Server fan modular energy saving array dynamic control method and system

Abstract

The application discloses a server fan modular energy-saving array dynamic control method and system, relates to the technical field of server heat dissipation control, and discloses the server fan modular energy-saving array dynamic control method and system, which realizes accurate prediction of load change and optimization of heat dissipation resource distribution through multidimensional data acquisition and federal learning cooperative optimization, combines dynamic speed regulation and redundancy control mechanism, and has the advantages of reducing energy consumption, improving heat dissipation efficiency and equipment operation stability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of server cooling control, in particular to a server fan modular energy-saving array dynamic control method and system. BACKGROUND

[0002] In the current server running environment, with the increasing demand for data processing, the power consumption and cooling problems of servers are becoming increasingly prominent. The traditional server fan control method often uses fixed speed or simple temperature feedback control. This method is difficult to dynamically adjust the fan speed according to the actual load of the server, resulting in energy waste when the fan is still running at a high speed under low load; while under high load, it may cause server performance degradation or even failure due to insufficient cooling.

[0003] In addition, most of the existing fan control schemes lack comprehensive analysis of multi-dimensional data such as vibration and airflow, making it difficult to accurately assess the cooling needs of the server and the running state of the fan. At the same time, in the face of multiple server clusters, there is a lack of effective cooperative control mechanism, resulting in inconsistent running state of each server fan, further affecting the overall cooling effect and energy utilization efficiency.

[0004] The above content is only used to assist in understanding the technical solutions of the present application, and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0005] The main purpose of the present application is to provide a server fan modular energy-saving array dynamic control method and system, which has the advantages of precise speed regulation, energy saving, improved cooling efficiency and running stability.

[0006] To achieve the above purpose, the present application provides a server fan modular energy-saving array dynamic control method, which comprises:

[0007] The sensor group in the modular energy-saving array collects original sensor data, including temperature data, vibration data and airflow data;

[0008] The original sensor data is preprocessed to obtain standardized sensor data;

[0009] The standardized sensor data is subjected to load prediction processing to obtain predicted load data;

[0010] Based on the predicted load data, pre-set regional temperature difference data and historical power consumption data, multi-objective dynamic speed regulation processing is performed to generate fan control instruction data;

[0011] Through a federated learning cooperative mechanism, the local model parameter data of multiple edge nodes is aggregated in a cloud server to generate global optimization strategy data;

[0012] generating a final control signal according to the global optimization strategy data and the fan control instruction data, and controlling the operation state of the fan modular array.

[0013] In an embodiment, the step of data pre-processing the original sensor data to obtain standardized sensor data comprises:

[0014] wavelet transform denoising processing is performed on vibration data in the original sensor data to obtain denoised vibration data;

[0015] Kalman filtering processing is performed on temperature data and airflow data in the original sensor data to obtain filtered temperature data and filtered airflow data;

[0016] The denoised vibration data, the filtered temperature data and the filtered airflow data are normalized to obtain standardized sensor data.

[0017] In an embodiment, the step of performing load prediction processing on the standardized sensor data to obtain predicted load data comprises:

[0018] historical load feature data is extracted from the standardized sensor data;

[0019] The historical load feature data is input into a pre-trained long short-term memory prediction model for time series prediction processing, and predicted load data is output.

[0020] In an embodiment, the step of generating fan control instruction data based on the predicted load data, pre-set regional temperature difference value data and historical power consumption data comprises:

[0021] current load rate data is calculated based on the predicted load data;

[0022] The current load rate data, the pre-set regional temperature difference value data and the historical power consumption data are input into a multi-agent collaborative optimization model;

[0023] optimization calculation is performed through the multi-agent collaborative optimization model, and target function value data is output;

[0024] Fan control instruction data is generated according to the target function value data.

[0025] In an embodiment, the step of performing optimization calculation through the multi-agent collaborative optimization model to output target function value data comprises:

[0026] A fan control agent generates fan action data based on temperature data and vibration data in the standardized sensor data;

[0027] The DVFS control agent generates voltage-frequency combination action data based on load data and core temperature data;

[0028] The fan action data and the voltage-frequency combination action data are jointly executed.

[0029] In an embodiment, the step of jointly executing the fan action data and the voltage-frequency combination action data comprises:

[0030] After executing the fan action data and the voltage-frequency combination action data, real-time power consumption data is collected;

[0031] Noise sensor data is collected and converted into linearized noise indicator data;

[0032] The standard deviation of the vibration data is calculated to obtain stability indicator data;

[0033] The real-time power consumption data, the linearized noise indicator data, and the stability indicator data are combined according to weights to calculate target function value data.

[0034] In an embodiment, after the step of collecting raw sensor data through a sensor group in the modular energy-saving array, the method further comprises:

[0035] The vibration data in the raw sensor data is subjected to frequency spectrum analysis processing to obtain vibration frequency spectrum feature data;

[0036] When the vibration frequency spectrum feature data exceeds a preset fault threshold, a fault fan unit identification signal is generated;

[0037] Based on the fault fan unit identification signal, a hot-plug redundancy control mechanism is triggered to activate a backup fan unit.

[0038] In an embodiment, through a federated learning collaborative mechanism, the cloud server aggregates local model parameter data of multiple edge nodes to generate global optimization strategy data.

