Control method and device of heat dissipation equipment

Abstract

The invention discloses a control method and device for heat dissipation equipment, and relates to the technical field of computers, and the method comprises the steps: employing a load dynamic perception and prediction model cooperative control technology, building a dynamic correlation model of a load and heat dissipation, and achieving the mode upgrade of a heat dissipation system from lagging regulation to advanced response. A control instruction is generated by predicting a load change trend in real time, and a dual-mode protection mechanism is combined, so that energy efficiency loss caused by traditional temperature control delay is avoided, system stability and energy efficiency performance are optimized, an efficient and energy-saving active heat dissipation solution is provided for high-load equipment, and the service life of the system is prolonged. The problems of low control efficiency of the heat dissipation equipment and energy efficiency loss caused by delayed adjustment in related technologies are solved, and the technical effects of improving the control efficiency of the heat dissipation equipment and reducing energy consumption are achieved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a control method and apparatus for heat dissipation equipment. Background Art

[0002] In the field of server heat dissipation control, traditional temperature control methods generally rely on feedback data from temperature sensors for delayed adjustment. Existing technologies usually use baseboard management controllers to periodically collect hardware temperature data and adjust the operating status of heat dissipation components according to the temperature. This control mode has a significant delay effect. Hardware temperature changes are essentially the result of component load fluctuations, and temperature sensors need to go through multiple links such as heat conduction, data acquisition, and signal processing to reflect the actual working conditions. When core components such as processors and memory experience a sudden increase in load, there is often a delay of several seconds from the detection of the temperature anomaly to the effectiveness of the heat dissipation response. During this period, the components may cause performance reduction or even trigger protection mechanisms due to high temperature. In addition, to deal with the potential risk of temperature overshoot, traditional methods often adopt redundant speed strategies, maintaining a high heat dissipation intensity even during low-load periods, resulting in unnecessary energy loss. Therefore, there are technical problems in related technologies such as low control efficiency of heat dissipation equipment and energy efficiency loss caused by delayed adjustment. Summary of the Invention

[0003] The present application provides a control method and device for a heat dissipation device, so as to at least solve the problems in the related art of low control efficiency of the heat dissipation device and energy efficiency loss caused by delayed adjustment.

[0004] The present application provides a control method for a heat dissipation device, comprising: obtaining target load parameters corresponding to functional components of a target device; inputting the target load parameters into a target model to generate target control parameters, wherein the target model represents a model obtained by training an initial model based on historical load parameters corresponding to the functional components and historical control parameters corresponding to the heat dissipation components of the target device, and the initial model represents a relationship model between control parameters corresponding to the heat dissipation components and load parameters of the functional components; and sending the target control parameters to the heat dissipation components so that the heat dissipation components operate according to the target control parameters.

[0005] The present application also provides a control device for a heat dissipation device, including: an acquisition module for acquiring target load parameters corresponding to functional components of a target device; a generation module for inputting the target load parameters into a target model to generate target control parameters, wherein the target model represents a model obtained by training an initial model based on historical load parameters corresponding to the functional components and historical control parameters corresponding to the heat dissipation components of the target device, and the initial model represents a relationship model between the control parameters corresponding to the heat dissipation components and the load parameters of the functional components; a sending module for sending the target control parameters to the heat dissipation components so that the heat dissipation components operate according to the target control parameters.

[0006] The present application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for implementing the steps of any of the above-mentioned control methods for heat dissipation devices when executing the computer program.

[0007] The present application also provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned control methods for heat dissipation devices are implemented.

[0008] The present application also provides a computer program product, including a computer program, which implements the steps of any of the above-mentioned control methods for heat dissipation devices when executed by a processor.

[0009] This application utilizes a technical approach that combines real-time multi-dimensional load parameter acquisition with intelligent model prediction. By establishing a dynamic mapping mechanism between functional component load characteristics and heat dissipation control parameters, it achieves the goal of proactively predicting heat dissipation requirements and precisely regulating device operating status, effectively resolving the multi-second delay issue inherent in traditional temperature feedback control. By deploying a predictive control model based on load characteristics and establishing a nonlinear correlation model between the rate of change of load parameters and the rate of increase in heat dissipation requirements, it achieves the technical goal of optimizing the energy efficiency of heat dissipation components, thereby significantly reducing overall server energy consumption. By establishing a closed-loop model optimization mechanism that continuously collects information on the deviation between actual temperature data and predicted values, it achieves the technical effect of dynamically correcting model parameters to maintain prediction accuracy, thereby improving system stability in long-term operating environments. By implementing a dual-mode seamless switching strategy and establishing intelligent switching logic between predictive control and the traditional temperature control mode of the baseboard management controller, it achieves both fault tolerance and system reliability, thus meeting the high availability requirements in critical business scenarios. Therefore, it addresses the problems of low control efficiency of heat dissipation devices and energy efficiency loss caused by delayed adjustment in related technologies, achieving the technical effect of improving control efficiency of heat dissipation devices and reducing energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0011] Figure 1 A schematic diagram of an application environment of an optional heat dissipation device control method provided in an embodiment of the present application;

[0012] Figure 2 A schematic flow chart of an optional method for controlling a heat dissipation device provided in an embodiment of the present application;

[0013] Figure 3 A schematic diagram of a specific flow chart of an optional control method for a heat dissipation device provided in an embodiment of the present application;

[0014] Figure 4 A schematic diagram of the effect of an optional control method for a heat dissipation device provided in an embodiment of the present application;

[0015] Figure 5 A schematic structural diagram of a control device for an optional heat dissipation device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0016] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0017] It should be noted that, in the description of this application, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. The terms "first," "second," etc., in this application are used to distinguish similar objects, and are not used to describe a particular order or sequence.

[0018] In order to enable those skilled in the art to better understand the present application, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0019] In conjunction with the specific application environment architecture or specific hardware architecture on which the execution of the control method for the heat dissipation device depends, the specific application environment architecture or specific hardware architecture is described herein.

[0020] Pre-training is a strategy for training deep learning models. Its core is to initially train the model using large-scale datasets, enabling it to learn common feature representations. This process is similar to the basic learning stage humans go through before acquiring new knowledge, accumulating experience through extensive reading and observation.

[0021] Fine-tuning refers to conducting small-scale training on the basis of a pre-trained model for specific task objectives (downstream tasks) and task data (downstream data), making slight adjustments to the parameters of the pre-trained model, and ultimately obtaining a model that is adapted to specific tasks and data.

[0022] OS (Operating System) refers to the basic software platform running on a server or terminal device, responsible for coordinating the interaction between hardware resources and application programs.

[0023] PCIE (Peripheral Component Interconnect Express) is a high-speed data transmission channel connecting the CPU and expansion devices, and is used for communication between the processor and components such as memory, network card, and graphics card.

[0024] LPC (Low-Pin-Count), a low-speed control bus on the motherboard, is mainly used for command transmission between the baseboard management controller and complex programmable logic devices.

[0025] The CPU (Central Processing Unit) is the core computing unit of the device and undertakes the main computing tasks. Its load parameters are important input indicators for heat dissipation control.

[0026] CPLD (Complex Programmable Logic Device), a programmable hardware control chip, is responsible for receiving control instructions and driving fans and other heat dissipation components to perform specific operations.

[0027] BMC (Baseboard Management Controller), an embedded management chip that runs independently of the operating system, is used to monitor hardware status and implement basic temperature control strategies.

[0028] I2C (Inter-Integrated Circuit), a serial communications protocol, is the physical interface between a complex programmable logic device and multiple fans.

[0029] FAN (cooling fan, Fan), the executive component of the cooling system, changes the air flow efficiency inside the equipment by adjusting the speed.

[0030] Raid card (disk array controller, Redundant Array of Independent Disks Controller), a storage control expansion card, is connected to the CPU through the PCIE bus and manages the read and write operations of the hard disk array.

[0031] The network card (NIC) is a network communication hardware module that transmits network data packets through the PCIE bus. Its bandwidth occupancy affects the heat dissipation requirements.

[0032] The application structure is as follows Figure 1 As shown in the figure, a closed-loop heat dissipation management system consisting of data acquisition, model prediction and control execution is constructed. The system realizes the intelligent linkage between functional component load perception and heat dissipation control through the collaborative work of the operating system application layer (OS application), driver layer and hardware component layer. Figure 1 Detailed description:

[0033] S1, functional component data collection:

[0034] Deploy data monitoring services at the operating system application layer (OS application) and use the high-speed expansion interface of the driver layer (such as Figure 1 The PCIE interface in the middle) collects the dynamic load parameters of the functional components in real time.

[0035] The specific process includes:

[0036] Processor and memory monitoring: Captures central processing unit (CPU) core utilization, memory bandwidth usage, and L3 cache hit rate through kernel modules;

[0037] Storage device monitoring: Use the driver layer storage protocol interface to obtain the hard disk read and write queue depth, solid-state storage devices (such as Figure 1 I / O latency indicators of hard disks and Raid cards;

[0038] Network and graphics monitoring: collect network throughput through the network card driver and the graphics processor (such as Figure 1 Graphics card) interface to obtain rendering load data;

[0039] After the collected data is standardized by the driver layer, it is temporarily stored in the memory database of the application layer to form a multidimensional data set with a timestamp (such as Figure 1 data flow from the OS application to the component layer).

[0040] S2, control parameter association modeling:

[0041] An initial relationship model is built at the operating system application layer. This model uses artificial intelligence (AI) algorithms to analyze the association patterns in historical data:

[0042] Data preprocessing: The historical load parameters of functional components (such as CPU utilization, memory bandwidth) and the historical control parameters of heat dissipation components (such as Figure 1 Perform timing alignment on the fan speed recorded by the CPLD in the image processing module to eliminate abnormal sensor data.

[0043] Feature Engineering: Screening core indicators that are strongly related to heat dissipation requirements, such as dynamic features such as the processor load mutation rate and the continuous read and write duration of storage devices;

[0044] Model training: A regression algorithm is used to establish a mapping relationship between load parameters and heat dissipation control parameters. For example, for every 10% increase in processor load, the model predicts the fan speed increase range.

