An intelligent monitoring method for AI computing power data centers

Abstract

The present invention relates to the field of load monitoring and analysis technology, and specifically to an intelligent monitoring method for an AI computing power data center. The method includes: obtaining an average load data sequence and a temperature sequence; obtaining the predicted load at the next moment using an autoregressive moving average algorithm; calculating the high-frequency predicted load at the next moment under a high-frequency sampling scale; calculating the medium-frequency predicted load at the next moment under a medium-frequency sampling scale; calculating the low-frequency predicted load at the next moment under a low-frequency sampling scale; obtaining the weights of the high-frequency predicted load, the medium-frequency predicted load, and the low-frequency predicted load, and performing weighted summation to obtain the precise predicted load at the next moment; calculating the temperature threshold at the next moment using the predicted load and the precise predicted load; predicting the temperature data at the next moment based on the temperature sequence, recorded as the predicted temperature; and issuing a reminder if the predicted temperature is greater than or equal to the temperature threshold. The present application can accurately identify temperature anomalies in a computing power data center.

Description

Technical Field

[0001] The present invention relates to the field of load monitoring and analysis technology, and specifically to an intelligent monitoring method for an AI computing power data center. Background Art

[0002] With the rapid development of artificial intelligence, big data, and cloud computing, AI computing data centers are shouldering increasingly high computing loads, and servers are operating in a state of high power consumption and high heat output for a long time. Server overload and local overheating are among the major risks affecting data center stability, energy efficiency, and hardware lifespan. In severe cases, they can lead to compute node downtime, downtime, and even hardware damage. Therefore, how to accurately and in real time monitor server operating status, provide early warnings before problems occur, and optimize scheduling are key challenges in intelligent data center management.

[0003] Currently, intelligent monitoring in data centers primarily relies on sensors to collect key operational data such as CPU load, temperature, fan speed, and power consumption, and uses fixed threshold methods for anomaly detection. Traditional processing methods use pre-set fixed thresholds to determine whether the CPU is overloaded or the temperature is too high. Autoregressive moving average algorithms can also be used to predict CPU load and temperature, issuing warnings based on fixed thresholds. However, these are fixed threshold warnings for a single indicator and lack relevance in complex scenarios. For example, monitoring only the CPU load can lead to false alarms; under certain high-load conditions, the server remains within a safe temperature range. Monitoring only the temperature also has a lag effect; an alarm is only issued when the system overheats, after the critical point has been exceeded, making it difficult to optimize scheduling in a timely manner. Furthermore, irreversible damage to the hardware may occur. Summary of the Invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide an intelligent monitoring method for AI computing power data centers. The technical solutions adopted are as follows:

[0005] One embodiment of the present invention provides an intelligent monitoring method for an AI computing data center, the method comprising:

[0006] Collect the average load data of the CPU and the temperature data in the server and pre-process them to obtain the average load data series and temperature series;

[0007] The autoregressive moving average algorithm is used to predict the average load data sequence based on the current time to obtain the average load data of the next moment, which is recorded as the predicted load; different sampling scales are set, including high-frequency sampling scale, medium-frequency sampling scale and low-frequency sampling scale;

[0008] At the high-frequency sampling scale, the high-frequency predicted load at the next moment is calculated based on the average load data before the next moment; at the medium-frequency sampling scale, the medium-frequency predicted load at the next moment is calculated based on the average load data before the next moment; at the low-frequency sampling scale, the low-frequency predicted load at the next moment is calculated based on the average load data before the next moment;

[0009] Obtain the weights of the high-frequency predicted load, the medium-frequency predicted load, and the low-frequency predicted load based on the high-frequency predicted load and the medium-frequency predicted load, and perform weighted summation to obtain the accurate predicted load at the next moment;

[0010] The temperature threshold at the next moment is calculated using the predicted load and the accurate predicted load; the temperature data at the next moment is predicted based on the temperature sequence and recorded as the predicted temperature; if the predicted temperature is greater than or equal to the temperature threshold, an alert is issued.

[0011] Preferably, the average load data of the CPU and the temperature data in the server are collected and preprocessed to obtain an average load data sequence and a temperature sequence, including:

[0012] Preprocessing mainly includes data cleaning and normalization. Data cleaning includes missing value filling, outlier identification and duplicate data processing. The preprocessed average load data and temperature data in the server are respectively composed of the average load data sequence and temperature sequence.

