Display card component performance monitoring method, device and equipment

Abstract

The invention relates to the technical field of graphics card performance monitoring, and discloses a graphics card component performance monitoring method, device and equipment, and the method comprises the steps: setting a temperature sensor array for a GPU chip, a video memory, a power management module and a radiator of a graphics card, and carrying out the temperature data collection and preprocessing, and obtaining a temperature distribution matrix; establishing a thermal network topological structure and performing finite element discretization calculation to obtain a thermal network state space equation; partial least square regression analysis is carried out to obtain m temperature control areas; the temperature state vector and the load state vector of the temperature control area are input into a mixed Q network structure, an epsilon-greedy strategy and a Q learning updating rule are executed, and a heat dissipation control parameter sequence is obtained.By means of the method, accurate calculation of a temperature coupling coefficient and design of a temperature decoupling controller are achieved, the problem of temperature coupling between assemblies is effectively solved, and the reliability of heat dissipation control is improved. And the heat dissipation control strategy can be adaptively adjusted to adapt to different working load conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of graphics card performance monitoring, and particularly to a method, device, and equipment for monitoring the performance of graphics card components. Background Art

[0002] Traditional graphics card cooling solutions mainly rely on fixed fan speed curves and simple temperature threshold control, and cannot accurately capture the temperature coupling relationship between different components such as GPU chips, video memories, and power management modules, resulting in unsatisfactory cooling effects and affecting the overall performance and service life of the graphics card.

[0003] Currently, the commonly used graphics card temperature monitoring methods in the industry often only focus on single-point temperature data and lack a holistic analysis of the temperature distribution of multiple components. This method ignores the heat conduction effect between components of the graphics card and cannot establish an accurate thermal network model, resulting in low temperature prediction accuracy and difficult optimization of the heat dissipation control strategy. At the same time, traditional temperature control methods mostly use simple PID control or look-up table methods and cannot cope with complex temperature dynamic change processes. Summary of the Invention

[0004] The present invention provides a method, device, and equipment for monitoring the performance of graphics card components. The present invention realizes the accurate calculation of the temperature coupling coefficient and the design of the temperature decoupling controller, effectively solves the problem of temperature coupling between components, enables the heat dissipation control strategy to be adaptively adjusted, and adapts to different working load conditions.

[0005] In a first aspect, the present invention provides a method for monitoring the performance of graphics card components. The method for monitoring the performance of graphics card components includes: Setting a temperature sensor array for the GPU chip, video memory, power management module, and radiator of the graphics card to collect and preprocess temperature data, and obtaining a temperature distribution matrix; Establishing a thermal network topology structure including n nodes according to the temperature distribution matrix, and performing finite element discretization calculation on the thermal network topology structure to obtain a thermal network state space equation; Performing partial least squares regression analysis on the thermal network state space equation, calculating a temperature coupling coefficient matrix, constructing a temperature decoupling controller through pole placement method, and obtaining m temperature control regions; Inputting the temperature state vector and load state vector of the temperature control region into a hybrid Q-network structure, executing an ε-greedy strategy and Q-learning update rule, and obtaining a heat dissipation control parameter sequence.

[0006] In a second aspect, the present invention provides a device for monitoring the performance of graphics card components. The device for monitoring the performance of graphics card components includes: A temperature data acquisition module is used to collect and preprocess temperature data of a temperature sensor array set for the GPU chip, video memory, power management module, and radiator of a graphics card, and obtain a temperature distribution matrix; A discretization calculation module is used to establish a thermal network topology structure including n nodes according to the temperature distribution matrix, and perform finite element discretization calculation on the thermal network topology structure to obtain a thermal network state space equation; A regression analysis module is used to perform partial least squares regression analysis on the thermal network state space equation, calculate a temperature coupling coefficient matrix, construct a temperature decoupling controller through pole placement method, and obtain m temperature control regions; An execution module is used to input the temperature state vector and load state vector of the temperature control region into a hybrid Q-network structure, execute the ε-greedy strategy and Q-learning update rule, and obtain a heat dissipation control parameter sequence.

[0007] The third aspect of the present invention provides a graphics card component performance monitoring device, including: a memory and at least one processor, wherein instructions are stored in the memory; the at least one processor calls the instructions in the memory to enable the graphics card component performance monitoring device to execute the above-mentioned graphics card component performance monitoring method.

[0008] In the technical solution provided by the present invention, by arranging a multi-point temperature sensor array and establishing a temperature distribution matrix, the accurate monitoring of the temperatures of each component of the graphics card is realized. By using a lumped parameter thermal network model and a finite element discretization method, the heat conduction relationship between the graphics card components is accurately described, a complete thermal network state space equation is established, partial least squares regression analysis and pole placement method are introduced, the accurate calculation of the temperature coupling coefficient and the design of the temperature decoupling controller are realized, and the problem of temperature coupling between components is effectively solved; a hybrid Q-network structure is used for heat dissipation control, combining the advantages of the global value network and the local value network, realizing the intelligent optimization of the graphics card heat dissipation parameters. Through the ε-greedy strategy and the experience replay mechanism, the exploration ability and learning efficiency of the control system are enhanced, and the heat dissipation control strategy can be adaptively adjusted to adapt to different workload conditions.

[0009] Other features and advantages of the present invention will be described in the following specification, and part of them will become obvious from the specification, or be understood by implementing the present invention. The objectives and other advantages of the present invention are achieved and obtained by the structures specifically pointed out in the specification, claims, and drawings.

[0010] To make the above objectives, features, and advantages of the present invention more obvious and understandable, the following specific preferred embodiments are given, and in conjunction with the accompanying drawings, the detailed description is as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1Schematic diagram of an embodiment of the method for monitoring the performance of a graphics card component in an embodiment of the present invention; Figure 2 Schematic diagram of an embodiment of the device for monitoring the performance of a graphics card component in an embodiment of the present invention; Figure 3 Schematic diagram of an embodiment of the device for monitoring the performance of a graphics card component in an embodiment of the present invention. Detailed implementation manners

[0012] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0013] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device including a series of steps or units is not limited to the listed steps or units, but optionally further includes other steps or units not listed, or optionally further includes other steps or units inherent to these processes, methods, products, or devices.

[0014] For ease of understanding of this embodiment, a method for monitoring the performance of a graphics card component disclosed in the embodiments of the present invention will be introduced in detail first. As Figure 1 shown, this method includes the following steps: 101. Set up a temperature sensor array for the GPU chip, video memory, power management module, and radiator of the graphics card to collect and preprocess temperature data, and obtain a temperature distribution matrix; It can be understood that the execution subject of the present invention can be a device for monitoring the performance of a graphics card component, or a terminal or a server. Specifically, it is not limited here. The embodiments of the present invention will be described by taking the server as the execution subject as an example.

[0015] Specifically, the graphics card is divided into regions. In the GPU chip region, since it is the component with the most concentrated heat in the graphics card and requires a high monitoring accuracy, four temperature sensors are set in this region, which are respectively arranged at different positions on the chip surface to capture local temperature changes. Similarly, in the video memory region, high heat is generated during high-load operation, and too high a temperature of the video memory will affect the stability of data transmission. Four temperature sensors are also arranged in the video memory region to monitor the temperature change trend of the video memory module at different positions. The power management module is the power supply core of the graphics card, and its temperature change will affect the efficiency of the entire power supply system. Two temperature sensors are arranged in this region to monitor the heat generation of the power conversion module. The radiator is an important part of the graphics card temperature management, and its temperature change reflects the heat dissipation efficiency. Two temperature sensors are arranged in the radiator region to monitor whether the heat dissipation effect is uniform and efficient. By evenly arranging these temperature sensors on the key components of the graphics card, a temperature sensor array is formed. During the temperature data acquisition process, each temperature sensor samples the current temperature value at a fixed time interval and records it in the data storage unit. The collected raw temperature data is preprocessed. An outlier detection method is used to identify and eliminate temperature data that significantly exceeds the normal range. For example, if the change amplitude of a certain temperature point is abnormally large, or the temperature data differs greatly from that of adjacent sensors, it is judged as an outlier and eliminated or corrected. And the data is smoothed, for example, by the moving average method to reduce short-term fluctuations in the data to obtain more stable temperature trend data. The preprocessed temperature data is matrix-reconstructed according to the spatial distribution relationship of the sensors. Since the temperature sensors are fixedly distributed on the graphics card surface according to a certain layout, according to their spatial positions, the collected data is filled into the matrix in the corresponding order to form temperature spatial distribution data. Each matrix element corresponds to a specific spatial position, and its value represents the temperature at that position, so that the entire matrix intuitively reflects the temperature distribution on the graphics card surface. A temperature distribution function describing the temperature change trend is constructed by mathematical fitting of the temperature spatial distribution data. Through polynomial fitting by the least squares method, an optimal mathematical model is found to approximate the temperature measurement data to the greatest extent. The fitted temperature distribution function more accurately describes the temperature conditions of each region on the graphics card surface, and is not only used for the current temperature assessment, but also for predicting future temperature change trends. The temperature distribution function is substituted into the temperature field equation to solve the temperature distribution of the entire graphics card. Since the heat conduction inside the graphics card is a complex process, affected by various factors such as material properties, power consumption distribution, and heat dissipation methods, it is difficult to directly analytically solve the temperature field equation. Therefore, numerical calculation methods, such as the finite difference method, are used to discretize the continuous temperature change process, and then the temperature distribution matrix is solved by iterative calculation.Through this process, the temperature distribution of different regions of the graphics card is obtained, and how heat is transferred between various components of the graphics card is analyzed, as well as which regions have problems of local overheating.