[0039] Each edge node trains a local strategy model based on standardized sensor data to generate local model parameter data;

[0040] Difference privacy noise is added to the local model parameter data to generate noise-added parameter data;

[0041] The noise-added parameter data is subjected to homomorphic encryption processing to generate encrypted model parameter data;

[0042] The multiple encrypted model parameter data are aggregated and subjected to strategy update processing to generate global optimization strategy data.

[0043] In an embodiment, the method further comprises:

[0044] collecting external environment temperature data;

[0045] when the external environment temperature data is lower than the low temperature threshold, controlling the fan to stop and activating the phase change material heat dissipation;

[0046] when the external environment temperature data is higher than the high temperature threshold, activating the redundant fan unit data and starting the liquid cooling auxiliary system.

[0047] In addition, to achieve the above-mentioned purpose, the application also provides a server fan modular energy-saving array dynamic control system, which comprises a memory, a processor, and a server fan modular energy-saving array dynamic control program stored in the memory and executable on the processor, and the server fan modular energy-saving array dynamic control program is configured to implement the steps of the server fan modular energy-saving array dynamic control method.

[0048] The server fan modular energy-saving array dynamic control method and system provided by the application realize accurate prediction of load changes and optimization of heat dissipation resource allocation through multi-dimensional data collection and federated learning collaborative optimization, combined with dynamic speed regulation and redundant control mechanism, and have the advantages of reducing energy consumption, improving heat dissipation efficiency and equipment operation stability. BRIEF DESCRIPTION OF DRAWINGS

[0049] The accompanying drawings, which are incorporated into and form part of the specification, illustrate embodiments consistent with the present application and, together with the specification, serve to explain the principles of the application.

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without creative labor.

[0051] Figure 1 a flowchart for an embodiment of the server fan modular energy-saving array dynamic control method of the application;

[0052] Figure 2 a detailed flowchart for step S200 in the application Figure 1

[0053] Figure 3 a detailed flowchart for step S400 in the application Figure 1

[0054] Figure 4 a detailed flowchart for step S500 in the application Figure 1 a detailed flowchart for step S500 in the application​​

[0055] Figure 5 The structural schematic diagram provided for an embodiment of the server fan modular energy-saving array dynamic control system.

[0056] Explanation of reference numerals:

[0057] 10, memory; 20, processor.

[0058] The purposes, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0059] The technical solutions in the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. The components of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work belong to the scope of protection of the present application.

[0060] It should be understood that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. Meanwhile, in the description of the present application, the terms "first", "second", etc. are only used for differentiation in description, and cannot be understood as indicating or implying relative importance.

[0061] In the prior art, the server heat dissipation control has long relied on fixed rotating speed or single temperature feedback mechanism, which cannot adapt to dynamic load changes. The traditional method continuously runs at high rotating speed at low load, resulting in energy waste, and insufficient heat dissipation at high load may cause equipment failure. At the same time, lack of comprehensive analysis of multi-dimensional data such as vibration and airflow makes it difficult to accurately evaluate the heat dissipation demand. In the multi-server cluster scenario, independent control of each node leads to low heat dissipation efficiency and increased overall energy consumption.

[0062] In order to solve the above problems, the present technology cannot dynamically adjust the fan rotating speed, which prompts the inventors to think about how to realize accurate prediction by combining multi-dimensional data. Considering the influence of load fluctuation on heat dissipation demand, introducing a time series prediction model to predict future load becomes a key. Multi-objective optimization demand drives the inventors to explore a cooperative speed regulation mechanism to balance temperature, power consumption and stability. The cross-node cooperation problem is solved through a federated learning framework to realize global strategy optimization.

[0063] Based on this, the embodiments of the present application provide a server fan modular energy-saving array dynamic control method, which is described with reference toFigure 1 The dynamic control method for the modular energy-saving array of server fans includes steps S100 to S600, wherein:

[0064] Step S100: Collect raw sensor data through the sensor group in the modular energy-saving array. The raw sensor data includes temperature data, vibration data, and airflow data.

[0065] Step S200: Perform data preprocessing on the raw sensor data to obtain standardized sensor data;

[0066] Step S300: Perform load prediction processing on the standardized sensor data to obtain predicted load data;

[0067] Step S400: Based on the predicted load data, preset regional temperature difference data and historical power consumption data, perform multi-objective dynamic speed regulation processing to generate fan control command data;

[0068] Step S500: Through the federated learning collaboration mechanism, the local model parameter data of multiple edge nodes are aggregated on the cloud server to generate global optimization strategy data;

[0069] Step S600: Generate a final control signal based on the global optimization strategy data and fan control command data to control the operating status of the fan modular array.