[0045] The trained initial model is deployed in the application layer model service module and is transmitted through the low pin count bus of the driver layer (such as Figure 1 LPC interface) to synchronize data with the baseboard management controller (BMC).

[0046] S3, real-time predictive control:

[0047] The system generates dynamic control instructions based on the target model. The specific process includes:

[0048] Load characteristic input: The target load parameters collected in the current cycle (such as Figure 1 The real-time data of CPU, memory and graphics card in the target model is input;

[0049] Control parameter generation: The model calculates the cooling requirements based on load trends and outputs target control parameters including speed gradient and adjustment timing.

[0050] Instructions are sent and executed: Control parameters are transmitted to the complex programmable logic device (CPLD) through the driver layer and transmitted through the serial communication interface (such as Figure 1 I2C interface) to drive multiple sets of fans (such as Figure 1 (FAN node in the middle) runs according to the instructions;

[0051] For example, when the model detects that the graphics card rendering load continues to increase, it generates phased speed-up instructions in advance to avoid insufficient heat dissipation caused by temperature feedback lag in the traditional temperature control mode.

[0052] S4, dual-mode fault tolerance mechanism:

[0053] The system ensures reliability through a dual-channel design using a baseboard management controller (BMC) and artificial intelligence (AI) predictive control:

[0054] Normal operation: Model predictive control mode is preferred, and CPLD directly executes target control parameters;

[0055] Abnormal switching: When the BMC detects that the deviation between the actual temperature and the predicted value exceeds the safety threshold, it immediately switches to the BMC preset temperature control strategy and triggers the following process:

[0056] Mark abnormal operating condition data and transmit it back to the application layer;

[0057] Start the incremental learning process of the model and optimize the parameter weights based on the new data;

[0058] Re-enable predictive control after the model accuracy is restored;

[0059] This mechanism is achieved through Figure 1 The LPC connection between BMC and CPLD enables fast state switching, ensuring the stability of the cooling system in the event of hardware failure or sudden load changes.

[0060] S5, dynamic model optimization:

[0061] The system deploys a continuous learning module at the application layer to maintain model validity through the following methods:

[0062] Closed-loop data collection: Compare the execution results of each control instruction (fan speed, actual temperature) with the predicted value to build an incremental training data set;

[0063] Parameter rolling update: Using online learning algorithms, model parameters are fine-tuned based on new data at set intervals to adapt to hardware aging or environmental changes.

[0064] Cross-device knowledge transfer: When deploying new functional components (such as Figure 1 When adding a new graphics card in the system, load the pre-trained model feature layer parameters to accelerate localization adaptation;

[0065] This process fully utilizes the local computing resources of the terminal or server, avoiding reliance on the cloud training platform, and conforming to Figure 1 The autonomous decision-making feature of the OS application layer in the architecture.

[0066] By directly collecting and model-predicting load parameters of functional components, the system overcomes the physical conduction delay limitations of traditional temperature feedback and achieves proactive response in heat dissipation control. The system can proactively adjust heat dissipation strategies based on load trends to avoid performance throttling caused by temperature fluctuations. The combination of model predictive control and a progressive adjustment strategy significantly reduces ineffective power consumption in heat dissipation components. By precisely matching load characteristics with heat dissipation requirements, the system reduces heat dissipation delays during high-load periods and redundant energy consumption during low-load periods while ensuring device stability. The layered architecture design is compatible with devices with varying hardware configurations. The incremental learning mechanism enables the model to adapt to changing heat dissipation requirements of new processors, expansion cards, and other components, reducing reliance on manual parameter adjustment. The dual control channel design allows for the collaborative operation of innovative algorithms and traditional hardware, enhancing the feasibility of the solution.

[0067] This application forms a complete control closed loop (such as Figure 1 Arrows):

[0068] Data uplink: functional components → driver layer (PCIE / LPC) → OS application layer → AI model;

[0069] Instruction downlink: AI model → driver layer (LPC / I2C) → CPLD → heat dissipation components;

[0070] Security feedback: BMC → Anomaly detection module → Model optimization module.

[0071] This framework fully leverages the advantages of localized computing, ensuring data privacy and real-time performance while providing an efficient and reliable heat dissipation management solution for high-density computing devices.

[0072] The embodiment of the present application provides a control method for a heat dissipation device, such as Figure 2 As shown, the method is described in detail in combination with the execution flow of the control method of the heat dissipation device.

[0073] S202, obtaining target load parameters corresponding to functional components of the target device;

[0074] Optionally, in an embodiment of the present application, the target device may include but is not limited to an electronic device system with computing capabilities and heat dissipation requirements, specifically referring to a server or similar hardware device with multiple components working together. The device usually includes a complete computing architecture consisting of a processor, memory, input and output interfaces, and a heat dissipation unit. Its function is to perform data processing, storage, or network communication tasks, and during operation, the heat generated by the power consumption of the components requires dynamic regulation of the heat dissipation system. For example, in a cloud computing scenario, the target device can be a multi-node server cluster deployed in a data center, including physical components such as a motherboard, power module, storage controller, and network adapter. The heat load generated by each functional module during operation needs to be temperature balanced by fan speed regulation.

[0075] Optionally, in an embodiment of the present application, the above-mentioned functional components may include but are not limited to electronic components that perform core computing or data transmission tasks in the target device, specifically referring to hardware units that generate heat inside the server. During operation, such components generate Joule heat due to the passage of current through semiconductor materials, and their heat load is positively correlated with the workload. For example, when executing instructions, the central processing unit increases the switching frequency of transistors according to the computational complexity, resulting in increased power consumption and temperature rise. Typical examples include the CPU chipset, memory module, hard disk controller, and network interface card on the server motherboard. These components generate heat sources of different intensities when working, and require differentiated heat dissipation control based on their respective load states.

[0076] Optionally, in an embodiment of the present application, the target load parameters may include, but are not limited to, a set of quantitative indicators that reflect the real-time working intensity of functional components, specifically a multi-dimensional data set such as processor usage, memory occupancy ratio, disk read and write rate, and network bandwidth utilization. For example, in the operation monitoring of a storage server, the target load parameters may include the IOPS (input and output operations per second) of the RAID controller, the L1 / L2 cache hit rate, and the throughput data of the PCIe channel. These parameters are collected at millisecond intervals through the operating system kernel module or dedicated sensors to form time series data for characterizing the instantaneous working status of the components and predicting future load trends.

[0077] It should be noted that there are many ways to obtain the target load parameters. From the data acquisition dimension, the current and temperature can be directly measured through embedded sensors, or the thread scheduling status can be extracted through the operating system kernel module, or the resource allocation ratio of the virtual device can be monitored through the virtualization layer. In the parameter type dimension, it may involve one or a combination of indicators such as the floating-point operation throughput of compute-intensive loads, the cache hit rate of storage-intensive loads, and the packet forwarding delay of network-intensive loads. In the time granularity dimension, the instantaneous peak value can be obtained based on millisecond-level real-time sampling, or the average value can be calculated based on a minute-level sliding window, or the trend characteristics can be statistically analyzed at the hourly level in combination with the business cycle characteristics. For example, in the on-board computing unit scenario, the load parameters may include the GPU shader utilization during autonomous driving algorithm inference, the memory bandwidth occupancy during multi-camera data stream decoding, and the number of concurrent connections of the Internet of Vehicles communication module. This application does not make specific restrictions on this.

[0078] S204: Input the target load parameters into a target model to generate target control parameters, wherein the target model represents a model obtained by training an initial model based on historical load parameters corresponding to the functional components and historical control parameters corresponding to the heat dissipation components of the target device, and the initial model represents a relationship model between the control parameters corresponding to the heat dissipation components and the load parameters of the functional components;

[0079] Optionally, in an embodiment of the present application, the above-mentioned target model may include but is not limited to a prediction and decision model constructed based on a machine learning algorithm, which is specifically implemented as a regression analysis model or a neural network model trained with historical data. The model converts input features such as CPU occupancy rate and memory bandwidth collected in real time into optimized control instructions for fan speed by establishing a nonlinear mapping relationship between functional component load parameters and heat dissipation control parameters. For example, when using the support vector regression (SVR) algorithm, the model will learn the optimal fan speed change curve corresponding to the disk array load rate increasing from 30% to 60% in the historical data during the training phase, thereby automatically generating a step-by-step speed regulation scheme according to the new load fluctuations during the prediction phase.

[0080] Optionally, in an embodiment of the present application, the aforementioned historical load parameters may include, but are not limited to, a data set of operating status recorded by functional components during past operating cycles, specifically including time-series metrics such as the number of processor instruction cycles, memory page swap frequency, and network packet processing volume marked with timestamps. For example, in a web server's operation and maintenance log, historical load parameters may store data on the number of Apache threads, MySQL query response time, and SSD write latency during peak hours every day over the past three months. This data is cleaned to form structured training samples for analyzing the correlation between load variation patterns and cooling requirements.

[0081] Optionally, in embodiments of the present application, the heat dissipation component may include, but is not limited to, a mechanical or electronic device for controlling device temperature, specifically a fan array and its drive control system within the server for forced convection cooling. This component regulates air flow by varying the impeller speed, thereby directing heat generated by electronic components out of the chassis. For example, in a blade server, the heat dissipation component typically includes a redundant heat dissipation module consisting of multiple axial fans. Each fan unit receives a speed command from a controller via a pulse width modulation (PWM) signal, enabling stepless speed regulation from 2000 RPM to 8000 RPM.

[0082] Optionally, in an embodiment of the present application, the historical control parameters may include, but are not limited to, a collection of actual operational records of the heat dissipation component during its historical operation, specifically control logs such as fan speed percentages, heat sink temperature thresholds, and speed control strategy switching times. For example, in a high-performance computing cluster, historical control parameters may include the liquid cooling pump power level, air duct baffle opening and closing angles, and auxiliary heat sink activation duration, recorded every five minutes for the past twelve hours. These data, along with the component load parameters for the same period, constitute a complete training data pair for optimizing the control model's predictive accuracy.