[0016] 102. Establish a thermal network topology structure containing n nodes based on the temperature distribution matrix, and perform finite element discretization calculation on the thermal network topology structure to obtain the state space equation of the thermal network; Specifically, construct a thermal network topology structure according to the temperature distribution matrix to characterize the heat transfer relationship between different components of the graphics card. Since the temperature distribution inside the graphics card has spatial correlation, the temperature data is divided into nodes according to the physical position relationship. The key components such as the GPU chip, video memory, power management module, and radiator of the graphics card are regarded as heat capacity nodes respectively, and other non-critical regions are further divided according to the gradient of temperature change to form a thermal network topology structure containing n nodes. In this topology structure, the temperature data of each node comes from the temperature distribution matrix, and the heat capacity of each node depends on its material properties and geometric dimensions, while the heat conduction path between adjacent nodes is represented by a thermal resistance connection to describe the heat transfer between different components. Establish a heat balance equation C i (dT i / d t )=P i +Σ(T j -T i ) / R ij, to describe the temperature change process of each node. The heat transfer of the graphics card component conforms to the first law of thermodynamics, that is, the temperature change of each heat capacity node is affected by its own heat input, heat capacity size, and heat exchange with adjacent nodes. Therefore, for any node i in the network, its temperature change rate is described by the heat balance equation, where the heat capacity of node i determines the inertia of temperature change, and the heat power input it receives comes from factors such as GPU load, video memory read / write power consumption, and power module conversion loss. At the same time, this node exchanges heat with adjacent nodes through the thermal resistance path. By traversing the entire thermal network topology, a complete set of heat balance equations is established, which describes the heat transfer and temperature evolution process inside the entire graphics card. The heat balance equations are discretized in the time domain for numerical solution. Time domain discretization decomposes the temperature change process into a series of discrete time steps. In each time step, the temperature change of the node is approximately calculated according to a certain numerical method. In this way, the continuous heat balance equations are converted into heat balance equations in discrete time format. The heat balance equations after time discretization are discretized in the spatial domain by the finite element method to optimize the stability and efficiency of the calculation. The finite element method decomposes the continuous temperature field into a finite number of discrete elements. The temperature change within each element is described by a specific interpolation function and is calculated jointly through the interaction between nodes. In the finite element discretization process of the thermal network, the temperature of each node is the main variable to be solved, while parameters such as heat capacity, heat power, and thermal resistance are organized into corresponding finite element coefficient matrices to form a discretized finite element equation set of the thermal network. The finite element equation set of the thermal network is reconstructed in matrix form to be organized into a standard state space expression form. All node temperatures are organized into a temperature state vector, while all heat capacities, heat power inputs, and thermal resistances respectively form the corresponding coefficient matrices. In this way, the finite element equation set of the thermal network is converted into a standardized state equation, making it more suitable for subsequent numerical analysis and control optimization. In the state space expression form, the temperature change inside the graphics card is described by matrix operations, making the overall calculation process clearer and combining it with existing control theory methods to optimize the thermal management strategy of the graphics card. The parameters of the thermal network state equation are calibrated, combined with experimental data, such as measuring the temperature response under different power inputs through thermal sensors, and adjusting parameters such as thermal resistance, heat capacity, and heat power input through optimization methods to make the calculation results match the actual measurement results, obtaining the thermal network state space equation.

[0017] The heat balance equations are discretized in the time domain for numerical calculation. Since the heat conduction process is a dynamic process that varies with time, directly solving the continuous differential equations is relatively complex in actual calculations. The numerical difference method is used to convert the time-continuous expression into a time-discrete form. The heat balance equations are expanded by difference in the time domain, and the time derivative is approximated by the finite difference method, so that the temperature change within each time step can be described by numerical calculation. During this process, the temperature terms at different time steps are discretely represented, and a system of equations in time-discrete form is constructed through the difference scheme. The terms of the time-discrete form equations are combined and sorted out to eliminate redundant terms and optimize the calculation structure, obtaining a normalized system of time-discrete equations. The condition number analysis of the coefficient matrix of the normalized time-discrete equations is carried out. The condition number of the coefficient matrix reflects the numerical stability of the equations. A matrix with a large condition number is prone to the accumulation of numerical errors during the solution process, affecting the reliability of the calculation results. By calculating the condition number, it is evaluated whether the heat balance equations after time discretization are suitable for spatial discretization, and the time step is adjusted or numerical preprocessing techniques are adopted when necessary to improve the solvability of the equations. After completing the condition number analysis, the time-discretized heat balance equations are obtained. The time-discretized heat balance equations are discretized by the finite element method in the spatial domain to accurately solve the temperature distribution in the graphics card thermal network model. The computational domain is meshed. To ensure the computational accuracy and adapt to complex geometric structures, tetrahedral meshes are used for spatial discretization, so that the heat transfer relationship between different components of the graphics card can be accurately modeled. After the meshing is completed, piecewise linear basis functions are constructed to interpolate and calculate the temperature distribution within each mesh element. The basis function is linear within the local element, thus better approximating the temperature change while maintaining the stability of the calculation. Through the construction of the basis function, the stiffness matrix and mass matrix of the local element are calculated, where the stiffness matrix is used to describe the interaction between nodes during the heat conduction process, and the mass matrix reflects the distribution of heat capacity. The local element stiffness matrix and mass matrix are globally assembled to form a complete finite element system of equations. During the assembly process, an assembly matrix is established according to the node number correspondence to ensure that the calculation data between different elements can be correctly merged. Since the finite element calculation of the thermal network involves a large number of computational nodes and coefficient matrices, the sparse matrix storage technology is adopted to reduce the storage overhead and improve the computational efficiency. The CSR format is used to store the global stiffness matrix and mass matrix, which can effectively compress the storage space and improve the speed of matrix operations. Through the above steps, the target finite element system of equations is obtained, which describes the temperature field distribution inside the graphics card and is suitable for large-scale numerical calculations. To solve the target finite element system of equations, an efficient numerical method is adopted. Among them, the preconditioned conjugate gradient method is a method for solving large-scale sparse matrix equations, which can effectively improve the convergence speed and reduce the computational cost.During the PCG solution process, a suitable preconditioning matrix is selected to improve the numerical stability of the equations. The incomplete Cholesky decomposition is used as the preconditioning matrix. This method reduces the condition number of the matrix while maintaining the computational efficiency, thereby improving the convergence rate of the iterative solution. During the specific iterative calculation process, appropriate convergence criteria are set to ensure the accuracy of the calculation results. The iterative convergence criterion is set to the residual norm being less than 10. -6 , that is, when the calculated residual is less than this threshold, the iterative process terminates, and it is considered that a sufficiently accurate solution has been obtained. Through the PCG solution process, the numerical solution of the finite element equations of the thermal network is finally obtained, and the temperature distribution results of different nodes inside the graphics card are obtained.