[0070] In this embodiment, the sensor group refers to temperature, vibration, and airflow detection devices distributed at different locations on the server. Specifically, it can be implemented using a combination of thermocouples, accelerometers, and flow meters for real-time monitoring of the equipment's operating environment. Data preprocessing includes noise reduction, filtering, and normalization operations. Specifically, wavelet transform can be used to process vibration signals, and Kalman filtering can be used to process temperature and airflow fluctuations, eliminating measurement noise interference. Load prediction processing refers to predicting future loads by analyzing historical data trends. Specifically, a long short-term memory neural network model can be used to process time-series data and capture load change patterns. Multi-objective dynamic speed regulation refers to a speed regulation strategy that simultaneously optimizes temperature control, energy saving, and equipment stability. Specifically, a multi-agent reinforcement learning model can be used to calculate the optimal speed combination. The federated learning collaborative mechanism refers to model training while protecting data privacy. Specifically, homomorphic encryption technology can be used to aggregate local model parameters from each node to form a global optimization strategy.

[0071] In this embodiment, after the temperature, vibration and airflow data are collected by the sensors, the mechanical vibration noise is eliminated by wavelet transform, the Kalman filter is used to smooth the temperature and airflow fluctuations, and the normalized processing forms the standardized data with unified dimensions. The long short-term memory model analyzes the load change trend in the standardized data and outputs the predicted load value in the future period. The multi-agent optimization model integrates the predicted load, regional temperature difference and power consumption record to calculate the fan speed combination that meets the cooling demand and has the lowest energy consumption. The cloud server collects the encrypted local model parameters of each edge node through the federated learning framework, decrypts and aggregates to update the global control strategy. The final control signal integrates the local speed regulation instruction and the global optimization strategy to dynamically adjust the start-stop state and speed parameters of each fan module.

[0072] In this embodiment, the scheme realizes forward-looking control through multi-source data fusion and load prediction, and realizes cluster-level cooling optimization under the premise of ensuring data security through the federated learning framework, and improves the state perception accuracy through multi-sensor data fusion. In this way, the present application effectively reduces the redundant energy consumption in the low load state and avoids the risk of insufficient cooling in the high load working condition. Multi-dimensional data fusion improves the perception ability of the equipment running state, making the cooling control more accurate and reliable. The federated learning mechanism realizes cross-node collaborative optimization while protecting data privacy, significantly improving the overall energy efficiency of large-scale server clusters. The dynamic speed regulation strategy balances the cooling efficiency and energy consumption indicators, prolonging the service life of the fan equipment.

[0073] In a feasible implementation manner, with reference to Figure 2 , the step S200 includes steps S210-S230, in which:

[0074] In step S210, the vibration data in the original sensor data is subjected to wavelet transform denoising processing to obtain denoised vibration data.

[0075] In step S220, the temperature data and airflow data in the original sensor data are subjected to Kalman filter processing to obtain filtered temperature data and filtered airflow data.

[0076] In step S230, the denoised vibration data, filtered temperature data and filtered airflow data are subjected to normalization processing to obtain standardized sensor data.

[0077] In this embodiment, the wavelet transform denoising processing refers to the process of decomposing the vibration signal by wavelet basis function and removing high-frequency noise components. Specifically, it can be implemented by combining discrete wavelet transform with threshold shrinkage algorithm, for example, using Daubechies wavelet basis to decompose the vibration signal for five layers and then performing soft threshold processing on the detail coefficients. This processing can effectively eliminate high-frequency interference noise in the mechanical vibration signal. The Kalman filter processing refers to the process of recursively optimally estimating the observation data of the dynamic system by the state space model. Specifically, it can use a linear Kalman filter to filter the time series data of the temperature sensor and the airflow sensor, for example, to establish the state transition equation and the observation equation for prediction and correction. This processing can eliminate random measurement errors in the temperature and airflow data. The normalization processing refers to the process of mapping sensor data of different dimensions to a unified numerical interval. Specifically, it can use the min-max normalization method to linearly transform the vibration, temperature and airflow data to the [0, 1] interval, for example, to perform scale conversion according to the maximum and minimum values of each sensor based on historical data. This processing can eliminate the dimensional differences between multi-source sensor data.

[0078] In this embodiment, the vibration data is first decomposed into approximation coefficients and detail coefficients by wavelet transform, and denoised vibration data is obtained by setting a threshold to remove high-frequency noise components. The temperature data and the airflow data are respectively predicted and corrected by Kalman filter, and the current measurement value is optimized using the state estimation value at the previous time to eliminate sensor drift error. After denoising and filtering, the three groups of data are normalized to make the vibration, temperature and airflow data comparable, forming standardized sensor data as input for subsequent load prediction.

[0079] In this embodiment, through the combination of wavelet transform and Kalman filter processing, the nonlinear characteristics of the vibration signal and the dynamic characteristics of the temperature and airflow are matched respectively, and the signal-to-noise ratio of multi-dimensional sensor data is significantly improved. In this way, the above scheme solves the problems of high-frequency noise, dynamic measurement error and non-uniform dimensions in the original sensor data, improves the data quality through targeted denoising and filtering processing, and ensures that multi-source data participates in the calculation under a unified scale through normalization processing, providing high-precision standardized input data for the subsequent load prediction model.

[0080] In a feasible implementation, the load prediction processing is performed on the standardized sensor data, and the step of obtaining predicted load data includes: extracting historical load feature data from the standardized sensor data; inputting the historical load feature data into a pre-trained long short-term memory prediction model for time series prediction processing, and outputting predicted load data.

