一种基于深度学习的服务器温度控制方法及系统

Abstract

This invention discloses a server temperature control method and system based on deep learning, belonging to the field of server temperature control technology. First, server hardware temperature, business load, and environmental parameters are simultaneously collected. Thermal imaging technology is used to capture heat conduction relationships, and a dynamic relationship network is constructed through spatiotemporal fusion processing. Based on this network, a heat flow prediction model is built. The model parameters are dynamically corrected by combining the thermal response characteristic attenuation coefficient generated from business load feature analysis to predict temperature distribution changes. A candidate control strategy set is generated accordingly. The optimal strategy is selected through composite scenario simulation, and control commands are generated. The actual temperature after the control command is executed is compared with the model's predicted temperature in real time. When the deviation exceeds a threshold, the system switches to a safe control mode. After the deviation recovers, incremental samples are extracted to update the model parameters, thereby achieving precise temperature control.

Description

Technical Field

[0001] This invention relates to the field of server temperature control technology, specifically a server temperature control method and system based on deep learning. Background Technology

[0002] In data center operations, temperature control of server clusters is a crucial aspect of ensuring stable equipment operation and extending hardware lifespan. With the rapid development of cloud computing and big data applications, the business load that servers need to handle exhibits significant dynamic fluctuations. For example, promotional activities on e-commerce platforms and peak transaction days in financial systems can lead to a sharp increase in load within a short period of time.

[0003] Existing server temperature control methods largely rely on real-time collected hardware parameters and temperature sensor data, using feedback adjustments to achieve cooling control. However, such methods have significant limitations: Firstly, there is an inherent lag between temperature changes and workload growth. When a sudden increase in load causes a rapid rise in temperature, traditional passive response control often lags behind the actual temperature change trend, easily leading to temperature overshoot. This can cause servers to trigger performance throttling due to overheating, affecting business continuity. Secondly, existing control logic adjusts only based on the current state, making it difficult to cope with sudden, periodic load fluctuations, resulting in energy waste or insufficient cooling. Furthermore, with the increase in server density and computing power demands, temperature anomalies in a single node can affect surrounding equipment through heat conduction, further amplifying the shortcomings of traditional control methods. Summary of the Invention

[0004] The purpose of this invention is to provide a server temperature control method and system based on deep learning to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a server temperature control method based on deep learning, comprising:

[0006] S100: Synchronously collects server hardware temperature data, business operation load data, and environmental parameters; captures the heat conduction relationship between devices through thermal imaging technology; and uses a spatiotemporal fusion algorithm to process multidimensional data and construct a dynamic relationship network that reflects the heat transfer path and intensity.

[0007] S200: Construct a heat flow prediction model based on a dynamic relationship network to predict temperature distribution changes in future time periods; construct a hardware state model based on business load characteristics to analyze the impact of load type and duration on server thermal response characteristics and output thermal response characteristic attenuation coefficient; input the thermal response characteristic attenuation coefficient as a dynamic correction term into the heat flow prediction model to adjust the parameters related to heat transfer in the model in real time.

[0008] S300: Based on the output temperature distribution change and the corrected heat transfer parameters, a set of candidate control strategies is generated; by simulating a combined scenario of business load fluctuation and local equipment anomalies, the comprehensive score of each candidate strategy in the scenario is calculated, and the optimal strategy is selected to generate control commands.

[0009] S400: Real-time comparison of actual temperature and predicted temperature after execution of control commands. When the local deviation exceeds the set threshold, it automatically switches to the preset safety control mode. When the deviation returns to the tolerance range, it extracts the operating data during the safety control period to form an incremental sample and updates the parameters of the heat flow prediction model and hardware status model.

[0010] Furthermore, S100 includes:

[0011] S110: Collects hardware temperature data through temperature sensor arrays deployed on key components of the server, obtains business load data through the server's built-in system monitoring interface, and collects data room temperature, humidity and airflow distribution parameters through environmental monitoring equipment.

[0012] S120: The infrared thermal imaging equipment is used to periodically scan the surface of the server to obtain spatial distribution images of thermal radiation intensity of each component of the server and adjacent devices, spatial topology of heat conduction path, and extract the temperature field gradient change rate as a quantitative indicator of heat conduction relationship.

[0013] S130: Nearest neighbor interpolation is used to timestamp and synchronize hardware temperature data, service load data, environmental parameters, and heat conduction relationships, unifying multidimensional data with different sampling frequencies into the same time grid to construct a multidimensional time-series dataset D={(x t 1 ,x t 2 ,...,x t n )}, t∈[1,T]; where x t n represents the feature value of the nth class of data at time t; T represents the total length of the time series; the multidimensional time series dataset is standardized and preprocessed, and the thermal imaging data is subjected to Gaussian smoothing and noise reduction processing;

[0014] A dynamic relationship network is constructed based on preprocessed multidimensional time-series data. The node set represents server components, with each node containing a feature vector of temperature, load characteristics, and environmental parameters. The edge set represents the heat conduction relationships between components, with edge weights w. ij Calculated from thermal imaging data, the specific calculation formula is: w ij =(▽T ij / ||▽T max ||)×α+β;where ▽Tij Represents the temperature gradient vector between nodes i and j, ||▽T max || represents the global maximum temperature gradient norm, α is the heat conduction sensitivity coefficient with a value range of (0,2], used to adjust the influence of the normalized temperature gradient on the edge weights; β is the basic thermal correlation bias term with a value range of [0.1,0.3]; α and β are both learnable adjustment parameters of the model, which are optimized by minimizing the heat flow prediction error during the model training phase;

[0015] The design employs a spatiotemporal attention mechanism to capture dynamic features of heat transfer. A spatial attention layer performs interactive calculations on the feature vectors of each component node to obtain the weight distribution of the heat conduction correlation between components at the current moment. A temporal attention layer performs temporal correlation analysis on the feature vectors of the same component at the current moment and historical moments to obtain the weight distribution of heat conduction correlation at different time scales. Finally, a spatiotemporal fusion layer weights the spatial and temporal correlation weights and updates the feature vectors of each node based on the fusion result.

