Server BMC intelligent management method and system based on edge computing

Abstract

The application discloses a kind of based on edge computing's server BMC intelligent management method and system, it is related to server hardware management technical field, based on edge computing's server BMC intelligent management method and system disclosed, by dynamic adjustment sensor sampling frequency, local AI inference and closed-loop control strategy, complete closed loop of data acquisition, processing and hardware control are realized in edge side, effectively improve data processing efficiency, reduce main processor load and network transmission stress, simultaneously, through predictive maintenance, improve system autonomy.

Description

Technical Field

[0001] This application relates to the field of server hardware management technology, and in particular to a server BMC intelligent management method and system based on edge computing. Background Technology

[0002] The Server Baseboard Management Controller (BMC), as a core component of server hardware management, plays a crucial role. In scenarios such as data centers, cloud computing, and enterprise server clusters, the BMC is a key technology for enabling remote server management, monitoring, and maintenance. It can operate independently of the server's main processor (CPU), collecting real-time server hardware status information such as temperature, voltage, fan speed, and memory errors via a dedicated network connection. It also supports remote power on / off, restart, and firmware upgrades, greatly improving the convenience and efficiency of server operation and maintenance.

[0003] With the continuous expansion of data center scale and the sustained growth in server performance demands, BMCs face numerous challenges. On the one hand, the increasing complexity of server hardware, the significant increase in the number of sensors and data volume, makes it increasingly difficult for traditional BMCs to meet the data processing needs. Large amounts of raw sensor data need to be transmitted to external management platforms for analysis, which not only consumes network bandwidth but also causes data processing delays, hindering timely responses to hardware anomalies. On the other hand, in terms of energy management, traditional BMCs lack flexible policy adjustment mechanisms. Their fixed sensor sampling frequencies and hardware control logic cannot adapt to dynamic changes in server load, resulting in energy waste and increased data center operating costs. Furthermore, faced with complex and ever-changing hardware failure scenarios, traditional BMCs mainly rely on preset threshold judgments and simple rule matching, lacking intelligent fault diagnosis and prediction capabilities, making it difficult to detect potential faults in advance and ensure stable server operation.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of this application is to provide a server BMC intelligent management method and system based on edge computing, which aims to improve data processing efficiency and reduce the load on the main processor.

[0006] To achieve the above objectives, this application proposes a server BMC intelligent management method based on edge computing, the method comprising:

[0007] Raw sensor data is acquired through a multi-protocol sensor interface; a server load index is calculated based on the raw sensor data; the sampling frequency is dynamically adjusted according to the server load index, and the adjusted sampling data is output.

[0008] The adjusted sampled data is subjected to denoising and normalization processing, and features are extracted based on the processing results to output a multidimensional feature vector;

[0009] The multidimensional feature vectors are input into a pre-trained lightweight AI model for analysis and reasoning to predict future hardware states and calculate prediction confidence. The output is inference data containing prediction results and confidence.

[0010] Based on the prediction results in the inference data, hardware control instructions are generated;

[0011] The hardware control instructions are converted into control signals and hardware control operations are executed, outputting execution feedback data.

[0012] Based on the execution feedback data, the execution strategy is optimized and the status feedback is updated, and the updated strategy library is output.

[0013] In one embodiment, the steps of acquiring raw sensor data through a multi-protocol sensor interface; calculating a server load index based on the raw sensor data; dynamically adjusting the sampling frequency according to the server load index; and outputting the adjusted sampling data include:

[0014] Acquire CPU utilization data, memory bandwidth usage data, and PCIe device throughput data through sensors;

[0015] The server load index is calculated by weighting the CPU utilization data, memory bandwidth usage data, and PCIe device throughput data.

[0016] When the server load index is lower than the first threshold, the sampling frequency is reduced and data compression is started, and the adjusted sampling data is output.

[0017] When the server load index is higher than the second threshold, the sampling frequency is increased and the real-time data stream mode is enabled, and the adjusted sampling data is output.

[0018] In one embodiment, the step of performing denoising and normalization processing on the adjusted sampled data, extracting features based on the processing results, and outputting a multidimensional feature vector includes:

[0019] The adjusted sampled data is subjected to denoising processing, and the median filtering algorithm is used to remove interference signals and output denoised data.

[0020] The denoised data is normalized to unify the data volume using a standardization method, and standardized data is output.

[0021] Based on the standardized data, a multidimensional feature vector is extracted, including the current sampled value, historical moving average, and standard deviation, and the multidimensional feature vector is output.

[0022] In one embodiment, the step of inputting the multidimensional feature vector into a pre-trained lightweight AI model for analysis and inference, predicting future hardware states and calculating prediction confidence, and outputting inference data containing prediction results and confidence levels includes:

[0023] Load a pre-trained lightweight AI model that matches the current scene on the server;

[0024] The multidimensional feature vector is input into the lightweight AI model to perform inference calculations and output the predicted future hardware state.

[0025] Calculate the prediction confidence of the future hardware state and output the prediction confidence.

[0026] The inference data is output by combining the future hardware state and the prediction confidence.

[0027] In one embodiment, the step of generating hardware control instructions based on the prediction results in the inference data includes:

[0028] Based on the temperature prediction results and temperature change rate in the inference data, a fan speed control command is generated.

[0029] Based on the power health prediction results in the inference data, power management instructions are generated.

[0030] Based on the load change prediction results in the inference data, a sampling frequency adjustment instruction is generated;

[0031] The fan speed control command, power management command, and sampling frequency adjustment command are combined to output hardware control commands.