[0013] Preferably, under the high-frequency sampling scale, the high-frequency predicted load at the next moment is calculated based on the average load data before the next moment, including:

[0014] A high-frequency sampling window is set under a high-frequency sampling scale. The high-frequency sampling window slides with a preset step size. After sliding to the current moment, the time series distance between an average load data in the high-frequency sampling window and the average load data at the current moment is obtained, which is recorded as the time series distance. An exponential function with a natural constant as the base is used to negatively correlate the product of the time series distance and the first adjustment coefficient to obtain the weight of the average load data. The weights of the average load data in the high-frequency sampling window are weighted and summed and averaged to obtain the high-frequency predicted load.

[0015] Preferably, under the intermediate frequency sampling scale, calculating the intermediate frequency predicted load at the next moment based on the average load data before the next moment includes:

[0016] The IF sampling window under the IF sampling scale is set. The IF sampling window slides with a preset step size. After sliding to the current moment, the average load data in the IF sampling window is used to calculate the IF predicted load at the next moment. The specific formula for calculating the IF predicted load is:

[0017] ,

[0018] in, Indicates the medium frequency predicted load at the next moment; It represents the mean of the preset number of average load data before the next moment; T represents the number of average load data in the intermediate frequency sampling window, , μ represents the window center, σ represents the smoothing coefficient; k and j represent the step index within the intermediate frequency sampling window; Represents the tjth average load data in the intermediate frequency sampling window; Represents the first data in the intermediate frequency sampling window; e represents a natural constant.

[0019] Preferably, at the low-frequency sampling scale, calculating the low-frequency predicted load at the next moment based on the average load data before the next moment includes:

[0020] Set the low-frequency sampling window under the low-frequency sampling scale. The low-frequency sampling window slides with a preset step size. After sliding to the current moment, the average load data in the low-frequency sampling window is used to calculate the low-frequency predicted load at the next moment. The specific formula for calculating the low-frequency predicted load is:

[0021] ,

[0022] in, represents the low-frequency predicted load at the next moment; A represents the number of average load data in the low-frequency sampling window; and They represent the tj-th and tj-1-th average load data in the low-frequency sampling window respectively; β represents the scaling factor; e represents the natural constant; Indicates the tth average load data in the low-frequency sampling window.

[0023] Preferably, obtaining the weights of the high-frequency predicted load, the medium-frequency predicted load, and the low-frequency predicted load according to the high-frequency predicted load and the medium-frequency predicted load includes:

[0024] An exponential function with a natural constant as the base is used to negatively map the difference between the high-frequency predicted load and the low-frequency predicted load to obtain a first mapping result, the first mapping result is multiplied by a first scaling factor and added to a preset value to obtain a first addition result, and the inverse of the first addition result is the weight of the high-frequency predicted load; an exponential function with a natural constant as the base is used to negatively map the difference between the low-frequency predicted load and the high-frequency predicted load to obtain a second mapping result, the second mapping result is multiplied by the first scaling factor and added to a preset value to obtain a second addition result, and the inverse of the second addition result is the weight of the low-frequency predicted load; the weight of the medium-frequency predicted load is obtained by subtracting the sum of the weight of the high-frequency predicted load and the weight of the low-frequency predicted load from the preset value.

[0025] Preferably, calculating the temperature threshold at the next moment using the predicted load and the accurate predicted load includes:

[0026] Multiply the difference between the precise predicted load and the predicted load and the first adjustment coefficient, and perform negative correlation mapping using an exponential function with a natural constant as the base to obtain a third mapping result; calculate the inverse of the third mapping result plus the preset value, calculate the difference between the inverse and the first adjustment coefficient, and multiply it by the second adjustment coefficient to obtain the multiplication result; the sum of the multiplication result and the temperature threshold at the current moment is the temperature threshold at the next moment.

[0027] Preferably, predicting the temperature data at the next moment based on the temperature sequence, recorded as predicted temperature, includes:

[0028] At the high-frequency sampling scale, the high-frequency sampling window slides with a preset step size. After sliding to the current moment, in the temperature sequence, the difference between the temperature data at the first moment in the high-frequency sampling window and the temperature data at the current moment is obtained, and compared with the number of temperature data in the high-frequency sampling window, the change rate of the average difference is obtained; the temperature data at the current moment is added to the average difference to obtain the temperature data at the next moment, which is recorded as the predicted temperature.

