of the present invention is used to execute all the process steps of the industrial motherboard computer heat dissipation optimization method based on AI intelligent control in the above embodiment. The working principles and beneficial effects of the two correspond one to one, so they will not be repeated here.

[0295] An embodiment of the present invention further provides an electronic device. The electronic device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor, such as an industrial motherboard computer heat dissipation optimization program based on AI intelligent control. When the processor executes the computer program, the steps in each of the above-mentioned embodiments of the industrial motherboard computer heat dissipation optimization method based on AI intelligent control are implemented, such as Figure 1 Alternatively, when the processor executes the computer program, the functions of the modules / units in the above-mentioned device embodiments are realized, such as the heat dissipation operation parameter module.

[0296] Exemplarily, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to implement the present invention. The one or more modules / units may be a series of computer program instruction segments capable of implementing specific functions, and the instruction segments are used to describe the execution process of the computer program in the electronic device.

[0297] The electronic device may be a computing device such as a desktop computer, notebook, PDA, or smart tablet. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will appreciate that the aforementioned components are merely examples of electronic devices and do not constitute a limitation of the electronic device. The electronic device may include more or fewer components than those described above, or a combination of certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, and the like.

[0298] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor. The processor is the control center of the electronic device and connects various parts of the entire electronic device using various interfaces and lines.

[0299] The memory can be used to store the computer programs and / or modules. The processor implements the various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and accessing the data stored in the memory. The memory may primarily include a program storage area and a data storage area. The program storage area may store an operating system and at least one application required for a function (such as a sound playback function or an image playback function); the data storage area may store data generated based on the use of the mobile phone (such as audio data, a phone book, etc.). Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as a hard disk, internal memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0300] If the module / unit integrated into the electronic device is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention can implement all or part of the process steps in the above-mentioned method embodiments by using a computer program to instruct the relevant hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium. It should be noted that the content of the computer-readable medium can be appropriately increased or decreased based on the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, based on legislation and patent practice, computer-readable media does not include electric carrier signals and telecommunication signals.

[0301] It should be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art can understand and implement the present invention without inventive effort.

[0302] The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for optimizing heat dissipation of industrial motherboard computers based on AI intelligent control, characterized in that: include: Get real-time temperature data of the motherboard; Performing time series preprocessing on the real-time temperature data of the mainboard to obtain distributed time series data; Perform feature extraction and classification recognition based on the distribution time series data to obtain the coordinate distribution of high energy consumption areas; Input the coordinate distribution of the high energy consumption area into a pre-trained classification model and output the classification result; Determine the areas exceeding a preset threshold in the classification results as high energy consumption areas, thereby obtaining priority regulation labels for the high energy consumption areas; Based on the priority control labels of the high-energy consumption areas, heat map generation and boundary extraction are performed to obtain boundary data of heat distribution perception; Performing thermal resistance analysis and bottleneck identification based on the boundary data of the thermal distribution perception to obtain the thermal resistance distribution of the bottleneck area; Optimizing and controlling fan parameters according to the thermal resistance distribution in the bottleneck area to obtain an optimized heat dissipation control strategy; Input the optimized heat dissipation control strategy into a pre-trained load model and output control parameters adapted to the dynamic load; According to the control parameters adapted to the dynamic load, real-time control and energy efficiency judgment are performed to obtain the final heat dissipation operating parameters; The step of performing thermal resistance analysis and bottleneck identification based on the boundary data of the thermal distribution perception to obtain the thermal resistance distribution of the bottleneck area includes: Performing finite element meshing processing based on the boundary data of the thermal distribution perception to obtain a set of heat conduction units; Perform node thermal resistance calculation and analysis based on the heat conduction unit set to obtain a thermal resistance distribution data set; Based on the cluster analysis algorithm, the bottleneck area clustering identification processing is performed on the thermal resistance distribution data set to obtain the thermal resistance distribution of the bottleneck area; The real-time control and energy efficiency judgment are performed based on the control parameters adapted to the dynamic load to obtain the final heat dissipation operating parameters, including: According to the control parameters adapted to the dynamic load, control is performed and response feedback is collected to obtain execution status data; Performing evaluation, regulation, and energy efficiency judgment based on the execution status data to obtain an energy efficiency judgment label; Performing final parameter screening and solidification based on the energy efficiency judgment label and the dynamic load adaptation control parameters to obtain final heat dissipation operating parameters; Among them, based on the execution status data, evaluation, regulation and energy efficiency judgment are carried out to obtain an energy efficiency judgment label, including: comparing the average temperature difference before and after regulation, if the temperature reduction ratio meets the set threshold, it is recorded as temperature control meeting the standard; comparing the total energy consumption change before and after regulation, if the energy consumption increase is less than 10%, it is recorded as energy consumption control meeting the standard; extracting the actual response time of the fan or other actuator, and comparing it with the standard response time. If it is within an acceptable range, it is recorded as timely response; when the three core indicators all meet the set standards, the system determines that the regulation result is valid, generates energy efficiency judgment label 1, and records it in the file.

