An aggregated intelligent control method and system for user flexible resources

Abstract

The application relates to the field of information technology and discloses an aggregation intelligent control method and system for user flexible resources. The method comprises the following steps: obtaining device real-time data from a resource pool and processing and generating resource morphing sequences; determining morphing trend descriptions based on the sequences; generating contour feature parameters by adopting K-means clustering; if there is nonlinear fluctuation, generating future morphing trend prediction values by using an LSTM network; comparing current state data to adjust aggregation timing parameters; combining speed parameters to calculate a capacity distribution uniformity index; if the index is lower than a threshold value, iteratively optimizing the speed parameters and generating aggregation control instructions to execute adjustment. Through multi-step data processing and intelligent optimization, the application solves the problem of lacking dynamic and accurate control of flexible resource aggregation, and improves aggregation stability and efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, in particular to an aggregation intelligent control method and system for user flexible resources. BACKGROUND

[0002] With the increasing demand for resource flexibility in digital business, user flexible resources (such as device performance and capacity) need to be adjusted in real time according to load fluctuations to achieve efficient aggregation and stable operation. The current core requirement in the field is to solve the problem of accurate control under dynamic changes of resources, ensure the balance of resource allocation and the real-time response, and support reliable execution of high concurrency and multiple types of tasks.

[0003] The existing technology has significant defects in the aggregation intelligent control of user flexible resources: first, traditional control strategies mostly rely on static parameter settings and cannot adapt to the nonlinear fluctuations of resource pool device performance and capacity. When facing sudden loads or hardware aging, the imbalance problem of some devices being overloaded and some devices being idle easily occurs, resulting in low resource utilization efficiency; second, the existing methods lack multi-dimensional data collaborative processing capability and can only capture resource changes through single time series analysis, which is difficult to extract typical features of trends and lacks accuracy in predicting future resource deformation trends, and cannot provide reliable basis for aggregation timing adjustment; third, the resource scheduling link lacks dynamic optimization mechanism and cannot update control parameters in real time combined with capacity distribution uniformity, often resulting in aggregation instruction lag or excessive adjustment, especially in large-scale resource pool scenarios, which easily causes system response delay and affects service quality. The above problems seriously restrict the stability and intelligent level of user flexible resource aggregation.

[0004] To overcome the above defects, the present application proposes to generate a resource deformation sequence by real-time data collection, extract features combined with time series analysis and K-means clustering, predict trends using LSTM network, and dynamically adjust parameters based on deviation and uniformity to generate aggregation control instructions, thereby realizing accurate intelligent control of flexible resources. SUMMARY

[0005] The present application proposes an aggregation intelligent control method and system for user flexible resources, aiming to solve the problems of lack of dynamic and accurate control, insufficient prediction accuracy and lack of scheduling optimization in flexible resource aggregation in the prior art. Through multi-step data processing, intelligent analysis and dynamic parameter adjustment, stable and efficient control of flexible resource aggregation is realized, and resource utilization efficiency and system service quality are improved.

[0006] In a first aspect, the present application provides an aggregation intelligent control method for user flexible resources, which comprises:

[0007] In step S101, real-time data of device performance and capacity are acquired from a resource pool, the real-time data are preprocessed, time series analyzed and fluctuation features extracted to generate resource deformation sequences;

[0008] In step S102, deformation trend descriptions are determined by smoothing processing, time series feature extraction and trend fitting according to the resource deformation sequences;

[0009] In step S103, the deformation trend descriptions are grouped by using a K-means clustering algorithm, and cluster center points are determined, features of the cluster center points are extracted as multi-dimensional vectors to form contour feature parameters;

[0010] In step S104, it is judged whether the contour feature parameters have nonlinear fluctuations, if yes, a future deformation trend prediction value is generated by using an LSTM network training;

[0011] In step S105, current state data of the resource pool are acquired, a deviation value of the future deformation trend prediction value and the current state data is calculated, and an aggregation timing parameter is adjusted according to a comparison result of the deviation value and a preset deviation threshold;

[0012] In step S106, current capacity distribution of a device is acquired based on the aggregation timing parameter and in combination with an initial setting of a speed parameter, the capacity distribution is statistically analyzed, and an evenness index is calculated;

[0013] In step S107, it is judged whether the evenness index is lower than a preset evenness threshold, if yes, a speed parameter is updated by iterative optimization, an aggregation control instruction is generated in combination with the aggregation timing parameter, and is output to the resource pool to perform dynamic adjustment.

[0014] Optionally, the step S101 comprises:

[0015] Through a sensor network deployed in the resource pool, performance and capacity fluctuations of devices in the resource pool are monitored, the real-time data are collected, and a median filtering method is applied to the real-time data to remove noise values;

[0016] The real-time data filtered of noise are normalized, and then sorted according to time stamps to generate time series data;

[0017] If a resource pool scene scale is a regular scale scene, an autoregressive moving average model is used to capture autocorrelation and moving average components of the time series data, and fluctuation amplitude and periodicity features are extracted as fluctuation features, if the resource pool scene scale is a large-scale resource pool scene, a Fourier transform method is used to convert the time series data from a time domain to a frequency domain, and a dominant frequency is extracted as a fluctuation feature;

[0018] The fluctuation features are combined into a multi-dimensional vector sequence to generate a resource deformation sequence representing the performance and capacity dynamic change trend of the device.

[0019] Optionally, the step S102 comprises:

[0020] The resource deformation sequence is smoothed by using a moving average method to generate a smoothed sequence.

[0021] Features are extracted from the smoothed sequence by using a time series analysis method, a trend slope and a periodic fluctuation are extracted as trend features by using an autoregressive moving average model to capture time series dependence.

[0022] The trend features are input into a linear regression model, a long-term trend line is fitted and a numerical indicator representing a long-term change direction is generated, and the deformation trend description representing the long-term change direction of the resource deformation sequence is constructed based on the numerical indicator.

[0023] Optionally, the step S103 comprises:

[0024] The deformation trend descriptions are grouped by using a K-means clustering algorithm to generate a plurality of stable clusters.

[0025] The average values of all data points in each cluster are extracted by analyzing the stable clusters, and the average values are used as the center points of the corresponding clusters, and the center points represent the concentrated features of the deformation trend in the cluster.

[0026] The fluctuation amplitude values of the deformation trend descriptions corresponding to the center points of each cluster are calculated, and the fluctuation amplitude values are the differences between the trend peak values and the valley values.

[0027] The average fluctuation frequencies of the deformation trend descriptions corresponding to the center points of each cluster in a unit time are counted as fluctuation frequency values.

[0028] The average durations of the deformation trend descriptions corresponding to the center points of each cluster from the start to the stability are measured as fluctuation duration values.

[0029] The fluctuation amplitude values, the fluctuation frequency values and the fluctuation duration values are combined into a multi-dimensional vector to generate the profile feature parameters, and the profile feature parameters are used to represent the typical features of the deformation trend.

[0030] Optionally, the step S104 comprises:

[0031] Second-order differences of the profile feature parameter sequence are calculated, and if the absolute values of the second-order differences exceed a preset fluctuation threshold at three or more consecutive time points, it is determined that the profile feature parameters are nonlinear fluctuations.

[0032] The profile characteristic parameters of multiple groups determined as nonlinear fluctuations in a historical time period are taken as feature inputs, and an LSTM network is used for training, a number of hidden layer units and a number of training times are set, and a mean square error is used as a loss function for back propagation optimization;

[0033] According to the trained LSTM network, the current profile characteristic parameters are input to generate a prediction sequence of multiple future time points, and the future deformation trend prediction value is obtained, wherein the future deformation trend prediction value represents the change trend of the future performance and capacity of the resource pool.

[0034] Optionally, the step S105 comprises:

[0035] The current state data of the resource pool is obtained, wherein the current state data is real-time performance and capacity data of the devices in the resource pool;

[0036] According to the future deformation trend prediction value, the performance and capacity data of the resource pool at the future time point are calculated;

[0037] The deviation value between the performance and capacity data at the future time point and the current state data is calculated;

[0038] It is judged whether the deviation value is greater than a preset deviation threshold value, if yes, the aggregation timing parameter needs to be adjusted to maintain the stability and balance of the resource pool, and if not, the aggregation timing parameter does not need to be adjusted.