[0029] The embodiments of the present invention have at least the following beneficial effects: the present application pre-processes the collected average load data of the CPU and the temperature data in the server to obtain an average load data sequence and a temperature sequence, which can improve data quality and ensure the accuracy of subsequent analysis; then, according to the autoregressive moving average algorithm, the predicted load at the next moment of the current moment is predicted based on the average load data sequence, and then different sampling scales are set. The collected average load data is analyzed at different sampling scales to obtain high-frequency predicted load, medium-frequency predicted load and low-frequency predicted load, which are synthesized to obtain an accurate predicted load, and the predicted load and the accurate predicted load are used to calculate the temperature threshold of the next moment, and the temperature data of the next moment is predicted at the same time, and the temperature threshold is adaptively adjusted, thereby accurately identifying data center anomalies. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0031] Figure 1 A flow chart of a method for intelligent monitoring of an AI computing power data center provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the following, in conjunction with the accompanying drawings and preferred embodiments, describes in detail the specific implementation, structure, features and effects of an intelligent monitoring method for an AI computing data center proposed by the present invention. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable form.

[0033] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0034] The following describes in detail a specific solution of an intelligent monitoring method for an AI computing power data center provided by the present invention with reference to the accompanying drawings.

[0035] Example:

[0036] The main application scenario of this invention is: AI computing power data center needs to monitor the CPU load at all times to ensure safe operation and smooth progress of the process.

[0037] See also Figure 1 , which shows a method flow chart of an intelligent monitoring method for an AI computing power data center provided by an embodiment of the present invention, the method comprising the following steps:

[0038] Step S1: collect the average load data of the CPU and the temperature data in the server and perform preprocessing to obtain an average load data sequence and a temperature sequence.

[0039] Collect server operating status data, including average CPU load data and internal server temperature data. Load average data refers to the average number of processes in the running (Running / Runnable) and uninterruptible states per unit time, that is, the average number of active processes. A smaller value indicates less system workload and lower load; conversely, a higher load indicates a higher workload. Average load data can be directly obtained from the system. For example, in Linux, use the top command to view average load data.

[0040] Furthermore, the average load data and the temperature data inside the server need to be preprocessed. The preprocessing mainly includes data cleaning and normalization. This is because missing values, outliers or duplicate data may appear during the data collection process, so data cleaning is required. Data cleaning includes missing value filling, outlier identification and duplicate data processing; at the same time, since the average load data and temperature data have different dimensions, normalization is required to facilitate subsequent processing.

[0041] For data cleaning, if there are short-term missing CPU load average or temperature data (e.g., missing load average or temperature data for a particular server within a certain minute), the data is filled in using the sliding window average. For outliers, we use box plot analysis (IQR) to detect abnormal values ​​in the CPU load average data and remove those that fall outside the normal range. We use the Z-score method to detect temperature data anomalies. If a temperature reading exceeds three standard deviations from the mean, it may indicate a sensor failure or an extreme condition, requiring further confirmation. For duplicate data, if multiple identical records exist within a short period of time (e.g., within one second), only the first record is retained to avoid duplicate calculations.

[0042] For normalization, the Min-Max normalization method is used to map the data to the [0, 1] interval for subsequent model calculations. Finally, the preprocessed average load data and server temperature data are combined into an average load data series and a temperature series, respectively. The last moment in each series is the current moment's data. The sampling frequency for both the average load data series and the temperature data series is once per second.

[0043] Step S2: Use the autoregressive moving average algorithm to predict the average load data of the next moment based on the average load data sequence, and record it as the predicted load; set different sampling scales, including high-frequency sampling scale, medium-frequency sampling scale and low-frequency sampling scale.

[0044] Using an autoregressive moving average algorithm, we adaptively model the server's load state, allowing us to dynamically adjust the temperature threshold distribution under varying workloads. This automatically relaxes the temperature detection threshold under high load conditions to avoid false positives, while appropriately lowering the threshold under low load conditions where thermal anomalies may exist, thereby improving the sensitivity of anomaly detection.

[0045] Because changes in average load data can be affected by a variety of factors, including sudden computing tasks and long-term trend fluctuations, traditional autoregressive moving average algorithms can only predict based on a single feature. Adjusting the temperature threshold based on CPU load begins with monitoring CPU load. This allows us to distinguish between normal high loads (such as batch computing tasks) and abnormally high loads (such as overloads caused by system resource exhaustion), thereby reducing false positives.