[0032] In one embodiment, the step of generating a fan speed control command based on the temperature prediction result and temperature change rate in the inference data includes:

[0033] Obtain the current CPU temperature data, temperature change rate data, and predicted temperature trend data from the inference data;

[0034] The current CPU temperature data and temperature change rate data are processed by a fuzzy control algorithm to output the fan PWM duty cycle base value.

[0035] The compensation value is calculated based on the predicted temperature trend data, and combined with the fan PWM duty cycle base value, the fan PWM duty cycle command is output.

[0036] In one embodiment, the step of generating power management instructions based on the power health prediction results in the inference data includes:

[0037] Obtain the voltage fluctuation parameters and current harmonic parameters from the inference data;

[0038] The voltage fluctuation parameters and current harmonic parameters are processed based on the hardware health model, and the power supply health is output.

[0039] When the power health level is below the threshold, a redundant power supply pre-start command is generated and output.

[0040] In one embodiment, the step of converting the hardware control instructions into control signals, executing hardware control operations, and outputting execution feedback data includes:

[0041] Convert the fan speed control command in the hardware control instructions into a PWM signal;

[0042] Convert the power management instructions in the hardware control instructions into switching signals;

[0043] The sampling frequency adjustment instruction in the hardware control instruction is converted into a sensor control signal;

[0044] The PWM signal, switching signal, and sensor control signal are sent to the hardware components via the bus.

[0045] Real-time acquisition of hardware status data after execution, and output of execution feedback data.

[0046] In one embodiment, the step of performing strategy optimization and status feedback based on the execution feedback data, and outputting an updated strategy library includes:

[0047] The execution feedback data is sent to the main processor via the internal communication interface;

[0048] When the prediction confidence in the inference data is lower than the threshold, abnormal data is sent to the management platform via an external communication protocol.

[0049] Based on the execution feedback data, the AI ​​model parameters and control strategies are updated, and the updated strategy library is output.

[0050] Furthermore, to achieve the above objectives, this application also proposes a server BMC intelligent management system based on edge computing. The server BMC intelligent management system based on edge computing includes: a memory, a processor, and an edge computing-based server BMC intelligent management program stored on the memory and executable on the processor. The edge computing-based server BMC intelligent management program is configured to implement the steps of the server BMC intelligent management method based on edge computing.

[0051] The server BMC intelligent management method and system proposed in this application realizes a complete closed loop of data acquisition, processing and hardware control at the edge by dynamically adjusting the sensor sampling frequency, localized AI inference and closed-loop control strategy. This effectively improves data processing efficiency, reduces the load on the main processor and network transmission pressure, and enhances the system autonomy through predictive maintenance. Attached Figure Description

[0052] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0053] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 A flowchart illustrating an embodiment of the server BMC intelligent management method based on edge computing provided in this application;

[0055] Figure 2 For this application Figure 1 A detailed flowchart of step S100;

[0056] Figure 3 For this application Figure 1 A detailed flowchart of step S200;

[0057] Figure 4 For this application Figure 1 Detailed flowchart of step S300;

[0058] Figure 5 For this application Figure 1 Detailed flowchart of step S400;

[0059] Figure 6 For this application Figure 5 A detailed flowchart of step S410;

[0060] Figure 7 For this application Figure 5 A detailed flowchart of step S420;

[0061] Figure 8 For this application Figure 1 A detailed flowchart of step S500;

[0062] Figure 9 For this application Figure 1A detailed flowchart of step S600;

[0063] Figure 10 This is a schematic diagram of a structural embodiment of the BMC intelligent management system for servers based on edge computing, as provided in this application.

[0064] Explanation of icon numbers:

[0065] 10. Memory; 20. Processor.

[0066] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0067] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of this application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0068] It should be understood that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0069] In existing technologies, server baseboard management controllers are responsible for hardware monitoring and remote management. However, traditional solutions suffer from low data processing efficiency, rudimentary energy management, and a lack of intelligence. As data centers expand, the surge in sensor data leads to network bandwidth pressure and response latency with traditional methods. Fixed sampling frequencies result in energy waste and missed anomaly detection under both high and low load scenarios, and static threshold rules cannot achieve fault prediction and dynamic adjustment. In a cloud computing center operation and maintenance scenario, server clusters frequently experience temperature anomalies with delayed processing, leading to hardware damage. Simultaneously, redundant sensors continuously sampling at high frequencies waste power resources, necessitating an autonomous intelligent management solution.

[0070] To address these issues, researchers discovered a core contradiction in traditional solutions: lengthy data processing paths and rigid decision-making mechanisms. Analysis revealed that embedding edge computing capabilities into the BMC (Blockchain Management Console) could shorten the data processing chain, but the challenge of real-time analysis under limited computing power remained. Further research showed that dynamically adjusting the sampling frequency could balance data accuracy and energy consumption, but this required establishing a load assessment model. The final technical approach was to construct a localized data processing pipeline, achieving autonomous decision-making through dynamic sampling and a lightweight model, forming a closed-loop control system.

[0071] Based on this, embodiments of this application provide a server BMC intelligent management method based on edge computing, referring to... Figure 1 The edge computing-based server BMC intelligent management method includes steps S100 to S600, wherein:

[0072] Step S100: Acquire raw sensor data through a multi-protocol sensor interface; calculate the server load index based on the raw sensor data; dynamically adjust the sampling frequency according to the server load index, and output the adjusted sampling data;

[0073] Step S200: Perform denoising and normalization processing on the adjusted sampled data, and extract features based on the processing results to output a multi-dimensional feature vector;

[0074] Step S300: Input the multidimensional feature vector into the pre-trained lightweight AI model for analysis and reasoning, predict the future hardware state and calculate the prediction confidence, and output reasoning data containing the prediction results and confidence.