[0018] 103. Perform partial least squares regression analysis on the state space equation of the thermal network, calculate the temperature coupling coefficient matrix, construct a temperature decoupling controller through the pole placement method, and obtain m temperature control regions; Specifically, analyze the thermal network state - space equation to extract the temperature transfer relationship between components inside the graphics card. Since the thermal network state - space equation is a dynamic system with multiple nodes, and its state matrix describes the temperature changes between different nodes, decompose this system matrix to obtain the temperature state - transfer matrix. This matrix reflects the degree of mutual influence between each temperature node inside the graphics card, especially the heat transfer path between different components, the influence of heat capacity on temperature changes, and the limitation of thermal resistance on the heat diffusion rate. By decomposing the state matrix, effectively extract the thermal coupling structure inside the graphics card. Input the temperature state - transfer matrix into the NIPALS algorithm to calculate the principal components, so as to reduce the dimension of the data and extract key features. Through principal component analysis, map the high - dimensional data to a low - dimensional feature space to extract the reduced - dimension feature matrix that can best represent the dynamic characteristics of the system. The NIPALS algorithm gradually extracts the main components in the data through an iterative calculation method and removes redundant information, so that the finally obtained reduced - dimension feature matrix accurately characterizes the temperature state changes inside the graphics card. Perform partial least - squares regression calculation on the reduced - dimension feature matrix to analyze the temperature coupling relationship between components of the graphics card. During the partial least - squares regression process, the input variable is the reduced - dimension feature matrix, and the output variable is the actual observed value of the temperature state. Through least - squares regression analysis, establish the temperature mapping relationship between components and calculate the regression coefficient matrix, which reflects the degree of thermal coupling between different components. Conduct a significance test on the regression coefficients to eliminate statistically insignificant coupling terms and obtain the final temperature coupling coefficient matrix, which describes the temperature transfer intensity between each component of the graphics card. Based on the temperature coupling coefficient matrix, analyze the temperature correlation degree between components of the graphics card and divide multiple initial control regions accordingly. The temperature correlation degree reflects the degree of thermal coupling between different components, that is, when the temperature of a certain component changes, the degree to which the temperatures of other components are affected. To determine the initial control regions, perform a clustering analysis on the temperature coupling coefficient matrix to identify combinations of components with a high degree of thermal coupling as a control region, so that the components within each control region have strong thermal connections, while the thermal influence between different control regions is weak. In this way, decompose the thermal management problem of the entire graphics card into multiple smaller sub - problems to improve the control accuracy and calculation efficiency. Construct a state - feedback controller based on multiple initial control regions, so that the closed - loop poles of the system are located in the left - half plane, thereby ensuring the stability of the system and optimizing the temperature regulation performance. During the pole - placement process, comprehensively consider factors such as the system's response speed, temperature control accuracy, and steady - state error to ensure that the final temperature controller can effectively reduce the temperature fluctuations of components of the graphics card and reduce local overheating phenomena. The temperature decoupling controller constructed by the pole - placement method realizes independent temperature management within each control region, while avoiding thermal interference between different regions and improving the overall heat dissipation efficiency. Conduct a Lyapunov stability analysis on the temperature decoupling controller.Lyapunov stability analysis is a mathematical method used to verify whether a dynamic system can remain stable under different initial conditions. A Lyapunov function is constructed and the trend of this function over time is analyzed to ensure that all state variables can converge to the equilibrium point within a finite time. In the temperature control problem, Lyapunov stability analysis is used to verify whether the temperature controller can stably regulate the temperatures of various components of the graphics card under external disturbances, thus ensuring the long-term stable operation of the system. After completing the stability analysis, the parameter settings of each control region are optimized to ensure that precise temperature regulation can be achieved in each initial control region, and finally m temperature control regions are obtained.

[0019] 104. Input the temperature state vector and load state vector of the temperature control region into the hybrid Q-network structure, execute the ε-greedy policy and Q-learning update rule to obtain a sequence of heat dissipation control parameters.

[0020] Specifically, the temperature state vector and load state vector of the temperature control region are concatenated and encoded to construct a state encoding vector. Since the thermal management of the graphics card is affected by multiple factors, including the current temperature distribution, the load conditions of the GPU and video memory, and the power consumption of the power management module, this information is integrated into a high-dimensional feature vector for input into the reinforcement learning network for optimization. The state encoding vector is input into the hybrid Q-network structure for calculation. The hybrid Q-network structure consists of a global value network and a local value network. The global value network is used to evaluate the overall heat dissipation control strategy, while the local value network independently models each temperature control region to improve the flexibility and accuracy of control. The global value network adopts a three-layer fully connected layer structure, where each layer contains multiple neurons and uses the ReLU activation function to enhance the network's non-linear expression ability, thereby more accurately estimating the overall heat dissipation control value of the graphics card. The local value network constructs an independent two-layer fully connected network for each temperature control region, and each local network only calculates the Q value for its corresponding temperature control region to ensure that the heat dissipation control strategies for different regions are optimized for local characteristics. In the entire network architecture, the outputs of the global value network and the local value network together constitute the Q-value evaluation result, which is used to measure the advantages and disadvantages of different heat dissipation control strategies in the current state. Based on the Q-value evaluation result, the ε-greedy strategy is executed to select an appropriate heat dissipation control action. The ε-greedy strategy is a reinforcement learning exploration strategy that selects the current optimal action with a certain probability to ensure that the heat dissipation strategy converges to the optimal solution, and randomly selects an action with the remaining probability to ensure that the algorithm can explore new heat dissipation strategies and avoid falling into local optima. Calculate the Q values of all current candidate actions and select the action with the highest Q value as the current optimal action, and randomly select a non-optimal action with a probability of ε to increase exploration. This strategy can conduct more exploration in the initial stage of training to discover possible optimal strategies, and gradually tend to utilize the known optimal strategies in the later stage of training to improve the stability of heat dissipation control. Calculate the reward for the control action selection result to guide the training of the network. The calculation of the reward value is based on multiple factors, including the reduction in the overall temperature of the graphics card, the balance of local temperatures, and the change in fan power consumption. For example, if a certain heat dissipation control strategy can effectively reduce the temperature of the graphics card core area without causing abnormal temperature increases in other areas, the reward value of this strategy should be relatively high; on the contrary, if a certain control strategy causes local overheating or too high fan speed, affecting the system power consumption, the reward value of this strategy should be relatively low. Through the reward mechanism, ensure that the reinforcement learning network gradually learns the optimal heat dissipation control strategy to optimize the energy efficiency of the heat dissipation system while ensuring the stability of the graphics card temperature. Store the reward value and state transition data in the experience replay buffer and randomly sample the samples in the buffer during network training.The role of the experience replay buffer is to store past training samples and randomly extract data from it during network training to avoid the over-reliance of the reinforcement learning algorithm on recent data, thereby improving the stability and generalization ability of training. During the experience replay process, a batch of samples is randomly selected from the buffer in each iteration and used to calculate the update gradient of the network parameters to optimize the performance of the Q-network. This method can effectively reduce the correlation of data, improve the convergence speed of training, and avoid the network falling into local optimal solutions. After completing the experience replay, the Q-learning update rule is used for network optimization and heat dissipation control parameter prediction. Q-learning is a model-free reinforcement learning algorithm that iteratively updates the Q-value function to gradually converge to the optimal policy. In the specific implementation, the error between the current Q-value and the target Q-value is calculated, and the mean squared error loss function is used for gradient backpropagation to adjust the weight parameters of the network. According to the calculated update gradient of the network parameters, the global value network and the local value network are optimized so that they can more accurately predict the Q-values of different heat dissipation control strategies. After multiple rounds of training, the Q-network gradually learns the optimal heat dissipation control strategy and dynamically adjusts the heat dissipation parameters according to different temperature and load states to ensure that the graphics card can maintain the best temperature state in various working environments. Through the above steps, the optimal heat dissipation control parameter sequence is obtained and applied to the heat dissipation system of the graphics card to optimize the fan speed, the working state of the liquid cooling system, and the adjustment method of the heat sink, realizing intelligent temperature management.