[0081] In this embodiment, the historical load feature data refers to a set of time series features associated with the server load state selected from the standardized sensor data. Specifically, the temperature change rate, vibration amplitude trend, and airflow fluctuation pattern in the past time period can be extracted using the sliding window technology to achieve this. These features can reflect the dynamic change law of the server load. The long short-term memory prediction model refers to a recurrent neural network structure with time series memory capability. Specifically, a multi-layer network architecture containing a forget gate, an input gate, and an output gate can be used to implement this model. By capturing the long-term dependencies in the historical load feature data, the model can predict the future load state.

[0082] In this embodiment, the standardized sensor data is extracted to form a historical load feature data set containing multi-dimensional time series after feature extraction. This data set is divided into a training set and a validation set for offline training of the long short-term memory prediction model. In the deployment phase, the real-time collected standardized sensor data is input into the trained model after the same feature extraction process for forward calculation, and the predicted load data in the future time period is output. The model adjusts the hidden layer neuron weights to make the mean square error of the prediction result and the real load data reach the convergence threshold.

[0083] In this embodiment, the long short-term memory prediction model is used to model multi-dimensional time series data, which can capture the periodicity, suddenness, and gradualness of load changes, significantly improving the prediction accuracy. For example, in the server burst computing task scenario, the model can predict the load rising trend based on the historical vibration amplitude mutation pattern, achieving high-precision load prediction based on time series features, so that the fan speed control can respond to load changes in advance. The prediction result is used as an input parameter for multi-objective optimization, effectively avoiding the problems of insufficient heat dissipation or excessive energy consumption caused by response lag in traditional methods, and reducing the system stability risk caused by load mutation.

[0084] In a feasible implementation manner, with reference to Figure 3 , step S400 includes steps S410-S440, in which:

[0085] Step S410, calculating current load rate data based on the predicted load data;

[0086] Step S420, inputting the current load rate data, pre-set regional temperature difference data, and historical power consumption data into the multi-agent collaborative optimization model;

[0087] Step S430, performing optimization calculation through the multi-agent collaborative optimization model to output target function value data;

[0088] Step S440, generating fan control instruction data according to the target function value data.

[0089] In this embodiment, the current load rate data refers to a real-time load state indicator determined by the ratio of predicted load data to the maximum carrying capacity of the system. Specifically, the ratio of the average load to the peak load in the period can be calculated using a sliding window statistical method to achieve this. It is used to reflect the intensity of the current processing task of the server. The regional temperature difference data refers to a quantitative indicator of the temperature difference between different heat dissipation regions. Specifically, it can be achieved by deploying temperature sensors at different positions of the server rack to collect data and calculate the absolute value of the temperature difference between adjacent regions to identify local areas with high heat dissipation demand. The historical power consumption data refers to the fan energy consumption information recorded by the system in the past running period. Specifically, it can be achieved by using a time series database to store and query by timestamp index, which is used to establish a correlation model between power consumption and load, temperature. The multi-agent collaborative optimization model refers to a distributed optimization framework composed of multiple decision units. Specifically, it can be achieved by training independent agents to control fan speed, voltage frequency and other parameters using reinforcement learning algorithm, which is used to coordinate the conflicts between different control objectives. The objective function value data refers to the optimization result score calculated by combining energy consumption, noise, heat dissipation efficiency and other multi-dimensional indicators. Specifically, it can be generated by weighted summation or Pareto frontier analysis method, which is used to determine the optimal control strategy.

[0090] In this embodiment, the predicted load data is compared with the maximum performance threshold after normalization to generate the current load rate data. This data, together with the real-time collected regional temperature difference data and the historical power consumption data in the running period, constitutes the input parameters of the multi-agent collaborative optimization model. The independent decision units in the multi-agent collaborative optimization model perform local optimization for heat dissipation efficiency, energy consumption control and mechanical stability respectively, and generate a global optimal solution through an interactive game mechanism. The objective function value data balances the priority between different optimization objectives through a dynamic weight adjustment mechanism. The final fan control instruction data includes speed adjustment amplitude, start-stop time interval and cooperative working mode parameters.

[0091] Compared with the prior art, the traditional method usually triggers a fixed speed regulation strategy based on a single temperature threshold, which cannot dynamically coordinate the complex relationship between load changes, regional temperature difference and historical energy consumption. While the present scheme realizes the collaborative processing of multi-dimensional data through a multi-agent model, which can reduce redundant power consumption while ensuring heat dissipation efficiency, and avoid equipment damage caused by local overheating or excessive vibration. Through the above technical scheme, the present application effectively solves the problems of response lag and target conflict in the traditional fan control strategy, and realizes the dynamic matching of load fluctuation and heat dissipation demand. By introducing a multi-dimensional data collaborative optimization mechanism, the overall energy consumption is reduced while the heat dissipation efficiency is improved, and the mechanical vibration and noise caused by sudden speed change are reduced.

[0092] In an implementable embodiment, the step of performing optimization calculation by the multi-agent collaborative optimization model and outputting target function value data comprises: a fan control agent generating fan action data based on temperature data and vibration data in the standardized sensor data; a DVFS control agent generating voltage-frequency combination action data based on load data and core temperature data; and jointly executing the fan action data and the voltage-frequency combination action data.