[0075] Step S400: Based on the prediction results in the inference data, generate hardware control instructions;

[0076] Step S500: Convert the hardware control command into a control signal and execute the hardware control operation, and output execution feedback data;

[0077] Step S600: Based on the execution feedback data, perform strategy optimization and status feedback, and output the updated strategy library.

[0078] In this embodiment, the multi-protocol sensor interface refers to a hardware interface module that supports I2C, SMBus, and PMBus protocols. Specifically, it can be implemented using a multiplexer in conjunction with a protocol conversion chip to ensure compatibility with different types of sensors. The server load index is a comprehensive indicator reflecting the current intensity of hardware resource usage. Specifically, it can be implemented using a weighted algorithm to process CPU utilization, memory bandwidth, and PCIe throughput data, providing a quantitative basis for dynamically adjusting the sampling frequency. Dynamically adjusting the sampling frequency refers to a mechanism that changes the data acquisition rate based on load intensity. Specifically, it can be implemented using a threshold comparator in conjunction with a clock generator, reducing the sampling frequency to decrease energy consumption under low load and increasing the sampling frequency to ensure data integrity under high load. The lightweight AI model refers to a neural network inference engine that has undergone model compression. Specifically, it can be implemented using an LSTM network optimized with knowledge distillation technology, enabling real-time inference in embedded environments. Policy optimization refers to the process of adjusting control parameters based on execution feedback. Specifically, it can be implemented using reinforcement learning algorithms to update the control strategy online, continuously improving the system's adaptive capabilities.

[0079] In this embodiment, the hardware sensor transmits raw data to the data processing module in real time via a multi-protocol interface. After load index calculation, a sampling frequency adjustment mechanism is triggered. The adjusted data stream undergoes median filtering to remove noise interference, and after standardization, it is input into the feature extraction module to generate a multi-dimensional vector. This vector is input into a pre-set lightweight model for inference, outputting a prediction result and confidence level including temperature trend and power health. The control command generation module dynamically adjusts parameters such as fan speed and power mode based on the prediction results, and the converted control signal is sent to the hardware execution unit via the bus. The executed status data is fed back to the strategy optimization module, which updates the model parameters and control rules through an online learning mechanism, forming a closed-loop optimization system.

[0080] Compared to existing technologies, traditional solutions require transmitting raw data to an external platform for processing, while this solution completes the entire data acquisition, processing, and decision-making process within the BMC, eliminating network transmission latency. Existing technologies use fixed sampling frequencies, leading to resource waste; this solution optimizes energy efficiency through a dynamic adjustment mechanism. Compared to control methods relying on preset thresholds, this solution proactively adjusts hardware status based on prediction results, proactively mitigating potential fault risks. Through these technical solutions, this application achieves localized real-time processing of hardware status data, effectively reducing main processor load and network bandwidth usage. The dynamic sampling mechanism balances data accuracy and energy efficiency, reducing power consumption while ensuring anomaly detection capabilities. Predictive maintenance strategies proactively identify hardware fault risks, shortening fault response time and improving server operational stability. A closed-loop feedback mechanism continuously optimizes the control strategy, enhancing the system's adaptability to different load scenarios.

[0081] In one feasible implementation, refer to Figure 2 Step S100 includes steps S110 to S140, wherein:

[0082] Step S110: Acquire CPU utilization data, memory bandwidth usage data, and PCIe device throughput data through sensors;

[0083] Step S120: Calculate the server load index based on the CPU utilization data, memory bandwidth usage data, and PCIe device throughput data using a weighted average.

[0084] Step S130: When the server load index is lower than the first threshold, reduce the sampling frequency and start data compression, and output the adjusted sampling data;

[0085] Step S140: When the server load index is higher than the second threshold, increase the sampling frequency and enable real-time data stream mode to output the adjusted sampling data.

[0086] In this embodiment, the multi-protocol sensor interface refers to a hardware interface module that supports communication protocols such as I2C, SMBus, and PMBus. Specifically, it can be implemented using a multiplexer combined with a protocol conversion chip, ensuring compatibility with sensor devices from different manufacturers and enabling unified data acquisition. The server load index is a quantitative indicator reflecting the real-time workload of the server. Specifically, it can be calculated by assigning weight coefficients to CPU utilization, memory bandwidth usage, and PCIe device throughput, and then performing a weighted summation to comprehensively assess the server's current load status. Dynamically adjusting the sampling frequency refers to automatically changing the sensor data acquisition cycle based on changes in the load index. This can be implemented using a programmable timer module, extending the sampling interval when the load index is below a set threshold and shortening it when it is above another threshold.

[0087] In this embodiment, during server operation, three key metrics—CPU utilization, memory bandwidth usage, and PCIe device throughput—are continuously collected via a multi-protocol interface. These three metrics are assigned different weighting coefficients; for example, CPU utilization is weighted at 0.5, memory bandwidth usage at 0.3, and PCIe throughput at 0.2. A load index value ranging from 0 to 100 is generated through weighted summation. When the index is below 30, a low-load state is identified, and the temperature sensor's sampling frequency is reduced from 1Hz to 0.2Hz, while the LZ4 compression algorithm is used to compress and store the collected data. When the index exceeds 70, a high-load state is identified, and the voltage sensor's sampling frequency is increased from 1Hz to 10Hz. Simultaneously, a real-time data stream transmission mode is enabled to ensure that abnormal data is uploaded without delay.

