Mainboard power consumption adjusting method and system

Abstract

The invention provides a mainboard power consumption adjusting method and system, and relates to the technical field of mainboards, and the mainboard power consumption adjusting method comprises the following steps: collecting current real-time data of a mainboard in real time, and carrying out fuzzification processing on the current real-time data to generate a multi-dimensional membership matrix; inputting a preset fuzzy rule base to generate an initial power supply strategy set; inputting the current real-time data and the historical load time sequence into a pre-constructed lightweight neural network, and outputting a rule weight correction amount and a membership function parameter offset amount; based on the initial power supply strategy set, calculating a rule weight correction amount and a membership function parameter offset amount through superposition, dynamically updating a preset weight of a fuzzy rule base and a membership function parameter, and generating an optimized power supply strategy set; and inputting the optimized power supply strategy set into a multi-target optimizer, generating a CPU frequency regulation instruction and a peripheral equipment power supply instruction, and issuing and executing the instructions, thereby realizing dynamic regulation of the power consumption of the mainboard, effectively reducing the power consumption of the system, and improving the stability and the reliability of the system.

Description

Technical Field

[0001] The present invention relates to the technical field of motherboards, and in particular to a motherboard power consumption adjustment method and system thereof. Background Art

[0002] With the rapid development of information technology, the performance of computer hardware has been continuously improved, but at the same time, the problem of motherboard power consumption has become increasingly prominent. High power consumption will not only increase energy consumption and increase operating costs, but may also cause hardware overheating and affect the stability and reliability of the system. Therefore, how to effectively adjust the power consumption of the motherboard to achieve energy saving and consumption reduction has become a technical problem that needs to be solved urgently.

[0003] Traditional motherboard power consumption adjustment methods are mostly based on static preset rules, which use fixed thresholds to determine the system status and then adjust the power supply strategy. However, this method often ignores the real-time and complexity of the system and cannot be flexibly adjusted according to the dynamic changes of the system, resulting in poor power consumption adjustment effects. In addition, although some advanced methods use machine learning technology, they are often difficult to promote in practical applications due to the high model complexity and large amount of calculation.

[0004] Therefore, it is necessary to provide a mainboard power consumption adjustment method and system to solve the above technical problems. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a motherboard power consumption adjustment method and system thereof, which combines the advantages of fuzzy logic and lightweight neural network to achieve dynamic and precise adjustment of motherboard power consumption.

[0006] In one aspect, the present invention provides a method for adjusting power consumption of a mainboard, the method comprising the following steps: Collecting the current real-time data of the motherboard in real time, and performing fuzzy processing on the current real-time data to generate a multi-dimensional membership matrix, wherein the current real-time data includes CPU utilization, memory load, motherboard spatial position and application behavior pattern; Inputting the multidimensional membership matrix into a preset fuzzy rule base to generate an initial power supply strategy set; Input the current real-time data and historical load time series into a pre-built lightweight neural network, and output a rule weight correction and a membership function parameter offset, wherein the historical load time series is an equally spaced sampling sequence of the CPU utilization and memory load in the past N seconds from the current moment; Based on the initial power supply strategy set, by superimposing and calculating the rule weight correction amount and the membership function parameter offset, the weight and membership function parameters of the preset fuzzy rule base are dynamically updated to generate an optimized power supply strategy set; Input the set of optimized power supply strategies into a multi-objective optimizer, generate CPU frequency adjustment instructions and peripheral device power supply instructions based on a dynamic weighted objective function, and issue them for execution.

[0007] Preferably, the fuzzification process includes: Apply a Gaussian membership function to the motherboard spatial location to calculate the membership degree A; Apply trapezoidal membership functions to the CPU utilization rate and memory load to calculate the membership degrees B and C respectively; Extract fuzzy classification labels of application behavior patterns based on process characteristics related to motherboard power consumption, and calculate the membership degree D of the application behavior pattern belonging to each fuzzy classification label through the fuzzy classifier; Combine the membership degrees A, B, C, and D into a multi-dimensional matrix according to the three-dimensional structure of spatial location, load level, and application mode to generate a multi-dimensional membership degree matrix.

[0008] Preferably, inputting the multi-dimensional membership degree matrix into a preset fuzzy rule base to generate an initial power supply strategy set includes: For the condition part of each fuzzy rule, take the minimum value of the membership degree values corresponding to the corresponding dimensions in the multi-dimensional membership degree matrix as the triggering strength of this rule; Divide the triggering strength of all fuzzy rules by the sum of the triggering strengths to obtain the normalized triggering strength; Perform weighted averaging on the preset policy output values of each fuzzy rule according to the normalized triggering strength to generate an initial power supply strategy set.