[0016] Furthermore, the S200 includes:

[0017] S210: A cascaded structure of a three-layer graph convolutional network (GCN) and a two-layer gated recurrent unit (GRU) is adopted, using the node feature matrix and edge weight matrix of the dynamic relational network as input; a heat conduction weighted aggregation function is designed in the GCN layer to incorporate heat conduction intensity into neighborhood feature aggregation, and the feature update formula for node i is:

[0018] ;

[0019] in, The feature vector of node i in the l-th layer is shown below. W represents the set of neighboring nodes of node i. (l) and b (l) Let be the trainable parameters of the l-th layer, and σ be the activation function. ;

[0020] The spatial feature sequence output by GCN is input into a two-layer gated recurrent unit (GRU). Historical temperature change information is selectively retained through update and reset gates, and the output is a feature vector containing time-dependent features. Parallel time-step prediction branches are designed to predict the temperature distribution in the next 15 minutes, 30 minutes, 1 hour, 2 hours and 3 hours, respectively. Each branch shares the bottom feature extraction layer, and the final prediction result is generated through a fully connected weight matrix.

[0021] S220: Convert workload data into a three-dimensional tensor B∈R T×L×CWhere T is the time step, L is the load type dimension, and C is the load feature dimension, the load type is encoded into a low-dimensional vector through an embedding layer, and fused with load intensity and duration to form the input features; an improved Transformer architecture is adopted, introducing a temperature-load two-factor attention mechanism, and a temperature-sensitive term is introduced into the self-attention calculation. The calculation of query Q, key K, and value V is as follows:

[0022] ;

[0023] Where A(Q,K,V) represents the attention output feature fused with the temperature factor, ΔT is the temperature change rate at the corresponding time step, and γ is a learnable sensitivity parameter. This represents element-wise multiplication; the fully connected layer maps the features output by the attention mechanism to a vector of thermal response attenuation coefficients, with each element corresponding to the degree of thermal response attenuation of a node in the dynamic relational network.

[0024] S230: The thermal response characteristic attenuation coefficient is used as a dynamic correction term input into the heat flux prediction model. Based on the edge weight matrix of the dynamic relationship network, the heat conduction centrality index of each node is calculated, as follows:

[0025] ;

[0026] Where c i σ represents the thermal centrality index of node i. ji σ represents the number of shortest paths from node j to i. ji (i) represents the number of shortest paths passing through node i;

[0027] Based on the node thermal conductivity center c i Design modified weights for different GCN layers, and adjust the weight matrix W for the l-th layer. (l) Make corrections:

[0028] W (l)’ =W (l) ×(1-μ i ×λ i );

[0029] Where μ i λ represents the normalized nodal thermal centrality index. i The corresponding node's thermal response attenuation coefficient is used; the corrected weight matrix is ​​applied to the graph convolutional layer of the heat flow prediction model, so that the model can adapt to the changes in thermal response characteristics caused by business load when predicting temperature distribution, thereby improving prediction accuracy.

[0030] Furthermore, the S300 includes:

[0031] S310: Taking the area exceeding the preset safety threshold in the temperature distribution change output by S200 as the core control object, and combining the operating parameter ranges of the cooling equipment and the auxiliary heat dissipation device, a multi-dimensional strategy parameter space is constructed; among them, the control parameters of the cooling equipment include the refrigeration power distribution ratio and the air flow circulation rate, and the control parameters of the auxiliary heat dissipation device include the fan speed level and the heat sink diversion angle. Each parameter is divided into several adjustment gradients according to its hardware characteristics;

[0032] S320: Based on the distribution characteristics of the heat conduction intensity in the heat transfer parameters corrected by S200, calculate the heat conduction comprehensive index s of each region i , representing the diffusion ability of region i in the dynamic network, and combining the difference ΔT between the predicted temperature of this region and the safety threshold i , construct the priority index p i =η×ΔT i +(1 - η)×s i , where p i represents the control priority of region i; η is the balance weight, and its value range is (0,1); according to the priority index p i set differential constraints on the parameter space: for the high-priority regions where p i is higher than the set threshold, limit the minimum values of the refrigeration power distribution ratio and the fan speed level; for the regions with a high heat conduction comprehensive index s i s i >=s0, where s0 is the heat conduction intensity threshold, and limit the lower limit of the air flow circulation rate; for the regions with a low heat conduction comprehensive index s i but a large ΔT i in the local high-temperature regions s i <s0 and ΔT i >=T0, where T0 is the temperature over-threshold determination threshold, and define a specific adjustment range for the heat sink diversion angle;

[0033] S330: Based on the differential constraint results, further establish the mapping relationship between the parameters and the region characteristics: for the regions with high heat conduction intensity, establish a positive correlation mapping between the refrigeration power distribution weight and s i : w c,i =s i / max(s); where, w c,i represents the refrigeration power weight assigned to region i, and max(s) represents the maximum value of s i among all regions;

[0034] For the regions with low heat conduction intensity but exceeding the temperature threshold, allocate the fan speed according to the amplitude of the temperature exceeding the threshold: f i =ε×(ΔT i / max(ΔT)); where, f iThis represents the relative fan speed of region i, with a value range of (0,1), and max(ΔT) represents the value of ΔT across all regions. i The maximum value of ε is the fan speed adjustment coefficient, used to adapt and map the temperature exceeding the threshold to the relative value of the fan speed, ensuring that the calculated relative fan speed value is within the hardware allowable range of the auxiliary cooling device. The value of ε is determined based on the maximum speed hardware parameter of the auxiliary cooling device, specifically the ratio of the relative value corresponding to the maximum fan speed to max(ΔT), calculated using the formula ε=1 / max(ΔT), ensuring that when ΔT... i When f reaches its maximum value i =1, meaning the fan speed reaches the maximum allowable value, when ΔT i For other values, f i The corresponding relative rotational speed is proportionally adapted; multiple candidate strategies are selected from the parameter space after the above constraints using the Latin hypercube sampling method, and each strategy is associated with the temperature change trend predicted by S200, thereby forming a set of candidate control strategies adapted to the temperature change trend.

[0035] S340: Construct a simulation model for a composite scenario of business load fluctuations and local equipment anomalies. Based on business load characteristic analysis, generate dynamic patterns of load type abrupt changes and load intensity step jumps, and convert them into corresponding thermal response characteristic attenuation coefficients through a hardware state model. For key nodes of the dynamic relationship network constructed in S100, simulate abnormal states such as cooling equipment efficiency attenuation and sensor signal offset. By adjusting the node edge weights and eigenvalues, parameterize the abnormal states to ensure that the heat transfer relationship of the simulation scenario is consistent with the topological logic of the dynamic relationship network. Input the candidate control strategy set into the simulation model, run the set simulation cycle, and obtain the execution data of each strategy in the composite scenario.