[0021] Input the state encoding vector into the hybrid Q-network structure, which consists of a global value network and a local value network, to evaluate the advantages and disadvantages of the overall heat dissipation control strategy and the local optimization solutions for different temperature control regions respectively. In the input data preparation stage, the state encoding vector is concatenated by the temperature state vector and the load state vector of the temperature control region to ensure that the input data can completely describe the thermal state and load conditions of the graphics card. Standardize this state encoding vector to eliminate the dimensional differences between different physical quantities and improve the stability of network training and the convergence speed of numerical calculations. This state encoding vector is input into the hybrid Q-network structure to calculate the Q-value evaluation results of different heat dissipation control strategies. In the hybrid Q-network structure, input the state encoding vector into the first fully connected layer of the global value network for non-linear transformation to extract the initial feature representation. The role of this first fully connected layer is to perform a linear transformation on the input vector through a weight matrix and introduce non-linear factors through the ReLU activation function to ensure that the network learns complex temperature and load relationships. Through this process, the state encoding vector is mapped to a high-dimensional feature space to generate the first-layer feature vector, which contains the initial feature description of the current thermal state and load conditions of the graphics card. Input the first-layer feature vector into the second fully connected layer of the global value network to extract deep features. The role of this layer is to enable the network to capture the implicit relationships between temperature and load states through more complex weight mappings, such as the thermal coupling effects of different components and the impact of load changes on temperature distribution. In this process, the second-layer feature vector is further optimized to enable the network to more accurately identify the key factors affecting the temperature management of the graphics card. Input the second-layer feature vector into the third fully connected layer of the global value network to calculate the global Q-value. This Q-value is used to evaluate the effectiveness of the overall heat dissipation control strategy and is used to select the optimal heat dissipation control scheme in the decision-making process of reinforcement learning. In the process of calculating the global Q-value, the network maps the second-layer feature vector to a scalar output space through linear transformation and non-linear activation, where each Q-value corresponds to an optional heat dissipation control action. The global value network outputs the global action evaluation result, which represents the overall value of different heat dissipation control strategies under the current temperature and load states. At the same time, to improve the fineness of heat dissipation control, group the features of the state encoding vector according to the temperature control region and input each group of feature vectors into the first fully connected layer of the corresponding local value network. Since different regions of the graphics card have different heat dissipation requirements, for example, the GPU core area requires more efficient heat dissipation, while the memory area pays more attention to temperature balance, the introduction of the local value network ensures that the heat dissipation control strategies for different regions can be optimized according to their specific requirements. In the calculation process of the first fully connected layer, the feature vector of each temperature control region is mapped to a new feature space to extract the local feature representation, thereby reflecting the temperature and load characteristics of this region.Input the local feature vector into the second fully connected layer of the local value network to calculate the local Q-value. The role of this layer is to further extract the key features of a specific temperature control area and calculate the values of different local heat dissipation control strategies. Different from the global Q-value, the local Q-value only focuses on the heat dissipation effect of a specific temperature control area and does not consider the overall heat dissipation optimization goal of the entire graphics card. This hierarchical calculation method ensures that the heat dissipation strategies of different areas can independently consider their own characteristics during optimization without being affected by the heat dissipation requirements of other areas, improving the accuracy and flexibility of heat dissipation control. Perform weighted fusion on the global action evaluation result and the local action evaluation result, and appropriately allocate weights to the global Q-value and the local Q-value according to the relative importance of the global heat dissipation optimization goal and the local temperature control requirements, and finally obtain the Q-value evaluation result.

[0022] In the embodiment of the present invention, by arranging a multi-point temperature sensor array and establishing a temperature distribution matrix, the accurate monitoring of the temperatures of each component of the graphics card is realized. The lumped parameter thermal network model and the finite element discretization method are used to accurately describe the heat conduction relationship between the components of the graphics card, and a complete thermal network state space equation is established. The partial least squares regression analysis and the pole placement method are introduced to realize the accurate calculation of the temperature coupling coefficient and the design of the temperature decoupling controller, effectively solving the problem of temperature coupling between components; the hybrid Q-network structure is used for heat dissipation control, combining the advantages of the global value network and the local value network, realizing the intelligent optimization of the heat dissipation parameters of the graphics card. Through the ε-greedy strategy and the experience replay mechanism, the exploration ability and learning efficiency of the control system are enhanced, enabling the heat dissipation control strategy to be adaptively adjusted to adapt to different workload conditions.

[0023] In a specific embodiment, the process of executing step 101 may specifically include the following steps: Divide the graphics card into regions, set 4 temperature sensors in the GPU chip region, 4 temperature sensors in the video memory region, 2 temperature sensors in the power management module region, and 2 temperature sensors in the radiator region to obtain a temperature sensor array; Perform temperature data sampling based on the temperature sensor array to obtain the original temperature data, and eliminate outliers from the original temperature data to obtain the preprocessed temperature data; Reconstruct the preprocessed temperature data into a matrix according to the spatial positions of the temperature sensors to obtain the temperature spatial distribution data; Perform least squares polynomial fitting on the temperature spatial distribution data to obtain the temperature distribution function coefficient matrix, substitute the temperature distribution function coefficient matrix into the temperature field equation, and use the finite difference method to solve the temperature field equation to obtain the temperature distribution matrix.

[0024] Specifically, the graphics card is divided into regions, and temperature sensors are reasonably arranged on key components to collect the temperature distribution inside the graphics card in real time. Since the GPU chip region is the core computing component of the entire graphics card, it has the highest heat density and the most drastic temperature changes. Four temperature sensors are arranged in this region to capture the temperature gradient changes around the GPU core. At the same time, the memory region is responsible for data storage and high-speed data exchange, and its temperature has an important impact on the stability of the overall system. Four temperature sensors are arranged in the memory region to monitor the temperature fluctuations of the memory module under high-load conditions. The power management module, as the power supply center of the graphics card, its temperature affects the power supply efficiency and the overall heat dissipation performance. Two temperature sensors are arranged on the power management module to monitor the temperature changes of the power supply module. The radiator is the core heat dissipation component of the entire graphics card, and its temperature distribution determines the uniformity and efficiency of heat dissipation. Two temperature sensors are arranged in the radiator region to evaluate the actual working state of the heat dissipation system. Through the above arrangements, a temperature sensor array is constructed. Temperature data sampling is performed based on the temperature sensor array to obtain the temperature information of the graphics card under different working conditions. Assume that the sampling time interval for each sensor is , and the original temperature data obtained by sampling is represented as , where represents the temperature value of the temperature sensor located at the position at the -th sampling moment, represents the sampling time point, and respectively represent the spatial position coordinates of the sensor. Outlier removal is performed on the original temperature data to ensure the accuracy of the data. The method of mean filtering or median filtering is used. For example, the mean value of the temperature data is calculated using the sliding window method: where is the average temperature value within the window, and is the sliding window size. If the temperature value of a certain sampling point exceeds a certain multiple of the standard deviation of : then this point is considered an outlier, and the temperature value of this point is replaced with the sliding window mean value. Among them, is the standard deviation of the temperature data, and is the rejection threshold. The processed temperature data is matrix-reconstructed according to the spatial position of the sensors to form temperature spatial distribution data. Assume that the surface of the graphics card is divided into a two-dimensional grid, then the temperature spatial distribution data is represented by a matrix as: where and respectively represent the number of rows and columns of the temperature sensor array, and each element represents the temperature data at that position. This matrix reflects the temperature distribution of each area of the graphics card. Perform least squares polynomial fitting on the temperature spatial distribution data to construct a temperature distribution function model. Assuming that the temperature distribution is represented by a binary polynomial, the following temperature distribution function is constructed: Among them, respectively represent the spatial coordinates on the surface of the graphics card, are the coefficients to be determined. By using the least squares method, solve these coefficients so that the polynomial fits the sampled temperature data as much as possible. After obtaining the temperature distribution function, substitute it into the temperature field equation to calculate the complete temperature field distribution inside the graphics card. The heat conduction process follows the heat conduction equation: Among them, is the thermal diffusivity, is the heat source term inside the graphics card. Since this equation cannot be directly solved analytically, the finite difference method is used for discretization. Using the central difference format, the partial derivatives are expressed as: Substitute these difference expressions into the temperature field equation and perform iterative calculations through numerical solution methods to obtain a complete temperature distribution matrix, which describes the spatial variation of the temperature on the surface of the graphics card.