[0009] Preferably, inputting the current real-time data and historical load time series into a pre-constructed lightweight neural network to output a rule weight correction amount and a membership function parameter offset amount includes: Take the current real-time data as the latest timestamp data and splice it with the historical load time series in chronological order to form a time series feature vector of length N + 1; Perform sliding window convolution on the time series feature vector through the one-dimensional convolutional layer of the lightweight neural network to extract local time series pattern features; Input the time series pattern features into the fully connected layer of the lightweight neural network, and output the rule weight correction amount after non-linear mapping through the ReLU activation function; Input the time series pattern features into the regression layer of the lightweight neural network, and fit the membership function parameter offset amount by the least squares method.

[0010] Preferably, based on the initial power supply strategy set, dynamically update the weights and membership function parameters of the preset fuzzy rule base by superimposing and calculating the rule weight correction amount and the membership function parameter offset amount to generate an optimized power supply strategy set, including: For the initial weight of each fuzzy rule in the fuzzy rule base, superimpose the corresponding rule weight correction amount to generate the updated rule weight, where the superimposition operation adopts an additive form and the weight value is restricted within a preset range of [0, 1]; For the initial parameters of each membership function in the fuzzy rule base, superimpose the corresponding membership function parameter offset to generate the updated membership function parameters, where the membership function parameter offset is constrained by L2 regularization; Based on the updated rule weights and membership function parameters, recalculate the triggering strength of each fuzzy rule; According to the recalculated triggering strength and the updated rule weights, regenerate the optimized power supply policy set through the weighted average method, and the optimized power supply policy set includes the dynamically adjusted CPU frequency range and the peripheral device power supply level.

[0011] Preferably, the recalculating the triggering strength of each fuzzy rule based on the updated rule weights and membership function parameters includes: Recalculate the membership degree values in the multi-dimensional membership degree matrix based on the updated membership function parameters; Take the minimum value of the recalculated membership degree values as the rule triggering strength; Normalize the triggering strengths of all rules.

[0012] Preferably, the inputting the optimized power supply policy set into a multi-objective optimizer, generating a CPU frequency adjustment instruction and a peripheral device power supply instruction based on a dynamic weighted objective function and issuing and executing them includes: Define a dynamic weighted objective function, including the following objectives: power consumption minimization objective, performance maximization objective, and heat dissipation efficiency optimization objective; Take the CPU frequency range and the peripheral device power supply level in the optimized power supply policy set as decision variables, substitute them into the dynamic weighted objective function, and calculate the comprehensive optimization objective value; Through the Pareto optimal solution screening algorithm, select the non-dominated solutions in the multi-objective conflict, generate the final CPU frequency adjustment instruction and peripheral device power supply instruction and execute them.

[0013] Preferably, the selecting the non-dominated solutions in the multi-objective conflict through the Pareto optimal solution screening algorithm includes: Use the NSGA-II algorithm to perform non-dominated sorting and crowding degree calculation on the optimized power supply policy set; Select the top K solutions with the largest crowding degree as the final instructions, where K is the preset number of candidate solutions.

[0014] On the other hand, the present invention also provides a motherboard power consumption adjustment system for executing a motherboard power consumption adjustment method, and the system includes: The membership degree matrix generation module is used to collect the current real-time data of the main board in real time, and perform fuzzy processing on the current real-time data to generate a multi-dimensional membership degree matrix, where the current real-time data includes CPU utilization rate, memory load, main board spatial position, and application behavior pattern; The initial policy generation module is used to input the multi-dimensional membership degree matrix into a preset fuzzy rule base to generate an initial power supply policy set; The dynamic rule evolution module is used to input the current real-time data and historical load time series into a pre-constructed lightweight neural network, and output a rule weight correction amount and a membership function parameter offset amount, where the historical load time series is an equally spaced sampling sequence of CPU utilization rate and memory load in the past N seconds before the current moment; The dynamic update module is used to dynamically update the weights and membership function parameters of the preset fuzzy rule base based on the initial power supply policy set by superimposing and calculating the rule weight correction amount and the membership function parameter offset amount, and generate an optimized power supply policy set; The instruction generation and execution module is used to input the optimized power supply policy set into a multi-objective optimizer, generate a CPU frequency adjustment instruction and a peripheral device power supply instruction based on a dynamic weighted objective function, and issue and execute them.

