Self-adaptive power consumption control method and system for CPU card chip

Abstract

The invention relates to the technical field of CPU card chips, and discloses a self-adaptive power consumption control method and system of a CPU card chip. The method comprises the following steps: collecting power consumption characteristic data of a CPU card chip in different working load states, and constructing a power consumption signal model; creating a generalized power consumption control operator, and performing matrix decomposition on the power consumption characteristic data of the CPU card chip to obtain a power consumption weight vector; performing dual-state monitoring and feature selection on the power consumption state vector and the performance state vector which are acquired in real time to obtain a low-power-consumption sparse control matrix; performing parameter optimization training through a five-layer fuzzy neural network to obtain a target inverse system model; the target inverse system model and the original power consumption control system are connected in series to form a pseudo-linear system, a linear closed-loop controller is added, and the decoupling control system is obtained, compared with a traditional method, the decoupling control system has the advantages that the energy efficiency ratio of a CPU card chip is remarkably increased, and guarantee is provided for efficient operation in various working environments.

Description

Technical Field

[0001] The present invention relates to the technical field of CPU card chips, and in particular to an adaptive power consumption control method and system for a CPU card chip. Background Art

[0002] Traditional CPU card chip power consumption control methods usually use fixed thresholds or simple linear adjustment strategies, which cannot accurately adapt to the changes in power consumption characteristics under different workloads, and are prone to control "blind spots", resulting in insufficient energy efficiency optimization or excessive performance fluctuations. In addition, existing power consumption control systems are sensitive to workload parameter estimation errors. When the actual working environment deviates from the preset conditions, the control effect will be significantly reduced, especially in complex scenarios such as temperature fluctuations and load mutations, and it is impossible to maintain a stable power consumption control effect.

[0003] The current mainstream power consumption control algorithms generally have problems such as insufficient explanation of the importance of power consumption characteristics and poor state transition smoothness. These algorithms often assign the same weight to different power consumption characteristics, ignoring the differentiated impact of each characteristic under different working states, and cannot achieve accurate power consumption regulation. At the same time, during the power consumption state switching process, it is difficult for the control system to maintain a smooth transition, which often leads to performance jitter and energy efficiency fluctuations, affecting user experience and system reliability. Especially in complex scenarios where multiple cores work together, improper handling of the power consumption coupling relationship between cores will cause a decrease in overall control efficiency. Summary of the invention

[0004] The present invention provides an adaptive power consumption control method and system for a CPU card chip. The present invention adopts a decoupling control system to significantly improve the energy efficiency ratio of the CPU card chip compared with the traditional method, and provides a guarantee for efficient operation in various working environments.

[0005] In a first aspect, the present invention provides an adaptive power consumption control method for a CPU card chip, the adaptive power consumption control method for a CPU card chip comprising: Based on the original power consumption control system, the power consumption characteristic data of the CPU card chip under different workloads is collected, and a power consumption signal model is constructed; Creating a generalized power consumption control operator based on the power consumption signal model, and performing matrix decomposition on the power consumption characteristic data of the CPU card chip according to the generalized power consumption control operator to obtain a power consumption weight vector; Using the power consumption weight vector, dual-state monitoring and feature selection are performed on the power consumption state vector and performance state vector collected in real time to obtain a low-power sparse control matrix; Inputting the low-power sparse control matrix into a five-layer fuzzy neural network for parameter optimization training to obtain a target inverse system model; The target inverse system model and the original power consumption control system are connected in series to form a pseudo-linear system, and a linear closed-loop controller is added to obtain a decoupling control system.

[0006] In a second aspect, the present invention provides an adaptive power consumption control system for a CPU card chip, wherein the adaptive power consumption control system for the CPU card chip comprises: The acquisition module is used to collect the power consumption characteristic data of the CPU card chip under different workloads based on the original power consumption control system and build a power consumption signal model; A creation module is used to create a generalized power consumption control operator based on the power consumption signal model, and perform matrix decomposition on the power consumption characteristic data of the CPU card chip according to the generalized power consumption control operator to obtain a power consumption weight vector; A feature selection module, used to use the power consumption weight vector to perform dual-state monitoring and feature selection on the power consumption state vector and performance state vector collected in real time to obtain a low-power sparse control matrix; A training module, used for inputting the low-power sparse control matrix into a five-layer fuzzy neural network for parameter optimization training to obtain a target inverse system model; The series connection module is used to connect the target inverse system model and the original power consumption control system in series to form a pseudo-linear system, and add a linear closed-loop controller to obtain a decoupled control system.

[0007] In the technical solution provided by the present invention, by establishing a multi-core power consumption signal model and introducing a multi-core cooperative matrix, the accuracy of power consumption control is significantly improved; the power consumption weight vector calculation method based on the singular value threshold enables the control system to automatically adjust the control parameters without accurately estimating the workload characteristics, thereby enhancing the adaptability of the system to different workloads; the designed segmented control strategy effectively solves the "blind spot" problem in traditional power consumption control, enabling the system to respond to small power consumption changes; regularization processing enhances the adaptability of the control system to ambient temperature changes; the combination of the fuzzy neural network inverse system and the improved differential evolution algorithm realizes the decoupling control of power consumption and performance, thereby reducing power consumption while ensuring performance; the state-keeping regularization on the low-dimensional power consumption data and the power consumption representation coefficient matrix ensures the smoothness of the power consumption state conversion process, thereby avoiding performance fluctuations and power consumption oscillations during state switching; the intelligent switching control mechanism and the adaptive control parameter sequence enhance the system's dynamic response capability to load mutations; the overall decoupling control system significantly improves the energy efficiency ratio of the CPU card chip compared to the traditional method, thereby providing a guarantee for efficient operation in various working environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0009] Figure 1 A schematic diagram of a flow chart of a method for adaptive power consumption control of a CPU card chip provided in an embodiment of the present application; Figure 2 A schematic block diagram of the structure of the adaptive power consumption control system of the CPU card chip provided in an embodiment of the present application. DETAILED DESCRIPTION

[0010] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0011] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change based on actual conditions.

[0012] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0013] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0014] In conjunction with the accompanying drawings, some embodiments of the present application are described in detail below. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0015] See also Figure 1 , Figure 1 A flow chart of the adaptive power consumption control method of the CPU card chip provided in the embodiment of the present application is shown as follows: Figure 1As shown, the adaptive power consumption control method of the CPU card chip provided in the embodiment of the present application includes steps S100 to S500.

[0016] Step S100: collecting power consumption characteristic data of the CPU card chip under different workloads based on the original power consumption control system, and constructing a power consumption signal model; It is understandable that the execution subject of the present invention may be an adaptive power consumption control system of a CPU card chip, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0017] Specifically, based on the original power consumption control system, the operating frequency, voltage, temperature, core usage rate, memory access frequency and instruction execution type of the CPU card chip are collected to obtain power consumption characteristic data. In the process of collecting data, the original power consumption control system works in coordination with multiple monitoring units, including a temperature sensor array, a current monitoring unit, a voltage monitoring unit and a workload analysis unit. These components together constitute a power consumption signal acquisition system to monitor the power consumption changes of the CPU card chip in real time under different load conditions. The collected data is digitized using a high-speed sampling circuit, and a suitable sampling frequency is set to ensure that the transient change characteristics of the power consumption signal are captured. On this basis, the digitized power consumption data undergoes a data preprocessing step to eliminate noise, normalize the data range and detect outliers, thereby ensuring the accuracy of the modeling process. Data denoising processing uses a filtering method to remove random noise introduced during the collection process, while normalization processing uses methods such as standardization or minimum-maximum normalization to map data of different dimensions to the same scale interval, thereby avoiding the introduction of deviations due to different variable scales during the model parameter estimation process. Outlier detection uses statistical analysis methods or outlier detection algorithms based on machine learning to eliminate data anomalies caused by external interference to improve the stability of modeling. A mathematical power consumption model is established based on the preprocessed data. The model is based on the mathematical formula Indicates that Represents the instantaneous power consumption of the chip, Represents the operating frequency of the CPU, Represents the CPU operating voltage, Represents the temperature of the chip, represents the workload characteristic vector, and is the corresponding weight coefficient matrix, which is used to measure the contribution of different influencing factors to power consumption. is an error term, which is used to represent random fluctuations or measurement errors that cannot be accurately described by the model. The establishment of this model is based on the physical characteristics of the chip and empirical data, so that it can accurately describe the power consumption changes of the CPU card chip under different working conditions. However, since the CPU card chip adopts a multi-core architecture, the power consumption of different cores will affect each other. Therefore, the multi-core synergy matrix is ​​introduced in the establishment of the basic power consumption model. Describe the power consumption coupling relationship between the cores. Multi-core coordination matrix The construction of is based on factors such as workload distribution of different cores, power management strategy, and task scheduling strategy, and is used to modify the original power consumption mathematical model to more accurately describe the power consumption characteristics in a multi-core environment. The modified power consumption mathematical model is expressed as ,in It is derived through historical data statistics and power consumption analysis, and can be dynamically adjusted to adapt to different workload modes. In order to improve the accuracy of the model, the parameters of the modified power consumption mathematical model are optimized. This process uses the least squares method and cross-validation method to ensure the optimal estimation of the model parameters. The least squares method is used to minimize the sum of squares of the errors between the model prediction value and the actual measurement value, so as to obtain a set of optimal parameter estimates, while the cross-validation method is used to evaluate the generalization ability of the model, that is, to ensure that the model not only performs well on the training data, but also maintains high prediction accuracy on unknown data. In the parameter optimization process, optimization algorithms such as gradient descent are used for iterative calculations to minimize the model error, while ensuring that the prediction error of the power consumption model is within 95% and does not exceed 3% within the confidence interval. The optimized power consumption signal model is obtained.