[0046] A brief overview of the autoregressive moving average algorithm is as follows:

[0047] The autoregressive part is based on past observations and uses a linear combination of these historical data to predict the data at the next moment. The autoregressive model relies on the values ​​at previous time points to make predictions;

[0048] The moving average part uses the error terms of the current moment and the previous moment to make predictions, correcting the errors caused by the autoregressive part. The moving average model uses a linear combination of the error terms to improve the prediction;

[0049] Combining autoregression and moving average to capture both linear trends and random fluctuations in the data provides more accurate forecasts for time series.

[0050] Set the data in the average load data series to: ,in The autoregressive moving average algorithm (known technology) combines autoregressive and moving average components to represent the inherent laws of time series data. The specific formula can be expressed as:

[0051] ,

[0052] in, It represents the average load data predicted at the next moment, recorded as the predicted load, μ represents the constant term, the initial value is set to 0, p and q are the autoregressive order and moving average order of the model respectively, It represents the coefficient of autoregression, which indicates the influence of the average CPU load data of the previous p moments on the predicted average load data. i represents the distance between the historical data and the current moment, such as the distance between the moment t-1 and the current moment is t; The coefficient of the moving average represents the impact of the error term at the past q moments on the current load; is the white noise error term, which represents the unexplained part of the observations, Represents the error of white noise at time tj. Through historical load data, first estimate the parameters of autoregressive and moving average terms and , use the least squares method to estimate the parameters and update the model parameters. This can get the predicted average load data for the next moment after the current moment, which is also called the predicted load.

[0053] Due to the limitations of using the autoregressive moving average algorithm for prediction, the predicted load may be somewhat different from the actual data. Therefore, it is necessary to introduce progressive multi-scale analysis on this basis to conduct in-depth modeling of load-temperature change trends. The system can distinguish between short-term fluctuations (such as temperature fluctuations caused by instantaneous computing tasks) and long-term trends (such as decreased heat dissipation efficiency caused by continuous high temperatures of the server), thereby accurately identifying true anomalies rather than falsely reporting instantaneous temperature peaks.

[0054] In this application, multi-scale analysis refers to setting different sampling scales, collecting average load data at different sampling scales, and then analyzing them. The sampling scales include high-frequency sampling scales, medium-frequency sampling scales, and low-frequency sampling scales. Preferably, in the embodiment of this application, the high-frequency sampling scale is once per second, the medium-frequency sampling scale is once per minute, and the low-frequency sampling scale is once per hour. The implementer can also adjust it according to actual circumstances.

[0055] Step S3, under the high-frequency sampling scale, calculate the high-frequency predicted load at the next moment based on the average load data before the next moment; under the medium-frequency sampling scale, calculate the medium-frequency predicted load at the next moment based on the average load data before the next moment; under the low-frequency sampling scale, calculate the low-frequency predicted load at the next moment based on the average load data before the next moment.

[0056] This solution uses multi-scale sampling at three different scales: high, medium, and low. This multi-scale analysis ensures that the predicted data is closer to the actual value and can be compared with the average load data predicted by the autoregressive moving average algorithm. The comparison results are used to analyze whether the CPU is at risk of overload in the future. The high-frequency sampling scale represents instantaneous load changes with a high sampling frequency, the medium-frequency sampling scale represents the average load changes over a short period of time, and the low-frequency sampling scale represents the cumulative load changes over a longer period of time. The temperature threshold is adjusted based on the CPU load data at these multiple scales. The load level for a period of time is predicted based on the historical load series, and the temperature threshold is adjusted based on the predicted results.

[0057] The high-frequency sampling scale under the multi-scale mainly represents the instantaneous fluctuation of the CPU load at each moment. Under the high-frequency sampling scale, the high-frequency predicted load at the next moment is calculated based on the average load data before the next moment. The high-frequency sampling window under the high-frequency sampling scale is set to N. The reference value is 30 seconds, which can also mean that the amount of data in the window is 30, that is, the time difference between two adjacent moments is 1 second. The high-frequency sampling window is the average load data of the 30 moments before the next moment.

[0058] The high-frequency sampling window slides with a preset step size. After sliding to the current moment, the time series distance between an average load data in the high-frequency sampling window and the average load data at the current moment is obtained, which is recorded as the time series distance; an exponential function with a natural constant as the base is used to negatively correlate the product of the time series distance and the first adjustment coefficient to obtain the weight of the average load data; the weights of the average load data in the high-frequency sampling window are weighted and summed and averaged to obtain the high-frequency predicted load.