[0039] Optionally, the adjustment of the aggregation timing parameter comprises:

[0040] A deviation adjustment coefficient is calculated based on the deviation value, the deviation adjustment coefficient is obtained by first calculating the ratio of the absolute value of the deviation to the preset deviation threshold value, and then comparing the ratio with the natural number 1 to take the minimum value of the two;

[0041] A new aggregation timing parameter is calculated based on the deviation adjustment coefficient, the new aggregation timing parameter = current aggregation timing parameter x (1-adjustment coefficient x deviation adjustment coefficient), wherein the adjustment coefficient is preset within 0-1.

[0042] Optionally, the step S106 comprises:

[0043] According to the aggregation timing parameter, the initial setting of the speed parameter is combined to obtain the current capacity distribution of the devices in the resource pool;

[0044] The current capacity distribution is statistically analyzed to calculate the mean, standard deviation and maximum-minimum difference to generate distribution feature data;

[0045] According to the distribution characteristic data, a uniformity index representing the capacity distribution uniformity is calculated, and the uniformity index is obtained by calculating a ratio of the standard deviation to the average value.

[0046] Optionally, the step S107 comprises:

[0047] It is judged whether the uniformity index is lower than the preset uniformity threshold value, and if so, the speed parameter value is updated by a step adaptive iterative optimization method to narrow the gap between the uniformity index and the preset uniformity threshold value.

[0048] According to the updated speed parameter and the aggregation timing parameter, an aggregation control instruction containing the values of the two parameters is generated.

[0049] The aggregation control instruction is output to the resource pool, instructing the resource pool to perform a resource aggregation operation according to a rate defined by the speed parameter value and a timing defined by the aggregation timing parameter, so as to realize stable aggregation.

[0050] In a second aspect, the present application provides an aggregation intelligent control system for user flexible resources, comprising:

[0051] A data acquisition module is configured to acquire real-time data of device performance and capacity from a resource pool, perform preprocessing, time series analysis and fluctuation feature extraction, and generate a resource deformation sequence.

[0052] A trend extraction module is configured to perform smoothing processing, time series feature extraction and trend fitting on the resource deformation sequence, and generate a deformation trend description representing a long-term change direction.

[0053] A feature clustering module is configured to generate stable clusters by using a K-means clustering algorithm based on the deformation trend description, extract a center point feature and construct a multi-dimensional contour feature parameter.

[0054] A trend prediction module is configured to train and generate a future deformation trend prediction value by using an LSTM network when the contour feature parameter has a nonlinear fluctuation.

[0055] A deviation calculation module is configured to compare the future deformation trend prediction value with current state data, calculate a deviation value and adjust the aggregation timing parameter accordingly.

[0056] A uniformity calculation module is configured to calculate a uniformity index based on the aggregation timing parameter and the speed parameter, and statistically calculate the current capacity distribution.

[0057] An instruction output module is configured to update the speed parameter by iterative optimization when the uniformity index is lower than a preset uniformity threshold value, and generate and output an aggregation control instruction in combination with the aggregation timing parameter.

[0058] The application provides an aggregation intelligent control method and system for user flexible resources, which is suitable for aggregation control of user flexible resources (such as device performance and capacity) in a digital service scenario, and can solve the problems of lack of dynamic and accurate control, insufficient prediction accuracy and lack of scheduling optimization in the aggregation of flexible resources in the prior art. Compared with the prior art, the application has at least the following advantages:

[0059] First, the nonlinear fluctuation of the performance and capacity of the adaptable resource pool device can avoid the imbalance problem of part of the device overload and part of the device idling in the traditional static parameter control strategy when facing a sudden load or hardware aging, and improve the resource utilization efficiency.

[0060] Second, it has multi-dimensional data collaborative processing capability, can effectively extract the typical characteristics of resource change trend, improve the prediction accuracy of future resource deformation trend, and provide reliable basis for aggregation opportunity adjustment.

[0061] Third, a dynamic optimization mechanism is established in the resource scheduling link, the control parameters can be updated in real time in combination with the capacity distribution uniformity, the aggregation instruction lag or excessive adjustment can be avoided, especially in the large-scale resource pool scene, the system response delay can be reduced, and the service quality can be guaranteed. BRIEF DESCRIPTION OF DRAWINGS

[0062] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor based on these drawings.

[0063] Figure 1 It is a flowchart of the aggregation intelligent control method for user flexible resources in the application;

[0064] Figure 2 It is a process diagram of the aggregation intelligent control method for user flexible resources in the application;

[0065] Figure 3 It is a structure diagram of the aggregation intelligent control system for user flexible resources in the application. DETAILED DESCRIPTION

[0066] The embodiments of the present application provide a method and system for flexible resource aggregation intelligent control. The terms "first", "second", "third", "fourth" and the like (if any) in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "comprise" or "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device comprising a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0067] For ease of understanding, the specific processes of the embodiments of the present application are described below. Please refer to Figure 1 One embodiment of the method for flexible resource aggregation intelligent control in the embodiments of the present application includes the following steps.

[0068] In step S101, real-time data of device performance and capacity is obtained from a resource pool, and the real-time data is preprocessed, time-series analyzed and fluctuation feature extracted to generate a resource deformation sequence.

[0069] In a specific embodiment, the process of step S101 can specifically include the following steps.

[0070] Through a sensor network deployed in the resource pool, performance and capacity fluctuations of devices in the resource pool are monitored, the real-time data is collected, and a median filtering method is applied to the real-time data to remove noise values;

[0071] The real-time data filtered by noise is normalized, and then sorted by time stamp to generate time series data;

[0072] If it is a regular scale scene, an autoregressive moving average model is used to capture the autocorrelation and moving average components of the time series data, and the fluctuation amplitude and periodicity features are extracted as fluctuation features. If it is a large-scale resource pool scene, a Fourier transform method is used to convert the time series data from time domain to frequency domain, and the dominant frequency is extracted as the fluctuation feature;

[0073] The fluctuation features are combined into a multi-dimensional vector sequence to generate a resource deformation sequence representing the dynamic change trend of device performance and capacity.

[0074] Specifically, in a scenario of dynamic resource pool adjustment, a sensor network deployed within the resource pool continuously monitors fluctuations in device performance (such as CPU utilization and memory usage) and capacity (such as remaining storage space and bandwidth usage). It acquires real-time data at a preset collection frequency (e.g., once per second). For example, in a server cluster, sensors collect data on each server's CPU utilization (range 0%–100%), memory usage (range 0%–100%), and remaining storage space (range 0GB–1000GB) every second. For the collected real-time data, a median filtering method is applied to remove noise. Specifically, a window is formed by taking a preset number (e.g., 5) of adjacent data points before and after a given data point, and the median value within the window replaces that data point. For example, if a server's CPU utilization values ​​for five consecutive data points are 25%, 28%, 150%, 26%, and 27%, the 150% value represents noise caused by network interference. After median filtering, this data point is replaced with 27%. This operation solves the problem of data distortion caused by noise interference in real-time data, ensuring the reliability of subsequent data processing.

[0075] The real-time data after noise filtering is normalized using the min-max normalization formula. (in x This is the original data. This is the minimum value of the indicator. This is the maximum value of the indicator. (This is for normalized data), unifying the performance and capacity indicators of different devices to a range of 0-1. For example, the original CPU utilization data range is 0% to 100%, and a server's CPU utilization is 50%, which is normalized to 0.5; the original storage space data range is 0GB to 1000GB, and a server's storage space is 500GB, which is normalized to 0.5. After normalization, the data is sorted according to the timestamp of data collection to generate time-series data, such as data arranged by timestamp. (in , The timestamp is used as the data source, and the subsequent values ​​are the normalized CPU utilization, memory usage, and remaining storage space, respectively. This process solves the problem that different devices have large differences in the range of values ​​of different indicators, making it impossible to perform collaborative analysis directly, and provides a standardized data format for subsequent time series analysis.