[0018] Step S200: creating a generalized power consumption control operator based on the power consumption signal model, and performing matrix decomposition on the power consumption characteristic data of the CPU card chip according to the generalized power consumption control operator to obtain a power consumption weight vector; Specifically, an energy consumption optimization objective function including a power consumption function and a control vector is defined according to the power consumption signal model. The function not only considers the power consumption characteristics of the CPU card chip, but also needs to balance the relationship between energy consumption and performance, and introduces a balance factor for adjustment, so that the optimization target can not only reduce power consumption, but also ensure that the computing performance meets the minimum threshold constraint. The optimization objective function includes the instantaneous power consumption described by the power consumption signal model, and adds a control vector to adjust the power consumption optimization process, so that it can not only meet the demand for energy consumption reduction, but also maintain the computing power of the CPU card chip not less than the set minimum performance threshold to ensure the normal operation of the chip. Based on the energy consumption optimization target and the minimum threshold constraint, a Lagrangian function containing Lagrangian multipliers is constructed, and the optimization target and constraint conditions are converted into mathematical optimization problems, so that it can take into account the minimization of power consumption and the maintenance of computing performance in the solution process. The Lagrangian function consists of an optimization objective function and a computing performance constraint, wherein the Lagrangian multiplier is used to adjust the adaptability of the optimization target to the constraint conditions, so that the computing performance always meets the set minimum requirements. When solving the optimization problem, the KKT condition is applied, that is, by analyzing the first-order conditions of the optimization problem, the optimal control vector is found to achieve a reasonable balance between power consumption and computing performance. The optimal control vector obtained by solving describes the optimization control relationship between power consumption and performance, and establishes the initial power consumption control operator. The initial power consumption control operator is designed as an adaptive structure combining the basic control part and the dynamic adjustment part, in which the basic control part provides the basic power consumption optimization strategy, and the dynamic adjustment part compensates the error through real-time feedback to enhance the adaptability of the control. The dynamic adjustment part consists of a proportional term and an integral term, in which the proportional term is used to quickly respond to the power consumption deviation, and the integral term is used to eliminate the steady-state error to ensure that the system always maintains the power consumption optimization target during long-term operation. Under the adaptive control structure, the power consumption control operator consists of a basic control operator and a dynamic adjustment term, in which the dynamic adjustment term depends on the real-time power consumption deviation and can be adjusted accordingly according to the operating state of the chip to improve the system's adaptability to different workloads. On this basis, a segmented control strategy is designed to ensure that the most appropriate control method is adopted within different power consumption deviation ranges. When the power consumption deviation is small, the linear control law is used for adjustment to keep the power consumption optimization process stable. When the power consumption deviation is large, the nonlinear control law is used to enhance the ability of rapid convergence under large deviation conditions. In order to determine the optimal parameters of the control strategy, the quadratic programming solution method is used to estimate the optimal parameters of the power consumption optimization target to ensure stable power consumption control performance under different load conditions. By optimizing the parameters, the generalized power consumption control operator can run stably under different load changes, thereby improving the power consumption control accuracy of the CPU card chip. Based on the generalized power consumption control operator, the power consumption characteristic data of the CPU card chip is matrix decomposed to calculate the power consumption weight vector.Construct a power consumption data matrix, calculate its feature correlation matrix, and then perform singular value decomposition on it to extract the main power consumption features and retain the key singular values ​​to reduce the computational complexity. Through this decomposition process, the weights of each power consumption feature are calculated to obtain a power consumption weight vector, which reflects the degree of influence of different power consumption features on the overall power consumption of the CPU chip. In the optimization process, a sparse regularization term is introduced to improve the stability of the power consumption weights and ensure that the final calculated weights can be applied to different workload scenarios. In the actual solution process, an optimization algorithm is used for iterative updates to ensure the convergence of the solution process and improve the accuracy of the final calculation results. The power consumption weight vector is obtained by the above method to describe the power consumption characteristics of the CPU card chip.

[0019] The power consumption characteristic data of the CPU card chip is calculated, and a power consumption characteristic correlation matrix is ​​constructed, which reflects the relationship between different power consumption characteristics, so that the main power consumption characteristics can be extracted through the matrix decomposition method in the future. In the calculation process, the power consumption characteristic data of the CPU card chip under different workload states, including frequency, voltage, temperature, core usage rate and other information, are collected, and the correlation between each feature is calculated based on these data to form a power consumption characteristic correlation matrix, which is used to describe the influence of different power consumption factors in the whole system. The power consumption characteristic correlation matrix is ​​subjected to singular value decomposition operation to obtain a diagonal matrix composed of singular values ​​and corresponding eigenvectors, wherein the singular values ​​represent the main characteristic components of the data matrix, and the eigenvectors are used to characterize the importance of each power consumption characteristic in different dimensions. An adaptive singular value threshold is set based on the distribution of singular values ​​to determine which singular values ​​are of practical significance for power consumption modeling. The setting of the adaptive singular value threshold is based on a statistical analysis method, such as adopting a median adjustment strategy, so that the threshold can change dynamically according to the data distribution, thereby adapting to different workload conditions. When the singular value is less than the set adaptive singular value threshold, these smaller singular values ​​are truncated to remove the features that contribute less to the power consumption modeling, thereby reducing the computational complexity and improving the robustness of the power consumption model. In this process, by retaining the main singular values ​​and ignoring the smaller singular values, the dimension can be effectively reduced and the truncated singular value matrix can be obtained. The mapping function is constructed using the truncated singular value matrix and the generalized power consumption control operator. The mapping function is used to describe the relationship between the power consumption characteristic data and the power consumption optimization control strategy, and the initial weight vector corresponding to the target power consumption vector is calculated by the reverse solution method. In this process, the generalized power consumption control operator is used to transform the reduced-dimensional data to ensure its adaptability under different load conditions. At the same time, the contribution of different power consumption characteristics to the overall power consumption optimization strategy is derived by the reverse mapping method. The initial weight vector obtained by this method is used to represent the importance of different power consumption characteristics. Sparse regularization constraints are introduced to the initial weight vector to enhance the interpretability of the power consumption weight calculation and reduce overfitting. By adding sparsity constraints to the optimization objectives, the power consumption weight vector obtained can more accurately reflect the impact of each power consumption feature on the overall power consumption control. The alternating direction multiplier method is used for iterative optimization and solution. The weight vector and constraint variables are updated separately in each iteration through step-by-step optimization, so as to gradually converge to the optimal solution. During the iterative optimization process, each update is fine-tuned based on the power consumption weight vector obtained by the current calculation to ensure the stability of the calculation results and reduce the deviation caused by data noise. Through multiple rounds of iterative optimization, the optimal power consumption weight vector is obtained, which is used to describe the role of different power consumption features in the energy consumption optimization process of the CPU card chip.

[0020] Step S300: using the power consumption weight vector, dual-state monitoring and feature selection are performed on the power consumption state vector and the performance state vector collected in real time to obtain a low-power sparse control matrix; Specifically, a state monitoring system including a power consumption state monitor and a performance state monitor is constructed to collect the power consumption state vector and performance state vector of the CPU card chip in real time. The power consumption state monitor mainly collects power consumption related parameters such as voltage, current, temperature, etc. of the CPU card chip, while the performance state monitor monitors computing performance indicators such as instruction execution rate, cache hit rate, and branch prediction accuracy. Through the collaborative work of these two monitors, the power consumption characteristics of the CPU under different workloads and their relationship with performance are obtained. On this basis, a dual-state fusion matrix is ​​designed based on the power consumption weight vector to map and transform the collected power consumption state vector and performance state vector, so that they are converted from the original data space to the control space, and the control space state vector is obtained to characterize the mapping relationship between power consumption and performance in the control space, and is used to optimize the construction of the control matrix. According to the control space state vector, a control matrix optimization problem with row sparse constraints is constructed to ensure that the trade-offs of different features in the power consumption optimization process can be adjusted under structured constraints. By imposing structured constraints on the control parameters, redundant parameters are effectively reduced, and the optimization process is made more stable, and an initial sparse control matrix is ​​obtained. Regularization terms are introduced into the initial sparse control matrix, and the power consumption error component is restricted to suppress abnormal changes in the power consumption control process and ensure the robustness of the optimization calculation. By introducing regularization terms, the complexity of the optimization model is reduced, the stability of the optimization results is improved, and the intermediate control matrix is ​​obtained. In order to enhance the adaptability of the control matrix, state-preserving regularization is applied to the low-dimensional power consumption data and the power consumption representation coefficient to ensure the smoothness of the state transition during the optimization process and reduce the mutation problem when the power consumption state changes. In the process of constructing the sparse control matrix, by applying state-preserving regularization, the control system can maintain the continuity of the power consumption optimization strategy when facing different workload changes, and reduce the mutation effect caused by control switching, so as to obtain a more stable sparse control structure. In order to finally solve the optimized control matrix, an alternating optimization algorithm is adopted, and the iterative method of optimizing the remaining variables by fixing some variables is adopted, so that the optimization problem can gradually converge to the optimal solution. In the iterative process, each update combines the power consumption weight vector and the sparse control structure to ensure the accuracy of the optimization calculation and avoid the optimization deviation caused by parameter estimation errors. Through the alternating optimization method, the complexity of the optimization calculation can be effectively reduced, and the convergence speed of the optimization process can be improved to obtain the final low-power sparse control matrix.