2. The method for optimizing heat dissipation of industrial motherboard computers based on AI intelligent control according to claim 1 is characterized in that: The method of performing time series preprocessing on the real-time temperature data of the mainboard to obtain distributed time series data includes: Perform smoothing and filtering processing on the real-time temperature data of the mainboard to obtain a stable temperature change sequence; According to the stable temperature change sequence, abnormal points are identified and structured to obtain distribution time series data.

3. The method for optimizing heat dissipation of industrial motherboard computers based on AI intelligent control according to claim 1 is characterized in that: The feature extraction and classification identification are performed based on the distribution time series data to obtain the coordinate distribution of the high energy consumption area, including: Extracting and analyzing the temperature slope according to the distribution time series data to obtain a set of thermal characteristic values; The thermal feature value set is used to perform threshold judgment and area identification to obtain a high energy consumption area identification result; Based on the high energy consumption area identification result, spatial mapping and coordinate positioning are performed to obtain the coordinate distribution of the high energy consumption area.

4. The method for optimizing heat dissipation of industrial motherboard computers based on AI intelligent control according to claim 1 is characterized in that: The step of inputting the coordinate distribution of the high energy consumption area into a pre-trained classification model and outputting the classification result includes: The classification model is obtained by training the MLP neural network model; Through the input layer of the classification model, the coordinate distribution of the high energy consumption area is converted into a unified format and dimensional encoded to obtain a standardized input vector; Performing weighted transformation and nonlinear activation on the standardized input vector through the hidden layer of the classification model to obtain an intermediate expression vector; The intermediate expression vector is classified and judged through the output layer of the classification model to obtain a classification result.

5. The method for optimizing heat dissipation of industrial motherboard computers based on AI intelligent control according to claim 1 is characterized in that: The step of generating a heat map and extracting boundaries based on the priority control labels of the high energy consumption areas to obtain boundary data of heat distribution perception includes: Constructing a spatial interpolation heat map based on the priority control labels of the high energy consumption areas to obtain two-dimensional heat map data; Performing image clustering and segmentation processing based on the two-dimensional thermal map data to obtain high-temperature area segmentation data; Based on the high-temperature area segmentation data, edge detection and boundary extraction analysis are performed to obtain boundary data for thermal distribution perception.

6. The method for optimizing heat dissipation of industrial motherboard computers based on AI intelligent control according to claim 1 is characterized in that: The fan parameters are optimized and regulated according to the thermal resistance distribution in the bottleneck area to obtain an optimized heat dissipation control strategy, including: According to the thermal resistance distribution of the bottleneck area, thermal resistance threshold screening and boundary fitting processing are performed to obtain high energy consumption area positioning data; Based on the high energy consumption area positioning data, fan status collection and parameter iterative optimization are performed to obtain optimized fan control parameter data; According to the optimized fan control parameter data, a regional control strategy is formulated to obtain an optimized heat dissipation control strategy.

7. The method for optimizing heat dissipation of industrial motherboard computers based on AI intelligent control according to claim 1 is characterized in that: The optimized heat dissipation control strategy is input into a pre-trained load model to output control parameters adapted to the dynamic load, including: The load model is obtained by training the LSTM neural network model; Performing numerical normalization and vector encoding on the optimized heat dissipation control strategy through the input layer of the load model to obtain a time-structured input tensor; Performing recursive gating processing on the time structured input tensor through the time recursive layer of the load model to obtain a state memory representation sequence; Performing state aggregation processing on the state memory representation sequence through the state fusion layer of the load model to obtain a compact control feature vector; Through the output layer of the load model, the compacted control feature vector is subjected to full connection calculation and nonlinear mapping to output multiple control variables, thereby obtaining control parameters adapted to the dynamic load.

8. An industrial motherboard computer heat dissipation optimization system based on AI intelligent control, characterized in that: A method for optimizing heat dissipation of an industrial motherboard computer based on AI intelligent control according to any one of claims 1 to 7, comprising: Data acquisition module, used to obtain real-time temperature data of the mainboard; A time series module is used to perform time series preprocessing based on the real-time temperature data of the mainboard to obtain distribution time series data; A coordinate distribution module is used to extract features and classify the distribution time series data to obtain the coordinate distribution of high energy consumption areas; A classification module, configured to input the coordinate distribution of the high energy consumption area into a pre-trained classification model and output a classification result; A priority regulation label module is used to determine the areas exceeding a preset threshold in the classification results as high energy consumption areas, thereby obtaining priority regulation labels for the high energy consumption areas; A heat distribution perception module is used to generate a heat map and extract boundaries based on the priority control labels of the high-energy consumption areas to obtain boundary data for heat distribution perception; A thermal resistance distribution module is used to perform thermal resistance analysis and bottleneck identification based on the boundary data of the thermal distribution perception to obtain the thermal resistance distribution of the bottleneck area; a heat dissipation control module, configured to optimize and control fan parameters according to the thermal resistance distribution in the bottleneck area to obtain an optimized heat dissipation control strategy; A control parameter module is used to input the optimized heat dissipation control strategy into a pre-trained load model and output control parameters adapted to dynamic loads; The heat dissipation operation parameter module is used to perform real-time control and energy efficiency judgment according to the control parameters adapted to the dynamic load to obtain the final heat dissipation operation parameters.

Images