[0015] Compared with the related technologies, a main board power consumption adjustment method and system provided by the present invention have the following beneficial effects: The present invention collects the current real-time data of the main board in real time, including CPU utilization rate, memory load, main board spatial position, and application behavior pattern, and performs fuzzy processing on these data to generate a multi-dimensional membership degree matrix. Then, the multi-dimensional membership degree matrix is input into a preset fuzzy rule base to generate an initial power supply policy set. However, the initial power supply policy set may be affected by various factors, and its accuracy needs to be further improved. Therefore, a lightweight neural network is introduced, with the current real-time data and historical load time series as inputs, outputting a rule weight correction amount and a membership function parameter offset amount. After obtaining the optimized power supply policy set, it is input into a multi-objective optimizer, generating a CPU frequency adjustment instruction and a peripheral device power supply instruction based on a dynamic weighted objective function, and issuing and executing them. Description of the Drawings

[0016] Figure 1 It is a flowchart of a main board power consumption adjustment method provided by the present invention; Figure 2 It is a module structure diagram of a main board power consumption adjustment system provided by the present invention. Detailed Embodiments

[0017] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the present invention, rather than limiting the present invention. In addition, it should be noted that for the convenience of description, only parts related to the present invention rather than all structures are shown in the drawings. In addition, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0018] In addition, it should be noted that for the convenience of description, only parts related to the present invention rather than all content are shown in the drawings. Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations can be implemented in parallel, concurrently, or simultaneously. In addition, the order of the operations can be rearranged. When the operations are completed, the process can be terminated, but there can also be additional steps not included in the drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, and so on.

[0019] Embodiment 1 On the one hand, the present invention provides a method for regulating the power consumption of a main board. Referring to Figure 1 as shown, the method includes the following steps: S1: Real-time collect the current real-time data of the main board, and perform fuzzy processing on the current real-time data to generate a multi-dimensional membership matrix, where the current real-time data includes CPU utilization rate, memory load, the spatial position of the main board, and application behavior patterns.

[0020] Among them, the fuzzy processing in step S1 specifically includes: S11: Apply a Gaussian membership function to the spatial position of the main board and calculate the membership degree A.

[0021] In this embodiment, first, obtain the real-time position information of the main board through a Bluetooth or Wi-Fi signal strength positioning module, and convert it into discrete states (including but not limited to indoor, mobile, unknown). Define a Gaussian membership function for each state, and its mathematical expression is: Among them, is the state center value (for example, the indoor state corresponds to a signal strength threshold of -50 dBm, and the mobile state corresponds to -60 dBm), is the standard deviation (set to 10 dBm). The real-time collected signal strength value is input into the Gaussian function to calculate the membership degree. For example, when the signal strength is -55 dBm, the indoor membership degree is 0.88, the mobile membership degree is 0.61, and the unknown membership degree is 0.05. Finally, output the membership degree vector of the three states 。

[0022] S12: Apply trapezoidal membership functions to the CPU utilization rate and memory load, and calculate the membership degrees B and C respectively.

[0023] In this embodiment, the calculation of the membership degree B of the CPU utilization rate includes: Define the parameters of the trapezoidal membership function: Low load interval: 0% ≤ CPU utilization rate < 30%, the membership degree linearly increases from 0 to 1; Medium load interval: 30% ≤ CPU utilization rate ≤ 70%, the membership degree remains 1; High load interval: 70% < CPU utilization rate ≤ 100%, the membership degree linearly decreases from 1 to 0.

[0024] Input the real-time CPU utilization rate into the trapezoidal function to calculate the membership degree B.

[0025] Exemplarily, when the CPU utilization rate is 50%, the membership degree B = 1.0; when the utilization rate is 80%, B = 0.67.

[0026] The calculation of the membership degree C of the memory load includes: The parameter definition is similar to that of the CPU utilization rate, and three intervals of low (0 - 40%), medium (40% - 80%), and high (80% - 100%) are divided; Input the real-time memory occupancy rate into the trapezoidal function. Exemplarily, when the memory occupancy rate is 60%, the membership degree C = 1.0; when the occupancy rate is 90%, C = 0.5.

[0027] S13: Extract the fuzzy classification labels of the application behavior patterns based on the process characteristics related to the motherboard power consumption, and calculate the membership degree D of the application behavior patterns belonging to each fuzzy classification label through the fuzzy classifier.

[0028] In this embodiment, the GPU video memory occupancy rate of the process (through the NVIDIA SMI interface), the number of active threads in the thread pool (parsed through / proc / / status), and the I / O request frequency (monitored through system calls) are collected in real time, and the above features are normalized.