[0025] In a specific embodiment, the process of executing step 102 may specifically include the following steps: Divide the temperature data in the temperature distribution matrix according to the physical position relationship, establish corresponding heat capacity nodes for each graphics card component, and set the heat conduction paths between adjacent nodes as thermal resistance connections to obtain a thermal network topology structure containing n nodes; Establish a heat balance equation C i (dT i / d t ) = P i +Σ(T j -T i ) / R ij for each node i in the thermal network topology structure, where C i is the heat capacity of node i, T i is the temperature of node i, P i is the heat power input of node i, and R ij is the thermal resistance between node i and node j, to obtain a system of heat balance equations; The heat balance equations are discretized in the time domain to obtain the time-discretized heat balance equations, and the time-discretized heat balance equations are discretized by the finite element method in the space domain to obtain the finite element equations of the heat network; The finite element equations of the heat network are matrix-reconstructed, and the node temperatures, heat capacities, heat powers, and thermal resistances are respectively organized into a state vector and a coefficient matrix to obtain the state equations of the heat network, and the state equations of the heat network are parameter-calibrated to obtain the state space equations of the heat network.

[0026] Specifically, the temperature data in the temperature distribution matrix are node-divided according to the physical position relationship to establish the heat network topology. Inside the graphics card, the temperature data of key components such as the GPU chip, video memory, power management module, and radiator are collected through a temperature sensor array and form a temperature distribution matrix. Due to the spatial coupling characteristics of heat conduction inside the graphics card, the entire graphics card is divided into multiple heat capacity nodes, and each heat capacity node represents a computing unit with an independent heat capacity. Suppose the graphics card is divided into nodes, and the heat capacity of each node is denoted as , and its physical meaning is the amount of heat that the node can store when the temperature changes by one unit, that is: where represents the material density of node , is the specific heat capacity of the material, is the volume of the node. To accurately describe the heat conduction process inside the graphics card, adjacent heat capacity nodes are connected by thermal resistances, and the thermal resistance reflects the heat conduction ability between node and node , and its calculation method is: where is the distance between node and , is the thermal conductivity of the material, is the effective heat conduction area between the two nodes. In this way, a heat network topology containing nodes is established and ensured to accurately reflect the heat conduction path inside the graphics card. After establishing the heat network topology, a heat balance equation is established for each heat capacity node to describe the temperature change process of the node. According to the first law of thermodynamics, the temperature change of node is affected by its own heat input, heat capacity, and heat exchange with adjacent nodes, so its heat balance equation is expressed as: Among them, is the temperature of the node , is the heat power input of this node, represents the set of neighbor nodes directly connected to the node . This equation describes the change of the temperature of the node over time. Among them, the first term represents the direct heating of the node by the external heat source, and the second term represents the heat exchange relationship between this node and neighboring nodes through the thermal resistance. In order to solve the above differential equation in numerical calculation, the heat balance equation set is discretized in the time domain to convert it into a discrete time format. The forward difference format is used for time discretization, and the time derivative is approximately expressed as: Among them, represents the temperature of the node at the time step , is the time step size. After transformation, we can get: This time-discretized heat balance equation set is iteratively solved on the computer to predict the temperature distribution of the graphics card at different time steps. After completing the time domain discretization, the finite element discretization in the spatial domain is carried out to optimize the calculation accuracy. The calculation domain is divided into a finite number of elements. For example, the surface of the graphics card is divided by tetrahedral meshes, and the temperature change inside each element is approximated by interpolation using piecewise linear basis functions. Based on the finite element method, the local element stiffness matrix and the mass matrix are constructed. Among them, the stiffness matrix is used to describe the heat conduction behavior, and the mass matrix is used to describe the heat capacity characteristics. The local element equation is expressed as: Among them, is the node temperature vector, is the heat power input vector, and are calculated through the heat conduction relationship between nodes. By globally assembling the stiffness matrices and mass matrices of all elements and optimizing the storage using the sparse matrix storage format, the complete finite element equation set of the thermal network is obtained. The finite element equation set of the thermal network is matrix-reconstructed to construct a standard state space equation. The node temperature vector , the heat capacity matrix , the heat power input vector and the thermal resistance matrix are organized into a state vector and a coefficient matrix to obtain the following thermal network state equation: Among them, is the diagonal heat capacity matrix, is the thermal resistance matrix, is the temperature vector of adjacent nodes. This state equation describes the dynamic characteristics of the entire graphics card thermal network and is used for temperature prediction and control optimization. Calibrate the parameters of the state equation. The process of parameter calibration needs to combine experimental data, measure the temperature response under different power inputs, and use methods such as the least squares method or Kalman filter to optimize the thermal resistance, heat capacity, and thermal power input parameters, so that the calculation results match the actual measurement results. The thermal network state space equation after parameter calibration more accurately describes the temperature changes of the graphics card under different loads.

[0027] In a specific embodiment, the process of performing time-domain discretization on the heat balance equations to obtain time-discretized heat balance equations and performing space-domain finite element discretization on the time-discretized heat balance equations to obtain the finite element equations of the thermal network can specifically include the following steps: Perform a difference expansion of the heat balance equations in the time domain to obtain a time-discretized form of the equations, and perform term combination and arrangement on the time-discretized form of the equations to obtain a normalized time-discretized equation set; Perform a condition number analysis on the coefficient matrix of the normalized time-discretized equation set to obtain the time-discretized heat balance equations; Discretize the time-discretized heat balance equations in the space domain, divide the computational domain into tetrahedral meshes, construct piecewise linear basis functions, and obtain local element stiffness matrices and mass matrices; Perform global assembly on the local element stiffness matrices and mass matrices, establish an assembly matrix according to the corresponding relationship of node numbers, and store the global stiffness matrix and mass matrix in the sparse matrix compressed storage format CSR to obtain the target finite element equations; Solve the target finite element equations using the preconditioned conjugate gradient method, select the incomplete Cholesky decomposition as the preconditioning matrix, and set the iterative convergence criterion to the residual norm being less than 10-6 to obtain the finite element equations of the thermal network.

[0028] Specifically, perform a difference expansion of the heat balance equations in the time domain to convert them from the continuous time form to the discrete time form for solving in numerical calculations. The basic form of the heat balance equation is expressed as: Among them, represents the heat capacity of the heat capacity node describing the temperature change inertia when the node absorbs or releases heat; is the temperature of node ; $Q_{in}$ is the heat power of the external heat source acting on the node, representing the heat absorbed by the node due to electrical power consumption or environmental heat sources; is the node and the adjacent node The thermal resistance between them represents the heat conduction capacity of this path. The smaller the value, the higher the heat transfer efficiency; represents the node The set of adjacent nodes. To discretize the time variable, the forward Euler difference method is adopted, and the time derivative is approximately expressed as: where and represent the temperatures at time steps and respectively, and is the time step size. Rearranging the above equation gives the heat balance equation in time-discrete form: This equation represents the update rule of the temperature within each time step. The first term represents the temperature state of the previous step, the second term represents the temperature change caused by the external heat source, and the third term represents the temperature change due to heat conduction. The condition number analysis of the coefficient matrix of the time-discrete form equation is performed to evaluate its numerical stability. The condition number of the coefficient matrix is defined as: where and are the largest and smallest singular values of the matrix respectively. If the condition number is too large, it indicates that the matrix is relatively ill-conditioned, which will cause the amplification of numerical errors during the solution process, thus affecting the accuracy of the results. The numerical stability of the matrix is improved through scaling or preprocessing methods. After condition number analysis and adjustment, the final time-discretized heat balance equations are obtained. Based on the time discretization, the computational domain is discretized in the spatial domain by the finite element method to establish a computational grid that conforms to the geometric structure of the graphics card. The tetrahedral mesh is used to divide the computational domain, and the temperature distribution of each tetrahedral element is approximated by interpolation using piecewise linear basis functions. Assume that the interpolation function of the temperature field is expressed as: where is the linear interpolation basis function inside the tetrahedral element, and are the temperature values at the vertices of the element. Based on this interpolation relationship, the local element stiffness matrix and the mass matrix are constructed, and their computational expressions are respectively: Among them, represents the unit volume, is the thermal conductivity, is the material density. The stiffness matrix reflects the influence of the temperature gradient on heat conduction, while the mass matrix reflects the influence of the heat capacity on the system response. The local element stiffness matrix and mass matrix are globally assembled to form the finite element equations of the entire graphics card. During the assembly process, an assembly matrix is established according to the correspondence of the global node numbers, so that the calculation data between different elements can be correctly merged together. Since the coefficient matrix of the heat conduction equation is usually sparse, a sparse matrix storage format, such as CSR, is adopted to reduce the storage requirements and improve the calculation efficiency. The finally obtained finite element equations are: Among them, is the temperature vector, is the heat source term vector, is the mass matrix, is the stiffness matrix. To solve this system of equations, a numerical method is adopted, and the preconditioned conjugate gradient method is an efficient iterative solution method. The preconditioned conjugate gradient method constructs a preconditioning matrix to make the solution of the original equation more stable and fast. The incomplete Cholesky decomposition is selected as the preconditioning matrix, so that the original equation is transformed into an equivalent form that is easier to solve: During the solution process of the preconditioned conjugate gradient method, the temperature distribution is calculated iteratively, and the iterative convergence criterion is set to the residual norm less than , that is: When the residual satisfies this condition, the iteration terminates, and it is considered that a sufficiently accurate temperature field solution has been obtained. The numerical solution of the finite element equations of the thermal network is obtained, so as to calculate the temperature distribution in different regions inside the graphics card and further optimize the heat dissipation strategy of the graphics card.