[0076] When processing time series data using time series analysis methods, it is necessary to first identify the scale of the resource pool scenario, and then select the corresponding method based on the scale of the resource pool. In typical scenarios (such as resource pools with fewer than 50 devices), an autoregressive moving average model is used. ,in p Let the order be the autoregressive order. qThe model processes the data by formula wherein is t time series data at time t, is an autoregressive coefficient, is a moving average coefficient, is white noise) captures the autocorrelation and moving average components of the time series data, and after model fitting, the fluctuation amplitude (such as the standard deviation of the model residual) and periodicity characteristics (such as the interval at which peaks appear in the autocorrelation function graph) of the data are extracted. For example, after a certain conventional resource pool time series data is processed by an ARMA(2, 1) model, the fluctuation amplitude is 0.05, and the periodicity characteristic is that a peak appears every 30 minutes. In a large-scale resource pool scenario (such as a resource pool with more than or equal to 50 devices), the Fourier transform method wherein is time series data in the time domain, is frequency domain data, is frequency, t is time) converts time series data in the time domain into frequency domain data, and by analyzing the power spectral density of the frequency domain data, the dominant frequency (i.e., the frequency corresponding to the maximum power spectral density) is extracted as the fluctuation characteristic. For example, after a certain large-scale resource pool is Fourier transformed, the dominant frequency is 0.0005 Hz (corresponding to a period of about 33 minutes). This processing process solves the problem that different sizes of resource pool data have large differences in characteristics, and a single analysis method cannot effectively extract fluctuation characteristics, ensuring that key information about data fluctuations can be accurately obtained in different scenarios.

[0077] According to the extracted fluctuation characteristics, each fluctuation characteristic at the same time point is combined into a multi-dimensional vector. For example, in a conventional scenario, the fluctuation amplitude at a certain time point is 0.05, and the periodicity characteristic corresponds to a period of 30 minutes (converted to a numerical value of 0.00056 Hz). Therefore, the multi-dimensional vector at the time point is ; in a large-scale scenario, the dominant frequency at a certain time point is 0.0005 Hz, and the multi-dimensional vector is (0.0005), which is a resource deformation sequence representing the dynamic change trend of device performance and capacity. This process solves the problem that the fluctuation characteristic sequence has short-term fluctuations and cannot clearly reflect long-term dynamic change trends, providing smooth and continuous data basis for subsequent determination of deformation trend description.

[0078] Step S102, according to the resource deformation sequence, the process is determined by smoothing, time sequence feature extraction and trend fitting, and the deformation trend description.

[0079] In a specific embodiment, the process of performing step S102 can specifically include the following steps:

[0080] Smooth the resource deformation sequence by using a moving average method to generate a smooth sequence;

[0081] Extract features from the smooth sequence by using a time series analysis method, use an autoregressive moving average model to capture time series dependence, and extract trend slope and periodic fluctuations as trend features;

[0082] Input the slope and fluctuation pattern in the trend features into a linear regression model, fit a long-term trend line and generate a numerical indicator representing the long-term change direction, and then build the deformation trend description representing the long-term change direction of the resource deformation sequence based on the numerical indicator.

[0083] Specifically, for the generated resource deformation sequence (such as a multi-dimensional vector sequence of CPU utilization rate and memory occupancy rate normalization fluctuation features of a certain server cluster every 5 minutes, example data is , , , , ), a moving average method is used for smoothing. The moving average method uses a sliding window mechanism, and the window size is set according to the collection frequency of the resource deformation sequence. If the collection frequency is 5 minutes / time, the window size is set to 3, that is, the arithmetic mean value of the resource deformation sequence data at the continuous 3 time points is calculated, and the formula is (where is the moving average value at time t, and are the resource deformation sequence data at time t, respectively). Taking the example data, the moving average value at time t is , , , , and so on to generate the smooth sequence. This processing solves the interference problem of short-term random noise (such as abnormal jump of CPU utilization rate at a certain time caused by instantaneous network fluctuation) on trend analysis, and makes the sequence more consistent with the real change law of device performance and capacity.

[0084] After smoothing, features are extracted from the smooth sequence by using a time series analysis method, and an autoregressive moving average model ( ) is selected to capture time series dependence. The model order parameters p (arithmetic order) and q (moving average order) are adjusted according to the length of the smooth sequence. If the smooth sequence contains 100 time point data, the AIC values of different combinations are calculated by AIC information criterion, and the combination with the minimum AIC value is selected, such as determining p =2, q =1, and the model formula is (wherein is the smoothed sequence data at time t, is the autoregressive coefficient, is the moving average coefficient, is the smoothed sequence data at time t, is the white noise at time t). The smoothed sequence data is input into the model for training, and the coefficients are estimated by least squares method, such as . Based on the trained model, the trend slope and periodic fluctuation are extracted as trend features, the trend slope is calculated by the first derivative of the fitted curve, if the first derivative of the fitted curve is 0.002 / 5 minutes in a certain period, it means that the smoothed sequence data increases by 0.002 on average every 5 minutes; the periodic fluctuation is analyzed by the autocorrelation function (ACF) of the model residual, if the ACF has a peak at lag 12 time points (corresponding to 60 minutes), it means that there is a 60-minute periodic fluctuation. This process solves the problem that a single data point cannot reflect the time series correlation and it is difficult to extract long-term change features, providing key feature support for subsequent trend fitting. According to the extracted trend features, the trend slope and fluctuation mode are input into the linear regression model for trend fitting. The linear regression model takes the timestamp as the independent variable

[0085] , and the smoothed sequence data t as the dependent variable, the model formula is (wherein is the regression coefficient, corresponding to the average level of the trend slope, k is the intercept). The trend slope in the trend features (such as 0.002 / 5 minutes) is converted into the unit time slope (0.0004 / minute), and the time node data corresponding to the fluctuation mode (such as the peak data every 60 minutes) is input into the model, and b and k are solved by least squares method, if b =0.0004 and k =0.4, the fitted long-term trend line is b . Based on the trend line, a numerical indicator representing the long-term change direction is generated, such as the predicted value of the next 120 minutes , compared with the current value (such as at time t , =0.43), the numerical indicator is "0.024 per hour". Based on the numerical indicator, a deformation trend description is constructed, such as "the smoothed sequence of resource pool device CPU utilization and memory occupancy shows a stable growth trend of 0.024 per hour, accompanied by 60-minute periodic fluctuations", which directly represents the long-term change direction of the resource deformation sequence, solves the problem of lack of clear trend expression and inability to provide direction guidance for subsequent resource scheduling, and provides clear trend input for the feature clustering module.

[0086] In step S103, the K-means clustering algorithm is used to group the deformation trend descriptions, and the cluster center points are determined. The feature combinations of the cluster center points are extracted as multi-dimensional vectors to form the contour feature parameters.

[0087] In an embodiment, the process of step S103 can specifically include the following steps:

[0088] The K-means clustering algorithm is used to group the deformation trend descriptions to generate a plurality of stable clusters;

[0089] The plurality of stable clusters are analyzed, and the average values of all data points in each cluster are extracted. The average values are taken as the center points of the corresponding clusters, and the center points represent the concentrated features of the deformation trends in the clusters;

[0090] The fluctuation amplitude values of the deformation trend descriptions corresponding to the center points of each cluster are calculated. The fluctuation amplitude values are the differences between the peak values and the valley values of the trends;

[0091] The average fluctuation frequencies of the deformation trend descriptions corresponding to the center points of each cluster in a unit time are counted as the fluctuation frequency values;

[0092] The average durations of the deformation trend descriptions corresponding to the center points of each cluster from the start to the stability are measured as the fluctuation duration values;

[0093] The fluctuation amplitude values, the fluctuation frequency values, and the fluctuation duration values are combined as multi-dimensional vectors to generate the contour feature parameters, wherein the contour feature parameters are used to represent the typical features of the deformation trends.