[0036] The load type mutation indicates that the type of business load changes in a short period of time, causing a sudden change in the thermal response characteristics of different hardware components of the server, thereby changing the distribution of heat conduction parameters; the load intensity step indicates that the intensity of the business load suddenly increases or decreases at a certain moment, causing the power consumption of the server hardware to change rapidly, resulting in a step-like fluctuation in the temperature field.

[0037] S350: Based on the output strategy execution data, calculate the comprehensive score of each candidate strategy; construct a multi-dimensional evaluation index system including temperature control accuracy, server energy consumption, and regulation stability, where temperature control accuracy is the deviation rate between the actual temperature and the safety threshold after strategy execution, server energy consumption is the total energy consumption of cooling equipment and auxiliary heat dissipation devices, and regulation stability is the temperature fluctuation variance; for the strategy execution effect in a single composite scenario, the score is obtained by multiplying the normalized score of temperature control accuracy, the normalized score of energy consumption, and the normalized score of stability by their respective weight coefficients and then summing them, with the three weight coefficients ranging from (0,1) and their sum being 1; the comprehensive score is the weighted average of the scores of each scenario, with high-risk scenarios having a higher weight than low-risk scenarios; the high-risk scenario is a composite scenario where the load fluctuation exceeds a set proportion or the equipment abnormality reaches a set level; the low-risk scenario is a composite scenario where the load fluctuation is lower than a set proportion and the equipment abnormality is lower than a set level.

[0038] S360: Sort the output comprehensive scores, select a preset number of strategies and delete strategies with scores below a set threshold; for the verified strategies, extract the control parameters of the cooling equipment and auxiliary heat dissipation devices, convert them into instruction formats that conform to the equipment control protocol, and add execution priority labels according to the temperature predicted by S200; if there are strategies with the same score, select the strategy with lower energy consumption as the final output through secondary screening by energy consumption index.

[0039] Furthermore, the S400 includes:

[0040] S410: Real-time acquisition of server hardware temperature data after control command execution; matching the actual temperature data with the concurrent predicted temperature data output by S200 through timestamp alignment to construct a deviation analysis dataset {(T p,i ,T r,i )}, where T p,i Let T be the predicted temperature for region i. r,i The actual temperature of region i;

[0041] S420: Calculate the temperature deviation value e for each region i =|T r,i -T p,i | When the deviation value in any region exceeds the preset threshold e th When this threshold is set based on the server hardware safe temperature difference range, it triggers the safety control mode switching mechanism: immediately invokes the preset safety control policy, which fixes the cooling power allocation ratio of the cooling equipment to more than 80% of the rated value, and adjusts the fan speed of the auxiliary heat dissipation device to the highest level; at the same time, it suspends the output of the regular control commands generated by S300, records the time, area and corresponding load characteristics of the deviation, and forms an abnormal event log.

[0042] S430: Continuously monitors temperature changes under the safety control mode. When the deviation values ​​of all areas return to the tolerance range and the stabilization time reaches the set duration, it automatically exits the safety control mode and resumes the execution of regular control commands. It extracts complete operating data during the safety control mode, including actual temperature sequences, load change curves, equipment control parameters, and environmental parameters. It standardizes and synchronizes the data with time according to the preprocessing standard of S130 to form an incremental sample set. It supplements the temperature-load correlation data in the incremental sample set into the hardware state model training set, retrains the model, and updates the thermal response characteristic attenuation coefficient.

[0043] A server temperature control system based on deep learning includes a multi-source data sensing module, a dual-model dynamic correction and prediction module, a strategy generation and composite scenario simulation module, and a control feedback module.

[0044] The multi-source data sensing module synchronously collects server hardware temperature data, business operation load data, and environmental parameters, and captures the heat conduction relationship between devices through thermal imaging technology; it uses a spatiotemporal fusion algorithm to process multidimensional data and construct a dynamic relationship network that reflects the heat transfer path and intensity.

[0045] The dual-model dynamic correction prediction module constructs a heat flow prediction model based on a dynamic relationship network to predict temperature distribution changes in future time periods; it constructs a hardware state model based on business load characteristics to analyze the impact of load type and duration on server thermal response characteristics and outputs a thermal response characteristic attenuation coefficient; the thermal response characteristic attenuation coefficient is input as a dynamic correction term into the heat flow prediction model to adjust the parameters related to heat transfer in the model in real time.

[0046] The strategy generation and composite scenario simulation module generates a set of candidate control strategies based on the output temperature distribution changes and the corrected heat transfer parameters. By simulating a composite scenario of sudden fluctuations in business load and local equipment anomalies, it calculates the comprehensive score of each candidate strategy in the scenario and selects the optimal strategy to generate control instructions.

[0047] The control feedback module compares the actual temperature with the predicted temperature after the control command is executed in real time. When the local deviation exceeds the set threshold, it automatically switches to the preset safety control mode. When the deviation returns to the tolerance range, it extracts the operating data during the safety control period to form an incremental sample and updates the parameters of the heat flow prediction model and the hardware status model.

[0048] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0049] This invention constructs a dynamic relationship network through a spatiotemporal fusion algorithm to accurately capture the heat conduction path and intensity between server components. It also combines graph convolutional networks and gated loop units to construct a heat flow prediction model. The model parameters are dynamically corrected by introducing the thermal response characteristic attenuation coefficient generated by business load feature analysis, thereby achieving accurate prediction of temperature distribution in multiple time periods in the future. This solves the problem that traditional temperature control methods cannot fully reflect the dynamic changes in heat transfer and have low prediction accuracy.

[0050] Meanwhile, candidate control strategies are evaluated and screened through multi-dimensional scoring in simulated composite scenarios. Combined with a weighted mechanism for high-risk scenarios, the reliability of the optimal strategy under complex load fluctuations and equipment anomalies is ensured. Furthermore, a closed loop of prediction, control, feedback, and iteration is formed through real-time deviation monitoring and incremental sample update mechanisms. This improves the accuracy and stability of temperature control while reducing system energy consumption, significantly enhancing the adaptive capability and safety of server temperature control. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 This is a flowchart of a server temperature control method based on deep learning. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Please see Figure 1 This invention provides a technical solution: a server temperature control method based on deep learning, comprising:

[0055] S100: Synchronously collects server hardware temperature data, business operation load data, and environmental parameters; captures the heat conduction relationship between devices through thermal imaging technology; and uses a spatiotemporal fusion algorithm to process multidimensional data and construct a dynamic relationship network that reflects the heat transfer path and intensity.