[0088] In this embodiment, a dynamic adjustment mechanism reduces invalid data collection and transmission during low-load phases. For example, when the server is idle, the fan speed sensor collects data only 3 times per hour, compared to the traditional method of collecting data once per minute, thus reducing the amount of data processing. During high-load phases, by increasing the sampling frequency of PCIe device throughput, it is possible to capture instantaneous peak data that may be missed by traditional solutions, solving the problem of mismatch between sensor data collection and the actual server load.

[0089] In one feasible implementation, refer to Figure 3 Step S200 includes steps S210 to S230, wherein:

[0090] Step S210: Perform denoising processing on the adjusted sampled data, using a median filtering algorithm to remove interference signals, and output denoised data;

[0091] Step S220: Perform normalization processing on the denoised data, unify the data volume using a standardization method, and output standardized data;

[0092] Step S230: Extract a multidimensional feature vector based on the standardized data, including the current sampled value, historical moving average and standard deviation, and output the multidimensional feature vector.

[0093] In this embodiment, the median filtering algorithm is a nonlinear signal processing technique. Specifically, it uses a sliding window approach to sort the sampled data and take the median value to eliminate impulse noise and sudden interference. This algorithm removes outliers while preserving data mutation characteristics, avoiding signal distortion caused by traditional mean filtering. Standardization methods refer to converting sensor data with different dimensions into a unified numerical range. Specifically, Z-score standardization or maximum / minimum scaling can be used to improve the convergence speed of subsequent model training by eliminating dimensional differences. Multidimensional feature vectors are sets of vectors containing time-series statistical features. The current sampled value reflects the real-time state, the historical moving average reflects trend changes, and the standard deviation characterizes the degree of data fluctuation. By combining multidimensional features, the hardware operating status can be comprehensively described, providing effective input for subsequent model inference.

[0094] In this embodiment, under edge computing conditions, the raw data collected by the sensors, even after dynamically adjusting the sampling frequency, may still contain noise interference and dimensional differences. The denoising stage employs a median filtering algorithm, for example, setting a sliding window with a window length of 5, sorting each sampling point and its two adjacent data points, and taking the median value to effectively filter out instantaneous spike noise. The normalization stage standardizes different types of sensor data, such as temperature and voltage, for example, linearly mapping temperature data to the 0-1 interval and converting voltage data to a Z-score distribution. The feature extraction stage calculates three-dimensional features from the standardized data: the sensor reading at the current moment, the moving average of the past 10-second window, and the standard deviation of the past 30-second window. These three features are combined into a vector form and used as input to a lightweight AI model.

[0095] In this embodiment, median filtering preserves effective signal edge features, standardization eliminates dimensional differences in multi-source data, and multidimensional feature extraction constructs a more comprehensive state representation, significantly improving data quality. Denoising effectively eliminates the interference of random noise on feature extraction, normalization ensures the comparability of data from different types of sensors, and multidimensional feature combination enhances the hardware state representation capability, providing high-quality input data for subsequent AI models, thereby improving the accuracy of fault prediction.

[0096] In one feasible implementation, refer to Figure 4 Step S300 includes steps S310 to S340, wherein:

[0097] Step S310: Load a pre-trained lightweight AI model that matches the current scene on the server;

[0098] Step S320: Input the multi-dimensional feature vector into the lightweight AI model to perform inference calculation and output the predicted future hardware state;

[0099] Step S330: Calculate the prediction confidence of the future hardware state and output the prediction confidence.

[0100] Step S340: Combine the future hardware state and prediction confidence to output inference data.

[0101] In this embodiment, the pre-trained lightweight AI model refers to a machine learning model trained offline and suitable for edge computing environments. Specifically, it can be implemented using a pruned neural network model. This model reduces computational complexity by removing redundant neuron layers, enabling real-time inference under the limited computing resources of the BMC. The multidimensional feature vector refers to a composite data structure containing the current sampled value, historical moving average, and standard deviation. Specifically, it can be implemented by extracting temporal features using a sliding window algorithm. This vector can reflect the dynamic changing trend of the hardware state. The prediction confidence is a quantitative indicator of the reliability of the model's output results. Specifically, it can be implemented using probability distribution calculations or the confidence score output by the model itself. This indicator is used to evaluate the accuracy of the prediction results and provide a reference for subsequent control decisions.

[0102] In this embodiment, during server operation, a corresponding lightweight AI model version is selected and loaded based on the current hardware configuration and load type. The standardized multidimensional feature vectors are input into the model, and forward propagation is performed through the model's internal computation graph to output predicted hardware state values ​​within a specific future time window. Prediction confidence is calculated by analyzing the probability distribution or activation intensity of the model's output layer. For example, if the standard deviation of the predicted temperature value's probability distribution exceeds a set range, the confidence level is marked as low. The final output inference data includes hardware state parameters such as predicted temperature and voltage deviation trends, along with their corresponding confidence scores, forming structured data for subsequent control module use.

[0103] In this embodiment, a lightweight AI model is deployed locally on the BMC to achieve millisecond-level real-time inference computation, while confidence metrics are used to enhance decision reliability. Compared to traditional threshold judgment mechanisms, this solution can identify complex features such as temperature change rate and load fluctuation trend, predict abnormal hardware states in advance, and solve the problem that traditional BMCs cannot autonomously perform complex fault prediction. Data analysis is completed at the edge to avoid main processor resource occupation, and the reliability of hardware control commands is improved through a confidence evaluation mechanism, enabling early warning and precise control of server hardware status.