[0083] Optionally, in an embodiment of the present application, the initial model may include, but is not limited to, an untrained basic algorithm framework, specifically a linear regression model with a preset parameter relationship or an unoptimized decision tree model. Prior to training, the model only has a theoretical mathematical expression. For example, when using polynomial regression, the initial model may be set to a quadratic function with initial values of its coefficient matrix being random numbers. This model needs to be gradually adjusted during training using a gradient descent algorithm to accurately fit the nonlinear relationship between GPU core temperature and fan speed.

[0084] It should be noted that there are multiple technical paths for the construction and reasoning process of the target model. From the model architecture dimension, a time series prediction model can be used to process the load fluctuation periodic characteristics, a reinforcement learning model can be used to achieve dynamic strategy optimization, or an integrated learning model can be built to integrate multiple heat dissipation control strategies. In the training data dimension, historical data may come from the full life cycle operation log of the same device, the horizontal migration data set of similar devices, or simulation data generated by digital twin technology. In the control parameter generation dimension, a direct speed setting value for a single-point optimal solution, a multi-step speed regulation sequence under a rolling time window, or a multi-objective optimization scheme that includes priority weights and fault tolerance mechanisms may be generated. For example, in the industrial Internet of Things scenario, the model may predict the torque load trend of the injection molding machine spindle based on the LSTM network, and generate a collaborative control matrix of the cooling tower valve opening and the circulating pump speed in combination with the ambient temperature and humidity sensor data. This application does not make specific restrictions on this.

[0085] S206: Send the target control parameters to the heat dissipation component, so that the heat dissipation component operates according to the target control parameters.

[0086] Optionally, in an embodiment of the present application, the above-mentioned target control parameters may include, but are not limited to, a set of heat dissipation device control instructions generated by model calculation, specifically including control variables such as fan speed target value, speed gradient change rate, and effective time window. For example, in a thermal management scenario of a virtualization platform, the target control parameters may include differentiated speed control strategies for servers in different rack positions: the server located at the top of the cabinet needs to increase the base speed by 15% due to the rising effect of hot air, and at the same time, it is set to start a two-stage speed increase protocol when the CPU load exceeds 75%. These parameters are transmitted to the heat dissipation controller via the IPMI protocol for real-time control.

[0087] It should be noted that there are multiple implementation methods for the transmission and execution of target control parameters. From the communication protocol dimension, the Modbus protocol can be used to implement register reading and writing with the industrial fan, send control frames to the vehicle-mounted cooling module through the CAN bus, or use the OPC UA protocol to synchronize and control parameters in the intelligent manufacturing system. In the control mode dimension, feedforward control can be implemented to adjust the cooling intensity in advance based on the prediction model, feedback control can be used to dynamically correct parameters according to the real-time temperature difference, or a hybrid control can be used to integrate device state estimation and disturbance compensation mechanism. In the actuator dimension, it is possible to drive the PWM speed control module of the centrifugal fan, adjust the opening of the heat exchange valve of the liquid cooling system, or control the current polarity of the semiconductor refrigeration plate to achieve active heat absorption. For example, in the edge computing gateway scenario, the control parameters may be sent to the distributed cooling node through LoRa wireless communication, while coordinating the thermal conductivity of the heat pipe and the speed ratio of the micro-turbo fan. This application does not make specific restrictions on this.

[0088] In an exemplary embodiment, it is assumed that the operation process of the server heat dissipation control system is taken as an example, including but not limited to the following process:

[0089] S1. The operating system periodically acquires the dynamic load characteristics of computing components through the kernel-mode performance monitoring module. For the central processing unit, instruction cycle utilization and cache miss rate are collected; for storage units, read and write queue depth and data transmission bandwidth are recorded; and for network interfaces, packet processing latency and buffer occupancy are monitored. Simultaneously, the baseboard management controller's sensor interface acquires historical operation logs of cooling components, including fan speed curves and ambient temperature fluctuation data. Data collection methods can include at least one of polling-based active query, event-triggered passive capture, or sliding window-based incremental statistics.

[0090] In step S2, the real-time load dataset is fed into a pre-trained hybrid prediction model. This model utilizes a long-short-term memory network module to process temporal correlation features, combined with a random forest regressor to analyze multivariate nonlinear relationships. The model outputs predicted cooling requirements for the next three control cycles and generates corresponding control instruction sequences. The control parameter generation process considers the device's thermal capacity characteristics, cooling efficiency attenuation coefficient, and ambient temperature and humidity compensation factors. For example, to address the cooling requirements of high-density computing nodes, the model may generate a composite control scheme that combines a step-by-step speed increase strategy with a gradual speed reduction strategy, while also reserving a safety redundancy speed threshold to prevent the risk of thermal runaway.

[0091] S3, through the relay control interface of the programmable logic device, converts the optimized control parameters into drive signals recognizable by the underlying hardware. For multi-fan parallel cooling systems, the instruction distribution module calculates the collaborative working parameters of each fan unit based on the airflow dynamics model, including but not limited to phase difference adjustment between the master and slave fans, standby switching strategy for redundant fans, and load balancing schemes under abnormal conditions. The control signal transmission protocol can use any transmission mechanism including pulse width modulation encoding, serial peripheral interface synchronous communication, or a custom binary instruction set. The specific implementation depends on the electrical characteristics of the cooling components and the real-time requirements of the system.

[0092] In step S4, the baseboard management controller continuously monitors the actual temperature curves of key components and calculates the deviation between the predicted and actual temperatures using a differential comparison algorithm. When the cumulative deviation exceeds a preset threshold, the model's online update mechanism is triggered. The model's weight matrix is dynamically adjusted using an incremental learning algorithm, using the current load characteristics, control parameter execution results, and environmental parameter changes as new training samples. A sliding time window mechanism is incorporated into the optimization process to ensure that the model can adapt to long-term operating mode changes while preventing overfitting due to short-term noise data.

[0093] By integrating multi-dimensional load characteristics with dynamic environmental parameters, a hybrid prediction model is constructed to predict cooling needs and generate proactive control strategies. This significantly shortens the heat conduction delay window in traditional temperature feedback mechanisms and effectively suppresses temperature fluctuations. By modeling the correlation between load characteristics and cooling efficiency, the system reduces redundant cooling energy consumption, improves fan operating efficiency, and mitigates the risk of electronic device performance degradation due to sudden temperature changes. A closed-loop optimization mechanism enables the model to continuously adapt to varying hardware configurations and environmental conditions, achieving refined lifecycle management of cooling components while ensuring cooling safety.

[0094] This application utilizes a technical approach that combines real-time multi-dimensional load parameter acquisition with intelligent model prediction. By establishing a dynamic mapping mechanism between functional component load characteristics and heat dissipation control parameters, it achieves the goal of proactively predicting heat dissipation requirements and precisely regulating device operating status, effectively resolving the multi-second delay issue inherent in traditional temperature feedback control. By deploying a predictive control model based on load characteristics and establishing a nonlinear correlation model between the rate of change of load parameters and the rate of increase in heat dissipation requirements, it achieves the technical goal of optimizing the energy efficiency of heat dissipation components, thereby significantly reducing overall server energy consumption. By establishing a closed-loop model optimization mechanism that continuously collects information on the deviation between actual temperature data and predicted values, it achieves the technical effect of dynamically correcting model parameters to maintain prediction accuracy, thereby improving system stability in long-term operating environments. By implementing a dual-mode seamless switching strategy and establishing intelligent switching logic between predictive control and the traditional temperature control mode of the baseboard management controller, it achieves both fault tolerance and system reliability, thus meeting the high availability requirements in critical business scenarios. Therefore, it addresses the problems of low control efficiency of heat dissipation devices and energy efficiency loss caused by delayed adjustment in related technologies, achieving the technical effect of improving control efficiency of heat dissipation devices and reducing energy consumption.

[0095] As an optional solution, before inputting the target load parameters into the target model and generating the target control parameters, the above method further includes:

[0096] receiving a real-time control parameter of the heat dissipation component sent by a baseboard management controller, wherein the real-time control parameter represents a control parameter generated in real time by the baseboard management controller based on a real-time temperature parameter of the heat dissipation component detected;

[0097] Obtain real-time load parameters of functional components;

[0098] An initial model is established based on the changes in real-time load parameters and real-time control parameters.

[0099] Optionally, in an embodiment of the present application, the baseboard management controller may include but is not limited to an independent microcontroller unit embedded in the server motherboard to implement hardware monitoring and management, which establishes a communication link with the temperature sensor, power module and heat dissipation component via a dedicated bus. The controller has the functions of real-time acquisition of hardware status data, execution of preset control strategies and abnormal alarms, including but not limited to reading the CPU thermistor temperature through the I2C interface, parsing the cooling fan speed feedback signal, and dynamically adjusting the output voltage of the power supply module. For example, during the operation of a distributed storage server, the baseboard management controller can simultaneously monitor the temperature gradients of multiple hard disk backplanes and send differentiated speed control instructions to the redundant fan array according to the preset strategy.

[0100] Optionally, in an embodiment of the present application, the above-mentioned real-time control parameters may include, but are not limited to, a set of dynamic adjustment instructions that characterize the current operating state of the heat dissipation component, specifically quantifiable parameters such as fan speed percentage, heat sink cooling power level, or liquid cooling pump flow adjustment coefficient. These parameters are generated by the closed-loop control algorithm of the baseboard management controller. For example, when a transient rise in GPU core temperature is detected, the controller may generate a composite control instruction set including fan acceleration slope limit, adjacent fan coordinated start and stop timing, and heat sink contact pressure compensation value to balance heat dissipation efficiency and mechanical loss.

[0101] Optionally, in embodiments of the present application, the aforementioned real-time temperature parameters may include, but are not limited to, a set of measurement data reflecting the current thermodynamic state of the heat dissipation components and associated hardware, specifically including information such as the absolute value of the temperature, the rate of change, and spatial distribution characteristics. For example, in a liquid-cooled heat dissipation system, real-time temperature parameters may include the temperature difference between the inlet and outlet of the cold plate, the change in the specific heat capacity of the coolant, and the temperature distribution diagram of the heat exchanger surface. These data are collected by fusing a distributed thermocouple array with an infrared thermal imaging module.