[0094] Specifically, the morphing trend description is presented in the form of multi-dimensional data, for example, the morphing trend description of different device clusters in a certain cloud computing resource pool includes the unit time change rate and the periodic fluctuation duration, which constitute the input data set, and each data point corresponds to a vector composed of the above dimensions. When applying the K-means clustering algorithm, first, the number of data points corresponding to the number of preset clusters is randomly selected from the input data set as the initial center point according to the data dimension of the morphing trend description and the device scale of the resource pool. Then, the Euclidean distance of each data point in the input data set to each initial center point is iteratively calculated, and each data point is assigned to the cluster where the nearest center point is located according to the distance. After all data points are assigned, the average value of each dimension of all data points in each cluster is calculated to update the cluster center point. Repeat the process of distance calculation, data point assignment, and center point update until the center point coordinate difference between the previous and subsequent iterations is less than the preset threshold, and generate multiple stable clusters. This process solves the problem of lack of multi-dimensional data collaborative processing capability and difficulty in extracting typical trend characteristics in the prior art, and classifies similar morphing trends through clustering, laying a foundation for subsequent feature extraction.

[0095] After generating stable clusters, analyze each cluster and extract the average value of each dimension of all data points in the cluster. The average value is used as the center point of the corresponding cluster. The values of each dimension of the center point directly represent the concentrated characteristics of the morphing trend in the cluster. For example, the unit time change rate of a certain cluster center point is 0.023, and the periodic fluctuation duration is 59 minutes, which means that the morphing trend of the devices in the cluster as a whole presents a change rate of 0.023 per unit time and a periodic fluctuation every 59 minutes. Based on the morphing trend description corresponding to the center point, calculate the fluctuation amplitude value. Specifically, extract the peak value and valley value from the morphing trend data corresponding to the center point, and the difference between the two is the fluctuation amplitude value. For example, the peak value of a certain center point is 0.08, and the valley value is 0.03, so the fluctuation amplitude value is 0.05. When calculating the fluctuation frequency value, take the unit time as the statistical period, and calculate the average number of fluctuations of the morphing trend description corresponding to the center point in the unit time. If the unit time is 60 minutes, the trend fluctuates 12 times in 60 minutes, and the fluctuation frequency value is 12 times / hour. When measuring the fluctuation duration, record the duration from the starting time to the stable state of the morphing trend corresponding to the center point, and take the average value of multiple measurements as the fluctuation duration. For example, the multiple measurement results are 175 minutes, 180 minutes, and 185 minutes, and the fluctuation duration is 180 minutes.

[0096] The fluctuation amplitude value, the fluctuation frequency value and the fluctuation duration time are combined in a fixed order to form a multi-dimensional vector, and the vector is the profile characteristic parameter. For example, the fluctuation amplitude value of a certain cluster is 0.05, the fluctuation frequency value is 12 times / hour, and the fluctuation duration time is 180 minutes. The multi-dimensional vector (0.05, 12, 180) formed by combination is the profile characteristic parameter of the deformation trend of the cluster. The profile characteristic parameter integrates the key typical characteristics of the deformation trend, solves the problem that only single time series analysis is used to capture resource changes in the prior art, and the prediction accuracy of the future resource deformation trend is insufficient, provides standardized feature input for subsequent judgment of nonlinear fluctuation and generation of future deformation trend prediction value, so that the subsequent prediction process can be based on more comprehensive and typical trend characteristics, improve the prediction accuracy, and then provide a reliable basis for aggregated timing adjustment, avoiding the imbalance problem of traditional static control strategy when facing resource nonlinear fluctuation.

[0097] In step S104, it is judged whether the profile characteristic parameter has nonlinear fluctuation. If yes, LSTM network is trained and future deformation trend prediction value is generated.

[0098] In an embodiment, the process of step S104 can specifically include the following steps:

[0099] The second-order difference of the profile characteristic parameter sequence is calculated. If the absolute value of the second-order difference exceeds a preset fluctuation threshold at 3 or more consecutive time points, it is determined that the profile characteristic parameter has nonlinear fluctuation.

[0100] A plurality of groups of profile characteristic parameters determined to have nonlinear fluctuation in a historical time period are taken as feature input, and LSTM network is used to train the same. The number of hidden layer units and the number of training times are set, and mean square error is used as a loss function for back propagation optimization.

[0101] According to the trained LSTM network, the current profile characteristic parameter is input to generate a prediction sequence of future time points, and the future deformation trend prediction value is obtained. The future deformation trend prediction value represents the change trend of the future performance and capacity of the resource pool.

[0102] Specifically, the profile characteristic parameter includes a fluctuation amplitude value, a fluctuation frequency value and a fluctuation duration time. These parameters are arranged in time stamp order to form a profile characteristic parameter sequence. For example, a resource pool collects data every 10 minutes to form a parameter sequence including a plurality of time points (e.g. When it is judged whether the sequence has nonlinear fluctuation, the second-order difference of each dimension in the sequence is calculated. Taking the fluctuation amplitude value dimension as an example, the first-order difference of the fluctuation amplitude values of adjacent two time points is calculated, i.e. (wherein to The second-order difference is calculated based on the first-order difference, i.e. , the fluctuation frequency value, and the fluctuation duration dimension. A preset fluctuation threshold is set for each dimension, such as a second-order difference threshold of the fluctuation amplitude value of 0.02, a second-order difference threshold of the fluctuation frequency value of 1 times / hour, and a second-order difference threshold of the fluctuation duration of 10 minutes. If the absolute value of the second-order difference of any dimension exceeds the corresponding threshold for 3 or more consecutive time points, it is determined that the profile characteristic parameter has nonlinear fluctuation. This process solves the problem that the traditional control strategy in the prior art cannot adapt to the nonlinear fluctuation of the performance and capacity of the resource pool device. Through multi-dimensional second-order difference analysis, the nonlinear characteristics of resource changes are accurately identified, and the fluctuation misjudgment caused by single dimension or first-order difference analysis is avoided, thereby providing accurate triggering conditions for subsequent prediction.

[0103] If it is determined that there is nonlinear fluctuation, the multiple groups of profile characteristic parameters determined as nonlinear fluctuation in the historical time period are taken as the feature input of the LSTM network. The time step of the input sequence is consistent with the parameter collection period, such as 10 minutes for the collection period, 10 for the time step, and 3 feature dimensions (fluctuation amplitude value, fluctuation frequency value, and fluctuation duration) for each time step. The number of hidden layer units of the LSTM network is set according to the data size. If the historical data contains 1000 time steps, the number of hidden layer units is set to 64, and the number of training times is set to 100 rounds. When constructing the training sample, the profile characteristic parameter sequence of consecutive N time steps is taken as the input sample, and the profile characteristic parameter of the N+1 time step is taken as the label. For example, the parameter sequence of is taken as the input, is taken as the label; the parameter sequence of is taken as the input, is taken as the label, and a batch of training samples is generated in this way. The mean square error is used as the loss function, and the loss function calculation formula is (wherein k is the number of samples, is the actual value of the label, is the network prediction value), and the network weight and bias are adjusted through the back propagation algorithm to make the loss function value gradually converge to a preset range (such as less than 0.001). This training process solves the problem of insufficient accuracy of future resource deformation trend prediction in the prior art. With the ability of the LSTM network to capture long sequence dependencies, the network fully learns the change rule of the profile characteristic parameter under nonlinear fluctuation, thereby providing reliable model support for subsequent prediction.

[0104] After training, the profile characteristic parameters of the current time and the consecutive time steps before the current time are input into the LSTM network, for example, the current time is , and the input The parameter sequence of 10 time steps, the network processes the input sequence through the internal memory unit, and outputs a prediction sequence of multiple future time points (such as ), each prediction time point corresponds to a set of prediction vectors containing fluctuation amplitude value, fluctuation frequency value, and fluctuation duration, and these prediction vectors collectively constitute the future deformation trend prediction value. Since the dimensions of the contour feature parameters are directly related to the performance and capacity changes of the resource pool, such as the fluctuation amplitude value corresponding to the performance fluctuation range, the fluctuation frequency value corresponding to the capacity fluctuation frequency, and the fluctuation duration corresponding to the resource instability duration, the future deformation trend prediction value can directly represent the future performance and capacity change trend of the resource pool. This prediction process provides a reliable basis for adjusting the aggregation timing, avoiding the problem of aggregation instruction lag due to the lack of accurate trend reference in the prior art, and ensuring that the resource aggregation can adapt to future resource changes in advance to maintain stable operation of the resource pool.