[0093] In this embodiment, the fan control agent refers to an independent decision-making module for dynamically adjusting fan speed based on temperature data and vibration data, which can be implemented by a policy network based on reinforcement learning. The fan control agent generates corresponding speed adjustment instructions by real-time sensing of heat dissipation demand and device vibration state. The DVFS control agent refers to an optimization module for dynamically adjusting processor voltage and frequency according to server load changes, which can be implemented by a dynamic voltage and frequency adjustment algorithm. The DVFS control agent generates optimal power consumption configuration parameters by analyzing core temperature and computing task amount. The fan action data includes fan speed adjustment instructions, which can be implemented by a pulse width modulation signal to drive a motor controller. The voltage-frequency combination action data includes voltage and frequency combination parameters of the processor core, which can be implemented by a mainboard power management unit for dynamic configuration. Joint execution refers to synchronously applying the two types of action data to the server system, which can be implemented by a parallel control signal distribution mechanism to ensure the synergy of heat dissipation and power consumption adjustment.

[0094] In this embodiment, temperature data and vibration data are input to the fan control agent, and fan speed adjustment instructions are generated by a policy network calculation, such as increasing the speed by a certain percentage when the temperature is higher than a set threshold. Load data and core temperature data are input to the DVFS control agent, and voltage and frequency combination parameters are generated by a dynamic voltage and frequency adjustment algorithm, such as reducing core voltage to reduce power consumption in a low load state. Subsequently, the fan action data and the voltage-frequency combination action data are synchronously sent to the corresponding execution units. The fan module adjusts air flow intensity according to the speed instruction, and the processor adjusts the running state according to the voltage-frequency parameters. The two work together to achieve balanced optimization of heat dissipation efficiency and energy consumption.

[0095] In this embodiment, the present scheme separately handles fan control and processor frequency adjustment by independent agents, and introduces a joint execution mechanism, so that the goals of improving heat dissipation capacity and reducing power consumption can be achieved synchronously, avoiding the decline in system stability caused by unilateral optimization. The present scheme can solve the energy efficiency imbalance problem caused by the fragmentation of fan speed and processor frequency adjustment strategies in traditional methods. In a high load scenario of a server, the present scheme avoids excessive core temperature through collaborative control, while in a low load scenario, the present scheme reduces overall power consumption, improves the dynamic response capability of the heat dissipation system, and improves energy utilization efficiency.

[0096] In an implementable embodiment, the step of jointly executing the fan action data and the voltage-frequency combination action data comprises: collecting real-time power consumption data after executing the fan action data and the voltage-frequency combination action data; collecting noise sensor data and converting the noise sensor data into linearized noise index data; calculating a standard deviation of vibration data to obtain stability index data; combining the real-time power consumption data, the linearized noise index data, and the stability index data according to weights to calculate target function value data.

[0097] In the embodiment, the real-time power consumption data refers to actual power consumption data after executing the control action, which can be obtained by multiplying real-time collected voltage and current using a power meter or a current sensor, and is used to quantify the energy consumption level. The linearized noise index data refers to data obtained by converting non-linear acoustic signals collected by a sound pressure level sensor into linear decibel values, which can be obtained by logarithmic transformation or A-weighting network processing, and is used to standardize noise evaluation. The stability index data refers to a statistical quantity reflecting the intensity of mechanical vibration, which can be obtained by calculating the standard deviation or root mean square value of vibration signals, and is used to represent fan running stability. The weight combination refers to weighted summation of proportional coefficients given to different performance indexes, which can be determined by using the analytic hierarchy process or entropy weight method, and is used to balance the conflict relationship in multi-objective optimization.

[0098] In the embodiment, after generating the fan control instruction, the system synchronously executes fan speed adjustment and processor voltage frequency adjustment actions, and then collects actual power consumption, noise, and vibration data through a sensor network. The noise signal is linearized to eliminate frequency response deviation, and the vibration data is quantified by calculating the standard deviation to represent mechanical stability. The real-time power consumption, linearized noise index, and stability index are weighted according to preset weights to calculate a comprehensive target function value, which is used to evaluate the pros and cons of the current control strategy and guide the subsequent optimization direction.

[0099] In the embodiment, by quantifying multi-dimensional running parameters and determining weights by using a scientific method, multi-objective dynamic balance is achieved, and the contradiction between heat dissipation demand in a high load scenario and energy saving and noise reduction in a low load scenario is effectively solved. Through the above technical solution, the application can monitor the comprehensive influence of the fan control strategy on energy consumption, noise, and mechanical stability in real time, dynamically evaluate the control effect through the target function, and ensure to reduce overall running noise and prolong the service life of the equipment on the premise of meeting the heat dissipation demand. The scheme solves the problems of energy waste or equipment loss caused by ignoring the multi-parameter coupling relationship in the traditional method, and improves the reliability and environmental adaptability of the server cluster.

[0100] In a feasible implementation, after the step of collecting raw sensor data by the sensor group in the modular energy-saving array, the method further comprises: performing frequency spectrum analysis processing on vibration data in the raw sensor data to obtain vibration spectrum feature data; generating a fault fan unit identification signal when the vibration spectrum feature data exceeds a preset fault threshold; triggering a hot-swap redundancy control mechanism based on the fault fan unit identification signal to activate a backup fan unit.