[0102] It should be noted that there are many implementation forms for the baseboard management controller to send real-time control parameters. From the data source dimension, it may include simple control instructions triggered by a single temperature threshold, a composite control strategy generated by multi-sensor fusion, or adaptive fault-tolerant parameters based on historical failure modes. In the transmission protocol dimension, the standard data format of the IPMI specification, a custom binary encoding structure, or a real-time streaming transmission mechanism based on a time-sensitive network can be used. From the parameter type dimension, it may involve discrete switch control signals, continuous proportional-integral-differential adjustment parameters, or a multi-objective optimization solution set containing priority tags. For example, in a high-availability cluster, the control parameters may integrate temperature balancing strategies and fault switching plans from multiple nodes, which is not specifically limited in this application.

[0103] Optionally, in embodiments of the present application, the aforementioned real-time load parameters may include, but are not limited to, a set of dynamic indicators that quantify the resource usage intensity of server functional modules, specifically core performance parameters such as processor instruction pipeline saturation, memory bus contention rate, and storage controller queue depth. For example, in a virtualized computing node, real-time load parameters may include the context switching frequency of the virtual machine monitor, the packet processing latency of the virtual network interface, and the hit rate curve of the distributed storage cache.

[0104] It should be noted that the method of collecting real-time load parameters of functional components can be expanded according to differences in system architecture. In the data granularity dimension, nanosecond-level precision processor instruction cycle counts, millisecond-level statistics of memory page errors, or minute-level aggregated storage throughput indicators can be collected. In the collection object dimension, it is possible to monitor the thread scheduling status of the physical core, the resource quota allocation ratio of the virtualization layer, or the number of concurrent service requests at the application layer. In the data processing dimension, direct uploading of raw data, sliding window mean filtering, or frequency domain feature extraction based on Fourier transform can be used. For example, in a heterogeneous computing scenario, the load parameters may include both the floating-point operation occupancy rate of the general-purpose processor and the tensor calculation load ratio of the accelerator card. This application does not make specific restrictions on this.

[0105] Optionally, in an embodiment of the present application, the initial model may include, but is not limited to, an untrained mathematical framework that describes the relationship between heat dissipation control parameters and load parameters, and may be implemented as a theoretical model based on the physical heat conduction equation or a machine learning architecture with adjustable parameters. For example, in an air-cooled heat dissipation scenario, the initial model may be constructed as a set of partial differential equations containing airflow rate, heat source power density, and ambient temperature variables, with boundary conditions calibrated using experimental data to predict heat dissipation requirements.

[0106] It should be noted that the construction method of the initial model can be flexibly adjusted according to the requirements of the application scenario. In the model input dimension, time series load fluctuation characteristics, spatially distributed temperature gradient data and environmental noise interference factors may be integrated. In the training data dimension, hybrid training can be performed in combination with laboratory calibration data, field operation history records and digital twin simulation results. In the model optimization dimension, offline batch training, online incremental learning or distributed parameter updates under federated learning can be used. For example, in an edge computing device, the initial model may fuse the local heat dissipation efficiency curve of the device with the typical load pattern characteristics sent from the cloud, and this application does not make specific restrictions on this.

[0107] By dynamically correlating real-time load parameters with heat dissipation control parameters, the initial model is constructed, overcoming the limitations of traditional temperature hysteresis control and enabling proactive matching of heat dissipation performance with computing load. Through multi-dimensional data fusion analysis, the system significantly improves the foresight and accuracy of heat dissipation resource scheduling, effectively suppressing local hotspots caused by sudden load changes. The model optimization mechanism based on real-time feedback enhances the system's adaptability to complex operating conditions, reducing ineffective heat dissipation energy consumption while extending the service life of key components. Furthermore, the flexible model architecture design provides the technical foundation for compatibility with different heat dissipation topologies, supporting scalable deployment from single machines to distributed clusters.

[0108] As an optional solution, an initial model is established based on the changes in real-time load parameters and real-time control parameters, including:

[0109] Continuously detect whether real-time load parameters have changed;

[0110] In the case where a change in the real-time load parameter is detected at a first time point, determining the real-time load parameter corresponding to the first time point as a sample load parameter, and continuously detecting whether the real-time control parameter changes;

[0111] In the case where a change in the real-time control parameter is detected at the second time point, determining the real-time control parameter corresponding to the second time point as the sample control parameter;

[0112] An initial model is established based on at least one set of sample data, wherein the set of sample data includes matched sample load parameters and sample control parameters.

[0113] Optionally, in an embodiment of the present application, the above-mentioned continuous detection of whether the real-time load parameters have changed may include but is not limited to a periodic or event-driven monitoring mechanism for the dynamic working status of the server functional module, specifically identifying significant fluctuations in the load parameters through a preset threshold comparison algorithm or a statistical process control method. The detection process may involve variance analysis within a sliding time window, differential calculation between adjacent sampling points, or abnormal pattern recognition based on machine learning. For example, in a virtualized environment, when the CPU scheduling delay of the virtual machine manager exceeds two standard deviations of the baseline value, a load change flag is triggered, and the trend change of the memory allocation fragmentation index is recorded at the same time.

[0114] It should be noted that the detection method of real-time load parameter changes can be expanded in multiple dimensions according to system requirements. In the detection frequency dimension, fixed-period polling detection, event-driven trigger detection, or adaptive detection interval adjustment based on load sensitivity can be adopted. In the judgment standard dimension, it is possible to set absolute value threshold comparison, relative change rate threshold monitoring, or logical combination judgment of compound conditions. In the data processing dimension, the original data can be subjected to sliding average filtering, wavelet denoising, or principal component analysis dimensionality reduction. For example, in a streaming computing scenario, the system may dynamically adjust the detection frequency according to changes in data throughput, and at the same time combine the variance analysis and business priority weights within the sliding window to make a compound judgment. This application does not make specific restrictions on this.

[0115] Optionally, in an embodiment of the present application, the first time point mentioned above may include but is not limited to the time when a monitoring event triggers a substantial change in the server load characteristics, which is specifically determined by the critical state determination logic marked by the baseboard management controller through the timestamp. The basis for determining this time point may include the duration of the load parameter continuously exceeding the threshold, the satisfaction of a multi-parameter combination alarm condition, or a specific business cycle switching event. For example, in a distributed database service, when the growth rate of the number of concurrent query requests exceeds a critical threshold within a preset time window, and the page error rate of the storage engine increases synchronously, the system marks the timestamp of this state mutation as the first time point.

[0116] Optionally, in an embodiment of the present application, the sample load parameters may include, but are not limited to, representative snapshot data of the working status captured when a load mutation event occurs, specifically including instantaneous measurement values and short-term statistical features in multiple dimensions. For example, in an edge computing node, the sample load parameters may include the peak utilization rate of the GPU shader unit, the instantaneous value of the packet loss rate of the network interface, and the queue depth distribution histogram of the solid-state storage device. These data are normalized to form a standardized feature vector for model training.

[0117] Optionally, in embodiments of the present application, the aforementioned second time point may include, but is not limited to, the time at which the decision to adjust the heat dissipation control parameters in response to load changes takes effect, which is determined by the execution delay of the heat dissipation control strategy and the physical heat conduction characteristics. Determining this time point may involve compensating for control instruction transmission delays, calibrating the response time of heat dissipation components, or filtering ambient temperature interference. For example, in a liquid cooling system, after the cooling pump power adjustment instruction is transmitted via the communication link and the actuator is activated, the system marks the actual flow rate change as the second time point.

[0118] It should be noted that there are multiple ways to implement the method of determining the first time point and the second time point. In the time synchronization dimension, precise alignment of hardware clocks, software timing protocol compensation, or logical timing marking based on event causality may be used. In the delay compensation dimension, transmission link delay estimation, actuator response model, or physical process transfer function calculation may be introduced. In the exception handling dimension, network jitter retransmission mechanism, sensor fault shielding strategy, or data integrity verification need to be considered. For example, in a cross-cabinet heat dissipation collaboration scenario, the system may synchronize the monitoring clocks of multiple nodes through a precise time protocol, and calculate the transmission delay of the heat dissipation instruction based on a fluid mechanics model. This application does not make specific limitations on this.

[0119] Optionally, in embodiments of the present application, the sample control parameters may include, but are not limited to, a quantitative expression of the control strategy implemented by the cooling system in response to load changes, specifically expressed as a fan speed adjustment gradient, a change in the opening of a liquid cooling valve, or a semiconductor refrigeration plate current polarity switching instruction. For example, in a redundant cooling module of a blade server, the sample control parameters may include the trigger conditions for switching between active and standby fans, the switching sequence of multiple speed control gears, and the speed reduction protection strategy for abnormal vibration conditions. These parameters are encoded via a control bus protocol to form a traceable operation record.

[0120] It should be noted that the process of determining the sample load parameters and control parameters can be optimized in multiple dimensions. In the feature extraction dimension, time domain statistics (such as mean, variance), frequency domain features (such as power spectral density) or spatiotemporal correlation features (such as covariance matrix) may be used. In the data association dimension, it is necessary to process data fusion of multi-source sensors, alignment matching of heterogeneous time series or interpretable analysis of causal relationships. In the quality control dimension, outlier removal, data confidence weighting or sample importance sampling can be implemented. For example, in a heterogeneous computing architecture, the system may fuse the differentiated load characteristics of CPU, GPU and FPGA, and perform correlation modeling of multimodal control parameters through the attention mechanism, which is not specifically limited in this application.

[0121] Optionally, in an embodiment of the present application, the at least one set of sample data may include, but is not limited to, spatiotemporal correlation data pairs of load events and cooling responses, specifically comprising a multi-dimensional combination of a load mutation feature vector, a control parameter adjustment sequence, and an environmental interference compensation factor. For example, in a high-performance computing cluster, a set of sample data may integrate the floating-point register occupancy curve of a computing node during a peak matrix operation period, the cooling fan acceleration strategy at the corresponding moment, and the air outlet temperature compensation coefficient of the computer room air conditioning system to form a multi-source heterogeneous joint training sample.