[0059] The specific calculation formula is:

[0060] ,

[0061] in, Indicates the high-frequency predicted load at the next moment of the current moment; N indicates the number of average load data in the high-frequency sampling window; i indicates the time series distance between the moment corresponding to an average load data and the current moment, that is, the time series distance. The closer the average load data is to the moment t, the greater the weight. Indicates the weight of the average load data corresponding to the ti-th moment, This value represents the instantaneous average CPU load data, which can capture instantaneous changes in the average load data. α represents the first adjustment coefficient. The reference value is 0.5, and the reference value of the preset step size is 1 second.

[0062] Furthermore, under the intermediate frequency sampling scale, the intermediate frequency predicted load at the next moment is calculated based on the average load data before the next moment. Specifically, the intermediate frequency sampling window under the intermediate frequency sampling scale is set, and the intermediate frequency sampling window slides with a preset step size. After sliding to the current moment, the intermediate frequency predicted load at the next moment is calculated using the average load data within the intermediate frequency sampling window. The calculation formula is as follows:

[0063] ,

[0064] in, Indicates the medium frequency predicted load at the next moment; It represents the mean of the preset number of average load data before the next moment; T represents the number of average load data in the intermediate frequency sampling window, , μ represents the window center, σ represents the smoothing coefficient; k and j represent the step index within the intermediate frequency sampling window; Represents the tjth average load data in the intermediate frequency sampling window; Represents the first data in the intermediate frequency sampling window; e represents a natural constant.

[0065] The IF predicted load is the average CPU load data predicted within the IF sampling window, balancing the noise interference in a short time window. Represents Gaussian weight, which represents Gaussian weight distribution and is used to control the smoothing effect of time series, so that the weight of data points close to the center is higher, while the weight of data points far away is gradually reduced. Indicates the difference in the average CPU load data at time tj The weight of Represents the weighted average difference within the intermediate frequency sampling window, giving closer data points higher weights. As the CPU load changes over time, it is necessary to add a short-term residual term to the baseline average to make the predicted value more accurate. It represents the mean of a preset number of average load data points before the next moment, where the preset number is 5. It can also be called the mean of the CPU average load data points of the last 5 sampling points in the window. Because the data obtained at the IF sampling scale can represent short-term fluctuations, the average of the most recent points is taken as the benchmark value. The length of the IF sampling window is 30 minutes, which means that the window can accommodate 30 average load data points at the IF sampling scale.

[0066] Finally, at the low-frequency sampling scale, the low-frequency predicted load at the next moment is calculated based on the average load data before the next moment.

[0067] Specifically, a low-frequency sampling window is set under a low-frequency sampling scale. The low-frequency sampling window slides with a preset step size. After sliding to the current moment, the average load data in the low-frequency sampling window is used to calculate the low-frequency predicted load at the next moment. The specific calculation formula is:

[0068] ,

[0069] in, represents the low-frequency predicted load at the next moment; A represents the number of average load data in the low-frequency sampling window; and They represent the tj-th and tj-1-th average load data in the low-frequency sampling window respectively; β represents the scaling factor; e represents the natural constant; Indicates the t-th average load data in the low-frequency sampling window, that is, the average load data at the t-th moment (the current moment).

[0070] in the formula Indicates the predicted CPU load in a low-frequency sampling window. It represents the average CPU load data accumulated over a longer window and can indicate the long-term trend of average CPU load data. It represents the weight of the average load data of adjacent CPUs. The smaller the difference between the two, the more stable the data, so the greater the weight is given to it. It represents the weighted difference of the adjacent data at time tj. After summing, and then adding the actual value at the current moment, it can be said that at this moment, the closer the data point is to time t, the greater the difference is, making the closer data points more representative and more accurate in predicting the actual data of the average load data of the CPU at time t+1.

[0071] Through multi-scale analysis, combining short-term, medium-term, and long-term time windows, we can obtain the predicted values ​​of the average load data for the next moment at the high-frequency sampling scale, the medium-frequency sampling scale, and the low-frequency sampling scale. Subsequent analysis can then yield more accurate load prediction data than that obtained using the autoregressive moving average algorithm.

[0072] Step S4: Obtain weights of the high-frequency predicted load, the medium-frequency predicted load, and the low-frequency predicted load according to the high-frequency predicted load and the medium-frequency predicted load, and perform weighted summation to obtain the accurate predicted load at the next moment.

[0073] In step S3, the high-frequency predicted load, medium-frequency predicted load, and low-frequency predicted load at the high-frequency sampling scale, medium-frequency sampling scale, and low-frequency sampling scale are obtained. It is further necessary to integrate these three and combine the average load data predicted at different scales to obtain a more accurate prediction value.