[0101] In this embodiment, the frequency spectrum analysis processing refers to an analysis method of converting a vibration signal from a time domain to a frequency domain, which can be implemented by using a fast Fourier transform algorithm, and is used to identify abnormal frequency components in the vibration signal. The vibration spectrum feature data refers to energy distribution data of each frequency band of the vibration signal extracted by frequency spectrum analysis, which can be implemented by using a frequency band energy proportion calculation method, and is used to quantify the vibration intensity in different frequency intervals. The preset fault threshold refers to a pre-set vibration spectrum abnormality determination standard, which can be obtained based on historical fault data statistics, and is used to distinguish between normal vibration and abnormal vibration caused by mechanical failure. The fault fan unit identification signal refers to a digital signal used to locate the position of the fault fan, which can be implemented by using a binary encoding method, and is used to accurately indicate the fan module that needs to be replaced. The hot-swap redundancy control mechanism refers to a control logic for replacing a hardware module under a continuous power supply state, which can be implemented by using a dual-power supply and bus isolation technology, and is used to ensure the continuous operation of the system during the fault replacement process. The backup fan unit refers to a pre-deployed redundant cooling device, which can be implemented by using a fan module of the same type, and is used to quickly take over the cooling task when a fault occurs.

[0102] In this embodiment, after the sensor group collects the vibration data, the vibration signal is decomposed into energy distribution data of multiple frequency bands by using fast Fourier transform. For example, 0-100Hz is divided into a low frequency band, 100-500Hz is divided into a medium frequency band, and 500Hz and above is divided into a high frequency band, and the energy proportion of each frequency band is calculated. When the energy proportion of a certain frequency band exceeds a preset threshold, for example, the energy proportion of the medium frequency band exceeds 60%, it is determined that the corresponding fan unit has a bearing wear or blade imbalance fault. At this time, an identification signal containing fault position information is generated, triggering the control circuit to cut off the power supply of the fault fan, and simultaneously activating the power supply line and communication interface of the backup fan. In this process, the bus isolation technology can ensure that the removal of the fault module does not affect the data transmission of other fans, and the dual-power supply design maintains the continuous power supply of the cooling array.

[0103] Compared with the prior art, the traditional method usually relies on manual inspection or single threshold alarm, and cannot accurately identify the type of vibration anomaly, and the fault switching needs to be maintained by stopping. The scheme can find early mechanical failure in advance through spectrum feature analysis, and the hot plug mechanism shortens the fault replacement time to milliseconds, avoiding the risk of server overheating caused by downtime maintenance. Therefore, the application can quickly locate the fault source and start the redundant cooling unit when the fan fails, effectively avoiding the local overheating problem caused by a single fan failure, while reducing the frequency of manual intervention and improving the continuous operation capability and reliability of the server cooling system.

[0104] In a feasible implementation, with reference to Figure 4 , step S500 includes steps S510-S540, wherein:

[0105] Step S510, training a local policy model based on standardized sensor data at each edge node to generate local model parameter data;

[0106] Step S520, adding differential privacy noise to the local model parameter data to generate noise-added parameter data;

[0107] Step S530, homomorphic encryption processing of the noise-added parameter data to generate encrypted model parameter data;

[0108] Step S540, aggregating multiple encrypted model parameter data and performing policy update processing to generate global optimization policy data.

[0109] In this embodiment, the federated learning collaboration mechanism refers to a distributed machine learning framework, which is specifically implemented by uploading the parameters of the locally trained model to the cloud for aggregation, which functions to achieve collaborative training of multiple nodes while protecting data privacy. The differential privacy noise refers to adding random noise to interfere with the original data, which can be implemented by using Laplace noise or Gaussian noise algorithm, which functions to prevent the local model parameter data from being inversely deduced to sensitive information. The homomorphic encryption processing refers to allowing data to be calculated in an encrypted state, which can be implemented by using Paillier encryption algorithm or full homomorphic encryption scheme, which functions to ensure that the model parameters cannot be stolen or tampered with during transmission and aggregation.

[0110] ​In this embodiment, after the standardized sensor data is distributed to each edge node, each node independently trains a local policy model based on local data and generates corresponding local model parameters. Before transmitting the parameters, differential privacy noise is injected into the model parameters to prevent the parameters from being inversely associated with a specific node. Subsequently, the noise-added parameters are converted into ciphertext form through a homomorphic encryption algorithm to ensure that they cannot be decrypted even if intercepted during transmission. After the cloud server receives multiple encrypted model parameters, it performs aggregation operations such as weighted average or gradient stacking in ciphertext state to update the global model parameters. The updated global optimization strategy data is distributed to each edge node to guide the next round of local model training iteration.

[0111] Compared with the prior art, the traditional multi-node collaborative control method usually directly shares raw data or plaintext parameters, which has the risk of data leakage and is difficult to optimize across nodes. The centralized training mode in the prior art easily forms a data island, and the transmission of plaintext parameters may be stolen by malicious nodes. The present scheme realizes data staying local through the federated learning framework, and combines differential privacy and homomorphic encryption technology to effectively prevent sensitive information leakage and external attacks while ensuring the effectiveness of multi-node collaborative training.