[0122] It should be noted that the combination of sample data can be flexibly configured according to the requirements of the model. In the time span dimension, instantaneous samples at adjacent time points, aggregated samples within a time window, or comparative samples across business cycles can be selected. In the spatial correlation dimension, it is possible to integrate multiple component data of the same device, associated node data across devices, or auxiliary data of the environmental monitoring system. In the data enhancement dimension, time series interpolation, adversarial sample generation, or synthetic expansion based on domain knowledge can be used. For example, in modular server design, the system may cross-domain associate the load parameters of the computing module with the heat dissipation control parameters of the power module to form a joint training sample set, which is not specifically limited in this application.

[0123] By establishing an accurate spatiotemporal correlation model of load changes and heat dissipation responses, this application effectively solves the problem of decreased control accuracy caused by monitoring delays and response lags in traditional methods. The system can capture subtle features of load mutations and trigger sample collection in a timely manner to ensure that the training data truly reflects the dynamic coupling relationship under actual working conditions. The multi-dimensional sample combination strategy enhances the model's generalization ability for complex heat dissipation scenarios, giving the initial model the potential to adapt to different hardware configurations and environmental interference. In addition, the refined time point determination mechanism reduces the interference of noise data on model training, improves the interpretability and reliability of the control strategy, and lays a data foundation for the continuous optimization of the intelligent cooling system.

[0124] As an optional solution, an initial model is established based on the changes in real-time load parameters and real-time control parameters, including:

[0125] Reinforcement learning is used to establish an initial model based on the changes in real-time load parameters and real-time control parameters.

[0126] Optionally, in an embodiment of the present application, the above-mentioned reinforcement learning method may include but is not limited to a technical path for establishing a dynamic decision-making model through an interaction mechanism between an intelligent agent and an environment, which is specifically manifested as a triple framework with the load parameters of the server cooling system as the state space, the control parameters as the action space, and the temperature stability and energy efficiency as the reward function. For example, in an air-cooled server scenario, the intelligent agent learns how to dynamically adjust the cooling strategy under different load fluctuation modes by monitoring the historical interaction data of the CPU core temperature and the fan speed, so as to maintain the temperature in a safe range while reducing energy consumption. The deep Q network (DQN) algorithm may be used in the reinforcement learning process to input the time series characteristics of the load parameters into the neural network, output the optimal fan speed adjustment strategy, and continuously optimize the strategy network weights through temperature deviation feedback.

[0127] Optionally, in an embodiment of the present application, the changes in the above-mentioned real-time parameters may include, but are not limited to, correlation analysis between the dynamic fluctuation characteristics of the working status of the functional components of the server and the response behavior of the cooling system, specifically including the fan acceleration delay time during the load surge phase, the cooling efficiency attenuation curve during the load stability period, and the inertial overcooling risk during a sudden load drop. For example, in an edge computing node, the real-time load parameter may manifest as an instantaneous jump in the GPU shader utilization of the image processing task, while the real-time control parameter corresponds to the step-by-step speed-up strategy of the cooling fan. The correlation analysis of the changes in the two can reveal the nonlinear relationship between the load change rate and the cooling response sensitivity, providing dynamic feature input for model construction.

[0128] It should be noted that the specific implementation plan for building the initial model based on reinforcement learning can be expanded in multiple dimensions according to system requirements. In the algorithm selection dimension, deep Q learning based on value iteration, Actor-Critic framework based on policy gradient, or hybrid learning method combined with model predictive control may be adopted. In the state space design dimension, the joint time series characteristics of real-time load parameters and control parameters can be encoded into a high-dimensional vector, or key features can be extracted through dimensionality reduction technology. In the reward function design dimension, multi-objective optimization indicators such as temperature stability, energy economy, equipment life loss rate and noise control can be integrated. For example, in a liquid-cooled server scenario, the system may use the correlation between coolant flow rate, pump power and load parameters as state inputs, and design a composite reward function that includes thermal balance rewards and pump wear penalties. This application does not make specific restrictions on this.

[0129] It should be noted that there are multiple technical paths for correlation modeling of real-time parameter changes. In the time correlation dimension, lagged cross-correlation analysis, dynamic time warping (DTW) algorithm or long short-term memory network (LSTM) can be used to capture timing dependencies. In the causal reasoning dimension, Granger causality test, Bayesian network or structural equation model may be applied to analyze the driving mechanism of load change on heat dissipation control. In the feature fusion dimension, the frequency domain transform can be combined to extract the load fluctuation period characteristics, wavelet analysis can be used to capture the control parameter mutation points, or the attention mechanism can be used to focus on the key correlation period. For example, in heterogeneous computing tasks, the system may perform multi-scale correlation analysis on the high-frequency fluctuation characteristics of the GPU tensor core utilization and the liquid cooling pump flow regulation strategy, which is not specifically limited in this application.

[0130] It should be noted that the training process of the reinforcement learning model can adjust the optimization strategy according to the actual scenario. In the exploration strategy dimension, the ε-greedy algorithm can be used to balance the use of experience and the exploration of unknown strategies, the curiosity-driven intrinsic reward mechanism, or the priority experience replay technology. In the model update dimension, synchronous parameter updates, asynchronous distributed training, or privacy-preserving training based on federated learning can be implemented. In the stability control dimension, it is necessary to design target network delay updates, gradient clipping, or entropy regularization constraints to prevent policy oscillations. For example, in a modular server cluster, the system may adopt a distributed reinforcement learning framework so that the cooling strategy of each computing node can be optimized independently and the global experience pool can be shared through the parameter server. This application does not make specific restrictions on this.

[0131] The dynamic correlation model constructed through reinforcement learning can effectively capture the complex nonlinear relationship between load changes and heat dissipation response, breaking through the static limitations of traditional threshold control strategies. Through a real-time interactive learning mechanism, the system adaptively adjusts the allocation strategy of heat dissipation resources to achieve a dynamic balance between temperature stability and energy efficiency. The model's multi-objective reward function design strengthens the comprehensive optimization capabilities of equipment life and operating noise, while the diversified feature modeling methods enhance the model's generalization to different hardware architectures and load patterns. In addition, the introduction of distributed training and privacy protection mechanisms provides a secure and reliable technical foundation for the optimization of collaborative heat dissipation strategies across nodes and computer rooms.

[0132] As an optional solution, before inputting the target load parameters into the target model and generating the target control parameters, the above method further includes:

[0133] Acquire historical load parameters, historical control parameters, historical temperature parameters, and predicted temperature parameters, wherein the predicted temperature parameters represent temperature values predicted by the baseboard management controller based on the historical control parameters;

[0134] The deviation between historical temperature parameters and predicted temperature parameters is used as the loss function, and the initial model is regressed and optimized based on historical load parameters and historical control parameters to obtain the target model.

[0135] Optionally, in embodiments of the present application, the aforementioned historical temperature parameters may include, but are not limited to, thermodynamic state data sets collected during the historical operation of the cooling system and associated hardware, specifically including temporal variations in absolute temperature values, spatial gradient distribution characteristics, and statistical indicators of heat conduction rate. For example, in an air-cooled server cabinet, historical temperature parameters may include the surface temperature curve of the processor package, the local hot spot distribution map of memory module particles, and temperature fluctuation records at the chassis inlet and outlet, collected through the fusion of a distributed thermocouple array and an infrared thermal imaging module.

[0136] Optionally, in embodiments of the present application, the predicted temperature parameters may include, but are not limited to, a set of expected temperature values calculated by the baseboard management controller based on the heat dissipation control strategy and thermodynamic model, specifically manifested as temperature change trend predictions within a future time window, thermal shock risk probability assessments, and temperature balance simulation results. For example, in an intelligent heat dissipation control scenario, the predicted temperature parameters may be generated using a finite element thermal simulation algorithm, including predicted peak values for the CPU core temperature under different fan speed strategies, a cloud diagram of the steady-state temperature distribution of the memory module, and an estimated curve of the thermal resistance change of the heat sink.

[0137] It should be noted that there are various implementation forms for the acquisition and integration of historical parameters. In the data source dimension, it may include operation and maintenance logs stored in local persistent storage, cluster history records in distributed databases, or standardized data sets exported by third-party monitoring platforms. In the time span dimension, single-day high-frequency sampling data, monthly periodic operation data, or multi-year equipment life cycle data can be selected. In the preprocessing dimension, it may involve missing value interpolation, outlier removal, data normalization, or interaction term construction in feature engineering. For example, in a multi-computer room collaborative heat dissipation scenario, the system may integrate the ambient temperature compensation parameters and load pattern characteristics of servers in different geographical areas to form cross-domain joint training samples, which is not specifically limited in this application.

[0138] Optionally, in an embodiment of the present application, the aforementioned loss function may include, but is not limited to, a mathematical evaluation index that measures the deviation between the predicted temperature and the actual temperature, and may be specifically implemented as a mean square error function, a weighted sum of absolute errors function, or a multi-objective optimization function that includes a temperature change rate constraint. For example, during model training, the loss function may be designed as the dynamic time-warped distance between historical temperature parameters and predicted temperature parameters, combined with a heat dissipation efficiency penalty term and a control parameter smoothness constraint to achieve a balanced optimization of temperature prediction accuracy and control stability.

[0139] Optionally, in an embodiment of the present application, the above-mentioned regression optimization may include but is not limited to a technical process of adjusting model parameters through an iterative algorithm to minimize the loss function, specifically using a gradient descent method, random forest regressor hyperparameter tuning, or parameter posterior distribution estimation under a Bayesian probability framework. For example, in nonlinear heat dissipation system modeling, regression optimization may adjust the neural network weight matrix through an adaptive moment estimation algorithm (Adam), while combining an early stopping mechanism to prevent overfitting, ultimately making the model output temperature prediction value close to the statistical distribution of the actual measured value.

[0140] Optionally, in embodiments of the present application, the target model may include, but is not limited to, a heat dissipation control decision model that has undergone regression optimization and is capable of practical deployment. Specifically, it may be a mathematical framework capable of analyzing the nonlinear relationship between load characteristics and heat dissipation efficiency, predicting temperature trends under the influence of multivariable coupling, and generating an optimal control strategy. For example, in a heterogeneous computing scenario, the target model may be deployed as a time series prediction network that includes an attention mechanism, capable of simultaneously handling the combined effects of GPU computing load surges, memory bandwidth bottlenecks, and liquid cooling circuit delays, and outputting an optimized control parameter sequence that balances heat dissipation efficiency and energy consumption.