[0104] In one feasible implementation, refer to Figure 5 Step S400 includes steps S410 to S440, wherein:

[0105] Step S410: Based on the temperature prediction results and temperature change rate in the inference data, generate a fan speed control command;

[0106] Step S420: Based on the power health prediction results in the inference data, generate power management instructions;

[0107] Step S430: Generate a sampling frequency adjustment instruction based on the load change prediction result in the inference data;

[0108] Step S440: Combine the fan speed regulation instruction, the power management instruction, and the sampling frequency adjustment instruction, and output a hardware control instruction.

[0109] In this embodiment, the temperature prediction result refers to the prediction data of the server hardware temperature trend by the AI model, which can be specifically implemented by processing historical temperature data using a time series prediction algorithm, and is used to pre-sense the risk of temperature anomalies. The temperature change rate refers to the rate of temperature change per unit time, which can be specifically implemented by calculating the temperature difference between adjacent sampling points through differentiation, and is used to judge the severity of the temperature fluctuation trend. The power health prediction result refers to the quantitative evaluation value of the stability of the power module, which can be specifically calculated by using voltage fluctuation parameters and current harmonic parameters as input features through a regression model, and is used to identify the risk of power aging or failure. The load change prediction result refers to the estimation of the future computing resource requirements of the server, which can be specifically implemented by analyzing historical load data using a sliding window statistical method, and is used to dynamically match the sensor sampling frequency with the actual requirements.

[0110] In this embodiment, when the inference data contains the temperature prediction result, the change trend of the heat dissipation requirement can be pre-judged by combining the temperature change rate, and then the corresponding fan speed regulation instruction can be generated. For example, if the predicted temperature shows an upward trend and the change rate exceeds the preset range, an instruction to increase the fan speed can be generated. When the power health prediction result is lower than the threshold, a redundant power supply start instruction is generated to ensure power supply stability. The load change prediction result is used to dynamically adjust the sensor sampling frequency. For example, when it is predicted that the future load will decrease, the sampling frequency is reduced to save energy consumption. The above three instructions are combined to form a comprehensive hardware control instruction set, and multi-dimensional collaborative control of the server hardware is achieved through parallel execution.

[0111] It can be understood that the existing solution can only generate a single fan control instruction based on the current temperature threshold, cannot perform dynamic adjustment by combining the temperature change trend, and lacks the linkage response to the power health status and load changes. However, this solution generates a composite instruction by integrating multiple prediction results, realizes the initiative and coordination of the hardware control strategy, and solves the problems of lagging control instructions and resource waste in traditional methods. In this way, this application can generate accurate hardware control instructions based on multi-dimensional prediction data, actively adjust the heat dissipation strategy before the temperature anomaly occurs, start redundant protection in advance when the power health deteriorates, and dynamically optimize the sensor sampling frequency according to the load change, so as to improve the hardware control response speed, reduce the anomaly handling delay, and reduce the ineffective energy consumption.

[0112] In a feasible implementation, refer to Figure 6Step S410 includes steps S411 to S413, wherein:

[0113] Step S411: Obtain the current CPU temperature data, temperature change rate data, and predicted temperature trend data from the inference data;

[0114] Step S412: Process the current CPU temperature data and temperature change rate data using a fuzzy control algorithm, and output the fan PWM duty cycle base value;

[0115] Step S413: Calculate the compensation value based on the predicted temperature trend data, and output the fan PWM duty cycle command in combination with the fan PWM duty cycle base value.

[0116] In this embodiment, the fuzzy control algorithm refers to a control method based on fuzzy logic rules. Specifically, it can be implemented using a fuzzy inference system with temperature deviation and rate of change as input variables. By setting the membership function and fuzzy rule table for temperature deviation and rate of change, continuous input quantities are converted into fuzzy sets for inference, and then the precise output value is obtained through defuzzification. This algorithm can handle nonlinear relationships in temperature control and avoids the dependence of traditional PID control on precise mathematical models. Predicted temperature trend data refers to the prediction results of the direction of temperature change over a future period by an AI model. Specifically, it can be implemented using the temperature change slope or trend classification label output by a time series prediction model. This data is used to predict the direction of temperature change in advance, providing a forward-looking compensation basis for fan speed adjustment. The PWM duty cycle base value refers to the basic speed adjustment value calculated by the fuzzy control algorithm, reflecting the basic heat dissipation requirements under the current temperature conditions. The compensation value refers to the dynamic correction amount of the base value based on the predicted temperature trend. Specifically, it can be implemented by multiplying the trend strength coefficient with the base value or by superimposing a fixed increment, and is used to adjust the fan speed in advance to cope with the upcoming temperature changes.

[0117] In this embodiment, the current CPU temperature data and temperature change rate data are input into the fuzzy control algorithm. After fuzzification, rule inference, and defuzzification processes, a PWM duty cycle base value matching the current heat dissipation requirements is generated. Simultaneously, the predicted temperature trend data is converted into a compensation coefficient by the trend intensity quantization module. This coefficient is then linearly superimposed or nonlinearly combined with the base value to ultimately generate an optimized PWM duty cycle command. For example, when the predicted temperature shows an upward trend, the compensation coefficient is positive, causing the duty cycle command to be higher than the current required base value, thus increasing the fan speed in advance to suppress the temperature rise. When the predicted temperature tends to stabilize, the compensation coefficient returns to zero, maintaining the fan speed level corresponding to the base value.

[0118] Understandably, existing solutions typically rely solely on threshold judgments based on the current temperature value. For example, they might directly switch fan speeds once the temperature exceeds a fixed threshold, leading to control lag and frequent speed changes. This solution, however, uses a fuzzy control algorithm to process the rate of temperature change, identifying the dynamic characteristics of temperature variations. Combined with forward-looking compensation based on predictive trend data, it synchronizes fan speed adjustments with the temperature change process, avoiding sudden speed changes and reducing temperature fluctuations. This solves the problems of lag and unstable speed regulation in traditional BMC temperature control, achieving refined fan control based on multi-dimensional temperature data. This ensures efficient heat dissipation while reducing fan energy consumption, extending fan lifespan, and effectively suppressing the impact of temperature fluctuations on server hardware stability.