[0105] Step S105, obtaining the current state data of the resource pool, calculating the deviation value between the future deformation trend prediction value and the current state data, and adjusting the aggregation timing parameter according to the comparison result between the deviation value and the preset deviation threshold.

[0106] In a specific embodiment, the process of performing step S105 can specifically include the following steps:

[0107] Obtaining the current state data of the resource pool, wherein the current state data is the real-time performance and capacity data of the devices in the resource pool;

[0108] According to the future deformation trend prediction value, calculating the performance and capacity data of the resource pool at the future time point;

[0109] Calculating the deviation value between the performance and capacity data at the future time point and the current state data;

[0110] Judging whether the deviation value is greater than the preset deviation threshold, if yes, the aggregation timing parameter needs to be adjusted to maintain the stability and balance of the resource pool, and if not, the aggregation timing parameter does not need to be adjusted.

[0111] Specifically, the acquisition of the current state data of the resource pool relies on the sensor network deployed in the resource pool, which continuously collects real-time performance data and capacity data of the devices. The performance data includes CPU utilization, memory occupancy, etc. of each device, and the capacity data covers storage remaining space, bandwidth occupancy, etc. The collection frequency is adapted to the dynamic change rate of the resource pool, for example, collecting once per second to ensure that the data can reflect the current running state of the device in real time. When calculating the performance and capacity data at the future time point according to the future deformation trend prediction value, it is necessary to establish a mapping relationship between the prediction value and the actual business indicators. The future deformation trend prediction value includes a multi-dimensional vector such as fluctuation amplitude value, fluctuation frequency value, and fluctuation duration. The fluctuation amplitude value corresponds to the change range of performance and capacity, the fluctuation frequency value corresponds to the number of changes per unit time, and the fluctuation duration corresponds to the time length from the start to the stability. Taking CPU utilization as an example, if the current CPU utilization is 50%, the fluctuation amplitude value at a certain time point in the future deformation trend prediction value is 0.05 (normalized), and combined with the original data range of CPU utilization 0%-100%, the predicted value of CPU utilization at that time point can be deduced as 50%+0.05x100%=55%; the current value of storage remaining space is 500GB, and the predicted fluctuation amplitude value is 0.1 (normalized, original range 0GB-1000GB), so the predicted value of storage remaining space at that time point is 500GB+0.1x1000GB=600GB. Through similar logic, the specific data of all performance and capacity indicators at future time points can be calculated.

[0112] When calculating the deviation value of the future time point data and the current state data, it is necessary to calculate each type of performance and capacity indicator respectively, and then obtain the comprehensive deviation value through weighted summation or Euclidean distance. Taking two core indicators of CPU utilization and storage remaining space as an example, let the current value of CPU utilization be A1, the future predicted value be B1, the current value of storage remaining space be A2, and the future predicted value be B2. First, calculate the single indicator deviation, the CPU utilization deviation is |B1-A1|, and the storage remaining space deviation is |B2-A2|. If the weights of the two types of indicators are ω1 and ω2 (set according to the importance of business, such as ω1=0.6, ω2=0.4), then the comprehensive deviation value is ω1x|B1-A1|+ω2x|B2-A2|. For example, A1=50%, B1=55%, A2=500GB, B2=600GB, the CPU utilization deviation is 5%, the storage remaining space deviation is 100GB, and the comprehensive deviation value is 0.6x5%+0.4x(100GB / 1000GB)=3%+4%=7% (the storage remaining space deviation is first normalized).

[0113] When judging the size relationship between the deviation value and the preset deviation threshold value, the preset deviation threshold value needs to be determined in combination with historical running stable data of the resource pool and service quality requirements. For example, according to historical data statistics, when the comprehensive deviation value exceeds 8%, the resource pool is prone to device overload or idling problems, and therefore the preset deviation threshold value is set to 8%. If the calculated comprehensive deviation value is 7%, which is less than 8%, it indicates that future resource changes are within the current aggregation strategy adaptation range, and there is no need to adjust the aggregation timing parameter. If the comprehensive deviation value is 9%, which is greater than 8%, the aggregation timing parameter needs to be adjusted. The adjustment process needs to dynamically adjust the aggregation interval based on the ratio of the deviation value to the threshold value. The aggregation timing parameter is usually represented by the aggregation interval time. Assuming that the current aggregation interval is T, the deviation value is D, the preset threshold value is D0, and the adjustment coefficient is k (0 k <1, such as k = 0.2), the new aggregation interval For example, the current aggregation interval T = 10 minutes, D = 9%, and D0 = 8%. After substitution, T' = 10 x (1-0.2 x (9%-8%) / 8%) = 10 x (1-0.0025) = 9.975 minutes. By shortening the aggregation interval, the resource aggregation operation is more frequent, and future resource changes are addressed in advance.

[0114] By obtaining the current state data in real time and comparing it with the future prediction value, the aggregation timing parameter is dynamically adjusted to avoid resource scheduling lag or excessive adjustment caused by fixed aggregation timing. At the same time, the mapping relationship between the prediction value and the actual business index and the multi-index comprehensive deviation calculation logic are established, solving the limitation of single-dimensional analysis of the existing method, ensuring more comprehensive and accurate deviation judgment, providing a reliable basis for aggregation timing adjustment, and further improving the resource pool running stability and resource utilization efficiency, and alleviating the imbalance problem of part of the device overload and part of the device idling.

[0115] In a specific embodiment, the process of adjusting the aggregation timing parameter can specifically include the following steps:

[0116] Based on the deviation value, a deviation adjustment coefficient is calculated. The deviation adjustment coefficient is obtained by first calculating the ratio of the deviation absolute value to the preset deviation threshold value, and then comparing the ratio with the natural number 1 to take the minimum value of the two.

[0117] Based on the deviation adjustment coefficient, a new aggregation timing parameter is calculated. The new aggregation timing parameter = current aggregation timing parameter x (1-adjustment coefficient x deviation adjustment coefficient), wherein the adjustment coefficient is preset within 0-1.

[0118] Specifically, to achieve the adjustment of the aggregation timing parameter, a deviation adjustment coefficient needs to be calculated based on the deviation value. This process needs to first determine the specific values of the deviation value and the preset deviation threshold value. The deviation value is derived from the difference calculation of the future time point performance and capacity data and the current state data. The preset deviation threshold value is set according to the historical stable operation data of the resource pool and the tolerance of the business to resource fluctuations. For example, in the server cluster resource pool scenario, if the comprehensive deviation value calculation result of CPU utilization rate and storage remaining space is 12%, and the preset deviation threshold value is set to 8%, the ratio of the deviation absolute value to the preset deviation threshold value is calculated first, and the deviation absolute value is 12%. The ratio is 12% ÷ 8% = 1.5. Then the ratio is compared with the natural number 1, and the minimum value of the two is taken as the deviation adjustment coefficient. Since 1.2 is greater than 1, the deviation adjustment coefficient is 1 at this time. If the deviation value is 6% and the preset deviation threshold value is still 8%, the ratio is 6% ÷ 8% = 0.75, which is less than 1, and the deviation adjustment coefficient is 0.75.

[0119] When calculating the new aggregation timing parameter, the current aggregation timing parameter and the adjustment coefficient need to be determined first. The current aggregation timing parameter is the interval time of the resource pool currently performing resource aggregation operation, for example, currently set to 30 minutes. The adjustment coefficient is preset in the range of 0-1, and its value needs to be set in combination with the sensitivity of the resource pool to parameter adjustment. If the number of resource pool devices is small and the load fluctuation is gentle, the adjustment coefficient can be set to 0.3. If the number of devices is large and the load fluctuation is frequent, the adjustment coefficient can be set to 0.6. Here, the adjustment coefficient is taken as 0.4 as an example. When the deviation adjustment coefficient is 0.75, the new aggregation timing parameter can be obtained by substituting the formula “new aggregation timing parameter = current aggregation timing parameter × (1- adjustment coefficient × deviation adjustment coefficient)” into the formula, which is 30 minutes × (1-0.4 × 0.75) = 30 minutes × 0.7 = 21 minutes.