[0141] It should be noted that the method for constructing the loss function can be expanded in multiple dimensions according to the differences in optimization objectives. In the error type dimension, different calculation methods such as absolute error, relative error or quantile error may be used. In the weight strategy dimension, a higher penalty coefficient can be given to the temperature deviation of key components, or the balance weight of heat dissipation efficiency and energy consumption can be dynamically adjusted according to the business scenario. In the regularization dimension, L2 norm constraints may be introduced to prevent parameter overfitting, smoothing terms for the rate of change of control parameters may be added, or prior knowledge of physical thermodynamic equations may be embedded. For example, in precision computing equipment, the loss function may impose an exponential penalty on the deviation of the core temperature of the processor from the standard, while imposing linear constraints on the frequent fluctuations in the fan speed. This application does not make specific restrictions on this.

[0142] It should be noted that there are multiple technical options for the implementation path of regression optimization. In the optimization algorithm dimension, the classical gradient descent method, the quasi-Newton method, or the genetic algorithm based on population intelligence can be used. In the training strategy dimension, full batch data training, small batch stochastic gradient descent, or online streaming learning can be implemented. In the model architecture dimension, it is possible to combine the integrated learning framework to stack multiple base models, use adversarial generative networks to enhance data diversity, or introduce transfer learning to reuse pre-trained parameters. For example, in the lightweight deployment scenario of edge computing nodes, regression optimization may use knowledge distillation technology to migrate the control strategy of complex models to streamlined models, which is not specifically limited in this application.

[0143] By driving model regression optimization based on the deviation between historical parameters and predicted temperature, this application effectively improves the physical rationality and environmental adaptability of the heat dissipation control strategy. By integrating multi-dimensional historical operating data, the target model can accurately capture the complex nonlinear relationship between load characteristics, control parameters and temperature response, significantly reducing the hysteresis effect of the traditional single temperature feedback mechanism. The system utilizes the multi-objective optimization characteristics of the loss function to achieve Pareto optimality of energy efficiency while ensuring heat dissipation safety, avoiding control strategy oscillations caused by local optimization. The dynamic adjustment mechanism in the regression optimization process enables the model to continuously adapt to long-term influencing factors such as hardware aging and environmental changes, providing reliable technical support for the intelligent upgrade of server cooling systems.

[0144] As an optional solution, the above method further includes:

[0145] In response to the functional component starting to execute the preset task, obtaining the expected load parameters and the change time point;

[0146] Input the expected load parameters into the target model to generate the expected control parameters;

[0147] In response to reaching the change time point, the expected control parameter is sent to the heat dissipation component, so that the heat dissipation component operates according to the expected control parameter.

[0148] Optionally, in an embodiment of the present application, the aforementioned functional components initiating the execution of preset tasks may include, but are not limited to, triggering specific business process operations by the server's internal computing module in response to scheduling instructions, specifically manifested as processor cores activating computing threads, storage controllers loading data sets, or network interface cards initializing data transmission channels. For example, in a distributed machine learning training scenario, a functional component may initiate parallel matrix operations on a GPU cluster, batch data preloading into a memory buffer, and parameter synchronization on a high-speed interconnect network in response to instructions assigned by a task scheduler, accompanied by phased changes in power consumption and thermal load.

[0149] Optionally, in an embodiment of the present application, the aforementioned expected load parameters may include, but are not limited to, forward-looking quantitative indicators of resource requirements during the execution of a preset task, specifically including a processor instruction throughput prediction value, a memory bandwidth occupancy estimation curve, and a spatiotemporal distribution model of storage device read and write operations. For example, in a high-performance computing task, the expected load parameters may be constructed based on a historical task feature library, including floating-point operation peak estimation during fluid dynamics simulation, a memory page swap frequency model during the finite element analysis phase, and a storage bandwidth demand prediction during the result output period, with a multi-dimensional load feature vector generated through a task decomposition algorithm.

[0150] Optionally, in an embodiment of the present application, the above-mentioned change time point may include but is not limited to the critical moment when the resource demand undergoes a phased transition during the execution of the preset task, which is specifically calibrated by the task decomposition logic and the hardware performance model. The basis for determining this time point may include the start signal of the computationally intensive stage, the moment when the data dependency is released, or the process branch switching condition triggered by an external event. For example, in a real-time stream processing task, the change time point may correspond to the trigger period of the window aggregation calculation, the threshold for the data backpressure mechanism to take effect, or the moment when the state synchronization is completed after fault recovery.

[0151] It should be noted that there are many ways to implement the prediction method of expected load parameters and the logic for determining the time point of change. In the dimension of prediction data source, it may be based on metadata parsing of task description files, pattern matching of historically similar tasks, or trend extrapolation of real-time resource monitoring data. In the dimension of time point determination, fixed cycle phase division, dynamic load threshold triggering, or topological sorting of task dependency graphs can be used. In the dimension of parameter modeling, it can be combined with statistical regression analysis, physical performance simulation, or machine learning time series prediction. For example, in a cloud computing elastic scaling scenario, the system may predict load fluctuations through the resource request pattern of the container orchestrator, and determine the key stage switching time points in combination with microservice call chain analysis. This application does not make specific restrictions on this.

[0152] Optionally, in embodiments of the present application, the aforementioned expected control parameters may include, but are not limited to, a cooling resource pre-allocation strategy generated based on predicted load characteristics, specifically manifested as a fan speed adjustment plan, a liquid cooling circuit flow control sequence, or a semiconductor refrigeration chip power allocation scheme. For example, in heterogeneous computing tasks, the expected control parameters may include fan acceleration gradients during the full-load computing phase of a GPU cluster, a liquid cooling pump flow step-up strategy during the in-memory database index building phase, and a dynamic adjustment curve for heat sink contact pressure during the network data encryption phase, thereby forming a control instruction set that matches the task phase.

[0153] It should be noted that the process of generating expected control parameters from the target model can be optimized for multiple objectives based on system constraints. In the model input dimension, it is possible to integrate task priority weights, heat dissipation component performance attenuation coefficients, and computer room ambient temperature compensation factors. In the control strategy dimension, multiple objectives such as maximizing heat dissipation efficiency, minimizing energy consumption, and balancing equipment life can be balanced. In the parameter form dimension, it is possible to generate discrete step-by-step control instructions, continuous smooth adjustment curves, or parallel control schemes with fault-tolerant redundancy. For example, in a green data center scenario, the model may combine a time-sharing optimization strategy for electricity price fluctuations to dynamically adjust the energy consumption allocation ratio of the cooling system during the task execution phase. This application does not make specific restrictions on this.

[0154] It should also be noted that the transmission and execution mechanism of the control parameters can be designed differently according to the cooling system architecture. In the dimension of instruction synchronization, global clock precision triggering, event-driven immediate execution, or dynamic advance compensation based on load status may be adopted. In the dimension of transmission protocol, it can be adapted to the IPMI standard instruction set, custom hardware driver interface, or real-time communication framework based on time-sensitive network. In the dimension of execution verification, it is necessary to implement instruction readback verification, closed-loop feedback of execution effect, and abnormal state rollback mechanism. For example, in a modular server design, the system may implement parallel control instruction distribution of the cooling module through a hot-swappable management bus, and use hash verification to ensure transmission integrity. This application does not make specific restrictions on this.

[0155] Through the load prediction and heat dissipation pre-control mechanism before task execution, this application significantly improves the response speed and resource utilization efficiency of the cooling system. The system can predict the cooling requirements of each stage of the computing task in advance, generate a control strategy that accurately matches the temporal and spatial distribution of the load characteristics, and effectively avoid local overheating or excessive cooling problems caused by traditional hysteresis control. The dynamic optimization allocation of expected control parameters realizes the coordinated scheduling of cooling resources and computing tasks, reducing ineffective energy consumption while ensuring stable operation of the equipment. The integrated design of multi-dimensional parameter modeling and diversified execution mechanisms enhances the system's adaptability to different task types, hardware configurations and environmental conditions, and provides a reliable technical path for the forward-looking control of intelligent cooling systems.

[0156] As an optional solution, the above method further includes:

[0157] Receiving real-time control parameters generated by the heat dissipation component based on the real-time temperature parameters sent by the baseboard management controller, and synchronously obtaining real-time load parameters of the functional components;

[0158] Based on the establishment of a temporal correlation between real-time load parameters and real-time control parameters, an initial model is constructed through reinforcement learning, wherein the initial model represents the dynamic mapping relationship between load parameters and control parameters;

[0159] Inputting historical load parameters into the initial model to generate historical control parameters, and sending the historical control parameters to the heat dissipation component to trigger it to adjust its working state according to the historical control parameters;

[0160] Detecting the temperature of the heat dissipation component, determining historical temperature parameters, and determining predicted temperature parameters based on the historical control parameters through the baseboard management controller;

[0161] The deviation between historical temperature parameters and predicted temperature parameters is used as the loss function to perform regression optimization on the initial model and generate the target model;

[0162] Obtaining target load parameters generated by functional components during operation, wherein the target load parameters include real-time collected computing utilization, data throughput, and communication bandwidth indicators;

[0163] Input the target load parameters into the target model to generate target control parameters;

[0164] Sending the target control parameters to the heat dissipation component to trigger it to adjust its working state according to the target control parameters;

[0165] When a functional component starts a preset task, it obtains the expected load parameters and change time points, and generates the expected control parameters through the target model;

[0166] In response to sending the expected control parameter to the heat dissipation component, adjusting the working state according to the expected control parameter at the change time point is triggered.

[0167] Optionally, in an embodiment of the present application, the above-mentioned timing correlation relationship may include, but is not limited to, describing the dynamic causal relationship between the load parameters and the control parameters over time, specifically by revealing the corresponding laws between the two in terms of delay, phase, and fluctuation amplitude through time series analysis technology. For example, when a server is running high-concurrency tasks, the time lag relationship between the sudden increase in CPU utilization and the fan speed increase instruction, and the correlation between the periodic fluctuation of memory bandwidth occupancy and the flow adjustment amplitude of the liquid cooling pump, can all constitute the basic data for timing correlation analysis, which is used to model the dynamic coupling mechanism of load and heat dissipation.