[0056] S200: Construct a heat flow prediction model based on a dynamic relationship network to predict temperature distribution changes in future time periods; construct a hardware state model based on business load characteristics to analyze the impact of load type and duration on server thermal response characteristics and output thermal response characteristic attenuation coefficient; input the thermal response characteristic attenuation coefficient as a dynamic correction term into the heat flow prediction model to adjust the parameters related to heat transfer in the model in real time.

[0057] S300: Based on the output temperature distribution change and the corrected heat transfer parameters, a set of candidate control strategies is generated; by simulating a combined scenario of business load fluctuation and local equipment anomalies, the comprehensive score of each candidate strategy in the scenario is calculated, and the optimal strategy is selected to generate control commands.

[0058] S400: Real-time comparison of actual temperature and predicted temperature after execution of control commands. When the local deviation exceeds the set threshold, it automatically switches to the preset safety control mode. When the deviation returns to the tolerance range, it extracts the operating data during the safety control period to form an incremental sample and updates the parameters of the heat flow prediction model and hardware status model.

[0059] S100 includes:

[0060] S110: Collects hardware temperature data through temperature sensor arrays deployed on key components of the server, obtains business load data through the server's built-in system monitoring interface, and collects data room temperature, humidity and airflow distribution parameters through environmental monitoring equipment.

[0061] S120: The infrared thermal imaging equipment is used to periodically scan the surface of the server to obtain spatial distribution images of thermal radiation intensity of each component of the server and adjacent devices, spatial topology of heat conduction path, and extract the temperature field gradient change rate as a quantitative indicator of heat conduction relationship.

[0062] S130: Nearest neighbor interpolation is used to timestamp and synchronize hardware temperature data, service load data, environmental parameters, and heat conduction relationships, unifying multidimensional data with different sampling frequencies into the same time grid to construct a multidimensional time-series dataset D={(x t 1 ,x t 2 ,...,x t n )}, t∈[1,T]; where x t n represents the feature value of the nth class of data at time t; T represents the total length of the time series; the multidimensional time series dataset is standardized and preprocessed, and the thermal imaging data is subjected to Gaussian smoothing and noise reduction processing;

[0063] A dynamic relationship network is constructed based on preprocessed multidimensional time-series data. The node set represents server components, with each node containing a feature vector of temperature, load characteristics, and environmental parameters. The edge set represents the heat conduction relationships between components, with edge weights w. ij Calculated from thermal imaging data, the specific calculation formula is: w ij =(▽T ij / ||▽T max ||)×α+β;where ▽T ij Represents the temperature gradient vector between nodes i and j, ||▽T max || represents the global maximum temperature gradient norm, α is the heat conduction sensitivity coefficient with a value range of (0,2], used to adjust the influence of the normalized temperature gradient on the edge weights; β is the basic thermal correlation bias term with a value range of [0.1,0.3]; α and β are both learnable adjustment parameters of the model, which are optimized by minimizing the heat flow prediction error during the model training phase;

[0064] The design employs a spatiotemporal attention mechanism to capture dynamic features of heat transfer. A spatial attention layer performs interactive calculations on the feature vectors of each component node to obtain the weight distribution of the heat conduction correlation between components at the current moment. A temporal attention layer performs temporal correlation analysis on the feature vectors of the same component at the current moment and historical moments to obtain the weight distribution of heat conduction correlation at different time scales. Finally, a spatiotemporal fusion layer weights the spatial and temporal correlation weights and updates the feature vectors of each node based on the fusion result.

[0065] S200 includes:

[0066] S210: A cascaded structure of a three-layer graph convolutional network (GCN) and a two-layer gated recurrent unit (GRU) is adopted, using the node feature matrix and edge weight matrix of the dynamic relational network as input; a heat conduction weighted aggregation function is designed in the GCN layer to incorporate heat conduction intensity into neighborhood feature aggregation, and the feature update formula for node i is:

[0067] ;

[0068] in, The feature vector of node i in the l-th layer is shown below. W represents the set of neighboring nodes of node i. (l) and b (l) Let be the trainable parameters of the l-th layer, and σ be the activation function. ;

[0069] The spatial feature sequence output by GCN is input into a two-layer gated recurrent unit (GRU). Historical temperature change information is selectively retained through update and reset gates, and the output is a feature vector containing time-dependent features. Parallel time-step prediction branches are designed to predict the temperature distribution in the next 15 minutes, 30 minutes, 1 hour, 2 hours and 3 hours, respectively. Each branch shares the bottom feature extraction layer, and the final prediction result is generated through a fully connected weight matrix.

[0070] S220: Convert workload data into a three-dimensional tensor B∈R T×L×C Where T is the time step, L is the load type dimension, and C is the load feature dimension, the load type is encoded into a low-dimensional vector through an embedding layer, and fused with load intensity and duration to form the input features; an improved Transformer architecture is adopted, introducing a temperature-load two-factor attention mechanism, and a temperature-sensitive term is introduced into the self-attention calculation. The calculation of query Q, key K, and value V is as follows:

[0071] ;

[0072] Where A(Q,K,V) represents the attention output feature fused with the temperature factor, ΔT is the temperature change rate at the corresponding time step, and γ is a learnable sensitivity parameter. This represents element-wise multiplication; the fully connected layer maps the features output by the attention mechanism to a vector of thermal response attenuation coefficients, with each element corresponding to the degree of thermal response attenuation of a node in the dynamic relational network.

[0073] S230: The thermal response characteristic attenuation coefficient is used as a dynamic correction term input into the heat flux prediction model. Based on the edge weight matrix of the dynamic relationship network, the heat conduction centrality index of each node is calculated, as follows:

[0074] ;

[0075] Where c i σ represents the thermal centrality index of node i. ji σ represents the number of shortest paths from node j to i. ji (i) represents the number of shortest paths passing through node i;

[0076] Based on the node thermal conductivity center c i Design modified weights for different GCN layers, and adjust the weight matrix W for the l-th layer. (l) Make corrections:

[0077] W (l)’ =W (l) ×(1-μ i ×λ i );

[0078] Where μi Denote the normalized node heat conduction centrality index, λ i is the heat response decay coefficient of the corresponding node; Apply the corrected weight matrix to the graph convolutional layer of the heat flow prediction model, so that when the model predicts the temperature distribution, it can adapt to the changes in the heat response characteristics caused by the business load and improve the prediction accuracy.