[0021] Step S400, inputting the low-power sparse control matrix into a five-layer fuzzy neural network for parameter optimization training to obtain a target inverse system model; Specifically, a five-layer fuzzy neural network structure including an input layer, a fuzzification layer, a rule layer, a defuzzification layer and an output layer is established. Among them, the input layer receives a low-power sparse control matrix and converts it into a power consumption target vector to ensure that the input data is compatible with the subsequent processing unit of the fuzzy neural network. Since the low-power sparse control matrix is ​​obtained through dual-state monitoring and feature selection, it contains the core control information of power consumption and performance. After conversion at the input layer, the adaptability and computational stability of the power consumption optimization strategy are effectively enhanced. Gaussian membership function nodes are constructed for the fuzzification layer of the five-layer fuzzy neural network structure to fuzzify the power consumption target vector. The input variables are mapped to fuzzy sets so that the power consumption state can be expressed in the form of membership. The design of the Gaussian membership function ensures the smoothness of the fuzzification process and improves the system's ability to describe the power consumption state, so that different power consumption characteristics are reasonably mapped to different fuzzy sets. In this process, each input variable calculates its membership value according to the predefined Gaussian membership function to form a representation of the power consumption state in the fuzzy space. After the fuzzification process is completed, based on the obtained membership value, the fuzzy rule set is implemented in the rule layer. The role of this rule set is to establish a mapping relationship between input fuzzy variables and output fuzzy variables to describe the correlation between different states in the power consumption control process. In the rule layer, each rule consists of multiple fuzzy variables, and the rule activation strength is calculated by the fuzzy reasoning method, that is, the contribution of each rule in the current input state. The activation strength of the rule depends on the membership of the input variable and the matching degree of the rule premise. Therefore, the fuzzy rules are reasonably designed according to the actual power consumption optimization requirements to ensure that the power consumption characteristics of different states can be accurately modeled. The output contribution weight of each rule is calculated in the defuzzification layer using the rule activation strength, and the rule output is weighted and combined to obtain the defuzzification result. The output contribution weight of each rule is determined by its corresponding rule activation strength, and the weighted combination process ensures that the comprehensive effect of all rules can be reflected in the final output result. Through reasonable defuzzification calculation, it is ensured that the reasoning results of the fuzzy neural network can meet the needs of power consumption optimization and can make accurate responses to different power consumption states. The defuzzification result is mapped to a control vector through the output layer, and combined with the nonlinear state equation of the original power consumption control system of the CPU card chip, a reversibility analysis is performed to build an initial inverse system model. Since the power consumption control system of the CPU card chip has nonlinear characteristics, when building the inverse system model, it is necessary to ensure the reversibility of the system, that is, the inverse system can accurately map the power consumption optimization target to the control parameter. During the reversibility analysis process, the relative order of the system is calculated, and it is verified whether the decoupling matrix of the system is reversible to ensure that the inverse system model can accurately describe the mapping relationship between the control strategy and the power consumption state. Through this process, the initial inverse system model is established.The differential evolution algorithm is used to iteratively optimize the parameters of the initial inverse system model, and the parameter population is updated through mutation operations, adaptive crossover operations, and selection operations to find the optimal control strategy. The optimization process of the differential evolution algorithm initializes the parameter population to cover a sufficiently wide solution space, and then performs a mutation operation in each round of iteration, that is, perturbing the existing parameter vector to generate a new candidate solution. Through the adaptive crossover operation, the mutated parameter vector is crossed with the original parameter vector to introduce diversity and enhance the optimization capability. Through the selection operation, the pros and cons of each candidate solution are evaluated according to the fitness function, and the parameter vector with higher fitness is retained to enter the next round of iteration. Through multiple rounds of iterative optimization, the optimal parameter combination is gradually approached, and finally the optimized target inverse system model is obtained.

[0022] Step S500: Connect the target inverse system model and the original power consumption control system in series to form a pseudo-linear system, and add a linear closed-loop controller to obtain a decoupled control system.

[0023] Specifically, the target inverse system model is connected in series with the original power consumption control system to form a pseudo-linear system with approximate unit matrix characteristics. Since the original power consumption control system has nonlinear characteristics, and the target inverse system model is an inverse mapping model constructed for the nonlinear characteristics, after the two are connected in series, their input-output relationship tends to be linearized, making the overall system more controllable in control calculation, while reducing complex nonlinear coupling problems. On this basis, in order to improve the stability and regulation ability of the system, a linear closed-loop controller including three terms of proportion, integration and differentiation is designed for the pseudo-linear system, and the controller gain matrix parameters are determined to obtain the basic closed-loop control structure. The proportional term is directly corrected according to the current error amount to improve the real-time response ability of the system, while the integral term is used to eliminate the steady-state error so that the system can accurately converge to the target power consumption level during long-term operation. At the same time, the differential term predicts the rate of change of the error signal to reduce the overshoot phenomenon that may occur during the power consumption adjustment process. In the process of determining the gain matrix parameters, the dynamic characteristics of the pseudo-linear system are combined for optimization to ensure that the controller can maintain stability under different power consumption states and can adapt to the dynamic load changes of the CPU card chip. After designing the basic closed-loop control structure, a multi-dimensional feedback control system including power consumption feedback loop and performance feedback loop is constructed to enhance the adaptive ability of the system and ensure that the computing performance will not be greatly affected during the power consumption optimization process. The power consumption feedback signal and performance feedback signal are calculated. The power consumption feedback signal is used to reflect the deviation between the actual power consumption and the target power consumption, while the performance feedback signal is used to monitor whether the computing performance of the system meets the preset requirements. Through the dual-loop feedback control network, a stable adaptive adjustment mechanism is formed inside the control system, so that it can take into account computing performance while optimizing power consumption and ensure the operation stability of the system. According to the error signal size of the dual-loop feedback control network, an adaptive switching strategy is designed so that the system activates conventional PID control or FNN inverse system control in different error ranges to realize an intelligent switching control mechanism. When the error is small, conventional PID control is used to maintain the stability of the system and reduce the computational complexity. When the error is large, FNN inverse system control is activated to optimize and adjust using more complex nonlinear mapping capabilities to speed up the convergence of the system and reduce the power consumption overshoot phenomenon under large error conditions. The introduction of the intelligent switching control mechanism enables the control system to dynamically adjust the control strategy under different working conditions, thereby improving the flexibility and energy efficiency optimization capabilities of the overall system. On this basis, in order to improve the adaptive ability of the system, an adaptive adjustment mechanism is implemented for the parameters of the linear closed-loop controller to dynamically adjust the control parameters according to the size, change rate and integral value of the error signal to obtain an adaptive control parameter sequence. This mechanism adjusts the proportional, integral and differential gain parameters by monitoring the changing trend of the error signal in real time to ensure that the control system can automatically optimize the control intensity according to different power consumption states.When the error signal is small, the role of the integral term is reduced to prevent the accumulation of integrals from causing the system to adjust excessively. When the error signal changes rapidly, the role of the differential term is enhanced to improve the dynamic response speed of the system. The adaptive adjustment mechanism can also dynamically correct the overall control gain according to the integral value of the historical error signal to improve the long-term stability of the system and ensure that the power consumption optimization strategy can always maintain high accuracy and stability under different working conditions. In order to improve the robustness of the control system and enhance its adaptability to external interference, a state observer and anti-disturbance mechanism of the CPU card chip are constructed to estimate the unmeasured state variables in real time and calculate the compensation control amount. The main function of the state observer is to infer the unmeasured variables inside the system based on the measured power consumption and performance parameters to improve the control accuracy of the system, while the anti-disturbance mechanism is used to detect the impact of external environmental changes on the CPU power consumption and calculate the corresponding compensation control amount to reduce the interference of external factors on the power consumption control effect. In this process, the combination of intelligent switching control mechanism and adaptive control parameter sequence enables the system to maintain stable operation in various complex environments and ensure the stability of computing performance while optimizing power consumption. Ultimately, a complete decoupled control system is obtained, which has efficient, stable and adaptive power consumption optimization capabilities.