[0029] Adopt the TSK fuzzy model, and the rule base includes but is not limited to the following rules: Rule 1: If the GPU video memory occupancy rate ≥ 60% and the number of threads ≥ 10, then the application behavior is the rendering mode (confidence level 0.9); Rule 2: If the I / O frequency ≥ 1000 times / second and the number of threads ≤ 5, then the application behavior is the idle mode (confidence level 0.8); Input the standardized features into a fuzzy classifier to calculate the membership degrees of the application behavior belonging to each label (including but not limited to idle, office, game, rendering). And ensure that the sum of the membership degrees is 1.0. Exemplarily, when the GPU video memory occupancy rate is 70% and the number of threads is 12, the membership degree of the rendering mode is 0.85, and the sum of the membership degrees of other labels is 0.15 S14: Combine the membership degrees A, B, C, and D into a multi-dimensional matrix according to the three-dimensional structure of spatial position, load level, and application mode to generate a multi-dimensional membership degree matrix.

[0030] In this embodiment, the matrix dimension definition is as follows: The first dimension: the number of main board spatial position states (3 categories: indoor, mobile, unknown); The second dimension: the number of load levels (3 levels: low, medium, high); The third dimension: the number of application behavior modes (4 categories: idle, office, game, rendering).

[0031] Matrix filling rule: Traverse all dimension combinations and calculate the joint membership degree of each combination.

[0032] Fill the calculation results into the corresponding positions of the three-dimensional matrix to generate a multi-dimensional membership degree matrix with a size of 3×3×4, and perform normalization processing on all elements in the matrix to ensure that the sum is 1.

[0033] S2: Input the multi-dimensional membership degree matrix into a preset fuzzy rule base to generate an initial power policy set.

[0034] Among them, step S2 specifically includes the following steps: S21: For the condition part of each fuzzy rule, take the minimum value of the membership degree values of the corresponding dimensions in the multi-dimensional membership degree matrix as the triggering intensity of this rule.

[0035] In this embodiment, according to the condition part of each rule in the preset fuzzy rule base, determine the corresponding dimension combination in the multi-dimensional membership degree matrix; extract the membership degree values of each dimension involved in the rule condition from the multi-dimensional membership degree matrix. For example, extract the membership degree values of the spatial position dimension, the load level dimension, and the application mode dimension; perform a minimum value operation on all the extracted dimension membership degree values to obtain the triggering intensity of this rule; if the dimension combination corresponding to the rule condition has no defined value in the matrix, skip this rule or set its triggering intensity to 0 to ensure logical completeness.

[0036] S22: Divide the triggering intensity of all fuzzy rules by the sum of the triggering intensities to obtain a normalized triggering intensity.

[0037] In this embodiment, traverse the triggering strengths of all rules in the fuzzy rule base, calculate their sum, perform a normalization operation on the triggering strength of each rule. If the sum of the triggering strengths of all rules is zero, evenly distribute the normalized strengths to avoid division-by-zero errors.

[0038] S23: Perform a weighted average on the preset policy output values of each fuzzy rule according to the normalized triggering strength to generate an initial power supply policy set.

[0039] In this embodiment, the power supply policies preset for each rule include the CPU frequency adjustment value and the peripheral device power supply level. Exemplarily, the policy of Rule 1 is "increase the CPU frequency to 3.0 GHz and fully supply power to the peripherals", and the policy of Rule 2 is "reduce the CPU frequency to 1.5 GHz and set the peripherals to low-power mode"; calculate the weighted average of the preset frequency values of all rules, with the weight being the normalized triggering strength. For non-numerical policies such as the power supply level, use the preset value corresponding to the rule with the highest normalized triggering strength; encapsulate the weighted calculation results into an initial power supply policy set, including the dynamically adjustable CPU frequency range and the peripheral device power supply level.

[0040] S3: Input the current real-time data and the historical load time series into a pre-constructed lightweight neural network to output a rule weight correction amount and a membership function parameter offset amount, where the historical load time series is an equally spaced sampling sequence of the CPU utilization rate and the memory load in the past N seconds from the current moment.

[0041] Among them, step S3 specifically includes the following steps: S31: Use the current real-time data as the latest timestamp data, and splice it with the historical load time series in chronological order to form a time series feature vector with a length of N + 1.

[0042] In this embodiment, the currently collected CPU utilization rate and memory load data are used as the latest time point data, and are spliced with the historical load time series (the CPU utilization rate and memory load sequences sampled at fixed intervals within the past N seconds) in chronological order. The sampling interval of the historical load time series is 100 milliseconds, and the time window length is N seconds (for example, when N = 5 seconds, it corresponds to 50 sampling points). The dimension of the spliced time series feature vector is [N + 1, 2], where the first column is the CPU utilization rate sequence and the second column is the memory load sequence. Before data splicing, perform a moving window mean filter (window size is 3) on the historical load time series to remove high-frequency noise, and normalize all data to the [0, 1] interval. After splicing, directly transmit the time series feature vector to the neural network input buffer through the DMA channel of the embedded hardware to avoid CPU intervention and reduce the transmission delay to within 0.5 milliseconds.