[0079] S300 includes:

[0080] S310: Take the area where the temperature distribution change output by S200 exceeds the preset safety threshold as the core control object, and construct a multi-dimensional policy parameter space in combination with the operating parameter ranges of the cooling equipment and the auxiliary heat dissipation device; Among them, the control parameters of the cooling equipment include the refrigeration power distribution ratio and the air flow circulation rate, and the control parameters of the auxiliary heat dissipation device include the fan speed level and the heat sink diversion angle. Each parameter is divided into several adjustment gradients according to its hardware characteristics;

[0081] S320: Calculate the comprehensive heat conduction index s of each region based on the distribution characteristics of the heat conduction intensity in the heat transfer parameters corrected by S200 i , which represents the diffusion ability of region i in the dynamic network, and combine the difference ΔT between the predicted temperature of this region and the safety threshold i , and construct the priority index p i =η×ΔT i +(1 - η)×s i , where p i represents the control priority of region i; η is the balance weight, and its value range is (0,1); According to the priority index p i Set differential constraints on the parameter space: For the high-priority regions where p i is higher than the set threshold, limit the minimum values of the refrigeration power distribution ratio and the fan speed level; For the regions where the comprehensive heat conduction index s i is high i >=s0, where s0 is the heat conduction intensity threshold, and limit the lower limit of the air flow circulation rate; For the local high-temperature regions where the comprehensive heat conduction index s i is low but ΔT i is large i <s0 and ΔT i >=T0, where T0 is the temperature over-threshold judgment threshold, and define a specific adjustment interval for the heat sink diversion angle;

[0082] S330: Based on the differential constraint results, further establish the mapping relationship between the parameters and the regional characteristics: For the regions with high heat conduction intensity, establish a positive correlation mapping between the refrigeration power distribution weight and s i : w c,i =s i / max(s); Among them, w c,iThis represents the cooling power weight assigned to region i, and max(s) represents the total cooling power weight across all regions s. i The maximum value;

[0083] For areas with low thermal conductivity but temperatures exceeding the threshold, fan speeds are allocated according to the magnitude of the temperature exceeding the threshold: f i =ε×(ΔT i / max(ΔT)); where f i This represents the relative fan speed of region i, with a value range of (0,1), and max(ΔT) represents the value of ΔT across all regions. i The maximum value of ε is the fan speed adjustment coefficient, used to adapt and map the temperature exceeding the threshold to the relative value of the fan speed, ensuring that the calculated relative fan speed value is within the hardware allowable range of the auxiliary cooling device. The value of ε is determined based on the maximum speed hardware parameter of the auxiliary cooling device, specifically the ratio of the relative value corresponding to the maximum fan speed to max(ΔT), calculated using the formula ε=1 / max(ΔT), ensuring that when ΔT... i When f reaches its maximum value i =1, meaning the fan speed reaches the maximum allowable value, when ΔT i For other values, f i The corresponding relative rotational speed is proportionally adapted; multiple candidate strategies are selected from the parameter space after the above constraints using the Latin hypercube sampling method, and each strategy is associated with the temperature change trend predicted by S200, thereby forming a set of candidate control strategies adapted to the temperature change trend.

[0084] S340: Construct a simulation model for a composite scenario of business load fluctuations and local equipment anomalies. Based on business load characteristic analysis, generate dynamic patterns of load type abrupt changes and load intensity step jumps, and convert them into corresponding thermal response characteristic attenuation coefficients through a hardware state model. For key nodes of the dynamic relationship network constructed in S100, simulate abnormal states such as cooling equipment efficiency attenuation and sensor signal offset. By adjusting the node edge weights and eigenvalues, parameterize the abnormal states to ensure that the heat transfer relationship of the simulation scenario is consistent with the topological logic of the dynamic relationship network. Input the candidate control strategy set into the simulation model, run the set simulation cycle, and obtain the execution data of each strategy in the composite scenario.

[0085] The load type mutation indicates that the type of business load changes in a short period of time, causing a sudden change in the thermal response characteristics of different hardware components of the server, thereby changing the distribution of heat conduction parameters; the load intensity step indicates that the intensity of the business load suddenly increases or decreases at a certain moment, causing the power consumption of the server hardware to change rapidly, resulting in a step-like fluctuation in the temperature field.

[0086] S350: Based on the output strategy execution data, calculate the comprehensive score of each candidate strategy; construct a multi-dimensional evaluation index system including temperature control accuracy, server energy consumption, and regulation stability, where temperature control accuracy is the deviation rate between the actual temperature and the safety threshold after strategy execution, server energy consumption is the total energy consumption of cooling equipment and auxiliary heat dissipation devices, and regulation stability is the temperature fluctuation variance; for the strategy execution effect in a single composite scenario, the score is obtained by multiplying the normalized score of temperature control accuracy, the normalized score of energy consumption, and the normalized score of stability by their respective weight coefficients and then summing them, with the three weight coefficients ranging from (0,1) and their sum being 1; the comprehensive score is the weighted average of the scores of each scenario, with high-risk scenarios having a higher weight than low-risk scenarios; the high-risk scenario is a composite scenario where the load fluctuation exceeds a set proportion or the equipment abnormality reaches a set level; the low-risk scenario is a composite scenario where the load fluctuation is lower than a set proportion and the equipment abnormality is lower than a set level.

[0087] S360: Sort the output comprehensive scores, select a preset number of strategies and delete strategies with scores below a set threshold; for the verified strategies, extract the control parameters of the cooling equipment and auxiliary heat dissipation devices, convert them into instruction formats that conform to the equipment control protocol, and add execution priority labels according to the temperature predicted by S200; if there are strategies with the same score, select the strategy with lower energy consumption as the final output through secondary screening by energy consumption index.

[0088] The S400 includes:

[0089] S410: Real-time acquisition of server hardware temperature data after control command execution; matching the actual temperature data with the concurrent predicted temperature data output by S200 through timestamp alignment to construct a deviation analysis dataset {(T p,i ,T r,i )}, where T p,i Let T be the predicted temperature for region i. r,i The actual temperature of region i;

[0090] S420: Calculate the temperature deviation value e for each region i =|T r,i -T p,i | When the deviation value in any region exceeds the preset threshold e th When this threshold is set based on the server hardware safe temperature difference range, it triggers the safety control mode switching mechanism: immediately invokes the preset safety control policy, which fixes the cooling power allocation ratio of the cooling equipment to more than 80% of the rated value, and adjusts the fan speed of the auxiliary heat dissipation device to the highest level; at the same time, it suspends the output of the regular control commands generated by S300, records the time, area and corresponding load characteristics of the deviation, and forms an abnormal event log.