[0119] In one feasible implementation, refer to Figure 7 Step S420 includes steps S421 to S423, wherein:

[0120] Step S421: Obtain the voltage fluctuation parameters and current harmonic parameters from the inference data;

[0121] Step S422: Process the voltage fluctuation parameters and current harmonic parameters based on the hardware health model, and output the power supply health status;

[0122] Step S423: When the power health is below the threshold, generate and output a redundant power supply pre-start command.

[0123] In this embodiment, the voltage fluctuation parameter refers to the variation amplitude of the power supply output voltage within a preset time window. Specifically, it can be calculated using real-time data collected by a voltage sensor and the standard deviation, reflecting the stability of the power supply output. The current harmonic parameter refers to the degree of distortion of non-fundamental frequency components in the AC current. Specifically, it can be implemented using a Fast Fourier Transform algorithm to perform spectral analysis on the current waveform, used to assess the interference level of the power supply line. The hardware health model is a power supply status assessment model trained based on historical fault data. Specifically, it can be implemented using a support vector machine or random forest algorithm to construct a classification model, used to predict power supply reliability by comprehensively considering voltage fluctuation and current harmonic parameters. The threshold is the critical health value that triggers the redundant power supply startup. Specifically, it can be determined through experimental testing or historical fault data analysis, for example, set to trigger the command when the health score is below 60. The redundant power supply pre-start command is the command to control the backup power module to enter standby mode. Specifically, it can be implemented by sending specific control codes through the power management bus, used to shorten the switching delay when the main power supply fails.

[0124] In this embodiment, during the power management command generation process, voltage fluctuation parameters and current harmonic parameters are first extracted from the inference data. These two parameters reflect the stability of the power output and the level of line interference, respectively. Subsequently, the hardware health model performs a comprehensive analysis of these two parameters and outputs a quantified power health score. When the score is lower than a preset threshold, it indicates that the main power supply has a high risk of failure, and at this time, a redundant power supply pre-start command is automatically generated. This command is sent to the backup power module through the power management bus, causing it to enter a low-power standby state in advance. Thus, in the event of a sudden failure of the main power supply, the backup power supply can complete the switchover within milliseconds, preventing the server from crashing due to power interruption.

[0125] In this embodiment, by integrating multi-dimensional electrical parameter analysis and combining it with a machine learning model to dynamically assess the power supply health status, potential fault risks can be detected earlier. Furthermore, by advancing the switching preparation phase through a pre-start mechanism, the fault response time is significantly shortened, thereby achieving intelligent assessment and proactive protection of the server power supply health status. This solves the problems of delayed power fault response and high false alarm rates in traditional solutions. By dynamically monitoring voltage fluctuations and current harmonic parameters, and combining them with predictive models to identify power supply degradation trends in advance, the redundant module preparation process can be initiated before the main power supply completely fails, effectively avoiding system downtime caused by power switching delays.

[0126] In one feasible implementation, refer to Figure 8 Step S500 includes steps S510 to S550, wherein:

[0127] Step S510: Convert the fan speed control command in the hardware control command into a PWM signal;

[0128] Step S520: Convert the power management instruction in the hardware control instruction into a switching signal;

[0129] Step S530: Convert the sampling frequency adjustment instruction in the hardware control instruction into a sensor control signal;

[0130] Step S540: Send the PWM signal, switching signal and sensor control signal to the hardware components via the bus;

[0131] Step S550: Collect the status data after hardware execution in real time and output the execution feedback data.

[0132] In this embodiment, the PWM signal refers to controlling the fan speed by adjusting the duty cycle of the pulse width modulation signal. Specifically, it can be implemented by using the timer module of the embedded system to generate square wave signals with different duty cycles. Its function is to convert the logic control instructions into physical signals that can be recognized by the fan drive circuit. The switching signal refers to the binary level signal used to control the power supply on and off. Specifically, it can be implemented by using optocouplers or relay circuits to achieve high-low level conversion. Its function is to convert the power management instructions into electrical signals that can trigger the power module state switching. The sensor control signal refers to the configuration instruction to adjust the sensor sampling frequency. Specifically, it can be implemented by modifying the sensor register parameters through the I2C or SPI bus protocol. Its function is to dynamically adjust the data acquisition frequency according to the load to balance energy consumption and monitoring accuracy. Bus transmission refers to the transmission of control signals through a standard communication interface. Specifically, it can be implemented using IPMB, SMBus, or a custom serial bus protocol. Its function is to transmit multiple types of control signals to the target hardware component in a unified manner. Execution feedback data refers to the real-time status information after the hardware executes the control. Specifically, it can be implemented by sensor feedback data or reading from the hardware status register. Its function is to provide closed-loop verification basis for subsequent strategy optimization.

[0133] In this embodiment, the fan speed control command, power management command, and sampling frequency adjustment command in the hardware control instructions are converted into corresponding physical control signals. For example, the fan speed control command is converted into a pulse signal with a specific duty cycle by the PWM generator of the embedded system, directly driving the fan motor to adjust its speed; the power management command is converted into a high / low level signal by a switching circuit to control the start or stop of the redundant power supply module; the sampling frequency adjustment command modifies the configuration register of the sensor chip through the bus protocol to dynamically change its data acquisition interval. After the above control signals are transmitted to the target hardware through the bus, the system collects feedback data such as the actual fan speed, the working status of the power supply module, and the latest sampling period of the sensor in real time, forming a complete control closed loop. Thus, the execution effect of the hardware control operation can be monitored in real time, providing data support for subsequent strategy optimization.