[0024] In the embodiment of the present invention, by establishing a multi-core power consumption signal model and introducing a multi-core cooperative matrix, the accuracy of power consumption control is significantly improved; the power consumption weight vector calculation method based on singular value threshold enables the control system to automatically adjust the control parameters without accurately estimating the workload characteristics, thereby enhancing the adaptability of the system to different workloads; the designed segmented control strategy effectively solves the "blind spot" problem in traditional power consumption control, enabling the system to respond to small power consumption changes; regularization processing enhances the adaptability of the control system to ambient temperature changes; the combination of the fuzzy neural network inverse system and the improved differential evolution algorithm realizes the decoupling control of power consumption and performance, thereby reducing power consumption while ensuring performance; the state-keeping regularization on the low-dimensional power consumption data and the power consumption representation coefficient matrix ensures the smoothness of the power consumption state conversion process, thereby avoiding performance fluctuations and power consumption oscillations during state switching; the intelligent switching control mechanism and the adaptive control parameter sequence enhance the system's dynamic response capability to load mutations; the overall decoupling control system significantly improves the energy efficiency ratio of the CPU card chip compared to the traditional method, thereby providing a guarantee for efficient operation in various working environments.

[0025] In a specific embodiment, the process of executing step S100 may specifically include the following steps: Based on the original power consumption control system, the operating frequency, voltage, temperature, core usage, memory access frequency and instruction execution type of the CPU card chip are collected to obtain power consumption characteristic data; A power consumption signal acquisition system is constructed based on a temperature sensor array, a current monitoring unit, a voltage monitoring unit, and a workload analysis unit to sample and process power consumption characteristic data to obtain digital power consumption data. De-noising, normalizing and detecting outliers are performed on the digital power consumption data to obtain pre-processed power consumption data; Based on the pre-processed power consumption data, the power consumption mathematical model P(t)=α·f(t)·V is established 2 (t)+β·T(t)+γ·L(t)+ε(t), where P(t) is the power consumption mathematical model, f(t) is the CPU frequency, V(t) is the operating voltage, T(t) is the chip temperature, L(t) is the workload characteristic vector, α, β, γ are the corresponding weight coefficient matrices, and ε(t) is the error term; According to the power consumption coupling relationship between different cores, a multi-core coordination matrix M is established, and the power consumption mathematical model is corrected based on the multi-core coordination matrix to obtain the corrected power consumption model P(t)=M·[α·f(t)·V 2 (t)+β·T(t)+γ·L(t)]+ε(t); The least square method and cross validation method are used to optimize the parameters of the modified power consumption model and obtain the power consumption signal model.

[0026] Specifically, the key operating parameters of the chip are monitored based on the original power consumption control system. These parameters include operating frequency, voltage, temperature, core usage, memory access frequency, and instruction execution type. These factors jointly determine the power consumption characteristics of the chip. A complete power consumption signal acquisition system is built, which consists of a temperature sensor array, a current monitoring unit, a voltage monitoring unit, and a workload analysis unit. The temperature sensor array is used to detect the temperature changes of the CPU chip under different workloads in real time, while the current monitoring unit and the voltage monitoring unit collect the current and voltage information of the chip respectively to calculate the instantaneous power consumption, and the workload analysis unit extracts load characteristic information such as core usage, memory access frequency, and instruction execution type by tracking the task scheduling of the CPU, so as to establish a corresponding relationship between power consumption and working status. In the data acquisition process, a high-speed sampling circuit is used to digitize the power consumption characteristic data to ensure that the transient change characteristics of power consumption can be captured, and the original signal is preprocessed to reduce the influence of measurement errors and environmental noise. Data denoising mainly uses filtering methods to remove high-frequency interference components, while normalization processing is to map data of different physical quantities to a unified scale interval by standardizing the collected data to avoid the influence of variable scale differences on modeling. In order to improve data quality, outlier detection is performed, such as using statistical analysis methods to identify abnormal data points, and correcting or eliminating them according to the set anomaly detection threshold, so as to ensure the accuracy of the subsequent modeling process. Based on the preprocessed data, a mathematical power consumption model is established to describe the power consumption characteristics of the CPU card chip under different operating conditions. An improved power consumption mathematical model is used to more accurately describe the power consumption characteristics of the CPU. The basic form of the model is expressed as: in, Represents the instantaneous power consumption of the CPU chip, is the CPU operating frequency, is the operating voltage of the chip, Represents the temperature of the chip, is the workload feature vector, which includes multiple feature information such as task parallelism, cache hit rate, memory access bandwidth, etc. is the corresponding weight coefficient matrix, which indicates the influence of different factors on power consumption. is an error term, which is used to represent random fluctuations or measurement errors that cannot be accurately described by the model. In a multi-core CPU architecture, there is a coupling relationship between the power consumption of different cores, so the multi-core coordination matrix is ​​introduced. Describes the power consumption coupling relationship between cores. This matrix is ​​calculated based on factors such as task scheduling, cache sharing, and thermal impact between cores, and is used to correct the original power consumption mathematical model so that it can reflect the power consumption changes in a multi-core environment. The corrected model is expressed as: in, is a core coupling relationship matrix, each element of which represents the degree of interaction between power consumption of different cores. If the load of a core increases, it will affect other cores with which it has a resource sharing relationship, causing their power consumption to rise synchronously. In order to optimize the corrected power consumption mathematical model and ensure its generalization ability on different data sets, the least squares method and cross-validation method are used for parameter optimization. The goal of the least squares method is to minimize the sum of square errors between the model prediction value and the actual measurement value, so as to solve the optimal weight coefficient matrix α, β, γ and the core coordination matrix M, so that the fitting accuracy of the corrected model is optimized. The introduction of the cross-validation method is to evaluate the applicability of the model on different data sets, that is, to divide the data set into a training set and a test set, which are used for parameter estimation and model verification respectively, and by adjusting the complexity of the model, the prediction error of the final model on different data sets is minimized, thereby ensuring the robustness of the model.

[0027] In a specific embodiment, the process of executing step S200 may specifically include the following steps: According to the power consumption signal model, an energy consumption optimization objective function including a power consumption function and a control vector is defined, and a balance factor is introduced for adjustment to obtain an energy consumption optimization target, and a minimum threshold constraint condition is set for the computing performance of the CPU card chip; Constructing a Lagrangian function including Lagrangian multipliers based on the energy consumption optimization objective and the minimum threshold constraint; According to the Lagrangian function, the KKT condition is applied to solve the optimal control vector, establish the mapping relationship between power consumption and performance, and obtain the initial power consumption control operator; The initial power consumption control operator is designed as an adaptive structure combining a basic control part and a dynamic adjustment part, wherein the dynamic adjustment part includes a proportional term and an integral term, thereby obtaining an adaptive control operator; According to the power consumption deviation and the adaptive control operator, a segmented control strategy is designed. Linear control law and nonlinear control law are used for different deviation ranges. The optimal parameter coefficient is determined by quadratic programming to obtain the generalized power consumption control operator. According to the generalized power consumption control operator, the power consumption characteristic data of the CPU card chip is subjected to singular value decomposition to obtain a power consumption weight vector.