[0043] S32: Perform sliding window convolution on the time series feature vector through the one-dimensional convolutional layer of the lightweight neural network to extract local temporal pattern features.

[0044] In this embodiment, the input layer of the lightweight neural network is a one-dimensional convolutional layer (Conv1D), and the configuration parameters are as follows: Number of convolutional kernels: 64.

[0045] Size of convolutional kernel: 3 (time step).

[0046] Sliding step: 1.

[0047] Activation function: ReLU.

[0048] Padding method: Same Padding (keep the output length the same as the input).

[0049] After the input time series feature vector (dimension [N + 1, 2]) passes through the convolutional layer, a local temporal feature map (dimension [N + 1, 64]) is output, and the short-term fluctuation patterns of CPU and memory load are captured through the convolution operation.

[0050] S33: Input the temporal pattern features into the fully connected layer of the lightweight neural network, and output the regular weight correction amount after non-linear mapping through the ReLU activation function.

[0051] In this embodiment, the output feature map of the one-dimensional convolutional layer is input into the fully connected layer after being flattened, and the structure of the fully connected layer is as follows: Input dimension: 64×(N + 1).

[0052] Number of hidden layer units: 128.

[0053] Activation function: ReLU.

[0054] Number of output layer units: M (M is the number of rules in the fuzzy rule base).

[0055] After calculation by the fully connected layer, the regular weight correction amount is output, and it is restricted to the interval [-0.5, 0.5] through the Sigmoid function.

[0056] S34: Input the temporal pattern features into the regression layer of the lightweight neural network, and fit the offset of the membership function parameters by the least squares method.

[0057] In this embodiment, the output feature map of the convolutional layer is simultaneously input into the regression layer, and the structure of the regression layer is: Input dimension: 64×(N + 1).

[0058] Number of hidden layer units: 64.

[0059] Activation function: linear activation.

[0060] Number of output layer units: P (P is the number of membership function parameters. For example, for a trapezoidal function, 2 boundary parameters need to be adjusted).

[0061] The regression layer fits the parameter offset by the least squares method, optimizes the parameter offset by the gradient descent method, and after outputting the parameter offset, performs boundary constraints on the membership function parameters.

[0062] S4: Based on the set of initial power supply strategies, by superimposing and calculating the rule weight correction amount and the membership function parameter offset, dynamically update the weights of the preset fuzzy rule base and the membership function parameters to generate an optimized set of power supply strategies.

[0063] Among them, step S4 specifically includes the following steps: S41: For the initial weight of each fuzzy rule in the fuzzy rule base, superimpose the corresponding rule weight correction amount to generate an updated rule weight, where the superimposition operation adopts an addition form and the weight value is restricted within a preset range of [0, 1].

[0064] In this embodiment, traverse each rule in the fuzzy rule base, obtain its initial weight value and the corresponding rule weight correction amount output by the lightweight neural network, update the weight through linear addition, and after the update, perform a clamping process on the weight value to ensure that it falls within the preset range of [0, 1]. All updated weights are stored in the non-volatile memory of the embedded system in floating-point form to ensure recovery after power failure.

[0065] S42: For the initial parameters of each membership function in the fuzzy rule base, superimpose the corresponding membership function parameter offset to generate an updated membership function parameter, where the membership function parameter offset is constrained by L2 regularization.

[0066] In this embodiment, for the membership function, obtain its initial parameters and the offset output by the neural network. During the neural network training stage, add an L2 regularization term to the loss function to perform logical verification on the updated parameters.

[0067] S43: Based on the updated rule weights and membership function parameters, recalculate the triggering strength of each fuzzy rule.

[0068] Among them, step S43 specifically includes: First, recalculate the membership degree values in the multi-dimensional membership degree matrix based on the updated membership function parameters.

[0069] In this embodiment, according to the updated membership function parameters, re-perform the fuzzification process in step S1 to generate a new multi-dimensional membership degree matrix. Secondly, take the minimum value of the recalculated membership degree value as the rule trigger strength.

[0070] In this embodiment, for each rule, extract the membership degree values of the corresponding dimensions from the new multi-dimensional membership degree matrix, and take the minimum value as the trigger strength. The specific operation is the same as that in step S21. Finally, normalize the trigger strengths of all rules.

[0071] S44: According to the recalculated trigger strength and the updated rule weights, regenerate the optimized power supply policy set by the weighted average method, and the optimized power supply policy set includes the dynamically adjusted CPU frequency range and the peripheral device power supply level.