[0074] An exponential function with a natural constant as the base is used to negatively map the difference between the high-frequency predicted load and the low-frequency predicted load to obtain a first mapping result, the first mapping result is multiplied by a first scaling factor and added to a preset value to obtain a first addition result, and the inverse of the first addition result is the weight of the high-frequency predicted load; an exponential function with a natural constant as the base is used to negatively map the difference between the low-frequency predicted load and the high-frequency predicted load to obtain a second mapping result, the second mapping result is multiplied by the first scaling factor and added to a preset value to obtain a second addition result, and the inverse of the second addition result is the weight of the low-frequency predicted load; the weight of the medium-frequency predicted load is obtained by subtracting the sum of the weight of the high-frequency predicted load and the weight of the low-frequency predicted load from the preset value.

[0075] The accurate predicted load at the next moment is obtained by weighting and summing the high-frequency predicted load, the medium-frequency predicted negative and the low-frequency predicted load using the weights of the high-frequency predicted load, the medium-frequency predicted negative and the low-frequency predicted load.

[0076] The specific calculation formula is:

[0077] ,

[0078] in, Indicates the accurate predicted load at the next moment; 、 and They represent high-frequency predicted load, low-frequency predicted negative and medium-frequency predicted load respectively.

[0079] represents the weight of high-frequency predicted load, represents the weight of the low-frequency predicted load, represents the weight of the intermediate frequency predicted load, e represents the natural constant, γ represents the first scaling factor, the preset value is 1, and the implementer can adjust it according to the actual situation; the reference value of the first scaling factor is 3. and When they are very close or equal, the weights of the high-frequency predicted load and the low-frequency predicted load can be around 0.25. This is because under the medium-frequency sampling scale, the collected data can better illustrate the changes in the data in the short term, and the trend changes in the short term described are more representative. Therefore, their weights should be the largest, close to 0.5.

[0080] and They represent the first mapping result and the second mapping result respectively. When the high-frequency predicted load is greater than the low-frequency predicted load, the first mapping result is less than the second mapping result. At this time, the weight of the high-frequency predicted load is greater than the weight of the low-frequency predicted load, which indicates that there is a data mutation. Amplifying its weight makes subsequent predictions more accurate. If the weight of the low-frequency predicted load is less than that of the high-frequency predicted load, it means that the trend of the short-term data collected at the high-frequency sampling scale (short-term trend) may be noise or misjudgment. Therefore, the weight of its low-frequency predicted load (long-term window) is increased, so that the final comprehensive predicted value is closer to the true value, thereby improving the sensitivity of the subsequent temperature threshold.

[0081] Multi-scale weighting is employed here to make the predicted CPU load more accurate. The core purpose is to balance short-term fluctuations with long-term trends, ensuring that the system can quickly respond to load changes while minimizing the impact of short-term anomalies on the forecast. The introduction of an exponential decay function further ensures smoothness, allowing information at different scales to dynamically adjust its contribution, improving forecast stability and accuracy.

[0082] Step S5, using the predicted load and the accurate predicted load to calculate the temperature threshold at the next moment; predicting the temperature data at the next moment based on the temperature sequence, recorded as the predicted temperature; if the predicted temperature is greater than or equal to the temperature threshold, issuing a reminder.

[0083] The predicted load obtained by using the autoregressive moving average algorithm is compared with the accurate predicted load obtained by multi-scale sampling frequency analysis to obtain the temperature threshold at the next moment.

[0084] The temperature threshold at the next moment is calculated using the predicted load and the precise predicted load. Specifically, the difference between the precise predicted load and the predicted load and the first adjustment coefficient are multiplied, and a negative correlation mapping is performed using an exponential function with a natural constant as the base to obtain a third mapping result; the inverse of the third mapping result plus the preset value is obtained, the difference between the inverse and the first adjustment coefficient is obtained, and the difference is multiplied by the second adjustment coefficient to obtain the multiplication result; the sum of the multiplication result and the temperature threshold at the current moment is the temperature threshold at the next moment.

[0085] The specific calculation formula is:

[0086] ,

[0087] in, Indicates the temperature threshold at the next moment, Indicates the temperature threshold at the current moment. The temperature threshold will be set at the initial time and will be continuously updated at each moment as the model is processed; λ represents the second adjustment coefficient, with a reference value of 0.1; α represents the first adjustment coefficient, with a reference value of 0.5. and They represent the precise predicted load and predicted load at the next moment, respectively, where the temperature threshold value ranges from 0 to 1.