[0028] Specifically, an optimization objective function is constructed based on the power consumption signal model of the CPU card chip, so that it can minimize power consumption while ensuring computing performance. In the design process of the optimization target, the energy consumption characteristics of the CPU card chip are considered, and both instantaneous power consumption and overall energy efficiency must be optimized. The energy consumption optimization objective function is defined as follows: in, represents the optimization goal, Represents the instantaneous power consumption of the CPU card chip, To control the vector, control the working state of the chip, The balancing factor is used to weigh the control cost and the power consumption minimization goal. The introduction of the balancing factor ensures that the control cost of the system will not be too high due to excessive control while optimizing power consumption. In order to ensure that the computing performance will not be greatly affected during the optimization process, a minimum threshold constraint is set for the computing performance, that is, to ensure the computing power of the CPU. Not less than the set minimum value , so that: in, Represents the current computing performance status, It is the minimum performance requirement to ensure that the CPU can maintain a certain computing power even during the power consumption optimization process. On this basis, in order to solve the optimal control vector, a Lagrangian function containing Lagrangian multipliers is constructed to combine the optimization objective and constraints and transform it into an unconstrained optimization problem in the form of: in, is the Lagrangian multiplier, which is used to adjust the balance between the optimization target and the computing performance, so that the computing performance is always kept above the minimum threshold during the optimization process. According to the Lagrangian function, the KKT condition is applied to solve the optimal control vector , thereby establishing a mapping relationship between power consumption and performance, and finally obtaining the initial power consumption control operator: in, It is the initial power consumption control operator, which defines the optimal mapping relationship between the power consumption state and computing performance of the CPU, so that it can dynamically adjust the power consumption allocation of the CPU under different load conditions. The initial power consumption control operator is designed as an adaptive structure combining the basic control part and the dynamic adjustment part. Among them, the basic control part provides the basic strategy for global power consumption optimization, while the dynamic adjustment part is used to compensate according to the real-time error to enhance the adaptability of the system to different power consumption states. The dynamic adjustment part adopts the form of proportional-integral (PI) control, in which the proportional term is used to quickly respond to power consumption deviations, and the integral term is used to eliminate steady-state errors, so that the system can maintain a stable power consumption optimization effect during long-term operation. The adaptive control operator is obtained: in, As the basic control operator, It is a dynamic adjustment item, which is adjusted according to the real-time power consumption error. The specific calculation method is as follows: in, is the deviation between the actual power consumption and the target power consumption, represents the power consumption deviation at all times τ from the initial time to the current time t, is the proportional coefficient matrix, which controls the immediate correction of the error. is the integral coefficient matrix to ensure that the power consumption deviation can be eliminated in the long-term optimization process. On this basis, in order to improve the control accuracy, a segmented control strategy is designed according to the size of the power consumption deviation and the adaptive control operator, so that the system can adopt different control methods within different deviation ranges. When the power consumption deviation is small, a linear control law is used to keep the power consumption optimization process stable, and when the power consumption deviation is large, a nonlinear control law is used to enhance the ability to converge quickly under large deviations. In order to determine the optimal parameters of the control strategy, the quadratic programming solution method is used, and the optimization objective function is as follows: Solve the optimal proportional gain by optimization and integral gain , so that the control system can maintain stable operation under different power consumption states, and obtain the generalized power consumption control operator. In order to optimize the power consumption management strategy, based on the generalized power consumption control operator, the power consumption characteristic data of the CPU card chip is decomposed by singular value to calculate the power consumption weight vector. Construct the power consumption data matrix and calculate its feature correlation matrix: Then Perform singular value decomposition and get the decomposition form: in, is a diagonal matrix of singular values, and are the left singular matrix and the right singular matrix respectively. In order to reduce the computational complexity, only the matrices larger than the set threshold are retained. The main singular values ​​of get the truncated singular value matrix , and calculate the power consumption weight vector based on the truncated matrix: in, is the target power consumption vector, Represents the weight distribution of each power consumption feature. The optimization algorithm is used for iterative update to ensure the convergence of the solution process and improve the accuracy of the final calculation result.

[0029] In a specific embodiment, the execution step performs singular value decomposition on the power consumption characteristic data of the CPU card chip according to the generalized power consumption control operator to obtain the power consumption weight vector, which may specifically include the following steps: The power consumption characteristic correlation matrix is ​​calculated based on the power consumption characteristic data of the CPU card chip, and a singular value decomposition operation is performed on the power consumption characteristic correlation matrix to obtain a diagonal matrix composed of singular values ​​and corresponding eigenvectors; An adaptive singular value threshold is set based on the distribution of singular values, and singular values ​​smaller than the adaptive singular value threshold are truncated to obtain a truncated singular value matrix after dimensionality reduction; The mapping function is constructed using the truncated singular value matrix and the generalized power consumption control operator, and the target power consumption vector is reversely solved to obtain the initial weight vector; A sparse regularization constraint term is introduced into the initial weight vector, and the power consumption weight vector is obtained by iterative optimization through the alternating direction multiplier method.

[0030] Specifically, the power consumption characteristic correlation matrix is ​​calculated based on the power consumption characteristic data of the CPU card chip. Each row of the matrix represents the power consumption characteristics within a time window, and each column represents a different power consumption influencing factor. Based on the data matrix, the power consumption characteristic correlation matrix is ​​calculated, which is defined as follows: in, is the power consumption feature correlation matrix, is the power consumption data matrix, each element Representative The power consumption characteristics and The correlation between the power consumption features. Since there is linear correlation or redundancy between different features, the matrix is ​​subjected to singular value decomposition to extract the main power consumption features and remove redundant information. Through singular value decomposition, the following form is obtained: in, and are the left and right singular matrices, respectively, containing eigenvector information, and is a diagonal matrix, where the diagonal elements are singular values Arranged in descending order, each singular value represents the energy contribution of the corresponding feature. An adaptive singular value threshold is set based on the distribution of singular values ​​to determine which singular values ​​need to be retained and which need to be truncated. The calculation method of the adaptive singular value threshold is based on the median adjustment strategy, for example: in, represents the singular value threshold, is the dimension of the power consumption data matrix. All singular values ​​smaller than this threshold will be truncated to reduce computational complexity and improve the generalization ability of the model. The truncated singular value matrix is ​​expressed as: Among them, only the former singular values, and the rest are set to zero, thereby reducing the data dimension and removing redundant features. After completing the dimensionality reduction, the mapping function is constructed using the truncated singular value matrix and the generalized power consumption control operator to establish a power consumption optimization model, and the target power consumption vector is reversely solved to calculate the initial power consumption weight vector. The construction of the mapping function is based on the characteristics of the generalized power consumption control operator, that is, it describes the relationship between power consumption characteristics and optimization control, so the initial weight vector is solved by the following calculation: in, represents the initial power consumption weight vector, and are the right and left singular matrices after truncation, respectively. is the target power consumption vector. Through this mapping process, a preliminary estimated power consumption weight is obtained. In order to optimize the initial power consumption weight vector, a sparse regularization constraint term is introduced to enhance its interpretability and reduce overfitting. The optimization objective is expressed as: Among them, the first term represents the fitting error, which ensures that the calculated power consumption weight can accurately predict the power consumption, and the second term is The regularization term keeps the weight vector sparse, so that only the most important power consumption features are selected. The third term is Regularization term is used to enhance the stability of weights and reduce noise interference. In order to solve this optimization problem, the alternating direction multiplier method is used for iterative optimization. Auxiliary variables are introduced in the optimization process and Lagrange multipliers , and perform step-by-step optimization. The iterative update process is as follows: in, is the adjustment parameter of the alternating direction multiplier method, It is a soft threshold operator used to enforce sparse constraints, making unimportant power consumption weights gradually approach zero. This iterative process continuously optimizes the power consumption weight vector, making it more sparse and stable while ensuring the fitting accuracy.

[0031] In a specific embodiment, the process of executing step S300 may specifically include the following steps: Build a status monitoring system including power consumption status monitor and performance status monitor, and collect power consumption status vector and performance status vector of CPU card chip in real time through the status monitoring system; Design a dual-state fusion matrix based on the power consumption weight vector, map and transform the power consumption state vector and the performance state vector to obtain the control space state vector; According to the control space state vector, a control matrix optimization problem with row sparse constraints is constructed, and structural constraints are imposed on the control parameters to obtain an initial sparse control matrix; A regularization term is introduced into the initial sparse control matrix, and the power consumption error component is restricted to obtain an intermediate control matrix. Based on the intermediate control matrix, a state-preserving regularization process is applied to the low-dimensional power consumption data and the power consumption representation coefficient to obtain a sparse control structure. The alternating optimization algorithm is used to solve the control matrix optimization problem. The iterative method of optimizing the remaining variables by fixing some variables is used. The feedback control law is designed by combining the power consumption weight vector and the sparse control structure to obtain a low-power sparse control matrix.