[0072] In this embodiment, for the preset frequency value of each rule, perform double weighting according to the normalized trigger strength and the updated rule weights, and select the level corresponding to the rule with the largest product of the normalized trigger strength and the rule weight.

[0073] If the calculated CPU frequency exceeds the hardware support range (such as exceeding the maximum turbo frequency of 3.5 GHz), it is clamped to the nearest valid value. For example, the calculated value of 4.0 GHz is corrected to 3.5 GHz.

[0074] Package the optimized CPU frequency range (such as [2.2 GHz, 3.2 GHz]) and the peripheral device power supply level (such as "turn off USB devices") into the optimized power supply policy set, and transfer it to the multi-objective optimizer through shared memory. S5: Input the optimized power supply policy set into the multi-objective optimizer, generate the CPU frequency adjustment instruction and the peripheral device power supply instruction based on the dynamic weighted objective function, and issue and execute them.

[0075] Among them, step S5 specifically includes the following steps: S51: Define the dynamic weighted objective function, which includes the following objectives: power consumption minimization objective, performance maximization objective, and heat dissipation efficiency optimization objective.

[0076] In this embodiment, the dynamic weighted objective function includes three core optimization objectives: power consumption minimization, performance maximization, and heat dissipation efficiency optimization.

[0077] Power consumption minimization: The power consumption calculation is based on the CPU frequency and the peripheral device power supply level. The higher the frequency or the higher the peripheral device power supply level, the greater the power consumption. Establish a power consumption model through actual measurement of hardware parameters, such as the cubic relationship of frequency and the linear superposition of peripheral power supply.

[0078] Performance maximization: The performance index is jointly determined by the CPU frequency and the application response latency. The frequency increase directly enhances the computing power, while the response latency is obtained by real-time monitoring of the process execution time. The lower the latency, the higher the performance score.

[0079] Optimization of heat dissipation efficiency: The heat dissipation efficiency is related to the current temperature and the fan speed. The lower the temperature and the higher the fan speed, the higher the heat dissipation score. The ambient temperature is introduced as a reference to dynamically adjust the heat dissipation weight.

[0080] Dynamic weight adjustment: The target weight is allocated in real time according to the system state. For example, in a high-temperature scenario, the heat dissipation weight is increased to 60%, and the performance weight is decreased to 20%; when the battery is low, the power consumption weight is set to 80%. All weights are normalized to ensure that the sum is 100%.

[0081] S52: Take the CPU frequency range and the peripheral device power supply level in the optimized power policy set as decision variables, substitute them into the dynamic weighted objective function, and calculate the comprehensive optimization objective value.

[0082] In this embodiment, the CPU frequency is discretized into multiple levels (for example, from 1.0 GHz to 3.5 GHz, with a step of 0.25 GHz), and the peripheral device power supply status (such as USB and network card) is represented in binary (0 - off, 1 - low power consumption, 2 - full power supply).

[0083] Combine the frequency level and the peripheral device status into a decision vector. For example, "frequency 2.5 GHz, USB in low power consumption, network card off" is encoded as [2.5, 1, 0].

[0084] Objective normalization: Normalize each optimization objective to eliminate the dimension difference. For example, map the power consumption, performance, and heat dissipation values to the 0 - 1 interval respectively to ensure fair weighting.

[0085] Calculate the comprehensive score according to the dynamic weight. The power consumption weight is inverted (the smaller the better), and the performance and heat dissipation weights are added positively. Finally, generate the comprehensive score for each decision for subsequent optimization.

[0086] S53: Through the Pareto optimal solution screening algorithm, select the non-dominated solutions in the multi-objective conflict, generate the final CPU frequency adjustment instruction and peripheral device power supply instruction and execute them.

[0087] In step S53, the NSGA-II algorithm process: Population initialization: Randomly generate candidate solutions (for example, 100 groups) covering all possible frequency and peripheral combinations to ensure diversity.

[0088] Non-dominated sorting: Evaluate the dominance relationship in the solution set. If solution A is not inferior to solution B in all objectives and is better than solution B in at least one objective, then A dominates B. Divide the solutions into different ranks according to the dominance relationship, and Rank 1 is the optimal non-dominated solution set.

[0089] Crowdedness calculation: In the target space, calculate the density of each solution and its adjacent solutions, and preferentially retain the solutions with sparse distribution to avoid local optima.

[0090] Evolution operation: Select solutions with high crowdedness as parents through tournament selection, perform simulated crossover (mixing parent parameters) and random mutation to generate a set of offspring solutions. After merging the parents and offspring, retain the top 100 optimal solutions to enter the next iteration.