[0029] In a specific embodiment, the process of executing step 103 may specifically include the following steps: Decompose the system matrix of the thermal network state space equation, extract the temperature transfer relationship between nodes, and obtain the temperature state transition matrix; Input the temperature state transition matrix into the NIPALS algorithm to calculate the principal components, obtain the dimensionality reduction feature matrix, perform partial least squares regression calculation on the dimensionality reduction feature matrix, and perform significance test on the regression coefficients to obtain the temperature coupling coefficient matrix; Calculate the temperature correlation degree between components based on the temperature coupling coefficient matrix to obtain multiple initial control regions; Construct a state feedback controller based on multiple initial control regions, configure the closed-loop poles in the left half plane, and obtain a temperature decoupling controller; Conduct Lyapunov stability analysis on the temperature decoupling controller, determine the control parameters of each initial control region, and obtain m temperature control regions.

[0030] Specifically, decompose the system matrix of the thermal network state space equation to extract the temperature transfer relationship between each component of the graphics card and construct a temperature state transition matrix. The thermal network state space equation is expressed as: Among them, is the heat capacity matrix, which describes the heat capacity of each node; is the temperature vector, representing the temperature of each node; is the thermal conductivity matrix, representing the heat conduction relationship between nodes; is the heat source input vector, which describes the heating effect of the external heat source on each node. Convert it into the state space form to obtain the following state equation: Among them, represents the temperature transfer characteristic matrix of the system, indicating the thermal coupling relationship between different nodes. By decomposing this matrix, the temperature state transition matrix is extracted: Among them, The element of the matrix reflects the influence degree of node on the temperature change of node . The diagonal element of this matrix is negative, indicating the natural decay of the temperature of this node over time, while the off-diagonal element represents the influence degree of node on the heat transfer to node . Perform dimensionality reduction on the temperature state transition matrix to extract key temperature transfer characteristics and remove redundant information. Since the temperature transfer in the graphics card thermal network has high-dimensional characteristics and the influence between some temperature nodes is weak, the NIPALS algorithm is used to calculate the principal components to reduce the data dimension. The NIPALS algorithm extracts the main change patterns in the data matrix through iterative calculation and retains the most important components. In the specific calculation, first perform singular value decomposition (SVD) on : Among them, and are the left and right singular vector matrices respectively, is the singular value matrix, where the larger singular values correspond to the main temperature transfer modes. By selecting the first principal components, a reduced-dimension feature matrix is constructed: where, and only contain the first important features, thus retaining the main modes of temperature transfer while reducing the computational complexity. Partial least squares regression calculation is performed on the reduced-dimension feature matrix to analyze the temperature coupling relationship between the components of the graphics card. During the partial least squares regression process, let the temperature input matrix be , and the temperature output matrix be , then the partial least squares regression model is expressed as: where, is the regression coefficient matrix, describing the linear relationship between the input variables (temperature states) and the output variables (temperature change trends), is the error term. When calculating , a significance test is performed on it to eliminate the statistically insignificant coupling terms, and the temperature coupling coefficient matrix is obtained: where, represents the significance level of the regression coefficient. If is lower than a certain threshold, it is considered that the temperature influence of node on node can be ignored. Based on the temperature coupling coefficient matrix , the temperature correlation degrees between the components of the graphics card are calculated, and the initial control regions are divided accordingly. The temperature correlation degree is calculated by normalizing the absolute value of the coupling coefficient matrix: For components with higher correlation degrees, they are grouped into the same initial control region, forming multiple initial temperature control regions. After obtaining the initial control regions, a temperature decoupling controller is constructed based on the state feedback control strategy. Assume that the temperature state of each control region is , and its dynamic equation is: where, is the temperature transfer matrix of the control region, is the control input matrix, For the control input, the pole placement method is used to design a state feedback controller so that the closed-loop poles are located in the left half-plane: where, is the feedback gain matrix, and its value is calculated by the pole placement method to ensure the stability and response speed of the temperature control system. To ensure the stability of the control system, Lyapunov stability analysis is performed on the temperature decoupling controller. Define the Lyapunov function: where, is a positive definite matrix. If for any , it satisfies: then it indicates that the system is asymptotically stable. By adjusting the feedback gain matrix , ensure that this condition holds to obtain a stable temperature control region. Obtain stable temperature control regions and optimize the temperature management strategy of the graphics card.

[0031] In a specific embodiment, the process of executing step 104 may specifically include the following steps: Concatenate and encode the temperature state vector and the load state vector of the temperature control region to obtain a state encoding vector; Input the state encoding vector into the global value network and the local value network of the hybrid Q-network structure. The global value network adopts a three-layer fully connected layer structure, and the local value network constructs an independent two-layer fully connected network for each temperature control region to obtain the Q-value evaluation result; Execute the ε-greedy policy based on the Q-value evaluation result to obtain the control action selection result, and calculate the reward for the control action selection result to obtain the reward value; Store the reward value and the state transition data in the experience replay buffer, randomly sample the samples in the experience replay buffer to obtain the network parameter update gradient; Adopt the Q-learning update rule, and perform network optimization and heat dissipation control parameter prediction according to the network parameter update gradient to obtain the heat dissipation control parameter sequence.

[0032] Specifically, concatenate and encode the temperature state vector and the load state vector of the temperature control region to construct a state encoding vector for use as the input of the reinforcement learning model. During the operation of the graphics card, the temperature states of different temperature control regions are affected by the current load situation. Concatenate the temperature state vector and the load state vector to form a high-dimensional feature vector: Among them, represents the temperature status of the th temperature control region, represents the th load status, where represents the current power consumption of the th component (such as GPU, video memory, power management module, etc.). Through the splicing operation, the complete state encoding vector is obtained to describe the current thermal load status of the graphics card. The state encoding vector Among them, is the weight matrix of the fully connected layer, is the bias vector, is the ReLU activation function, and represent the hidden layer outputs of the first and second layers respectively, and the final is the global action evaluation result. At the same time, the local value network constructs an independent two-layer fully connected network for each temperature control region, and each network only calculates the Q value for its corresponding temperature control region. For the th temperature control region, its local value network calculation formula is as follows: Among them, is the th region feature of the state encoding vector, is the weight matrix of the local value network, is the bias vector, and the final is the local action evaluation result. The outputs of the global value network and the local value network are weighted and fused to obtain the final Q value evaluation result: Among them, is the fusion weight coefficient, which controls the balance between the global and local strategies. Based on the Q-value evaluation result, the ε-greedy strategy is executed to select appropriate heat dissipation control actions. The ε-greedy strategy improves the efficiency of the learning process through the combination of random exploration and exploitation of the current optimal strategy. Under this strategy, with probability select a random action (explore new heat dissipation solutions), and with probability select the current optimal action (exploit the existing optimal heat dissipation strategy): where, is a random variable between [0, 1], is the selected control action. After determining the control action, the reward value is calculated to evaluate the effectiveness of this heat dissipation strategy. The reward function is defined based on the temperature reduction amplitude, fan power consumption optimization, and heat dissipation balance. For example: where, is the target temperature, is the fan power consumption, is the fan power consumption weight coefficient. If the temperature is close to the target value, the reward value is high; if the temperature deviates from the target value, the reward value is low. The reward value and state transition data are stored in the experience replay buffer, and random sampling is performed during the training process to reduce data correlation and improve the stability of training. The data stored in the experience replay buffer includes the state, action, reward, and the next state: During the training process, a batch of samples is randomly drawn from the buffer, and the network is optimized based on the Q-learning update rule. The Q-learning update rule is as follows: where, is the discount factor, which controls the importance of long-term rewards, is the learning rate, which controls the step size of parameter update, is the neural network parameter, is the target Q-value. Through multiple rounds of training, the Q-network learns the optimal heat dissipation control strategy and dynamically adjusts parameters such as the fan speed and the opening and closing state of the heat sink according to different temperature and load states.