[0120] The process avoids system instability caused by excessive adjustment amplitude when the deviation value is too large, for example, when the deviation value is far beyond the preset threshold, the deviation adjustment coefficient is 1, limiting the upper limit of the adjustment amplitude, preventing the sudden drop of the aggregation timing parameter from causing frequent resource scheduling; when the deviation value is close to the preset threshold, the deviation adjustment coefficient is calculated according to the actual ratio, realizing fine adjustment and ensuring that the parameter adjustment fits the actual situation of resource fluctuation. At the same time, the introduction of the adjustment coefficient can be flexibly adapted according to the characteristics of the resource pool scene, solving the problem of lack of dynamic optimization mechanism and single adjustment strategy in the existing method, for example, in a large-scale resource pool scene, by setting a larger adjustment coefficient, the response speed of the aggregation timing parameter is accelerated to cope with complex load fluctuations, in a conventional scale resource pool scene, a smaller adjustment coefficient is set to maintain parameter stability and avoid excessive adjustment, ultimately realizing dynamic and accurate control of the aggregation timing parameter, relieving the imbalance problem of some equipment overload and some equipment idle, and improving resource utilization efficiency and system service quality.

[0121] In step S106, based on the aggregation timing parameter, the current capacity distribution of the device is obtained in combination with the initial setting of the speed parameter, the capacity distribution is statistically analyzed, and the uniformity index is calculated.

[0122] In an embodiment, the process of performing step S106 can specifically include the following steps:

[0123] According to the aggregation timing parameter, in combination with the initial setting of the speed parameter, the current capacity distribution of the device in the resource pool is obtained;

[0124] The current capacity distribution is statistically analyzed, and the mean value, standard deviation and maximum-minimum difference value are calculated to generate distribution characteristic data;

[0125] According to the distribution characteristic data, a uniformity index representing the balance of the capacity distribution is calculated, and the uniformity index is obtained by calculating the ratio of the standard deviation to the mean value.

[0126] Specifically, when obtaining the current capacity distribution of the devices based on the initial settings of the aggregation timing parameter and the speed parameter, the aggregation timing parameter determines the time node for triggering capacity data collection. For example, if the aggregation timing parameter is set to perform data collection once every 20 minutes, the system will automatically start the data acquisition process after reaching this time node. The initial settings of the speed parameter limit the rate of data transmission and processing, such as setting it to process 15 device capacity data per second, to ensure that the full amount of device data collection is completed within the time window specified by the aggregation timing parameter. In practice, the system sends data collection commands to all devices in the resource pool based on the aggregation timing parameters. After responding to the commands, the devices provide real-time capacity data, such as the remaining storage space of the servers and the available memory of the computing nodes. Taking a resource pool containing 20 servers as an example, the remaining storage space data provided by each server are 45GB, 52GB, 48GB, 55GB, 49GB, 51GB, 47GB, 53GB, 50GB, 46GB, 54GB, 48GB, 50GB, 52GB, 49GB, 51GB, 47GB, 53GB, 48GB, and 50GB, respectively. These data are stored according to device identifiers to form the current capacity distribution dataset.

[0127] When performing statistical analysis on the current capacity distribution to calculate the mean, standard deviation, and maximum / minimum difference, the collected capacity data is first cleaned to remove obvious outliers (such as data exceeding the device's rated storage capacity). Then, the mean is calculated based on the cleaned data using the arithmetic mean method, with the formula: Mean = Sum of all device capacity data ÷ Number of devices. The average of the remaining storage space data for the 20 servers is 49.8GB, and the standard deviation is approximately 2.87GB. When calculating the maximum / minimum difference, the maximum value of 55GB and the minimum value of 45GB are selected from the capacity data, with a difference of 10GB. The mean, standard deviation, and maximum / minimum difference together constitute the distribution characteristic data.

[0128] When calculating the evenness index based on distribution characteristic data, the formula " Substituting the calculated standard deviation of 2.87 GB and mean of 49.8 GB into the formula, we first calculate the ratio of the standard deviation to the mean, i.e. Subtracting this ratio from 1 yields a uniformity index of approximately 0.9424.

[0129] The time effectiveness and completeness of the capacity data are ensured by the coordination of the aggregation opportunity parameter and the speed parameter, and the problem of data lag in the traditional static collection method is solved; the multi-dimensional distribution characteristic data is generated through statistical analysis, avoiding the limitation that a single data dimension cannot reflect the discrete degree of resource distribution; the abstract distribution characteristics are converted into intuitive and quantitative indicators of balance through the uniformity index calculation, solving the problem that the traditional method cannot determine whether the resource allocation is reasonable, for example, when the uniformity index is too low, the capacity distribution imbalance can be identified in time, providing a basis for subsequent parameter optimization, avoiding the problem of partial equipment overload and partial equipment idling, and improving the resource utilization efficiency.

[0130] In step S107, it is determined whether the uniformity index is lower than the preset uniformity threshold value. If yes, the speed parameter is updated through iterative optimization, and an aggregation control instruction is generated in combination with the aggregation opportunity parameter and output to the resource pool for dynamic adjustment.

[0131] In a specific embodiment, the process of performing step S107 can specifically include the following steps:

[0132] It is determined whether the uniformity index is lower than the preset uniformity threshold value. If yes, the speed parameter value is updated through a step-adaptive iterative optimization method to narrow the gap between the uniformity index and the preset uniformity threshold value.

[0133] According to the updated speed parameter and the aggregation opportunity parameter, an aggregation control instruction containing the values of the two parameters is generated.

[0134] The aggregation control instruction is output to the resource pool, instructing the resource pool to perform a resource aggregation operation according to the rate defined by the speed parameter value and the opportunity defined by the aggregation opportunity parameter, so as to realize stable aggregation.

[0135] Specifically, the preset uniformity threshold value needs to be determined in combination with historical stable operation data of the resource pool and the requirement of the business for capacity distribution balance. For example, in a resource pool scene containing 50 servers, if the historical data shows that the uniformity index is maintained at 0.9 or above, the resource allocation is balanced and the device runs stably, the preset uniformity threshold value can be set to 0.9. The uniformity index is obtained through step S106, and its value is derived from the statistical analysis of the current capacity distribution of the device. Specifically, it is 1 minus the ratio of the standard deviation to the average value of the capacity distribution. If the average value of the current capacity distribution of a resource pool is 50 GB and the standard deviation is 3 GB, then the uniformity index is 1-(3÷50)=0.94. When the uniformity index (such as 0.85) is lower than the preset uniformity threshold value (0.9), the iterative optimization update process of the speed parameter needs to be started.

[0136] The iterative optimization of the speed parameter adopts a step self-adaptive iterative optimization method, the core of which is to dynamically adjust the step according to the deviation of the uniformity index from the preset uniformity threshold, so as to realize the accurate update of the speed parameter. First, the deviation value is calculated, that is, the preset uniformity threshold minus the current uniformity index. If the preset threshold is 0.9 and the current uniformity index is 0.85, the deviation value is 0.05. Then, the initial step is determined according to the deviation value. The larger the deviation value, the larger the initial step is set. For example, when the deviation value is 0.05, the initial step is set to 0.02 (the unit of the speed parameter adjustment is GB / minute). Subsequently, the iterative process is entered. In the first iteration, the current speed parameter (for example, initially set to 2 GB / minute) is added to the initial step to obtain a new speed parameter candidate value of 2.02 GB / minute. Based on the candidate value, the capacity distribution of the resource pool device is re-acquired, and a new uniformity index is calculated. If the new uniformity index (for example, 0.87) and the deviation (0.03) of the preset threshold are smaller than the previous deviation (0.05), it indicates that the step direction is correct, and the step direction can be maintained in the next iteration and the step size can be adjusted according to the deviation change. The step size is adjusted to 0.015. If the new uniformity index and the deviation of the preset threshold increase, the step direction is adjusted in the opposite direction. Repeat this process until the calculated uniformity index reaches or approaches the preset uniformity threshold. At this time, the corresponding speed parameter is the updated speed parameter. For example, after 3 iterations, the speed parameter is updated to 2.04 GB / minute, and the corresponding uniformity index is improved to 0.902, which meets the deviation requirement of the preset threshold.