[0168] It should be noted that the synchronous acquisition method of real-time parameters can be adjusted according to the differences in system architecture. In the time synchronization dimension, hardware interrupt triggered acquisition, software timed polling or precise alignment based on the Network Time Protocol (NTP) may be used. In the data integration dimension, memory mapping fast access, distributed message queue buffering or time series database storage can be implemented. In the exception handling dimension, it is necessary to design a sensor fault degradation strategy, a data verification and retransmission mechanism or an outlier dynamic correction algorithm. For example, in an edge computing node, the system may capture load mutation events through the interrupt service routine of the real-time operating system (RTOS) and align them with the feedback signal of the heat dissipation controller at the microsecond level. This application does not make specific restrictions on this.

[0169] It should also be noted that there are many ways to implement the design of the reinforcement learning framework. In the state space definition dimension, it may include original load parameters, sliding window statistical features, or energy distribution after frequency domain transformation. In the action space design dimension, discrete speed gear selection, continuous speed percentage adjustment, or hybrid control instruction combination can be set. In the reward function construction dimension, it is necessary to balance multi-objective optimization requirements such as temperature stability, energy efficiency, equipment life loss, and noise control. For example, in a modular server scenario, the model may encode the load of each computing unit independently as a state vector, and design a hierarchical action space to achieve coordinated control of multiple cooling modules. This application does not make specific restrictions on this.

[0170] In addition, the application method of historical parameters can be flexibly configured according to training needs. In the data screening dimension, importance sampling may be used to prioritize the replay of high-deviation samples, uniform sampling may be used to ensure data balance, or a progressive difficulty improvement strategy based on curriculum learning. In the model update dimension, offline batch training, online incremental learning, or distributed parameter aggregation under federated learning can be implemented. In the optimization target dimension, it is possible to minimize temperature deviation, control parameter volatility, and mechanical loss of heat dissipation components at the same time. For example, in the green data center scenario, the system may dynamically weight energy consumption-related incentives in combination with carbon emission trading data to guide the model to generate control strategies that meet environmental protection requirements. This application does not make specific limitations on this.

[0171] Finally, there are multiple optimization paths for the generation and implementation of expected control strategies. In the time dimension, the control sequence of the complete task cycle can be generated in advance, the rolling time window strategy can be generated in stages, or the pre-generated plan can be adjusted dynamically in real time. In the fault tolerance dimension, it is necessary to design a backup strategy cache, real-time monitoring of execution effects, and an abnormal state fallback mechanism. In the collaborative control dimension, it is possible to integrate the switching logic of the main and standby cooling modules, the linkage rules of the multi-level cooling system, and the cross-device cooling resource sharing strategy. For example, in an immersion liquid cooling system, the expected control parameters may include the activation threshold of the phase change material, the dynamic switching strategy of the coolant circulation path, and the linkage adjustment instructions of the external cooling tower. This application does not make specific restrictions on this.

[0172] The time-series correlation model constructed through reinforcement learning can effectively capture the dynamic coupling laws of load fluctuations and heat dissipation responses, breaking through the limitations of traditional static control strategies. The system uses a collaborative mechanism of historical data playback and online optimization to continuously improve the model's adaptability to complex load patterns, achieving simultaneous optimization of heat dissipation control accuracy and energy efficiency. The pre-generation mechanism of expected control parameters significantly reduces the risk of temperature fluctuations in the early stages of task execution, ensuring that key components remain in the optimal operating temperature range when the load suddenly changes. Multi-dimensional parameter fusion and fault-tolerant design enhance the robustness of the system, enabling it to adapt to different heat dissipation architectures and environmental interference, providing a complete technical closed loop for the dynamic optimization of intelligent heat dissipation systems.

[0173] The following is a further explanation of this application with reference to specific examples:

[0174] Traditional server fan control methods typically rely on the baseboard management controller (BMC) to collect hardware temperature sensor data and adjust the fan speed based on preset temperature thresholds. However, this method has the following problems:

[0175] Response lag: Temperature changes are usually the result of changes in component load and resource utilization. The temperature sensor takes a certain amount of time to reflect the actual heat accumulation. In addition, the BMC generally uses polling to obtain temperature. Due to the performance limitations of the BMC chip, it will cause fan control response delays.

[0176] High energy consumption: To address potential high-temperature risks and avoid serious problems such as temperature overshoot, traditional methods often use higher fan speed redundant settings, thereby increasing unnecessary energy consumption.

[0177] Lack of refined management: Relying solely on temperature data cannot fully reflect the actual workload and cooling requirements of components, making it difficult to achieve precise fan control.

[0178] Therefore, there is an urgent need for a fan control device and method that can respond more quickly to component load changes, improve regulation efficiency, and reduce energy consumption.

[0179] This application consists of two mutually reinforcing parts: system architecture and OS application. The system architecture is responsible for the design of the overall framework and the coordination of hardware resources, while the OS application achieves precise monitoring and dynamic regulation of resource utilization through optimization at the operating system level and in two models. The deep integration of these two parts ensures this application's superior performance in improving efficiency and reducing energy consumption.

[0180] The system framework consists of the OS application layer, drivers, components that need to be monitored, CPLD, fan module, and BMC module.

[0181] OS applications are mainly used to call system resources and model calculations. In the early stage of OS applications, the load rate of each component and the fan speed transmitted by the BMC module are collected to generate a corresponding load rate and fan speed model through machine learning. The regression optimization model is then used to predict the component load rate and control the fan speed to achieve the purpose of precise control.

[0182] The OS application layer is responsible for invoking and managing system resources, executing model calculations, and interacting with hardware modules. The OS application layer collects real-time load data from each component and fan speed information from the BMC module, using machine learning to generate a relationship between load rate and fan speed.

[0183] Driver: Provides an interface between the hardware and the operating system, ensuring that OS applications can access system resources and effectively communicate with the fan control hardware.

[0184] Components that need to be monitored: mainly include system resources such as CPU, memory, disk I / O, network bandwidth, etc., monitor their utilization information in real time, and provide this data to OS applications for analysis and processing.

[0185] CPLD and fan module: As a hardware control unit, the CPLD is responsible for receiving fan speed adjustment commands from the OS application and achieving precise temperature control by controlling the fan speed.

[0186] BMC module: As part of a traditional fan control system, the BMC is responsible for monitoring hardware temperature through temperature sensors and transmitting fan speed information to OS applications for decision-making assistance and feedback.

[0187] like Figure 3 As shown, the specific implementation of this application mainly includes the following steps:

[0188] S1, start OS;

[0189] S2, load PCIE and CPLD access drivers;

[0190] In S3, the OS application layer regularly collects component load information, including CPU, memory, disk I / O, and network bandwidth utilization, using system monitoring tools such as top, vmstat, and iostat. Simultaneously, the BMC module transmits real-time fan speed data to the operating system. The OS application uses this data as input for further analysis and calculations.

[0191] In the initial system phase, the OS application uses machine learning algorithms (such as linear regression and support vector machines) to build a model that models the relationship between load factor and fan speed using collected component load data and fan speed information provided by the BMC. This model maps the relationship between changes in load factor and required fan speed, and further improves prediction accuracy through regression optimization.

[0192] In S5, using the optimized regression model, the OS application can predict system load changes over time based on real-time load rate data. Based on these predictions, the OS application controls the CPLD module to send precise fan speed adjustment commands to the fan module. Specifically, when a significant load increase is predicted, the OS application adjusts the fan speed in advance to prevent heat dissipation issues caused by temperature lag.

[0193] The fan module adjusts the fan speed based on instructions from the CPLD and transmits the current fan speed status back to the BMC module. The BMC module detects the actual temperature of each component using a temperature sensor and feeds the result back to the OS application. If there is a significant deviation between the actual temperature and the predicted temperature, the OS application further optimizes the model for more precise adjustment. As the system runs, the OS application continuously accumulates data on resource utilization and fan speed, using this data to update and optimize the machine learning model online, thereby continuously improving the accuracy of load rate prediction and fan control. Through long-term training and optimization, the model can more accurately respond to changes in different types of loads, improving system response speed and energy saving.

[0194] like Figure 4 As shown, the fan speed control based on BMC in the related art depends on the component temperature, and the component temperature is a manifestation of the component load change. Therefore, the BMC control is delayed behind the component load change; while the fan speed control based on the server hybrid model algorithm of the present application can predict the component load change in advance, and then control the fan in advance to achieve the purpose of precise control.

[0195] As system load patterns become more complex and diverse, traditional static fan adjustment methods are unable to respond to component load changes in a timely manner due to architectural design reasons, making it difficult to cope with rapidly changing demands in high-load scenarios. Therefore, this application innovatively proposes a device and method for the server OS to automatically adjust the fan speed based on load adjustment and predicted load changes. This device can directly and dynamically adjust the adjustment strategy based on different time periods, load changes, hardware performance, and ambient temperature changes, ultimately achieving a method that abandons the traditional BMC design of regulating fans based on temperature and directly allows the OS to regulate fan speed based on the load of each component. Its advantages are as follows:

[0196] Improved real-time fan adjustment: By adjusting fan speed directly based on component load, temperature control lag is avoided. The operating system can respond more quickly to load changes, especially under high load conditions and frequent load changes, and can make fan adjustments in advance to ensure timely and effective cooling.

[0197] Reduced energy waste: Because the fan speed is adjusted based on actual load demand, the system can avoid ineffective high-speed operation and reduce energy waste. Compared with traditional BMC control, the BMC has a certain delay in obtaining temperature. In a complex component utilization environment, the fan speed will be increased to ensure normal system operation, resulting in significant power consumption.

[0198] Noise reduction: By precisely adjusting the fan speed, the high-speed operation time of the fan is reduced, thereby effectively reducing noise pollution.

[0199] This application provides a single-board optimal fan control device and method under a hybrid model for server application. It is completely different from the traditional BMC temperature monitoring and fan control method, and is combined with the training model of the OS application layer. It has the advantages of fast response, precise adjustment, and high energy efficiency. It is suitable for high-load server environments such as high-performance computing and big data processing. It can effectively improve the energy efficiency of the server and reduce operation and maintenance costs. In theory, it can be applied to all servers and significantly improve fan control efficiency and reduce power consumption.