[0091] After the algorithm iteration is completed, select the top 3 solutions with the highest crowdedness from the Rank 1 solution set to ensure balanced coverage of multiple optimization goals, analyze the CPU target frequency and peripheral power supply instructions of the solutions, such as "frequency 3.0 GHz, turn off unnecessary peripherals".

[0092] Send instructions through the power management method of the ACPI interface.

[0093] Embodiment 2 On the other hand, the present invention also provides a motherboard power consumption regulation system for executing a motherboard power consumption regulation method. Refer to Figure 2 as shown, the system includes: A membership degree matrix generation module 100, which is used to collect the current real-time data of the motherboard in real time, and perform fuzzy processing on the current real-time data to generate a multi-dimensional membership degree matrix, where the current real-time data includes CPU utilization rate, memory load, motherboard spatial position, and application behavior pattern.

[0094] An initial policy generation module 200, which is used to input the multi-dimensional membership degree matrix into a preset fuzzy rule base to generate an initial power supply policy set.

[0095] A dynamic rule evolution module 300, which is used to input the current real-time data and historical load time series into a pre-constructed lightweight neural network, and output a rule weight correction amount and a membership function parameter offset amount, where the historical load time series is an equally spaced sampling sequence of CPU utilization rate and memory load in the past N seconds from the current moment.

[0096] A dynamic update module 400, which is used to dynamically update the weights and membership function parameters of the preset fuzzy rule base based on the initial power supply policy set by superimposing and calculating the rule weight correction amount and the membership function parameter offset amount, and generate an optimized power supply policy set.

[0097] An instruction generation and execution module 500, which is used to input the optimized power supply policy set into a multi-objective optimizer, generate CPU frequency adjustment instructions and peripheral device power supply instructions based on a dynamic weighted objective function, and issue and execute them.

[0098] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of flows and / or blocks in the flowchart and / or block diagram can also be implemented. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0099] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing relevant hardware through a program, and this program can be stored in a computer-readable storage medium. The storage medium includes read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM), or other optical disc memories, magnetic disk memories, tape memories, or any other medium that can be used to carry or store data and is computer-readable.

[0100] It should also be noted that the term "comprising", "including", or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, commodity, or device including a series of elements not only includes those elements but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, commodity, or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, commodity, or device including the element.

Claims

1. A method for regulating the power consumption of a main board, characterized in that, The method includes the following steps: Collect the current real-time data of the main board in real time, and perform fuzzification processing on the current real-time data to generate a multi-dimensional membership matrix, where the current real-time data includes CPU utilization, memory load, main board spatial position, and application behavior pattern; Input the multi-dimensional membership matrix into a preset fuzzy rule base to generate an initial power policy set; Input the current real-time data and historical load time series into a pre-constructed lightweight neural network, and output a rule weight correction amount and a membership function parameter offset amount, where the historical load time series is an equally spaced sampling sequence of CPU utilization and memory load in the past N seconds at the current moment; Based on the initial power policy set, dynamically update the weights and membership function parameters of the preset fuzzy rule base by superimposing and calculating the rule weight correction amount and the membership function parameter offset amount to generate an optimized power policy set; Input the optimized power policy set into a multi-objective optimizer, and generate a CPU frequency adjustment instruction and a peripheral device power supply instruction based on a dynamic weighted objective function and issue them for execution.

2. The method for regulating the power consumption of a main board according to claim 1, characterized in that The fuzzification processing includes: Apply a Gaussian membership function to the main board spatial position and calculate the membership degree A; Apply a trapezoidal membership function to CPU utilization and memory load, and calculate the membership degrees B and C respectively; Extract the fuzzy classification label of the application behavior pattern based on the process characteristics related to the main board power consumption, and calculate the membership degree D of the application behavior pattern belonging to each fuzzy classification label through the fuzzy classifier; Combine the membership degrees A, B, C, and D into a multi-dimensional matrix according to the three-dimensional structure of spatial position, load level, and application mode to generate a multi-dimensional membership matrix.

3. The method for regulating the power consumption of a main board according to claim 2, wherein The step of inputting the multi-dimensional membership matrix into a preset fuzzy rule base to generate an initial power policy set includes: For the condition part of each fuzzy rule, take the minimum value of the membership degree values corresponding to the corresponding dimension in the multi-dimensional membership matrix as the triggering intensity of the rule; Divide the triggering intensity of all fuzzy rules by the sum of the triggering intensities to obtain a normalized triggering intensity; Perform a weighted average on the preset policy output values of each fuzzy rule according to the normalized triggering intensity to generate an initial power policy set.