[0033] In a specific embodiment, the process of inputting the state encoding vector into the global value network and the local value network of the hybrid Q-network structure, where the global value network adopts a three-layer fully connected layer structure and the local value network constructs an independent two-layer fully connected network for each temperature control region to obtain the Q-value evaluation result, may specifically include the following steps: Input the state encoding vector into the first fully connected layer of the global value network in the hybrid Q-network structure for non-linear transformation to obtain the first-layer feature vector; Input the first-layer feature vector into the second fully connected layer of the global value network for feature extraction to obtain the second-layer feature vector; Input the second-layer feature vector into the third fully connected layer of the global value network to calculate the global Q value and obtain the global action evaluation result; Group the state encoding vector according to the temperature control regions, and input each group of feature vectors into the first fully connected layer of the corresponding local value network to obtain the local feature vectors; Input the local feature vectors into the second fully connected layer of the local value network to calculate the local Q value and obtain the local action evaluation result; Perform weighted fusion on the global action evaluation result and the local action evaluation result to obtain the Q value evaluation result.

[0034] Specifically, input the state encoding vector into the hybrid Q-network structure, and perform calculations separately in the global value network and the local value network to evaluate the Q values of different heat dissipation control strategies. In the hybrid Q-network structure, the global value network learns the overall heat dissipation optimization strategy of the graphics card, while the local value network independently models each temperature control region to ensure that different regions can be optimized and adjusted according to their own thermal dynamic characteristics. Extract features from the state encoding vector and construct feature expressions at different levels to ensure that the network can learn the global and local patterns of temperature control. Input the state encoding vector into the first fully connected layer of the global value network for non-linear transformation to extract basic feature information. The state encoding vector is composed of the temperature state vector and the load state vector , that is: Among them, represents the temperature states of temperature control regions in the graphics card, while represents the current load conditions of key components. During the calculation of the fully connected layer, perform linear transformation and then pass through the ReLU activation function to enhance the non-linear expression ability of the network: Among them, is the weight matrix, is the bias vector, is the ReLU activation function, is the feature vector of the first fully connected layer, and this vector contains a preliminary feature description of the current heat load state of the graphics card. Input the first-layer feature vector Input into the second fully connected layer of the global value network to further extract deep features. The calculation method of the second layer is as follows: Wherein, is the weight matrix of the second layer, is the bias vector, represents the feature vector of the second layer. The function of this layer is to learn the heat transfer relationship between different temperature control regions and the influence of the graphics card load status on the heat dissipation efficiency. After completing the feature extraction, the second-layer feature vector is input into the third fully connected layer of the global value network to calculate the global Q value, that is, the global action evaluation result. The calculation formula is as follows: Wherein, is the weight matrix of the third layer, is the bias vector, represents the Q value evaluation result of the global heat dissipation strategy, which is used to measure the influence of the current heat dissipation control strategy on the overall temperature optimization of the graphics card. At the same time, in order to improve the fineness of heat dissipation control, the state encoding vector is feature-grouped according to the temperature control regions, and each group of feature vectors is input into the first fully connected layer of the corresponding local value network. When feature-grouping, the state vector of each temperature control region is composed of the temperature information and related load information of this region to balance the global heat dissipation optimization target and the local temperature optimization target. Through this fusion method, it is ensured that the final heat dissipation control strategy can not only optimize the overall thermal management of the graphics card but also precisely control specific regions to avoid local overheating problems. The calculated Q value evaluation result is used in the decision-making process of reinforcement learning and guides the optimization of the graphics card's heat dissipation control strategy.

[0035] The method for monitoring the performance of a graphics card component in the embodiment of the present invention has been described above. Next, the device for monitoring the performance of a graphics card component in the embodiment of the present invention will be described. Please refer to Figure 2 , an embodiment of the device for monitoring the performance of a graphics card component in the embodiment of the present invention includes: A temperature data acquisition module 201, configured to perform temperature data acquisition and preprocessing on the GPU chip, video memory, power management module, and radiator of the graphics card by setting a temperature sensor array to obtain a temperature distribution matrix; A discretization calculation module 202, configured to establish a thermal network topology structure including n nodes according to the temperature distribution matrix, and perform finite element discretization calculation on the thermal network topology structure to obtain a thermal network state space equation; The regression analysis module 203 is used to perform partial least squares regression analysis on the thermal network state space equation, calculate the temperature coupling coefficient matrix, construct a temperature decoupling controller through the pole placement method, and obtain m temperature control regions. The execution module 204 is used to input the temperature state vector and load state vector of the temperature control region into the hybrid Q-network structure, execute the ε-greedy strategy and Q-learning update rule, and obtain a sequence of heat dissipation control parameters.

[0036] Through the collaborative cooperation of the above-mentioned various components, by arranging a multi-point temperature sensor array and establishing a temperature distribution matrix, the accurate monitoring of the temperatures of each component of the graphics card is realized. By using the lumped parameter thermal network model and the finite element discretization method, the heat conduction relationship between the components of the graphics card is accurately described, a complete thermal network state space equation is established, and partial least squares regression analysis and the pole placement method are introduced to realize the accurate calculation of the temperature coupling coefficient and the design of the temperature decoupling controller, effectively solving the problem of temperature coupling between components. The hybrid Q-network structure is used for heat dissipation control, combining the advantages of the global value network and the local value network, realizing the intelligent optimization of the heat dissipation parameters of the graphics card. Through the ε-greedy strategy and the experience replay mechanism, the exploration ability and learning efficiency of the control system are enhanced, enabling the heat dissipation control strategy to be adaptively adjusted to different workload conditions.

[0037] Above Figure 2 The graphics card component performance monitoring device in the embodiment of the present invention is described in detail from the perspective of modular functional entities. Next, the graphics card component performance monitoring device in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0038] Figure 3 FIG. is a schematic structural diagram of a graphics card component performance monitoring device provided by an embodiment of the present invention. The graphics card component performance monitoring device 300 may vary greatly due to configuration or performance, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage device terminals) for storing application programs 333 or data 332. Among them, the memory 320 and the storage media 330 may be transient storage or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations on the graphics card component performance monitoring device 300. Further, the processor 310 may be configured to communicate with the storage media 330 and execute a series of instruction operations in the storage media 330 on the graphics card component performance monitoring device 300 to implement the steps of the above-mentioned graphics card component performance monitoring method.

[0039] The graphics card component performance monitoring device 300 may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, and so on. Those skilled in the art can understand that Figure 3 The structure of the graphics card component performance monitoring device shown does not constitute a limitation on the graphics card component performance monitoring device provided by the present invention. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0040] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described systems, systems, and units can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0041] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0042] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for monitoring the performance of a graphics card component, characterized in that: The method comprises: A temperature sensor array is set for the GPU chip, video memory, power management module and radiator of the graphics card to collect and preprocess temperature data to obtain a temperature distribution matrix; Establishing a thermal network topology structure including n nodes according to the temperature distribution matrix, and performing finite element discretization calculation on the thermal network topology structure to obtain a thermal network state space equation; Performing partial least squares regression analysis on the thermal network state space equation, calculating the temperature coupling coefficient matrix, constructing a temperature decoupling controller by a pole placement method, and obtaining m temperature control areas; The temperature state vector and the load state vector of the temperature control area are input into a hybrid Q network structure, and an ε-greedy strategy and a Q learning update rule are executed to obtain a heat dissipation control parameter sequence.