[0091] S430: Continuously monitors temperature changes under the safety control mode. When the deviation values ​​of all areas return to the tolerance range and the stabilization time reaches the set duration, it automatically exits the safety control mode and resumes the execution of regular control commands. It extracts complete operating data during the safety control mode, including actual temperature sequences, load change curves, equipment control parameters, and environmental parameters. It standardizes and synchronizes the data with time according to the preprocessing standard of S130 to form an incremental sample set. It supplements the temperature-load correlation data in the incremental sample set into the hardware state model training set, retrains the model, and updates the thermal response characteristic attenuation coefficient.

[0092] Embodiments of the present invention:

[0093] The system collects the temperature of key components such as the server CPU and memory every 30 seconds using a sensor array, obtains real-time business load data, and simultaneously collects the temperature, humidity, and airflow parameters of the data center. Every 5 minutes, the server surface is scanned with an infrared thermal imaging device to extract the temperature field gradient change rate. After timestamp synchronization and standardized preprocessing, a dynamic relationship network is constructed: nodes represent server components and contain temperature, load, and environmental feature vectors; edge weights are calculated based on temperature gradients, with α initially set to 1.5 and β set to 0.2. Spatial and temporal correlation weights are calculated through a spatiotemporal attention mechanism, and the node features are updated after fusion.

[0094] A three-layer GCN and two-layer GRU cascaded structure is adopted as the heat flow prediction model. The GCN layer updates node features through a heat conduction weighted aggregation function, and the GRU captures time-series dependencies and predicts the temperature distribution for the next 15 minutes to 3 hours through parallel branches. The hardware state model transforms the load data into a three-dimensional tensor, which is then fused with load intensity and duration through an embedding layer to form input features. The input is a Transformer with a temperature-load dual-factor attention mechanism, and the output is a thermal response characteristic decay coefficient. Based on this coefficient and the node heat conduction centrality index, the GCN weight matrix is ​​modified to improve the prediction adaptability.

[0095] Taking the region exceeding the safety threshold (CPU 80℃, memory 70℃) as the core, and combining the parameter ranges of cooling equipment (cooling power 0-100%, airflow 1-5m / s) and auxiliary devices (fan level 1-5, heat dissipation angle 0-90°), a multi-dimensional strategy parameter space is constructed; the comprehensive heat conduction index s of the region is calculated. i With temperature difference ΔT i Construct priority p i (Balance weight η=0.6), and based on this, differentiated constraints are applied to the parameter space: for high-priority areas, the cooling power allocation ratio is limited to no less than 50%, and the fan speed level is limited to no less than level 3; for s i In areas ≥0.6, the airflow circulation rate is limited to no less than 3 m / s; for s i <0.6 and ΔTi For local high-temperature areas ≥5℃, the heat sink airflow angle is defined as 30-60°. Based on the above constraints, a mapping relationship between parameters and regional characteristics is established. Fifty candidate strategies are selected from the constrained parameter space using the Latin hypercube sampling method, and each strategy is adapted to the predicted temperature change trend. A composite scenario simulation model is constructed, including load type abrupt changes, load intensity step jumps, and local equipment anomalies (such as cooling equipment efficiency decay and sensor signal offset). The candidate strategies are input into the model and run for a set period. A comprehensive score is calculated based on temperature control accuracy (weight 0.5), server energy consumption (weight 0.3), and regulation stability (weight 0.2). The score weight for high-risk scenarios (load fluctuation exceeding the set proportion or equipment anomaly reaching the set level) is set to 0.7, and for low-risk scenarios it is set to 0.3.

[0096] Based on the comprehensive score ranking, strategies with scores below 80 are deleted. For validated strategies, execution priority labels are added according to the magnitude of predicted temperature exceeding the threshold. If the scores are the same, the strategy with lower energy consumption is selected to generate control commands. The actual temperature after the control command is executed is compared with the predicted temperature in real time. When the deviation value in any region exceeds 5℃, the safety control mode is triggered: the cooling power allocation ratio of the cooling equipment is fixed at more than 80%, the fan speed of the auxiliary heat dissipation device is adjusted to the highest level, the output of regular control commands is suspended, and the abnormal event log is recorded. Temperature changes are continuously monitored. When the deviation values ​​in all regions fall back to 0-2℃ and the stable time reaches 30 minutes, the safety control mode is automatically exited and regular control is restored. The operating data during the safety control period is extracted to form incremental samples, which are added to the hardware state model training set to update the thermal response characteristic decay coefficient, so as to achieve continuous optimization of model performance.

[0097] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A server temperature control method based on deep learning, characterized in that: The method includes: S100: Synchronously collects server hardware temperature data, business operation load data, and environmental parameters; captures the heat conduction relationship between devices through thermal imaging technology; and uses a spatiotemporal fusion algorithm to process multidimensional data and construct a dynamic relationship network that reflects the heat transfer path and intensity. S100 includes: S110: Collects hardware temperature data through temperature sensor arrays deployed on key components of the server, obtains business load data through the server's built-in system monitoring interface, and collects data room temperature, humidity and airflow distribution parameters through environmental monitoring equipment. S120: The infrared thermal imaging equipment is used to periodically scan the surface of the server to obtain spatial distribution images of thermal radiation intensity of each component of the server and adjacent devices, spatial topology of heat conduction path, and extract the temperature field gradient change rate as a quantitative indicator of heat conduction relationship. S130: Nearest neighbor interpolation is used to timestamp and synchronize hardware temperature data, service load data, environmental parameters, and heat conduction relationships, unifying multidimensional data with different sampling frequencies into the same time grid to construct a multidimensional time-series dataset D={(x t 1 ,x t 2 ,...,x t n )}, t∈[1,T]; where x t n represents the feature value of the nth class of data at time t; T represents the total length of the time series; the multidimensional time series dataset is standardized and preprocessed, and the thermal imaging data is subjected to Gaussian smoothing and noise reduction processing; A dynamic relationship network is constructed based on preprocessed multidimensional time-series data. The node set represents server components, with each node containing a feature vector of temperature, load characteristics, and environmental parameters. The edge set represents the heat conduction relationships between components, with edge weights w. ij Calculated from thermal imaging data, the specific calculation formula is: w ij =(▽T ij / ||▽T max ||)×α+β;where ▽T ij Represents the temperature gradient vector between nodes i and j, ||▽T max || represents the global maximum temperature gradient norm, α is the heat conduction sensitivity coefficient with a value range of (0,2], used to adjust the influence of the normalized temperature gradient on the edge weights; β is the basic thermal correlation bias term with a value range of [0.1,0.3]; α and β are both learnable adjustment parameters of the model, which are optimized by minimizing the heat flow prediction error during the model training phase; The design employs a spatiotemporal attention mechanism to capture dynamic characteristics of heat transfer. A spatial attention layer performs interactive calculations on the feature vectors of each component node to obtain the weight distribution of the heat conduction correlation between components at the current moment. A temporal attention layer performs temporal correlation analysis on the feature vectors of the same component at the current moment and historical moments to obtain the weight distribution of heat conduction correlation at different time scales. Finally, a spatiotemporal fusion layer weights the spatial and temporal correlation weights and updates the feature vectors of each node based on the fusion result. S200: Construct a heat flow prediction model based on a dynamic relationship network to predict temperature distribution changes in future time periods; construct a hardware state model based on business load characteristics to analyze the impact of load type and duration on server thermal response characteristics and output thermal response characteristic attenuation coefficient; input the thermal response characteristic attenuation coefficient as a dynamic correction term into the heat flow prediction model to adjust the parameters related to heat transfer in the model in real time. S300: Based on the output temperature distribution change and the corrected heat transfer parameters, a set of candidate control strategies is generated; by simulating a combined scenario of business load fluctuation and local equipment anomalies, the comprehensive score of each candidate strategy in the scenario is calculated, and the optimal strategy is selected to generate control commands. S400: Real-time comparison of actual temperature and predicted temperature after execution of control commands. When the local deviation exceeds the set threshold, it automatically switches to the preset safety control mode. When the deviation returns to the tolerance range, it extracts the operating data during the safety control period to form an incremental sample and updates the parameters of the heat flow prediction model and hardware status model.