[0134] In this embodiment, bidirectional interaction between control commands and hardware responses is achieved through bus signal conversion and feedback data acquisition. Furthermore, this solution completes the command-to-signal conversion within the BMC, reducing the intervention of the main processor. Simultaneously, the bus protocol adaptation mechanism ensures compatibility with multiple hardware interface standards, solving the problems of low signal conversion efficiency and invisible execution status in traditional BMC hardware control. By directly converting logical instructions into physical control signals, processing delays caused by main processor involvement are avoided. Real-time acquired feedback data allows the system to verify the execution effect of control commands; for example, immediately detecting whether the actual fan speed reaches the expected value after fan speed adjustment, thus promptly identifying hardware anomalies. A unified transmission mechanism for multiple types of control signals improves hardware compatibility; for example, the same bus can simultaneously transmit PWM signals and switching signals. The closed-loop control mechanism provides a data foundation for subsequent strategy optimization; for example, when feedback data shows excessively high power switching delay, the triggering timing of the pre-start command can be automatically adjusted.

[0135] In one feasible implementation, refer to Figure 9 Step S600 includes steps S610 to S630, wherein:

[0136] Step S610: The execution feedback data is sent to the main processor through the internal communication interface;

[0137] Step S620: When the prediction confidence in the inference data is lower than the threshold, abnormal data is sent to the management platform through an external communication protocol;

[0138] Step S630: Update the AI ​​model parameters and control strategy based on the execution feedback data, and output the updated strategy library.

[0139] In this embodiment, the execution feedback data refers to the actual state data after the hardware control operation is executed. Specifically, it can be implemented using real-time temperature, voltage, and fan speed data collected by sensors, used to verify the actual effect of the control commands. The internal communication interface refers to the data transmission channel between the BMC and the main processor, specifically implemented using PCIe or I2C bus protocols to ensure efficient data transmission. The prediction confidence threshold is the critical value for judging the reliability of the AI ​​model's inference results, specifically set through the validation set accuracy during model training, for example, set to 85%. The external communication protocol refers to the communication method between the BMC and the management platform, specifically implemented using IPMI or Redfish protocols, used to transmit abnormal data. Abnormal data refers to the set of information where the prediction result and the execution feedback data deviate significantly, specifically including the difference between the predicted temperature and the actual temperature, and power health deviation values, used to trigger the management platform's alarm mechanism. AI model parameters and control strategies refer to the weight parameters and hardware control rules of the lightweight AI model, specifically updated through online learning algorithms or incremental training methods to achieve dynamic optimization.

[0140] In this embodiment, after the execution feedback data is transmitted to the main processor via the internal communication interface, the main processor can synchronously update the global state view to ensure overall system coordination. When the prediction confidence level is lower than the threshold, the external communication protocol is activated, and the abnormal data is encapsulated into a specific format message and sent to the management platform, such as using JSON format containing a timestamp, device ID, and abnormality type. Simultaneously, locally on the BMC, the execution feedback data is input to the online learning module, which adjusts the AI ​​model weights using a gradient descent algorithm and corrects the parameters in the control strategy based on the feedback results, such as adjusting the fuzzy control rule coefficients of the fan speed control command. The updated strategy library is stored in non-volatile memory and takes effect in the next control cycle.

[0141] In some specific implementations, the internal communication interface can use a PCIe bus to achieve high-speed data transmission, the external communication protocol can be configured as the Redfish RESTful API to be compatible with the cloud management platform, and the prediction confidence threshold can be dynamically adjusted according to different hardware types, for example, the threshold is set to 80% for GPU servers and 90% for storage servers. The transmission of abnormal data can use an asynchronous queue mechanism to avoid blocking the real-time control process.

[0142] In this embodiment, by processing execution feedback data locally in real time within the BMC, model parameters and control strategies are directly updated, avoiding network transmission delays. Simultaneously, this solution triggers abnormal data reporting through a confidence threshold, achieving synergy between proactive early warning and local optimization. Thus, this application solves the problem of policy update delays caused by the reliance on external resources in traditional BMCs, improving the real-time performance of hardware control; through the dual mechanism of localized policy optimization and abnormal data reporting, it enhances the system's autonomous operation capability in the event of network outages or main processor failures; and by dynamically updating AI model parameters, it achieves continuous adaptation of control strategies to hardware state changes, improving energy management accuracy.

[0143] In the embodiments of this application, the server BMC intelligent management method based on edge computing achieves a complete closed loop of data acquisition, processing and hardware control at the edge by dynamically adjusting the sensor sampling frequency, localized AI inference and closed-loop control strategy. This effectively improves data processing efficiency, reduces the load on the main processor and the pressure on network transmission, and enhances system autonomy through predictive maintenance.

[0144] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the server BMC intelligent management method based on edge computing in this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0145] This application also provides a server BMC intelligent management system based on edge computing, see reference. Figure 10 The edge computing-based server BMC intelligent management system includes: a memory 10, a processor 20, and an edge computing-based server BMC intelligent management program stored on the memory 10 and executable on the processor 20. The edge computing-based server BMC intelligent management program is configured to implement the steps of the edge computing-based server BMC intelligent management method.