[0088] The third mapping result is ,exist Greater than hour, The average load data at time t+1 predicted under multi-scale sampling can better reflect the actual average load data of the CPU. Indicates the average CPU load data predicted in a short period of time. If the CPU load is large, it means that the actual average CPU load data may be large (the CPU volatility at time t+1 is relatively strong, which makes the autoregressive moving average data unable to predict this large fluctuation trend). In this way, there is a risk of CPU overload at time t+1, and it is necessary to lower the temperature threshold and optimize heat dissipation in advance. Less than or equal to When , it indicates that the CPU load tends to be stable or decreasing, the temperature threshold can be appropriately relaxed to reduce energy waste caused by overcooling.

[0089] The temperature threshold at the next moment is obtained from this. It is further necessary to predict the temperature data at the next moment, and then monitor and warn the CPU ambient temperature of the computing power data center.

[0090] At the high-frequency sampling scale, the high-frequency sampling window slides with a preset step size. After sliding to the current moment, in the temperature sequence, the difference between the temperature data at the first moment in the high-frequency sampling window and the temperature data at the current moment is obtained, and compared with the number of temperature data in the high-frequency sampling window, the change rate of the average difference is obtained; the temperature data at the current moment is added to the average difference to obtain the temperature data at the next moment, which is recorded as the predicted temperature.

[0091] The specific calculation formula is:

[0092] ,

[0093] in, Represents the temperature data at the next moment, which is also the predicted temperature; Indicates the temperature data at the current moment; Represents the first temperature data in the high-frequency sampling window, and N represents the number of temperature data in the high-frequency sampling window. is the rate of change of the average difference. From this, we can get the predicted temperature.

[0094] If the predicted temperature is greater than or equal to the temperature threshold, a reminder will be issued. The staff will add coolant according to the reminder, or increase the power of the cooling system to increase the circulation speed of the coolant to cope with the temperature increase and ensure the normal operation of the computing power data center.

[0095] It should be noted that in the embodiment of the present application, when the window moves at each scale, it slides with the sampling unit at the high-frequency sampling scale as the preset step size, that is, 1 second. The implementer can adjust it according to the actual situation to update the temperature threshold per second to ensure the normal operation of the computing center. At the same time, for the high-frequency sampling scale, the medium-frequency sampling scale and the low-frequency sampling scale, as well as their respective corresponding windows, experimental adjustments can be made based on the error between the actual value and the predicted value, and timely updates can be made to ensure the robustness of the model provided by this application, improve the accuracy of the prediction, and make the average load data predicted based on the multi-scale analysis closer to the true value.

[0096] It should be noted that the order in which the embodiments of the present invention are described above is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. Furthermore, the foregoing descriptions of specific embodiments of this specification are provided. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential sequence shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0097] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0098] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent monitoring method for an AI computing power data center, characterized in that: The method includes: Collect the average load data of the CPU and the temperature data in the server and pre-process them to obtain the average load data series and temperature series; The average load data of the next moment at the current moment is predicted based on the average load data sequence using the autoregressive moving average algorithm, which is recorded as the first predicted load; different sampling scales are set, including high-frequency sampling scale, medium-frequency sampling scale and low-frequency sampling scale; At the high-frequency sampling scale, the high-frequency predicted load at the next moment is calculated based on the average load data before the next moment; at the medium-frequency sampling scale, the medium-frequency predicted load at the next moment is calculated based on the average load data before the next moment; at the low-frequency sampling scale, the low-frequency predicted load at the next moment is calculated based on the average load data before the next moment; Obtaining the weights of the high-frequency predicted load, the medium-frequency predicted negative load, and the low-frequency predicted load according to the high-frequency predicted load and the medium-frequency predicted load includes: using an exponential function with a natural constant as the base to negatively correlate the difference between the high-frequency predicted load and the low-frequency predicted load to obtain a first mapping result, multiplying the first mapping result by a first scaling factor and adding the result to a preset value to obtain a first addition result, and the inverse of the first addition result is the weight of the high-frequency predicted load; using an exponential function with a natural constant as the base to negatively correlate the difference between the low-frequency predicted load and the high-frequency predicted load to obtain a second mapping result, multiplying the second mapping result by the first scaling factor and adding the result to a preset value to obtain a second addition result, and the inverse of the second addition result is the weight of the low-frequency predicted load; subtracting the sum of the weight of the high-frequency predicted load and the weight of the low-frequency predicted load from the preset value to obtain the weight of the medium-frequency predicted load, and performing weighted summation to obtain the precise predicted load at the next moment; The temperature threshold at the next moment is calculated using the first predicted load and the accurate predicted load; the temperature data at the next moment is predicted based on the temperature sequence and recorded as the predicted temperature; if the predicted temperature is greater than or equal to the temperature threshold, an alert is issued.