[0200] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus the necessary general hardware platform, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method.

[0201] The embodiment of the present application also provides a control device for a heat dissipation device, such as Figure 5 As shown, the device includes:

[0202] An acquisition module 502 is used to acquire target load parameters corresponding to functional components of a target device;

[0203] A generation module 504 is configured to input the target load parameters into a target model to generate target control parameters, wherein the target model represents a model obtained by training an initial model based on historical load parameters corresponding to the functional components and historical control parameters corresponding to the heat dissipation components of the target device, and the initial model represents a relationship model between the control parameters corresponding to the heat dissipation components and the load parameters of the functional components;

[0204] The sending module 506 is configured to send the target control parameters to the heat dissipation component so that the heat dissipation component operates according to the target control parameters.

[0205] As an optional solution, the above-mentioned device is also used to: input the target load parameters into the target model, and before generating the target control parameters, receive the real-time control parameters of the heat dissipation components sent by the baseboard management controller, wherein the real-time control parameters represent the control parameters generated in real time by the baseboard management controller based on the real-time temperature parameters of the heat dissipation components detected; obtain the real-time load parameters of the functional components; and establish an initial model based on the changes in the real-time load parameters and the changes in the real-time control parameters.

[0206] As an optional solution, the above-mentioned device is used to establish an initial model based on changes in real-time load parameters and changes in real-time control parameters in the following manner: continuously detecting whether the real-time load parameters have changed; when a change in the real-time load parameters is detected at a first time point, determining the real-time load parameters corresponding to the first time point as sample load parameters, and continuously detecting whether the real-time control parameters have changed; when a change in the real-time control parameters is detected at a second time point, determining the real-time control parameters corresponding to the second time point as sample control parameters; establishing an initial model based on at least one set of sample data, wherein a set of sample data includes matching sample load parameters and sample control parameters.

[0207] As an optional solution, the above-mentioned device is used to establish an initial model based on the changes in real-time load parameters and real-time control parameters in the following manner: using reinforcement learning to establish an initial model based on the changes in real-time load parameters and real-time control parameters.

[0208] As an optional solution, the above-mentioned device is also used to: input the target load parameters into the target model, and before generating the target control parameters, obtain historical load parameters, historical control parameters, historical temperature parameters and predicted temperature parameters, wherein the predicted temperature parameters represent the temperature values predicted by the baseboard management controller based on the historical control parameters; use the deviation between the historical temperature parameters and the predicted temperature parameters as the loss function, and perform regression optimization on the initial model based on the historical load parameters and historical control parameters to obtain the target model.

[0209] As an optional solution, the above-mentioned device is also used to: obtain expected load parameters and change time points in response to the functional component starting to execute a preset task; input the expected load parameters into the target model to generate expected control parameters; and send the expected control parameters to the heat dissipation component in response to reaching the change time point, so that the heat dissipation component operates according to the expected control parameters.

[0210] As an optional solution, the above device is also used to: receive real-time control parameters generated by the heat dissipation component based on the real-time temperature parameters sent by the baseboard management controller, and synchronously obtain the real-time load parameters of the functional components; establish a time series association relationship based on the real-time load parameters and the real-time control parameters, and build an initial model through reinforcement learning, wherein the initial model characterizes the dynamic mapping relationship between the load parameters and the control parameters; input the historical load parameters into the initial model to generate the historical control parameters, and send the historical control parameters to the heat dissipation component to trigger it to adjust its working state according to the historical control parameters; detect the temperature of the heat dissipation component, determine the historical temperature parameters, and determine the predicted temperature parameters according to the historical control parameters through the baseboard management controller; The deviation between the historical temperature parameters and the predicted temperature parameters is used as the loss function to perform regression optimization on the initial model and generate a target model; the target load parameters generated by the functional components during operation are obtained, wherein the target load parameters include the computing utilization, data throughput and communication bandwidth indicators collected in real time; the target load parameters are input into the target model to generate target control parameters; the target control parameters are sent to the heat dissipation component to trigger it to adjust the working state according to the target control parameters; when the functional component starts the preset task, the expected load parameters and the change time point are obtained, and the expected control parameters are generated through the target model; in response to sending the expected control parameters to the heat dissipation component, the working state is triggered to be adjusted according to the expected control parameters at the change time point.

[0211] For the description of the features in the embodiment corresponding to the control device of the heat dissipation device, reference can be made to the relevant description of the embodiment corresponding to the control method of the heat dissipation device, which will not be repeated here.

[0212] An embodiment of the present application further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any of the above-mentioned control method embodiments of the heat dissipation device.

[0213] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above-mentioned control method embodiments of the heat dissipation device when running.

[0214] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk, or an optical disk.

[0215] An embodiment of the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps in any of the above-mentioned control method embodiments of the heat dissipation device are implemented.

[0216] An embodiment of the present application also provides another computer program product, including a non-volatile computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps in any of the above-mentioned control method embodiments of the heat dissipation device.

[0217] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0218] The above is a detailed introduction to a control method and device for a heat dissipation device provided by the present application. This article uses specific examples to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core ideas of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

Claims

1. A method for controlling a heat dissipation device, characterized in that: include: Obtain target load parameters corresponding to functional components of the target device; Inputting the target load parameter into a target model to generate a target control parameter, wherein the target model represents a model obtained by training an initial model based on historical load parameters corresponding to the functional component and historical control parameters corresponding to the heat dissipation component of the target device, and the initial model represents a relationship model between the control parameters corresponding to the heat dissipation component and the load parameters of the functional component; The target control parameter is sent to the heat dissipation component, so that the heat dissipation component operates according to the target control parameter.

2. The control method of the heat dissipation device according to claim 1, characterized in that: Before inputting the target load parameters into the target model to generate the target control parameters, the method further includes: receiving a real-time control parameter of the heat dissipation component sent by a baseboard management controller, wherein the real-time control parameter represents a control parameter generated in real time by the baseboard management controller based on a real-time temperature parameter of the heat dissipation component detected; Obtaining real-time load parameters of the functional components; The initial model is established based on the change of the real-time load parameter and the change of the real-time control parameter.

3. The control method of the heat dissipation device according to claim 2, characterized in that: The establishing of the initial model based on the change of the real-time load parameter and the change of the real-time control parameter includes: Continuously detecting whether the real-time load parameter changes; In the case where a change in the real-time load parameter is detected at a first time point, determining the real-time load parameter corresponding to the first time point as a sample load parameter, and continuously detecting whether the real-time control parameter changes; In a case where a change in the real-time control parameter is detected at a second time point, determining the real-time control parameter corresponding to the second time point as a sample control parameter; The initial model is established based on at least one set of sample data, wherein the set of sample data includes the matched sample load parameters and the matched sample control parameters.

4. The control method of the heat dissipation device according to claim 2, characterized in that: The establishing of the initial model based on the change of the real-time load parameter and the change of the real-time control parameter includes: The initial model is established based on the changes in the real-time load parameters and the changes in the real-time control parameters using a reinforcement learning method.

5. The control method of the heat dissipation device according to claim 1, characterized in that: Before inputting the target load parameters into the target model to generate the target control parameters, the method further includes: Acquire the historical load parameter, the historical control parameter, the historical temperature parameter, and the predicted temperature parameter, wherein the predicted temperature parameter represents a temperature value predicted by the baseboard management controller according to the historical control parameter; The deviation between the historical temperature parameter and the predicted temperature parameter is used as a loss function, and regression optimization is performed on the initial model based on the historical load parameter and the historical control parameter to obtain the target model.

6. The control method of the heat dissipation device according to claim 1, characterized in that: The method further comprises: In response to the functional component starting to execute a preset task, obtaining expected load parameters and change time points; Inputting the expected load parameters into the target model to generate expected control parameters; In response to reaching the change time point, the expected control parameter is sent to the heat dissipation component, so that the heat dissipation component operates according to the expected control parameter.

7. The control method of the heat dissipation device according to claim 1, characterized in that: The method further comprises: receiving a real-time control parameter generated by the heat dissipation component based on the real-time temperature parameter and sent by the baseboard management controller, and synchronously acquiring a real-time load parameter of the functional component; Establishing a temporal association relationship between the real-time load parameters and the real-time control parameters, and constructing an initial model through reinforcement learning, wherein the initial model represents a dynamic mapping relationship between the load parameters and the control parameters; Inputting the historical load parameters into the initial model to generate historical control parameters, and sending the historical control parameters to the heat dissipation component to trigger it to adjust its working state according to the historical control parameters; detecting the temperature of the heat dissipation component, determining a historical temperature parameter, and determining a predicted temperature parameter according to the historical control parameter by the baseboard management controller; Using the deviation between the historical temperature parameter and the predicted temperature parameter as a loss function, performing regression optimization on the initial model to generate a target model; Acquiring the target load parameters generated by the functional components during operation, wherein the target load parameters include computing utilization, data throughput, and communication bandwidth indicators collected in real time; Inputting the target load parameters into the target model to generate target control parameters; Sending the target control parameter to the heat dissipation component to trigger it to adjust its working state according to the target control parameter; When the functional component starts a preset task, it obtains the expected load parameters and the change time point, and generates the expected control parameters through the target model; In response to sending the expected control parameter to the heat dissipation component, adjusting the working state according to the expected control parameter at the change time point is triggered.

8. A control device for a heat dissipation device, characterized in that: include: An acquisition module, used to obtain target load parameters corresponding to functional components of a target device; a generation module, configured to input the target load parameter into a target model to generate a target control parameter, wherein the target model represents a model obtained by training an initial model based on historical load parameters corresponding to the functional component and historical control parameters corresponding to the heat dissipation component of the target device, and the initial model represents a relationship model between the control parameters corresponding to the heat dissipation component and the load parameters of the functional component; The sending module is used to send the target control parameter to the heat dissipation component, so that the heat dissipation component works according to the target control parameter.

9. An electronic device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the method for controlling the heat dissipation device according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, the steps of the control method of the heat dissipation device according to any one of claims 1 to 7 are implemented.

Images