[0146] The edge computing-based server BMC intelligent management system provided in this application, employing the edge computing-based server BMC intelligent management method in the above embodiments, can improve data processing efficiency and reduce the load on the main processor. Compared with the prior art, the beneficial effects of the edge computing-based server BMC intelligent management system provided in this application are the same as those of the edge computing-based server BMC intelligent management method provided in the above embodiments, and other technical features in the edge computing-based server BMC intelligent management system are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0147] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0148] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. All equivalent structural transformations made under the technical concept of this application using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included within the scope of patent protection of this application.

Claims

1. A server BMC intelligent management method based on edge computing, characterized in that, The method includes: Raw sensor data is acquired through a multi-protocol sensor interface; a server load index is calculated based on the raw sensor data; the sampling frequency is dynamically adjusted according to the server load index, and the adjusted sampling data is output. The adjusted sampled data is subjected to denoising and normalization processing, and features are extracted based on the processing results to output a multi-dimensional feature vector; The multidimensional feature vectors are input into a pre-trained lightweight AI model for analysis and reasoning to predict future hardware states and calculate prediction confidence. The output is inference data containing prediction results and confidence. Based on the prediction results in the inference data, hardware control instructions are generated; The hardware control instructions are converted into control signals and hardware control operations are executed, outputting execution feedback data. Based on the execution feedback data, the execution strategy is optimized and the status feedback is processed, and an updated strategy library is output. The step of generating hardware control instructions based on the prediction results in the inference data includes: Based on the temperature prediction results and temperature change rate in the inference data, a fan speed control command is generated. Based on the power health prediction results in the inference data, power management instructions are generated. Based on the load change prediction results in the inference data, a sampling frequency adjustment instruction is generated; The fan speed control command, power management command, and sampling frequency adjustment command are combined to output hardware control commands; The step of generating fan speed control commands based on the temperature prediction results and temperature change rate in the inference data includes: Obtain the current CPU temperature data, temperature change rate data, and predicted temperature trend data from the inference data; The current CPU temperature data and temperature change rate data are processed by a fuzzy control algorithm to output the fan PWM duty cycle base value. The compensation value is calculated based on the predicted temperature trend data, and combined with the fan PWM duty cycle base value, the fan PWM duty cycle command is output. The step of generating power management instructions based on the power health prediction results in the inference data includes: Obtain the voltage fluctuation parameters and current harmonic parameters from the inference data; The voltage fluctuation parameters and current harmonic parameters are processed based on the hardware health model, and the power supply health is output. When the power health level is below the threshold, a redundant power supply pre-start command is generated and output.

2. The server BMC intelligent management method based on edge computing as described in claim 1, characterized in that, The raw sensor data is acquired through a multi-protocol sensor interface; The steps of calculating the server load index based on the raw sensor data, dynamically adjusting the sampling frequency according to the server load index, and outputting the adjusted sampling data include: Acquire CPU utilization data, memory bandwidth usage data, and PCIe device throughput data through sensors; The server load index is calculated by weighting the CPU utilization data, memory bandwidth usage data, and PCIe device throughput data. When the server load index is lower than the first threshold, the sampling frequency is reduced and data compression is started, and the adjusted sampling data is output. When the server load index is higher than the second threshold, the sampling frequency is increased and the real-time data stream mode is enabled, and the adjusted sampling data is output.

3. The server BMC intelligent management method based on edge computing as described in claim 1, characterized in that, The steps of performing denoising and normalization processing on the adjusted sampled data, extracting features based on the processing results, and outputting a multidimensional feature vector include: The adjusted sampled data is subjected to denoising processing, and the median filtering algorithm is used to remove interference signals and output denoised data. The denoised data is normalized to unify the data volume using a standardization method, and standardized data is output. Based on the standardized data, a multidimensional feature vector is extracted, including the current sampled value, historical moving average, and standard deviation, and the multidimensional feature vector is output.

4. The server BMC intelligent management method based on edge computing as described in claim 1, characterized in that, The steps of inputting the multidimensional feature vector into a pre-trained lightweight AI model for analysis and inference, predicting future hardware states, calculating prediction confidence, and outputting inference data containing prediction results and confidence levels include: Load a pre-trained lightweight AI model that matches the current scene on the server; The multidimensional feature vector is input into the lightweight AI model to perform inference calculations and output the predicted future hardware state. Calculate the prediction confidence of the future hardware state and output the prediction confidence. The inference data is output by combining the future hardware state and the prediction confidence.

5. The server BMC intelligent management method based on edge computing as described in claim 1, characterized in that, The step of converting the hardware control instructions into control signals, executing hardware control operations, and outputting execution feedback data includes: Convert the fan speed control command in the hardware control instructions into a PWM signal; Convert the power management instructions in the hardware control instructions into switching signals; The sampling frequency adjustment instruction in the hardware control instruction is converted into a sensor control signal; The PWM signal, switching signal, and sensor control signal are sent to the hardware components via the bus. Real-time acquisition of hardware status data after execution, and output of execution feedback data.

6. The server BMC intelligent management method based on edge computing as described in claim 1, characterized in that, The step of performing strategy optimization and status feedback based on the execution feedback data, and outputting an updated strategy library includes: The execution feedback data is sent to the main processor via the internal communication interface; When the prediction confidence in the inference data is lower than the threshold, abnormal data is sent to the management platform via an external communication protocol. Based on the execution feedback data, the AI ​​model parameters and control strategies are updated, and the updated strategy library is output.

7. A server BMC intelligent management system based on edge computing, characterized in that, The edge computing-based server BMC intelligent management system includes: a memory, a processor, and an edge computing-based server BMC intelligent management program stored on the memory and executable on the processor, wherein the edge computing-based server BMC intelligent management program is configured to implement the steps of the edge computing-based server BMC intelligent management method as described in any one of claims 1 to 6.

Images