[0032] Specifically, a status monitoring system is designed, which includes two main parts, namely, the power consumption status monitor and the performance status monitor. The power consumption status monitor is used to collect power consumption related information of the CPU card chip in real time, such as voltage, current, temperature, operating frequency, etc., while the performance status monitor is used to obtain the CPU's computing performance indicators, such as instruction execution rate, cache hit rate, memory access bandwidth, etc. Through the collaborative work of these two monitors, a CPU operation status monitoring framework is constructed, so that the system can obtain the CPU's power consumption state vector and performance state vector at different times, which are expressed as and After obtaining these state vectors, a dual-state fusion matrix is ​​designed based on the power consumption weight vector to map and transform the power consumption state vector and the performance state vector, and obtain the control space state vector. Due to the different influences of different state variables, the power consumption weight vector is used in the mapping process. Weighting is performed to ensure that the mapped state vector can accurately describe the power consumption optimization target. The calculation of the dual-state fusion matrix is ​​expressed as: in, represents the control space state vector, is a two-state fusion matrix, each element of which consists of a power consumption weight vector The calculation describes the degree of influence of different state variables on the overall control strategy. The construction of this matrix can ensure that the control space state vector not only contains power consumption state information, but also reflects the changes in CPU computing performance, so that the subsequent power consumption optimization process can take into account the dual goals of power consumption and performance. Based on the control space state vector, a row-sparse constrained control matrix optimization problem is constructed to optimize the control parameters and obtain the initial sparse control matrix. When constructing the optimization problem, structured constraints are imposed to ensure that the optimized control matrix is ​​sparse, thereby reducing the computational complexity and improving the interpretability of the optimization results. The objective function of the optimization problem is expressed as: in, is the power consumption weight matrix, is the control matrix to be optimized, is the target state matrix, is the sparsification parameter. Regulatory constraints , that is, for each row The norm is summed to ensure that the optimized control matrix is ​​sparse, so that the redundant control variables are automatically weakened. A regularization term is introduced into the initial sparse control matrix to restrict the power consumption error component to obtain an intermediate control matrix. The purpose of regularization is to suppress the influence of abnormal power consumption state on the optimization results and enhance the robustness of the model. The optimization objective is expressed as: in, is the power consumption error matrix, is the error regularization parameter, which ensures that the control strategy will not be offset due to a small amount of abnormal data during the optimization process. The introduction of this regularization term can ensure that the optimization process is more robust and reduce the sensitivity of the power control strategy to environmental noise. The intermediate control matrix is ​​optimized to ensure that the optimized control matrix maintains consistency in low-dimensional power consumption data and power consumption representation coefficients, thereby improving the stability of the power control strategy. During the optimization process, a state-preserving regularization constraint is imposed to keep the control matrix changes smooth under different states. The optimization goal is: in, is the low-dimensional power consumption data matrix, is the power consumption coefficient matrix, is the power consumption coefficient matrix of the previous moment, is the state holding parameter. This optimization goal can ensure the smoothness of power state transition and avoid performance fluctuations due to drastic adjustments of the control strategy. In order to solve the above optimization problem and obtain the final low-power sparse control matrix, an alternating optimization algorithm is used to optimize the remaining variables by fixing some variables, so as to gradually converge to the optimal solution. In the optimization process, each round of iteration optimizes the control matrix And power consumption coefficient , so that the optimization problem can be solved efficiently. The optimization process is expressed as: By continuously optimizing and , and finally the optimized control matrix is ​​obtained. The feedback control law is designed by combining the power consumption weight vector and the sparse control structure so that it can be adjusted according to the actual power consumption state. The final low-power sparse control matrix is ​​expressed as: In this embodiment, after obtaining the low-power sparse control matrix and before inputting the low-power sparse control matrix into the five-layer fuzzy neural network, the dynamic balanced allocation of multi-core power consumption of the CPU card chip is performed, including: performing power consumption distribution analysis among multiple cores on the low-power sparse control matrix, establishing a power consumption state vector group of each core, and obtaining a core power consumption distribution map; designing a power consumption imbalance detection threshold among cores based on the core power consumption distribution map, marking core pairs whose power consumption difference exceeds the threshold, and obtaining a power consumption imbalance core group; hierarchically sorting each core in the power consumption imbalance core group according to workload importance and thermal sensitivity, establishing a core power consumption priority queue, and obtaining a core power consumption adjustment priority sequence; constructing a dynamic power consumption migration algorithm according to the core power consumption adjustment priority sequence, calculating the power consumption that can be migrated among each core and the optimal migration path, and obtaining a power consumption migration mapping table; based on the power consumption The power migration mapping table is used to migrate some tasks of the high-load core, and the frequency and voltage of the low-load core are compensated to obtain the balanced core power consumption distribution; a temperature gradient model of each core in the CPU card chip is established, and the potential hot spot area is predicted in combination with the power consumption change trend, and the working frequency and task allocation strategy are dynamically adjusted to obtain the hot spot prevention control instruction; according to the real-time workload fluctuation and application characteristics, a multi-core collaborative working mode including high-performance mode, balanced mode and low-power mode is designed, and a smooth switching mechanism between modes is constructed to obtain a multi-mode switching control strategy; the balanced core power consumption distribution, hot spot prevention control instructions and multi-mode switching control strategy are comprehensively processed to generate a multi-core power consumption balancing control instruction sequence, and the multi-core power consumption balancing control instruction sequence is merged with the low-power sparse control matrix to obtain an enhanced low-power sparse control matrix.

[0033] In a specific embodiment, the process of executing step S400 may specifically include the following steps: A five-layer fuzzy neural network structure including an input layer, a fuzzification layer, a rule layer, a defuzzification layer and an output layer is established. The input layer is used to receive the low-power sparse control matrix and convert it into a power consumption target vector. Gaussian membership function nodes are constructed for the fuzzification layer of the five-layer fuzzy neural network structure, and the power consumption target vector is fuzzified to obtain the membership value of the power consumption state; Based on the membership value, the fuzzy rule set is realized at the rule layer, and the mapping relationship between the input fuzzy variables and the output fuzzy variables is established to obtain the rule activation strength; The output contribution weight of each rule is calculated in the defuzzification layer using the rule activation strength, and the rule output is weighted combined to obtain the defuzzification result. The defuzzification result is mapped into a control vector through the output layer, and the reversibility is analyzed by combining the nonlinear state equation of the original power consumption control system of the CPU card chip to obtain the initial inverse system model; The differential evolution algorithm is used to iteratively optimize the parameters of the initial inverse system model. The parameter population is updated through mutation operation, adaptive crossover operation and selection operation to obtain the target inverse system model.

[0034] Specifically, a five-layer fuzzy neural network is established, including an input layer, a fuzzification layer, a rule layer, a defuzzification layer, and an output layer. The input layer is used to receive the low-power sparse control matrix and convert it into the power consumption target vector , which contains the power consumption state parameters after power consumption optimization calculation. Since multiple state variables are involved in the power consumption optimization process, the input layer maps the low-power sparse control matrix to ensure that the input feature dimension matches the structure of the fuzzy neural network, so that it can efficiently handle the power consumption optimization problem. After the input layer completes the data conversion, the power consumption target vector is fuzzified in the fuzzification layer so that the uncertainty of the power consumption state can be more accurately described in the subsequent reasoning process. The key to fuzzification processing is to construct a Gaussian membership function node, map the input variable to the fuzzy set through the Gaussian membership function, and calculate its membership value. The expression of the Gaussian membership function is as follows: in, Represents input variables In fuzzy sets The membership degree under is the center of the Gaussian function, is the standard deviation of the Gaussian function. By calculating the membership value of each input variable, the fuzzified information is further used for reasoning in the rule layer. In the rule layer, a fuzzy rule set is established based on the membership value to describe the mapping relationship between the input fuzzy variables and the output fuzzy variables. In the actual calculation process, the rule layer calculates the rule activation strength according to the membership value of the input variable, and the calculation method is as follows: in, Representative The activation strength of the rule is the product of the membership of all input variables under the rule, which indicates the contribution of the rule under the current input state. After obtaining the rule activation strength, the output contribution weight of each rule is calculated in the defuzzification layer, and the rule output is weighted and combined to generate the final defuzzification result. Defuzzification calculates the weighted average based on the rule activation strength, so that the final output can accurately reflect the comprehensive effect of all rules. The calculation formula is as follows: in, represents the final defuzzification result, It's a rule The output, is the corresponding rule activation strength. In this way, the final output is ensured to be the weighted average of all rule contributions, thereby obtaining a stable control signal. After the defuzzification result is calculated, it is mapped to a control vector through the output layer, and combined with the nonlinear state equation of the original power consumption control system of the CPU card chip to analyze the reversibility and construct the initial inverse system model. Since the CPU power consumption control system has strong nonlinear characteristics, the construction of the inverse system needs to ensure that it can accurately describe the mapping relationship between the input power consumption control signal and the output power consumption state. The goal of the inverse system is to solve the control signal and power consumption state The inverse mapping relationship between them makes: in, is the inverse system model, which is used to convert the power consumption target Inverse mapping to control signal , to ensure that the system can optimize power consumption under different load conditions. In order to solve the inverse system model, a data-driven method is adopted, and the nonlinear approximation ability of the neural network is combined to build a model. In order to optimize the initial inverse system model, the differential evolution algorithm is used to iteratively optimize the parameters of the model, and the parameter population is continuously updated through mutation operations, adaptive crossover operations, and selection operations to find the optimal control strategy. In the optimization process of the differential evolution algorithm, the parameter population is initialized to cover a sufficiently wide solution space, and then the mutation operation is performed in each round of iteration, that is, the existing parameter vector is perturbed to generate a new candidate solution. The mutation formula is as follows: in, It is a mutated individual. is a randomly selected individual, are other randomly selected individuals, is the mutation factor, which is used to adjust the mutation amplitude. Through the adaptive crossover operation, the mutated parameter vector is crossed with the original parameter vector to introduce diversity and enhance the optimization ability. The calculation of the crossover operation is as follows: in, is the individual after the crossover, is the crossover probability, which is used to control the probability of parameter replacement after mutation. Through the selection operation, the quality of each candidate solution is evaluated according to the fitness function, and the parameter vector with higher fitness is retained to enter the next round of generation. The selection criteria are as follows: in, Represents the fitness function, and the individual with the minimum fitness is selected as the next generation parameter.