2. The intelligent monitoring method for an AI computing power data center according to claim 1, characterized in that: The collecting and preprocessing of the average load data of the CPU and the temperature data in the server to obtain the average load data sequence and the temperature sequence includes: Preprocessing mainly includes data cleaning and normalization. Data cleaning includes missing value filling, outlier identification and duplicate data processing. The preprocessed average load data and temperature data in the server are respectively composed of the average load data sequence and temperature sequence.

3. The intelligent monitoring method for an AI computing power data center according to claim 1, characterized in that: The step of calculating the high-frequency predicted load at the next moment based on the average load data before the next moment under the high-frequency sampling scale includes: A high-frequency sampling window is set under a high-frequency sampling scale. The high-frequency sampling window slides with a preset step size. After sliding to the current moment, the time series distance between an average load data in the high-frequency sampling window and the average load data at the current moment is obtained, which is recorded as the time series distance. An exponential function with a natural constant as the base is used to negatively correlate the product of the time series distance and the first adjustment coefficient to obtain the weight of the average load data. The weights of the average load data in the high-frequency sampling window are weighted and summed and averaged to obtain the high-frequency predicted load.

4. The intelligent monitoring method for an AI computing power data center according to claim 1, characterized in that: The calculation of the intermediate frequency predicted load at the next moment based on the average load data before the next moment under the intermediate frequency sampling scale includes: The IF sampling window under the IF sampling scale is set. The IF sampling window slides with a preset step size. After sliding to the current moment, the average load data in the IF sampling window is used to calculate the IF predicted load at the next moment. The specific formula for calculating the IF predicted load is: ， in, Indicates the medium frequency predicted load at the next moment; It represents the mean of the preset number of average load data before the next moment; T represents the number of average load data in the intermediate frequency sampling window, , μ represents the window center, σ represents the smoothing coefficient; k and j represent the step index within the intermediate frequency sampling window; Represents the tjth average load data in the intermediate frequency sampling window; Represents the first data in the intermediate frequency sampling window; e represents a natural constant.

5. The intelligent monitoring method for an AI computing power data center according to claim 1, characterized in that: The step of calculating the low-frequency predicted load at the next moment based on the average load data before the next moment under the low-frequency sampling scale includes: The low-frequency sampling window under the low-frequency sampling scale is set. The low-frequency sampling window slides with a preset step size. After sliding to the current moment, the low-frequency predicted load at the next moment is calculated using the average load data in the low-frequency sampling window. The specific formula for calculating the low-frequency predicted load is: ， in, represents the low-frequency predicted load at the next moment; A represents the number of average load data in the low-frequency sampling window; and They represent the tj-th and tj-1-th average load data in the low-frequency sampling window respectively; β represents the second scaling factor; e represents a natural constant; Indicates the tth average load data in the low-frequency sampling window.

6. The intelligent monitoring method for an AI computing power data center according to claim 1, characterized in that: The calculating the temperature threshold at the next moment by using the first predicted load and the accurate predicted load includes: The difference between the precise predicted load and the first predicted load is multiplied by the first adjustment coefficient, and a negative correlation mapping is performed using an exponential function with a natural constant as the base to obtain a third mapping result; the inverse of the third mapping result plus the preset value is calculated, the difference between the inverse and the first adjustment coefficient is calculated, and the difference is multiplied by the second adjustment coefficient to obtain a multiplication result; the sum of the multiplication result and the temperature threshold at the current moment is the temperature threshold at the next moment.

7. The intelligent monitoring method for an AI computing power data center according to claim 1, characterized in that: The step of predicting the temperature data at the next moment based on the temperature sequence, recorded as predicted temperature, includes: At the high-frequency sampling scale, the high-frequency sampling window slides with a preset step size. After sliding to the current moment, in the temperature sequence, the difference between the temperature data at the first moment in the high-frequency sampling window and the temperature data at the current moment is obtained, and compared with the number of temperature data in the high-frequency sampling window, the change rate of the average difference is obtained; the temperature data at the current moment is added to the average difference to obtain the temperature data at the next moment, which is recorded as the predicted temperature.

Images