4. The method for regulating the power consumption of a main board according to claim 3, characterized in that The step of inputting the current real-time data and historical load time series into a pre-constructed lightweight neural network and outputting a rule weight correction amount and a membership function parameter offset amount includes: Take the current real-time data as the latest timestamp data and splice it with the historical load time series in chronological order to form a time series feature vector with a length of N + 1; Perform sliding window convolution on the time series feature vector through the one-dimensional convolutional layer of the lightweight neural network to extract local time series pattern features; Input the time series pattern features into the fully connected layer of the lightweight neural network, and output a rule weight correction amount after non-linear mapping through the ReLU activation function; Input the time series pattern features into the regression layer of the lightweight neural network, and fit the membership function parameter offset amount by the least squares method.

5. A motherboard power consumption adjustment method according to claim 4, characterized in that, Based on the initial power policy set, by superimposing and calculating the rule weight correction amount and the membership function parameter offset amount, dynamically update the weights of the preset fuzzy rule base and the membership function parameters to generate an optimized power policy set, including: For the initial weight of each fuzzy rule in the fuzzy rule base, superimpose the corresponding rule weight correction amount to generate the updated rule weight, where the superimposing operation adopts an addition form and the weight value is restricted within the preset range of [0, 1]; For the initial parameters of each membership function in the fuzzy rule base, superimpose the corresponding membership function parameter offset amount to generate the updated membership function parameters, where the membership function parameter offset amount is constrained by L2 regularization; Based on the updated rule weights and membership function parameters, recalculate the triggering strength of each fuzzy rule; According to the recalculated triggering strength and the updated rule weights, regenerate the optimized power policy set by the weighted average method, and the optimized power policy set includes the dynamically adjusted CPU frequency range and the peripheral device power supply level.

6. A main board power consumption regulation method according to claim 5, characterized in that, The recalculating the triggering strength of each fuzzy rule based on the updated rule weights and membership function parameters includes: Recalculate the membership degree values in the multi-dimensional membership degree matrix based on the updated membership function parameters; Take the minimum value of the recalculated membership degree values as the rule triggering strength; Normalize the triggering strengths of all rules.

7. A motherboard power consumption adjustment method according to claim 6, characterized in that Inputting the optimized power policy set into a multi-objective optimizer, generating a CPU frequency adjustment instruction and a peripheral device power supply instruction based on the dynamic weighted objective function and sending them down for execution includes: Define a dynamic weighted objective function, which includes the following objectives: power consumption minimization objective, performance maximization objective, and heat dissipation efficiency optimization objective; Take the CPU frequency range and the peripheral device power supply level in the optimized power policy set as decision variables, substitute them into the dynamic weighted objective function, and calculate the comprehensive optimization objective value; Through the Pareto optimal solution screening algorithm, select the non-dominated solutions in the multi-objective conflict, generate the final CPU frequency adjustment instruction and peripheral device power supply instruction and execute them.

8. A motherboard power consumption adjustment method according to claim 7, characterized in that, The selecting the non-dominated solutions in the multi-objective conflict through the Pareto optimal solution screening algorithm includes: Use the NSGA-II algorithm to perform non-dominated sorting and crowding degree calculation on the optimized power policy set; Select the top K solutions with the largest crowding degree as the final instructions, where K is the preset number of candidate solutions.

9. A main board power consumption regulation system for implementing a main board power consumption regulation method according to any one of claims 1 to 8, characterized in that, The system includes: A membership degree matrix generation module, which is used to collect the current real-time data of the main board in real time, and perform fuzzy processing on the current real-time data to generate a multi-dimensional membership degree matrix, where the current real-time data includes CPU utilization rate, memory load, main board spatial position, and application behavior pattern; An initial policy generation module, which is used to input the multi-dimensional membership degree matrix into a preset fuzzy rule base to generate an initial power policy set; A dynamic rule evolution module, which is used to input the current real-time data and historical load time series into a pre-built lightweight neural network, and output a rule weight correction amount and a membership function parameter offset amount, where the historical load time series is an equally-spaced sampling sequence of CPU utilization rate and memory load in the past N seconds from the current moment; A dynamic update module, which is used to dynamically update the weights and membership function parameters of a preset fuzzy rule base based on the initial power supply policy set by superimposing and calculating the rule weight correction amount and the membership function parameter offset amount, and generate an optimized power supply policy set; An instruction generation and execution module, which is used to input the optimized power supply policy set into a multi-objective optimizer, generate CPU frequency adjustment instructions and peripheral device power supply instructions based on a dynamic weighted objective function, and issue and execute them.

Images