[0137] After completing the speed parameter update, the aggregation control instruction needs to be generated in combination with the updated speed parameter and the aggregation timing parameter adjusted in step S105. The aggregation timing parameter is the interval time for the resource pool to perform the aggregation operation, for example, 15 minutes after the deviation adjustment. The aggregation control instruction needs to contain the specific values of the speed parameter and the aggregation timing parameter, forming an instruction content such as "performing resource aggregation operation at a rate of 2.04 GB / minute every 15 minutes". The instruction needs to comply with the instruction format specification of the resource pool control system to ensure that the device can accurately parse the parameter meaning.

[0138] After outputting the aggregation control instruction to the resource pool, the control module in the resource pool receives the instruction and parses the speed parameter and the aggregation timing parameter therein. According to the rate defined by the speed parameter, the device adjusts the transmission rate of capacity allocation in the resource aggregation process, for example, transferring the capacity of idle devices to devices with higher load at a rate of 2.04 GB / minute. According to the timing defined by the aggregation timing parameter, the capacity redistribution operation in the whole resource pool range is triggered every 15 minutes, realizing the dynamic adjustment of the device capacity distribution.

[0139] The dynamic adjustment is started by judging the relationship between the uniformity index and the preset threshold value, solving the problem that the traditional static parameters cannot adapt to the capacity distribution change; the step adaptive iterative optimization method is applied, avoiding the over or insufficient in the speed parameter adjustment process, ensuring that the parameter update can accurately reduce the gap between the uniformity index and the threshold value, solving the problem of lack of flexibility in the parameter adjustment of the existing method; the aggregation control instruction contains specific speed and timing parameters, making the resource aggregation operation have clear execution basis, avoiding the resource imbalance caused by the lag of the aggregation instruction, in the large-scale resource pool scene, through reasonable aggregation timing and rate control, the system response delay can be reduced, the service quality can be guaranteed, and the problems of part of the equipment overload and part of the equipment idling are solved, improving the resource utilization efficiency.

[0140] The above describes a method for user flexible resource aggregation intelligent control in the embodiment of the application. Please refer to Figure 2 The user flexible resource aggregation intelligent control process 200 in the embodiment of the application includes seven sub-processes, which are sequentially connected and logically coherent, and the specific operations are as follows:

[0141] The data acquisition and preprocessing process 201 starts with sensor network monitoring, collects the performance and capacity data of each device in the resource pool in real time through the sensor network, and forms a time series of uniform scale after median filtering denoising and min-max normalization processing. Different feature extraction methods are selected according to the device size: less than 50 devices for conventional scene to analyze the fluctuation amplitude and period by ARMA model; more than 50 devices for large-scale scene to extract the dominant frequency in frequency domain by Fourier transform. Finally, these features are combined into a multi-dimensional vector sequence to construct the resource deformation sequence.

[0142] The deformation trend description generation process 202 takes the resource deformation sequence as input, first sets the sliding window size according to the collection frequency, uses the moving average method to take the arithmetic mean of the continuous multi-time point data, generates a smooth sequence to eliminate short-term noise. Then use the ARMA model to extract the characteristics, determine the model order according to the length of the smooth sequence through the AIC criterion, input the data for training, and estimate the coefficients by the least square method, extract the trend slope and periodic fluctuation as the trend characteristics. Finally, the trend characteristics are input into the linear regression model, the time stamp is used as the independent variable, and the smooth sequence data is used as the dependent variable, the least square method is used to solve the regression coefficient and intercept, the trend line is fitted, the numerical index representing the long-term change direction is generated, and the deformation trend description is constructed.

[0143] The profile feature parameter forming process 203 groups based on the deformation trend description using K-means clustering: preset the number of clusters according to the data dimension and the size of the device, select the initial center point, iteratively calculate the Euclidean distance of the data points to the center point and assign the cluster, update the center point (take the average of each dimension of the data in the cluster), until the difference between the center points is less than the threshold, and generate stable clusters. Extract the average of each dimension of the data in each cluster as the center point to represent the cluster deformation trend. Calculate the fluctuation amplitude value (peak-valley) of the center point deformation trend, count the average fluctuation frequency in a unit time as the fluctuation frequency value, and measure the average duration from the start to the stable as the fluctuation duration. Combine the three in a fixed order to form a multi-dimensional vector, and form the profile feature parameter.

[0144] The LSTM network predicts future deformation trend process 204 takes the profile feature parameter as the object, calculates the second-order difference of each dimension of the time series parameter, and if the difference value of any dimension at three or more consecutive time points exceeds the threshold, it is determined that there is a nonlinear fluctuation. If so, input the historical nonlinear fluctuation profile feature, set the LSTM network structure according to the data size (such as the time step and the collection period are consistent, and the number of hidden layer units is configured according to the number of steps), construct samples in the way of predicting the N+1 step from the continuous N steps input, and train until convergence using mean square error as the loss function. Finally, input the current continuous time step parameter into the trained LSTM, output the future multi-step prediction sequence, and obtain the deformation trend.

[0145] The adjustment of the aggregation timing parameter process 205 first acquires the real-time performance and capacity data of the device through the sensor network as the current state data. Then, combined with the mapping relationship between the future deformation trend prediction value and the actual business indicators (such as the normalized fluctuation amplitude to reverse the original indicators), the performance and capacity data at the future time point are calculated. Then calculate the deviation value: first calculate the single indicator deviation, then set the weight according to the business importance, and calculate the comprehensive deviation value by weighted summation. Judge whether the deviation value exceeds the preset threshold (combined with historical stable data and business requirements): if not, maintain the current aggregation timing parameter; if so, adjust it: first calculate the deviation adjustment coefficient (the smaller of the ratio of the absolute value of the deviation to the threshold and 1), then calculate the new aggregation timing parameter according to the formula "new aggregation timing parameter = current parameter × (1 - adjustment coefficient × deviation adjustment coefficient)", and the adjustment coefficient is preset to 0-1 (according to the resource pool sensitivity), to get the new parameter.

[0146] The calculation of the uniformity index process 206 acquires the current capacity distribution data of the device (such as the remaining storage space) according to the adjusted aggregation timing parameter (determines the collection time) and the initial setting of the speed parameter (limits the data transmission processing rate), and forms a data set according to the device identifier. After cleaning the data to remove outliers, calculate the average, standard deviation, maximum and minimum difference, and generate distribution feature data. Finally, calculate the uniformity index = 1 - (standard deviation / average) to get the uniformity index representing the balance of the capacity distribution.

[0147] The aggregation control instruction generation process 207 takes the uniformity index as the basis. If the current uniformity index value is lower than the preset uniformity index threshold, step-adaptive iterative optimization is adopted: the initial step is set according to the deviation, the speed parameter is adjusted through iteration (candidate values are generated and the uniformity index is recalculated to update the step), and the index is close to the threshold; otherwise, the current speed parameter is maintained. Subsequently, the final speed parameter and the aggregation timing parameter are assembled into an instruction (such as “at XX rate, aggregate every XX minutes”) according to the requirements of the control system. After the instruction is issued to the resource pool control module, the capacity allocation rate and the aggregation timing are adjusted according to the parameters, and dynamic resource regulation and control are realized.

[0148] The above describes a process for the flexible resource aggregation of users in the embodiments of the present application. Please refer to Figure 3 An embodiment of the flexible resource aggregation system for users in the embodiments of the present application 300 includes:

[0149] A data acquisition module 301 is configured to acquire real-time data of device performance and capacity from a resource pool, perform preprocessing, time series analysis, and fluctuation feature extraction, and generate a resource deformation sequence.

[0150] A trend extraction module 302 is configured to perform smoothing processing, time series feature extraction, and trend fitting on the resource deformation sequence, and generate a deformation trend description representing a long-term change direction.