[0035] In a specific embodiment, the process of executing step S500 may specifically include the following steps: The target inverse system model is connected in series with the original power consumption control system to form a pseudo-linear system with approximate unit matrix characteristics; Design a linear closed-loop controller containing proportional, integral and differential terms for the pseudo-linear system, determine the controller gain matrix parameters, and obtain the basic closed-loop control structure; Based on the basic closed-loop control structure, a multi-dimensional feedback control system including a power consumption feedback loop and a performance feedback loop is constructed, and the power consumption feedback signal and the performance feedback signal are calculated to obtain a dual-loop feedback control network; According to the error signal size of the dual-loop feedback control network, an adaptive switching strategy is designed, and conventional PID control or FNN inverse system control is activated in different error ranges to obtain an intelligent switching control mechanism. An adaptive adjustment mechanism is implemented for the parameters of the linear closed-loop controller, and the control parameters are dynamically adjusted according to the size, change rate and integral value of the error signal to obtain an adaptive control parameter sequence; The state observer and anti-disturbance mechanism of the CPU card chip are constructed to estimate the unmeasured state variables in real time and calculate the compensation control quantity. The decoupled control system is obtained by combining the intelligent switching control mechanism and the adaptive control parameter sequence.

[0036] Specifically, the target inverse system model is connected in series with the original power consumption control system to eliminate the nonlinear effects of the original system and make the dynamic response characteristics of the overall system tend to be linear. The original power consumption control system exhibits complex nonlinear dynamic characteristics, which makes it difficult for traditional linear control methods to accurately optimize power consumption. The role of the target inverse system model is to provide an equivalent inverse transformation, so that the entire system can achieve the desired effect under the control signal. and power output The relationship between them forms an approximate unit matrix, that is: in, represents the original power consumption control system, is the target inverse system model, is the unit matrix. In this way, the complex nonlinear dynamic characteristics of the system are approximately eliminated, so that the subsequent controller design can be carried out in the framework of a pseudo-linear system, thereby simplifying the implementation of the control strategy. After obtaining the pseudo-linear system, a linear closed-loop controller containing three terms, proportional, integral and differential, is designed for it to ensure that the system can maintain stability during the power consumption optimization process and respond quickly to dynamic load changes. The design of the PID controller requires determining the gain matrix parameters to optimize the dynamic characteristics of the system. The control law of the PID controller is expressed as: in, is the proportional gain, the response speed of the control system, is the integral gain, ensuring the minimum steady-state error, The differential gain is used to suppress overshoot caused by sudden changes in power consumption. In order to optimize the PID parameter matrix, the optimal gain parameter is solved based on the dynamic characteristics of the system to optimize the performance index of the closed-loop system. The optimization objective is expressed as: in, is the error signal, is a regularization factor used to control the amplitude of the gain parameter to prevent the controller from over-responding to power consumption changes and causing system instability. After obtaining the basic closed-loop control structure, a multi-dimensional feedback control system including a power consumption feedback loop and a performance feedback loop is constructed to ensure that computing performance is maintained while optimizing power consumption. The power consumption feedback loop is used to monitor the actual power consumption of the system and compare it with the target power consumption to calculate the power consumption feedback signal: in, is the target power consumption, is the actual measured power consumption. The performance feedback loop is used to ensure that the computing performance does not drop below the set threshold. The error is calculated as follows: in, To calculate the performance of the target, is the actual measured computing performance. Based on the power consumption feedback signal and the performance feedback signal, a dual-loop feedback control network is constructed, so that the system can dynamically adjust the control strategy according to the real-time feedback signal, thereby achieving a balance between energy efficiency optimization and computing performance stability. In order to enhance the intelligence of the system, an adaptive switching strategy is designed based on the error signal size of the dual-loop feedback control network, so that the system can activate conventional PID control or FNN inverse system control in different error ranges. When the error signal is small, conventional PID control is used to reduce the computational complexity and maintain the stability of the system. When the error signal is large, FNN inverse system control is activated, so that the fuzzy neural network can use nonlinear mapping capabilities to quickly adjust the control parameters to speed up the error convergence. The decision function of the switching strategy is expressed as: in, is the error threshold, represents a conventional PID controller, Represents fuzzy neural network controller. Through this intelligent switching control mechanism, the system can ensure that the control method with lower computational complexity is adopted within the small error range, and a more powerful intelligent control strategy is adopted within the large error range, thereby improving the control accuracy and reducing the computational burden. In order to optimize the dynamic performance of the controller, an adaptive adjustment mechanism is implemented for the parameters of the linear closed-loop controller to dynamically adjust the control parameters according to the size, change rate and integral value of the error signal to obtain an adaptive control parameter sequence. The calculation method of this adjustment mechanism is as follows: in, is the initial gain value, The adaptive adjustment coefficient ensures that the gain parameter can be automatically adjusted with the change of the error signal to improve the adaptability of the control system and reduce overshoot and oscillation in the energy consumption optimization process. In order to enhance the anti-interference ability of the system and improve the estimation accuracy of the unmeasured state variables, the state observer and anti-disturbance mechanism of the CPU card chip are constructed to calculate the compensation control quantity, and combined with the intelligent switching control mechanism and the adaptive control parameter sequence, a complete decoupling control system is finally formed. The role of the state observer is to use the measured data to infer the internal state variables of the system, such as the unmeasured temperature distribution, internal power consumption distribution, etc. Its state estimation equation is expressed as: in, is the estimated state, is the gain matrix, is the measured value. The anti-disturbance mechanism is used to detect external disturbances And calculate the compensation control amount: By combining state observers, anti-disturbance mechanisms, adaptive control parameter sequences, and intelligent switching control mechanisms, the system can achieve efficient power consumption optimization and ensure stable computing performance under different load conditions.

[0037] See also Figure 2 , Figure 2 A schematic block diagram of the structure of the adaptive power consumption control system 200 of the CPU card chip provided in the embodiment of the present application is shown as follows: Figure 2 As shown, the adaptive power consumption control system 200 of the CPU card chip includes: The acquisition module 210 is used to acquire the power consumption characteristic data of the CPU card chip under different workloads based on the original power consumption control system, and to construct a power consumption signal model; A creation module 220 is used to create a generalized power consumption control operator based on the power consumption signal model, and perform matrix decomposition on the power consumption characteristic data of the CPU card chip according to the generalized power consumption control operator to obtain a power consumption weight vector; The feature selection module 230 is used to use the power consumption weight vector to perform dual-state monitoring and feature selection on the power consumption state vector and performance state vector collected in real time to obtain a low-power sparse control matrix; A training module 240 is used to input the low-power sparse control matrix into a five-layer fuzzy neural network for parameter optimization training to obtain a target inverse system model; The series connection module 250 is used to connect the target inverse system model and the original power consumption control system in series to form a pseudo-linear system, and add a linear closed-loop controller to obtain a decoupled control system.

[0038] Through the cooperation of the above components, the accuracy of power consumption control is significantly improved by establishing a multi-core power consumption signal model and introducing a multi-core cooperative matrix; the power consumption weight vector calculation method based on singular value threshold enables the control system to automatically adjust the control parameters without accurately estimating the workload characteristics, thereby enhancing the system's adaptability to different workloads; the designed segmented control strategy effectively solves the "blind spot" problem in traditional power consumption control, enabling the system to respond to small power consumption changes; regularization processing enhances the control system's adaptability to ambient temperature changes; the combination of the fuzzy neural network inverse system and the improved differential evolution algorithm realizes the decoupling control of power consumption and performance, reducing power consumption while ensuring performance; the state-preserving regularization on the low-dimensional power consumption data and the power consumption representation coefficient matrix ensures the smoothness of the power consumption state transition process, avoiding performance fluctuations and power consumption oscillations during state switching; the intelligent switching control mechanism and the adaptive control parameter sequence enhance the system's dynamic response capability to load mutations; the overall decoupling control system significantly improves the energy efficiency of the CPU card chip compared to the traditional method, providing a guarantee for efficient operation in various working environments.

[0039] 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 aforementioned method embodiments and will not be repeated here.

[0040] 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 this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0041] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for adaptive power consumption control of a CPU card chip, characterized in that: include: Based on the original power consumption control system, the power consumption characteristic data of the CPU card chip under different workloads is collected, and a power consumption signal model is constructed; Creating a generalized power consumption control operator based on the power consumption signal model, and performing matrix decomposition on the power consumption characteristic data of the CPU card chip according to the generalized power consumption control operator to obtain a power consumption weight vector; Using the power consumption weight vector, dual-state monitoring and feature selection are performed on the power consumption state vector and performance state vector collected in real time to obtain a low-power sparse control matrix; Inputting the low-power sparse control matrix into a five-layer fuzzy neural network for parameter optimization training to obtain a target inverse system model; The target inverse system model and the original power consumption control system are connected in series to form a pseudo-linear system, and a linear closed-loop controller is added to obtain a decoupling control system.