[0112] Through the above technical solutions, the present application solves the contradictory problem of data privacy and collaborative optimization in a multi-server cluster environment, generates a global optimization strategy through encrypted aggregation under the premise of ensuring the safety of sensor data of each node, and improves the dynamic control precision of the fan array. At the same time, the introduction of differential privacy noise avoids the inverse analysis of local model parameters, and homomorphic encryption processing ensures the security of the data transmission process, so that the multi-node collaborative control mechanism can be reliably implemented in a complex network environment.

[0113] In a feasible implementation, the method further comprises: collecting external environment temperature data; when the external environment temperature data is lower than a low temperature threshold, controlling the fan to stop and activating the phase change material heat dissipation; when the external environment temperature data is higher than a high temperature threshold, activating the redundant fan unit data and starting the liquid cooling auxiliary system.

[0114] In this embodiment, the external environment temperature data refers to the real-time temperature information of the server room where the server is located, which can be collected by a temperature sensor deployed in the server room, and is used to reflect the temperature condition of the server running environment. The low temperature threshold refers to a critical temperature value that triggers the fan to stop, which can be 15°C for example, and the specific value can be adjusted according to the design standard of the server room, and is used to determine whether to enable the passive cooling mode. The high temperature threshold refers to a critical temperature value that triggers the active cooling enhancement measure, which can be 35°C for example, and is used to identify the cooling demand in an extreme high temperature environment. The phase change material cooling refers to passive cooling by using the heat absorption characteristic of phase change material in the phase change process, and specifically, a paraffin-based or metal-based phase change material can be packaged in the server cooling module to replace the fan for heat storage in a low temperature environment. The redundant fan unit data refers to the preconfigured backup fan device information, which can be stored by the device management module to store the fan position and performance parameters, and is used to quickly start additional cooling capacity in high temperature. The liquid cooling auxiliary system refers to a liquid circulation based enhanced cooling device, which can specifically include a pump body, a cooling liquid pipeline and a cooling plate, and is used to improve the cooling efficiency when the high temperature threshold is triggered.

[0115] In this embodiment, the external environment temperature data is periodically collected by temperature sensors deployed in different areas of the server room and transmitted to the control module for threshold comparison. When the temperature is detected to be lower than the low temperature threshold, the control module sends a stop instruction to the fan array and sends an activation signal to the phase change material cooling unit, so that the phase change material packaged in the cooling fin starts to absorb the heat generated by the server. When the temperature is detected to be higher than the high temperature threshold, the control module first calls the redundant fan unit database, selects the appropriate position of the standby fan according to the current load condition, and sends a start instruction to the liquid cooling system controller, so that the cooling liquid starts to circulate and carries away the heat accumulated inside the server.

[0116] In some specific embodiments, the phase change material cooling module can be designed as a replaceable slot structure, which facilitates the replacement of materials with different phase change points according to the temperature range of the server room environment. The start process of the liquid cooling auxiliary system can include a cooling liquid pre-circulation stage, which reduces the system delay by establishing the cooling liquid flow path in advance. The activation of the redundant fan unit can be combined with load prediction data to preferentially start the standby fan corresponding to the high load area.

[0117] Compared with the prior art, the traditional scheme only relies on a single fan speed regulation mechanism to cope with temperature changes, cannot effectively utilize passive heat dissipation resources in low temperature environments, and lacks multi-stage heat dissipation enhancement means in extremely high temperatures. The scheme realizes intelligent switching of active and passive heat dissipation modes by introducing an ambient temperature threshold judgment mechanism, and forms a multi-stage heat dissipation guarantee system through the cooperative control of redundant fans and liquid cooling systems. In this way, the application solves the problems of serious energy waste in low temperature environments and insufficient heat dissipation capacity in extremely high temperatures of the traditional fan control method. Through the mode switching mechanism triggered by the ambient temperature, the energy consumption in the low load low temperature scene is significantly reduced under the premise of ensuring the heat dissipation effect, and the overheat risk of the server in the high temperature scene is effectively prevented through the linkage control of the multi-stage heat dissipation system, improving the environmental adaptability and reliability of the heat dissipation system.

[0118] In the embodiments of the application, the server fan modular energy-saving array dynamic control method realizes accurate prediction of load changes and optimization of heat dissipation resource allocation through multi-dimensional data acquisition and federated learning cooperative optimization, combined with dynamic speed regulation and redundant control mechanism, has the advantages of reducing energy consumption, improving heat dissipation efficiency and equipment operation stability.

[0119] It should be noted that the above examples are only for understanding the application and do not constitute a limitation on the server fan modular energy-saving array dynamic control method of the application. Further simple transformations based on this technical concept are within the scope of protection of the application.

[0120] The application also provides a server fan modular energy-saving array dynamic control system, which comprises a memory 10, a processor 20, and a server fan modular energy-saving array dynamic control program stored on the memory 10 and executable on the processor 20. The server fan modular energy-saving array dynamic control program is configured to implement the steps of the server fan modular energy-saving array dynamic control method.