2. The server temperature control method based on deep learning according to claim 1, characterized in that: S200 includes: S210: A cascaded structure of a three-layer graph convolutional network (GCN) and a two-layer gated recurrent unit (GRU) is adopted, using the node feature matrix and edge weight matrix of the dynamic relational network as input; a heat conduction weighted aggregation function is designed in the GCN layer to incorporate heat conduction intensity into neighborhood feature aggregation, and the feature update formula for node i is: ； in, The feature vector of node i in the l-th layer is shown below. W represents the set of neighboring nodes of node i. (l) and b (l) Let be the trainable parameters of the l-th layer, and σ be the activation function. ; The spatial feature sequence output by GCN is input into a two-layer gated recurrent unit (GRU). Historical temperature change information is selectively retained through update and reset gates, and the output is a feature vector containing time-dependent features. Parallel time-step prediction branches are designed to predict the temperature distribution in the next 15 minutes, 30 minutes, 1 hour, 2 hours and 3 hours respectively. Each branch shares the bottom feature extraction layer, and the final prediction result is generated through a fully connected weight matrix.

3. The server temperature control method based on deep learning according to claim 1, characterized in that: The S200 further includes: S220: Convert workload data into a three-dimensional tensor B∈R T×L×C Where T is the time step, L is the load type dimension, and C is the load feature dimension, the load type is encoded into a low-dimensional vector through an embedding layer, and fused with load intensity and duration to form the input features; an improved Transformer architecture is adopted, introducing a temperature-load two-factor attention mechanism, and a temperature-sensitive term is introduced into the self-attention calculation. The calculation of query Q, key K, and value V is as follows: ； Where A(Q,K,V) represents the attention output feature fused with the temperature factor, ΔT is the temperature change rate at the corresponding time step, and γ is a learnable sensitivity parameter. This represents element-wise multiplication; the fully connected layer maps the features output by the attention mechanism to a vector of thermal response attenuation coefficients, with each element corresponding to the degree of thermal response attenuation of a node in the dynamic relational network. S230: The thermal response characteristic attenuation coefficient is used as a dynamic correction term input into the heat flux prediction model. Based on the edge weight matrix of the dynamic relationship network, the heat conduction centrality index of each node is calculated, as follows: ； Where c i σ represents the thermal centrality index of node i. ji σ represents the number of shortest paths from node j to i. ji (i) represents the number of shortest paths passing through node i; Based on the node thermal conductivity center c i Design modified weights for different GCN layers, and adjust the weight matrix W for the l-th layer. (l) Make corrections: W (l)’ =W (l) ×(1-μ i ×λ i ); Where μ i λ represents the normalized nodal thermal centrality index. i The corresponding node's thermal response attenuation coefficient is used; the corrected weight matrix is ​​applied to the graph convolutional layer of the heat flow prediction model, so that the model can adapt to the changes in thermal response characteristics caused by business load when predicting temperature distribution, thereby improving prediction accuracy.

4. The server temperature control method based on deep learning according to claim 1, characterized in that: The S300 includes: S310: Taking the area where the temperature distribution change output by S200 exceeds the preset safety threshold as the core control object, and combining the operating parameter range of the cooling equipment and auxiliary heat dissipation device, a multi-dimensional strategy parameter space is constructed; among them, the control parameters of the cooling equipment include the cooling power allocation ratio and the airflow circulation rate, and the control parameters of the auxiliary heat dissipation device include the fan speed level and the heat sink guide angle. Each parameter is divided into several adjustment gradients according to its hardware characteristics. S320: Calculate the comprehensive heat conduction index s for each region based on the distribution characteristics of the heat conduction intensity in the heat transfer parameters corrected by S200 i , representing the diffusion ability of region i in the dynamic network, and combine the temperature difference ΔT between the predicted temperature of this region and the safety threshold i to construct the priority index p i = η × ΔT i + (1 - η) × s i , where p i represents the regulation priority of region i; η is the balance weight, and its value range is (0, 1); according to the priority index p i set differential constraints on the parameter space: for high-priority regions where p i is higher than the set threshold, limit the minimum values of the refrigeration power distribution ratio and the fan speed level; for regions with a high comprehensive heat conduction index s i s i >= s0, where s0 is the heat conduction intensity threshold, and limit the lower limit of the air flow circulation rate; for regions with a low comprehensive heat conduction index s i but a large ΔT i in local high-temperature regions s i < s0 and ΔT i >= T0, where T0 is the temperature over-threshold determination threshold, and define a specific adjustment range for the diversion angle of the heat sink S330: Based on the results of differentiated constraints, a mapping relationship between parameters and regional characteristics is further established: for regions with high thermal conductivity, a weighting of cooling power allocation is established with respect to s. i Positive correlation mapping: w c,i =s i / max(s); where w c,i This represents the cooling power weight assigned to region i, and max(s) represents the total cooling power weight across all regions s. i The maximum value; For areas with low thermal conductivity but temperatures exceeding the threshold, fan speeds are allocated according to the magnitude of the temperature exceeding the threshold: f i =ε×(ΔT i / max(ΔT)); where f i This represents the relative fan speed of region i, with a value range of (0,1), and max(ΔT) represents the value of ΔT across all regions. i The maximum value of ε is the fan speed regulation coefficient. The Latin hypercube sampling method is used to screen multiple candidate strategies from the parameter space after the above constraints. Each strategy is associated with the temperature change trend predicted by S200, so as to form a set of candidate control strategies that are adapted to the temperature change trend.