2. The method for monitoring the performance of a graphics card component according to claim 1, characterized in that: The GPU chip, video memory, power management module and heat sink of the graphics card are provided with a temperature sensor array to collect and preprocess temperature data to obtain a temperature distribution matrix, including: The graphics card is divided into regions, 4 temperature sensors are set in the GPU chip region, 4 temperature sensors are set in the video memory region, 2 temperature sensors are set in the power management module region, and 2 temperature sensors are set in the radiator region to obtain a temperature sensor array; Sampling temperature data based on the temperature sensor array to obtain raw temperature data, and removing abnormal values ​​from the raw temperature data to obtain preprocessed temperature data; Reconstructing the preprocessed temperature data into a matrix according to the spatial position of the temperature sensor to obtain temperature spatial distribution data; The temperature spatial distribution data is fitted with a least squares polynomial to obtain a temperature distribution function coefficient matrix, and the temperature distribution function coefficient matrix is ​​substituted into a temperature field equation, and the temperature field equation is solved using a finite difference method to obtain a temperature distribution matrix.

3. The method for monitoring the performance of a graphics card component according to claim 1, characterized in that: The thermal network topology structure including n nodes is established according to the temperature distribution matrix, and the thermal network topology structure is subjected to finite element discretization calculation to obtain the thermal network state space equation, including: The temperature data in the temperature distribution matrix is ​​divided into nodes according to the physical position relationship, a corresponding heat capacity node is established for each graphics card component, and the heat conduction path between adjacent nodes is set as a thermal resistance connection to obtain a thermal network topology structure containing n nodes; For each node i in the thermal network topology, a heat balance equation C is established. i (dT i / d t )=P i +Σ(T j -T i ) / R ij , where C i is the heat capacity of node i, T i is the temperature of node i, P i is the thermal power input of node i, R ij is the thermal resistance between node i and node j, and the thermal balance equations are obtained; Performing time domain discretization processing on the heat balance equation group to obtain a time discretized heat balance equation group, and performing space domain finite element discretization processing on the time discretized heat balance equation group to obtain a thermal network finite element equation group; The thermal network finite element equation group is matrix reconstructed, and the node temperature, heat capacity, thermal power and thermal resistance are respectively organized into state vectors and coefficient matrices to obtain the thermal network state equation, and the thermal network state equation is parameterized to obtain the thermal network state space equation.

4. The method for monitoring the performance of a graphics card component according to claim 3, characterized in that: The heat balance equations are discretized in the time domain to obtain a time discretized heat balance equations, and the time discretized heat balance equations are discretized in the space domain to obtain a heat network finite element equations, including: Differentially expanding the heat balance equations in the time domain to obtain a time discrete form equation group, and merging and arranging the terms of the time discrete form equation group to obtain a normalized time discrete form equation group; Performing condition number analysis on the coefficient matrix of the normalized time-discrete equation group to obtain a time-discrete heat balance equation group; Discretizing the time discretized heat balance equations in the space domain, dividing the computational domain into tetrahedral grids, constructing piecewise linear basis functions, and obtaining a local unit stiffness matrix and a mass matrix; The local unit stiffness matrix and mass matrix are assembled as a whole, an assembly matrix is ​​established according to the corresponding relationship of the node numbers, and the overall stiffness matrix and mass matrix are stored in a sparse matrix compression storage format CSR to obtain a target finite element equation group; The target finite element equations are solved by the preconditioned conjugate gradient method. The preconditioned matrix uses incomplete Cholesky decomposition, and the iterative convergence criterion is set to the residual norm less than 10 -6 , and the finite element equations of the thermal network are obtained.

5. The method for monitoring the performance of a graphics card component according to claim 1, characterized in that: The thermal network state space equation is subjected to partial least squares regression analysis, the temperature coupling coefficient matrix is ​​calculated, and a temperature decoupling controller is constructed by a pole configuration method to obtain m temperature control areas, including: Decomposing the system matrix of the thermal network state space equation, extracting the temperature transfer relationship between nodes, and obtaining a temperature state transfer matrix; The temperature state transfer matrix is ​​input into the NIPALS algorithm to calculate the principal component, and a dimension reduction feature matrix is ​​obtained; and a partial least squares regression calculation is performed on the dimension reduction feature matrix, and a significance test is performed on the regression coefficient to obtain a temperature coupling coefficient matrix; Calculate the temperature correlation between components based on the temperature coupling coefficient matrix to obtain multiple initial control areas; Building a state feedback controller based on the multiple initial control regions, configuring the closed-loop poles in the left half plane, and obtaining a temperature decoupling controller; The temperature decoupling controller is subjected to Lyapunov stability analysis to determine control parameters of each initial control region, thereby obtaining m temperature control regions.

6. The method for monitoring the performance of a graphics card component according to claim 1, characterized in that: The temperature state vector and the load state vector of the temperature control area are input into the hybrid Q network structure, the ε-greedy strategy and the Q learning update rule are executed, and the heat dissipation control parameter sequence is obtained, including: Concatenating and encoding the temperature state vector and the load state vector of the temperature control area to obtain a state encoding vector; The state encoding vector is input into the global value network and the local value network of the hybrid Q network structure, wherein the global value network adopts a three-layer fully connected layer structure, and the local value network constructs an independent two-layer fully connected network for each temperature control area to obtain a Q value evaluation result; Execute the ε-greedy strategy based on the Q-value evaluation result to obtain a control action selection result, and perform reward calculation on the control action selection result to obtain a reward value; The reward value and state transition data are stored in an experience replay buffer, and samples in the experience replay buffer are randomly sampled to obtain a network parameter update gradient; The Q learning update rule is adopted to perform network optimization and heat dissipation control parameter prediction according to the network parameter update gradient to obtain a heat dissipation control parameter sequence.

7. The method for monitoring the performance of a graphics card component according to claim 6, characterized in that: The state encoding vector is input into the global value network and the local value network of the hybrid Q network structure, the global value network adopts a three-layer fully connected layer structure, and the local value network constructs an independent two-layer fully connected network for each temperature control area to obtain a Q value evaluation result, including: Input the state encoding vector into the first fully connected layer of the global value network in the hybrid Q network structure for nonlinear transformation to obtain a first-layer feature vector; Inputting the first layer feature vector into the second fully connected layer of the global value network to perform feature extraction to obtain a second layer feature vector; Input the second layer feature vector into the third fully connected layer of the global value network, calculate the global Q value, and obtain the global action evaluation result; The state encoding vector is feature grouped according to the temperature control area, and each group of feature vectors is input into the first fully connected layer of the corresponding local value network to obtain a local feature vector; Inputting the local feature vector into the second fully connected layer of the local value network, calculating the local Q value, and obtaining the local action evaluation result; The global action evaluation result and the local action evaluation result are weightedly fused to obtain a Q value evaluation result.

8. A graphics card component performance monitoring device, characterized in that: The device is used to execute the graphics card component performance monitoring method according to any one of claims 1 to 7, the device comprising: The temperature data acquisition module is used to set a temperature sensor array for the GPU chip, video memory, power management module and radiator of the graphics card to collect and preprocess the temperature data to obtain a temperature distribution matrix; A discretization calculation module, used for establishing a thermal network topology structure including n nodes according to the temperature distribution matrix, and performing finite element discretization calculation on the thermal network topology structure to obtain a thermal network state space equation; A regression analysis module, used for performing partial least squares regression analysis on the thermal network state space equation, calculating the temperature coupling coefficient matrix, constructing a temperature decoupling controller by a pole placement method, and obtaining m temperature control areas; The execution module is used to input the temperature state vector and the load state vector of the temperature control area into the hybrid Q network structure, execute the ε-greedy strategy and the Q learning update rule, and obtain the heat dissipation control parameter sequence.

9. A graphics card component performance monitoring device, characterized in that: The graphics card component performance monitoring device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instruction in the memory so that the graphics card component performance monitoring device executes the graphics card component performance monitoring method according to any one of claims 1-7.

Images