[0121] The server fan modular energy-saving array dynamic control system provided by the application adopts the server fan modular energy-saving array dynamic control method in the above embodiments, which can realize accurate speed regulation to reduce energy consumption, improve heat dissipation efficiency and operation stability. Compared with the prior art, the server fan modular energy-saving array dynamic control system provided by the application has the same beneficial effects as the server fan modular energy-saving array dynamic control method provided by the above embodiments, and other technical features in the server fan modular energy-saving array dynamic control system are the same as the features disclosed in the above method, which will not be repeated here.

[0122] It should be understood that portions of the application disclosed can be implemented in hardware, software, firmware, or combinations thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner within the scope of the present application.

[0123] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any equivalent structure transformation made by using the content of the present application specification and drawings, or direct / indirect application in other related technical fields is included in the patent protection scope of the present application.

Claims

1. A dynamic control method for a modular energy-saving array of server fans, characterized in that, The method includes: Raw sensor data, including temperature data, vibration data, and airflow data, is collected through a sensor array in a modular energy-saving array. The raw sensor data is preprocessed to obtain standardized sensor data. The standardized sensor data is subjected to load prediction processing to obtain predicted load data; Based on the predicted load data, the preset regional temperature difference data, and the historical power consumption data, multi-objective dynamic speed regulation processing is performed to generate fan control command data. Through a federated learning collaboration mechanism, local model parameter data from multiple edge nodes are aggregated on a cloud server to generate global optimization strategy data. The final control signal is generated based on the global optimization strategy data and fan control command data to control the operating state of the fan modular array; Based on the predicted load data, preset regional temperature difference data, and historical power consumption data, the steps for performing multi-objective dynamic speed regulation processing and generating fan control command data include: Calculate the current load rate data based on the predicted load data; The current load rate data, the preset regional temperature difference data, and the historical power consumption data are input into the multi-agent collaborative optimization model. The optimization calculation is performed through the multi-agent collaborative optimization model, and the objective function value data is output. Generate fan control command data based on the objective function value data; The step of performing optimization calculations through the multi-agent collaborative optimization model and outputting objective function value data includes: The fan control agent generates fan action data based on the temperature and vibration data from the standardized sensor data. The DVFS control agent generates voltage-frequency combined action data based on load data and core temperature data; The fan action data and voltage-frequency combined action data are executed in combination. The steps for jointly executing the fan action data and the voltage-frequency combined action data include: After executing the fan action data and voltage-frequency combined action data, real-time power consumption data is collected; Collect noise sensor data and convert it into linearized noise index data; The standard deviation of the vibration data is calculated to obtain the stability index data. The objective function value is calculated by combining the real-time power consumption data, linearized noise index data, and stability index data according to weights.

2. The dynamic control method for modular energy-saving array of server fans as described in claim 1, characterized in that, The steps for preprocessing the raw sensor data to obtain standardized sensor data include: The vibration data in the original sensor data is subjected to wavelet transform noise reduction processing to obtain noise-reduced vibration data. The temperature data and airflow data in the original sensor data are processed by Kalman filtering to obtain filtered temperature data and filtered airflow data. The noise-reduced vibration data, filtered temperature data, and filtered airflow data are normalized to obtain standardized sensor data.

3. The dynamic control method for modular energy-saving arrays of server fans as described in claim 1, characterized in that, The steps for performing load prediction processing on the standardized sensor data to obtain predicted load data include: Extract historical load characteristic data from the standardized sensor data; The historical load feature data is input into a pre-trained long short-term memory prediction model for time-series prediction processing, and the predicted load data is output.

4. The dynamic control method for modular energy-saving array of server fans as described in claim 1, characterized in that, Following the step of acquiring raw sensor data via a sensor array in a modular energy-efficient array, the method further includes: The vibration data in the original sensor data is subjected to spectral analysis to obtain vibration spectral characteristic data; When the vibration spectrum characteristic data exceeds the preset fault threshold, a faulty fan unit identification signal is generated; The hot-swap redundancy control mechanism is triggered based on the faulty fan unit identification signal to activate the backup fan unit.

5. The dynamic control method for modular energy-saving array of server fans as described in claim 1, characterized in that, The steps involved in generating global optimization strategy data by aggregating local model parameter data from multiple edge nodes on a cloud server through a federated learning collaborative mechanism include: Local policy models are trained at each edge node based on standardized sensor data to generate local model parameter data. Differential privacy noise is added to the local model parameter data to generate noise-added parameter data; The noise-added parameter data is homomorphically encrypted to generate encrypted model parameter data. The system aggregates parameter data from multiple encryption models and performs policy update processing to generate global optimization policy data.

6. The dynamic control method for modular energy-saving array of server fans as described in claim 1, characterized in that, The method further includes: Collect external ambient temperature data; When the ambient temperature is below the low temperature threshold, the fan is stopped and the phase change material is activated to dissipate heat. When the external ambient temperature exceeds the high temperature threshold, the redundant fan unit data is activated and the liquid cooling auxiliary system is started.

7. A dynamic control system for a modular energy-saving array of server fans, characterized in that, The server fan modular energy-saving array dynamic control system includes: a memory, a processor, and a server fan modular energy-saving array dynamic control program stored in the memory and executable on the processor. The server fan modular energy-saving array dynamic control program is configured to implement the steps of the server fan modular energy-saving array dynamic control method as described in any one of claims 1 to 6.

Images