2. The method for adaptive power consumption control of a CPU card chip according to claim 1, characterized in that: The original power consumption control system collects power consumption characteristic data of the CPU card chip under different workload states and constructs a power consumption signal model, including: Based on the original power consumption control system, the operating frequency, voltage, temperature, core usage, memory access frequency and instruction execution type of the CPU card chip are collected to obtain power consumption characteristic data; A power consumption signal acquisition system is constructed based on a temperature sensor array, a current monitoring unit, a voltage monitoring unit and a workload analysis unit, and the power consumption characteristic data is sampled and processed to obtain digital power consumption data; Performing denoising, normalization and outlier detection processing on the digitized power consumption data to obtain pre-processed power consumption data; Based on the pre-processed power consumption data, a power consumption mathematical model P(t)=α·f(t)·V is established 2 (t)+β·T(t)+γ·L(t)+ε(t), where P(t) is the power consumption mathematical model, f(t) is the CPU frequency, V(t) is the operating voltage, T(t) is the chip temperature, L(t) is the workload characteristic vector, α, β, γ are the corresponding weight coefficient matrices, and ε(t) is the error term; A multi-core coordination matrix M is established according to the power consumption coupling relationship between different cores, and the power consumption mathematical model is corrected based on the multi-core coordination matrix to obtain a corrected power consumption model P(t)=M·[α·f(t)·V 2 (t)+β·T(t)+γ·L(t)]+ε(t); The least square method and cross validation method are used to optimize the parameters of the modified power consumption model to obtain a power consumption signal model.

3. The method for adaptive power consumption control of a CPU card chip according to claim 1, characterized in that: The generalized power consumption control operator is created based on the power consumption signal model, and the power consumption characteristic data of the CPU card chip is matrix decomposed according to the generalized power consumption control operator to obtain a power consumption weight vector, including: An energy consumption optimization objective function including a power consumption function and a control vector is defined according to the power consumption signal model, and a balance factor is introduced for adjustment to obtain an energy consumption optimization target, and a minimum threshold constraint condition is set for the computing performance of the CPU card chip; Constructing a Lagrangian function including Lagrangian multipliers based on the energy consumption optimization target and the minimum threshold constraint condition; Applying the KKT condition according to the Lagrangian function to solve the optimal control vector, establish a mapping relationship between power consumption and performance, and obtain an initial power consumption control operator; The initial power consumption control operator is designed as an adaptive structure combining a basic control part and a dynamic adjustment part, wherein the dynamic adjustment part includes a proportional term and an integral term, to obtain an adaptive control operator; According to the power consumption deviation and the adaptive control operator, a segmented control strategy is designed, and a linear control law and a nonlinear control law are respectively adopted for different deviation ranges, and the optimal parameter coefficient is determined by quadratic programming to obtain a generalized power consumption control operator; According to the generalized power consumption control operator, the power consumption characteristic data of the CPU card chip is subjected to singular value decomposition to obtain a power consumption weight vector.

4. The method for adaptive power consumption control of a CPU card chip according to claim 3, characterized in that: The method of performing singular value decomposition on the power consumption characteristic data of the CPU card chip according to the generalized power consumption control operator to obtain a power consumption weight vector includes: Calculating a power consumption characteristic correlation matrix based on the power consumption characteristic data of the CPU card chip, and performing a singular value decomposition operation on the power consumption characteristic correlation matrix to obtain a diagonal matrix composed of singular values ​​and corresponding eigenvectors; An adaptive singular value threshold is set based on the distribution of singular values, and singular values ​​smaller than the adaptive singular value threshold are truncated to obtain a truncated singular value matrix after dimensionality reduction; A mapping function is constructed using the truncated singular value matrix and the generalized power consumption control operator, and a target power consumption vector is reversely solved to obtain an initial weight vector; A sparse regularization constraint term is introduced into the initial weight vector, and an iterative optimization solution is performed through an alternating direction multiplier method to obtain a power consumption weight vector.

5. The method for adaptive power consumption control of a CPU card chip according to claim 1, characterized in that: The method uses the power consumption weight vector to perform dual-state monitoring and feature selection on the power consumption state vector and the performance state vector collected in real time to obtain a low-power sparse control matrix, including: Construct a status monitoring system including a power consumption status monitor and a performance status monitor, and collect the power consumption status vector and performance status vector of the CPU card chip in real time through the status monitoring system; Designing a dual-state fusion matrix based on the power consumption weight vector, mapping and transforming the power consumption state vector and the performance state vector to obtain a control space state vector; Constructing a control matrix optimization problem with row sparse constraints according to the control space state vector, applying structural constraints to control parameters, and obtaining an initial sparse control matrix; A regularization term is introduced into the initial sparse control matrix, and a power consumption error component is restricted to obtain an intermediate control matrix; Based on the intermediate control matrix, a state-preserving regularization process is applied to the low-dimensional power consumption data and the power consumption representation coefficient to obtain a sparse control structure; An alternating optimization algorithm is used to solve the control matrix optimization problem. The remaining variables are optimized iteratively by fixing some variables. The power consumption weight vector and the sparse control structure are combined to design a feedback control law to obtain a low-power sparse control matrix.

6. The method for adaptive power consumption control of a CPU card chip according to claim 1, characterized in that: The low-power sparse control matrix is ​​input into a five-layer fuzzy neural network for parameter optimization training to obtain a target inverse system model, including: Establishing a five-layer fuzzy neural network structure including an input layer, a fuzzification layer, a rule layer, a defuzzification layer and an output layer, wherein the input layer is used to receive the low-power sparse control matrix and convert it to obtain a power consumption target vector; Constructing Gaussian membership function nodes for the fuzzification layer of the five-layer fuzzy neural network structure, performing fuzzification processing on the power consumption target vector, and obtaining a membership value of the power consumption state; Implementing a fuzzy rule set at the rule layer based on the membership value, establishing a mapping relationship between input fuzzy variables and output fuzzy variables, and obtaining rule activation strength; The output contribution weight of each rule is calculated in the defuzzification layer using the rule activation strength, and the rule output is weighted combined to obtain a defuzzification result; The defuzzification result is mapped into a control vector through the output layer, and the reversibility is analyzed by combining the nonlinear state equation of the original power consumption control system of the CPU card chip to obtain an initial inverse system model; The parameters of the initial inverse system model are iteratively optimized using a differential evolution algorithm, and the parameter population is updated through mutation operations, adaptive crossover operations, and selection operations to obtain a target inverse system model.

7. The method for adaptive power consumption control of a CPU card chip according to claim 1, characterized in that: The target inverse system model is connected in series with the original power consumption control system to form a pseudo-linear system, and a linear closed-loop controller is added to obtain a decoupling control system, including: Connecting the target inverse system model and the original power consumption control system in series to form a pseudo-linear system with approximate unit matrix characteristics; Designing a linear closed-loop controller including proportional, integral and differential terms for the pseudo-linear system, determining controller gain matrix parameters, and obtaining a basic closed-loop control structure; Based on the basic closed-loop control structure, a multi-dimensional feedback control system including a power consumption feedback loop and a performance feedback loop is constructed, and a power consumption feedback signal and a performance feedback signal are calculated to obtain a dual-loop feedback control network; Design an adaptive switching strategy according to the error signal size of the dual-loop feedback control network, activate conventional PID control or FNN inverse system control in different error ranges, and obtain an intelligent switching control mechanism; An adaptive adjustment mechanism is implemented for the parameters of the linear closed-loop controller, and the control parameters are dynamically adjusted according to the size, change rate and integral value of the error signal to obtain an adaptive control parameter sequence; A state observer and anti-disturbance mechanism of the CPU card chip are constructed, unmeasured state variables are estimated in real time and compensation control quantities are calculated, and a decoupling control system is obtained by combining the intelligent switching control mechanism and the adaptive control parameter sequence.

8. An adaptive power consumption control system for a CPU card chip, characterized in that: The method for executing the adaptive power consumption control method of the CPU card chip according to any one of claims 1 to 7, wherein the adaptive power consumption control system of the CPU card chip comprises: The acquisition module is used to collect the power consumption characteristic data of the CPU card chip under different workloads based on the original power consumption control system and build a power consumption signal model; A creation module is used to create a generalized power consumption control operator based on the power consumption signal model, and perform matrix decomposition on the power consumption characteristic data of the CPU card chip according to the generalized power consumption control operator to obtain a power consumption weight vector; A feature selection module, used to use the power consumption weight vector to perform dual-state monitoring and feature selection on the power consumption state vector and performance state vector collected in real time to obtain a low-power sparse control matrix; A training module, used for inputting the low-power sparse control matrix into a five-layer fuzzy neural network for parameter optimization training to obtain a target inverse system model; The series connection module is used to connect the target inverse system model and the original power consumption control system in series to form a pseudo-linear system, and add a linear closed-loop controller to obtain a decoupled control system.

Images