5. The server temperature control method based on deep learning according to claim 4, characterized in that: The S300 also includes: S340: Construct a simulation model for a composite scenario of business load fluctuations and local equipment anomalies. Based on business load characteristic analysis, generate dynamic patterns of load type abrupt changes and load intensity step jumps, and convert them into corresponding thermal response characteristic attenuation coefficients through a hardware state model. For key nodes of the dynamic relationship network constructed in S100, simulate abnormal states such as cooling equipment efficiency attenuation and sensor signal offset. By adjusting the node edge weights and eigenvalues, parameterize the abnormal states to ensure that the heat transfer relationship of the simulation scenario is consistent with the topological logic of the dynamic relationship network. Input the candidate control strategy set into the simulation model, run the set simulation cycle, and obtain the execution data of each strategy in the composite scenario. The load type mutation indicates that the type of business load changes in a short period of time, causing a sudden change in the thermal response characteristics of different hardware components of the server, thereby changing the distribution of heat conduction parameters; the load intensity step indicates that the intensity of the business load suddenly increases or decreases at a certain moment, causing the power consumption of the server hardware to change rapidly, resulting in a step-like fluctuation in the temperature field. S350: Based on the output strategy execution data, calculate the comprehensive score of each candidate strategy; construct a multi-dimensional evaluation index system including temperature control accuracy, server energy consumption, and regulation stability, where temperature control accuracy is the deviation rate between the actual temperature and the safety threshold after strategy execution, server energy consumption is the total energy consumption of cooling equipment and auxiliary heat dissipation devices, and regulation stability is the temperature fluctuation variance; for the strategy execution effect in a single composite scenario, the score is obtained by multiplying the normalized score of temperature control accuracy, the normalized score of energy consumption, and the normalized score of stability by their respective weight coefficients and then summing them, with the three weight coefficients ranging from (0,1) and their sum being 1; the comprehensive score is the weighted average of the scores of each scenario, with high-risk scenarios having a higher weight than low-risk scenarios; the high-risk scenario is a composite scenario where the load fluctuation exceeds a set proportion or the equipment abnormality reaches a set level; the low-risk scenario is a composite scenario where the load fluctuation is lower than a set proportion and the equipment abnormality is lower than a set level. S360: Sort the output comprehensive scores, select a preset number of strategies and delete strategies with scores below a set threshold; for strategies that pass the verification, extract the control parameters of the cooling equipment and auxiliary heat dissipation devices, convert them into instruction formats that conform to the equipment control protocol, and add execution priority labels according to the temperature predicted by S200; if there are strategies with the same score, select the strategy with lower energy consumption as the final output through secondary screening by energy consumption index.

6. The server temperature control method based on deep learning according to claim 1, characterized in that: The S400 includes: S410: Real-time acquisition of server hardware temperature data after control command execution; matching the actual temperature data with the concurrent predicted temperature data output by S200 through timestamp alignment to construct a deviation analysis dataset {(T p,i ,T r,i )}, where T p,i Let T be the predicted temperature for region i. r,i The actual temperature of region i; S420: Calculate the temperature deviation value e for each region i =|T r,i -T p,i | When the deviation value in any region exceeds the preset threshold e th When this threshold is set based on the server hardware safe temperature difference range, it triggers the safety control mode switching mechanism: immediately invokes the preset safety control policy, which fixes the cooling power allocation ratio of the cooling equipment to more than 80% of the rated value, and adjusts the fan speed of the auxiliary heat dissipation device to the highest level; at the same time, it suspends the output of the regular control commands generated by S300, records the time, area and corresponding load characteristics of the deviation, and forms an abnormal event log. S430: Continuously monitors temperature changes under the safety control mode. When the deviation values ​​of all areas return to the tolerance range and the stabilization time reaches the set duration, it automatically exits the safety control mode and resumes the execution of regular control commands. It extracts complete operating data during the safety control mode, including actual temperature sequences, load change curves, equipment control parameters, and environmental parameters. It standardizes and synchronizes the data with time according to the preprocessing standard of S130 to form an incremental sample set. It supplements the temperature-load correlation data in the incremental sample set into the hardware state model training set, retrains the model, and updates the thermal response characteristic attenuation coefficient.

7. A server temperature control system based on deep learning, applied to the server temperature control method based on deep learning according to any one of claims 1-6, characterized in that: The system includes a multi-source data perception module, a dual-model dynamic correction and prediction module, a strategy generation and composite scenario simulation module, and a control feedback module. The multi-source data sensing module synchronously collects server hardware temperature data, business operation load data, and environmental parameters, and captures the heat conduction relationship between devices through thermal imaging technology; it uses a spatiotemporal fusion algorithm to process multidimensional data and construct a dynamic relationship network that reflects the heat transfer path and intensity. The dual-model dynamic correction prediction module constructs a heat flow prediction model based on a dynamic relationship network to predict temperature distribution changes in future time periods; it also constructs a hardware state model based on business load characteristics to analyze the impact of load type and duration on server thermal response characteristics and outputs thermal response characteristic attenuation coefficient. The thermal response characteristic attenuation coefficient is used as a dynamic correction term and input into the heat flow prediction model to adjust the parameters related to heat transfer in the model in real time. The strategy generation and composite scenario simulation module generates a set of candidate control strategies based on the output temperature distribution change and the corrected heat transfer parameters; it simulates a composite scenario of sudden fluctuations in business load and local equipment anomalies, calculates the comprehensive score of each candidate strategy in the scenario, and selects the optimal strategy to generate control instructions. The control feedback module compares the actual temperature with the predicted temperature after the control command is executed in real time. When the local deviation exceeds the set threshold, it automatically switches to the preset safety control mode. When the deviation returns to the tolerance range, it extracts the operating data during the safety control period to form an incremental sample and updates the parameters of the heat flow prediction model and the hardware status model.

Images