[0151] A feature clustering module 303 is configured to generate stable clusters based on the deformation trend description using a K-means clustering algorithm, extract center point features, and construct multi-dimensional contour feature parameters.

[0152] A trend prediction module 304 is configured to train and generate future deformation trend prediction values using an LSTM network when there is nonlinear fluctuation in the contour feature parameters.

[0153] A deviation calculation module 305 is configured to compare the future deformation trend prediction values with current state data, calculate a deviation value, and adjust the aggregation timing parameter accordingly.

[0154] An uniformity calculation module 306 is configured to calculate an uniformity index based on the aggregation timing parameter and the speed parameter, and to calculate the current capacity distribution.

[0155] An instruction output module 307 is configured to update the speed parameter through iterative optimization when the uniformity index is lower than the preset uniformity threshold, and to generate and output an aggregation control instruction in combination with the aggregation timing parameter.

[0156] The above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An aggregated intelligent control method for user flexible resources, characterized in that, The method comprises: Step S101, obtaining real-time data of device performance and capacity from a resource pool, the device performance including CPU utilization and memory occupancy, and the device capacity including storage remaining space and bandwidth occupancy, preprocessing, time series analysis and fluctuation feature extraction are performed on the real-time data to generate a resource deformation sequence; Step S102, according to the resource deformation sequence, a deformation trend description is determined through smoothing processing, time series feature extraction and trend fitting; Step S103, the K-means clustering algorithm is used to group the deformation trend description, and the cluster center point is determined, the feature combination of the cluster center point is extracted as a multi-dimensional vector to form a contour feature parameter; Step S104, it is judged whether the contour feature parameter has nonlinear fluctuation, if so, the LSTM network is trained and future deformation trend prediction value is generated; Step S105, the current state data of the resource pool is obtained, the deviation value of the future deformation trend prediction value and the current state data is calculated, and the aggregation opportunity parameter is adjusted according to the comparison result of the deviation value and the preset deviation threshold; Step S106, based on the aggregation opportunity parameter, and combined with the initial setting of the speed parameter, the current capacity distribution of the device is obtained, the capacity distribution is statistically analyzed, and the uniformity index is calculated; Step S107, it is judged whether the uniformity index is lower than the preset uniformity threshold, if so, the speed parameter is updated through iterative optimization, the speed parameter is defined as the transmission rate of capacity allocation in the resource aggregation process, and the aggregation control instruction is generated combined with the aggregation opportunity parameter and output to the resource pool for dynamic adjustment.

2. The method of claim 1, wherein, The step S101 comprises: Through the sensor network deployed in the resource pool, the performance and capacity fluctuation of the device in the resource pool are monitored, the real-time data is collected, and the median filtering method is applied to the real-time data to remove noise values; The real-time data filtered by noise is normalized, and then sorted by timestamp to generate time series data; If it is a regular scale scene, the autoregressive moving average model is used to capture the autocorrelation and moving average components of the time series data, and the fluctuation amplitude and periodicity features are extracted as fluctuation features, if it is a large-scale resource pool scene, the Fourier transform method is used to convert the time series data from time domain to frequency domain, and the dominant frequency is extracted as the fluctuation feature; The fluctuation features are combined into a multi-dimensional vector sequence to generate a resource deformation sequence representing the dynamic change trend of device performance and capacity.

3. The method of claim 1, wherein, The step S102 comprises: The moving average method is applied to the resource deformation sequence for smoothing processing to generate a smoothing sequence; The smoothing sequence is feature-extracted through a time series analysis method, the autoregressive moving average model is used to capture the time series dependence, and the trend slope and periodic fluctuation are extracted as trend features; The slope and fluctuation mode in the trend features are input into a linear regression model to fit a long-term trend line and generate a numerical index representing the long-term change direction, and the deformation trend description representing the long-term change direction of the resource deformation sequence is constructed based on the numerical index.

4. The method of claim 1, wherein, Step S103 includes: The K-means clustering algorithm was used to group the deformation trend descriptions, generating multiple stable clusters. Analyze multiple stable clusters, extract the average value of all data points in each cluster, and use the average value as the center point of the corresponding cluster. The center point represents the concentrated characteristics of the deformation trend within the cluster. Calculate the fluctuation amplitude value of the deformation trend description corresponding to the center point of each cluster, where the fluctuation amplitude value is the difference between the trend peak and the trough value; The average number of fluctuations in the deformation trend description corresponding to each cluster center point per unit time is used as the fluctuation frequency value. The deformation trend at each cluster center point is measured to describe the average time from the start to the stabilization point, which is taken as the duration of the fluctuation. The fluctuation amplitude value, fluctuation frequency value, and fluctuation duration are combined into a multi-dimensional vector to generate the contour feature parameters, wherein the contour feature parameters are used to characterize typical features of deformation trends.

5. The method of claim 1, wherein, Step S104 includes: Calculate the second difference of the contour feature parameter sequence. If the absolute value of the second difference exceeds a preset fluctuation threshold for three or more consecutive time points, the contour feature parameter is determined to be nonlinear fluctuation. The contour feature parameters that are determined to be nonlinear fluctuations within a historical time period are used as feature inputs. An LSTM network is used to train the network, and the number of hidden layer units and the number of training iterations are set. The mean squared error is used as the loss function for backpropagation optimization. Based on the trained LSTM network, the current contour feature parameters are input to generate a prediction sequence for multiple future time points, thereby obtaining the predicted value of the future deformation trend, wherein the predicted value of the future deformation trend characterizes the future performance and capacity change trend of the resource pool.

6. The method of claim 1, wherein, Step S105 includes: Obtain the current status data of the resource pool, wherein the current status data is the real-time performance and capacity data of the devices in the resource pool; Based on the predicted future deformation trend, calculate the performance and capacity data of the resource pool at future points in time. Calculate the deviation between the performance and capacity data at the future time point and the current state data; Determine whether the deviation value is greater than a preset deviation threshold. If so, the aggregation timing parameter needs to be adjusted to maintain the stability and balance of the resource pool. If not, the aggregation timing parameter does not need to be adjusted.

7. The method of claim 6, wherein the method further comprises: The adjustment of the aggregation timing parameters includes: The deviation adjustment coefficient is calculated based on the deviation value. The deviation adjustment coefficient is obtained by first calculating the ratio of the absolute value of the deviation to the preset deviation threshold, and then comparing the ratio with the natural number 1, taking the minimum value between the two. The new aggregation timing parameter is calculated based on the deviation adjustment coefficient. The new aggregation timing parameter = current aggregation timing parameter × (1 - adjustment coefficient × deviation adjustment coefficient), wherein the adjustment coefficient is preset to be within the range of 0 to 1.

8. The method of claim 1, wherein, Step S106 includes: Based on the aggregation timing parameters and the initial setting of the speed parameters, the current capacity distribution of devices in the resource pool is obtained; Statistical analysis is performed on the current capacity distribution to calculate the mean, standard deviation, and maximum and minimum difference, generating distribution characteristic data; The uniformity index characterizing the balance of capacity distribution is calculated based on the distribution characteristic data. The uniformity index is obtained by subtracting the standard deviation from the mean value from the natural number 1.

9. The method of claim 1, wherein, Step S107 includes: Determine whether the uniformity index is lower than the preset uniformity threshold. If it is lower, update the speed parameter value through a step-size adaptive iterative optimization method to reduce the gap between the uniformity index and the preset uniformity threshold. Based on the updated speed parameters and aggregation timing parameters, an aggregation control command containing the values ​​of both is generated; The aggregation control command is output to the resource pool, instructing the resource pool to perform resource aggregation operations according to the rate defined by the speed parameter value and the timing defined by the aggregation timing parameter, so as to achieve stable aggregation.

10. An aggregated intelligent control system for user flexible resources, characterized in that, For implementing the aggregated intelligent control method for user flexible resources as described in any one of claims 1 to 9, the aggregated intelligent control system for user flexible resources comprises: The data acquisition module is used to obtain real-time data on equipment performance and capacity from the resource pool, perform preprocessing, time series analysis and fluctuation feature extraction, and generate resource deformation sequences. The trend extraction module is used to smooth the resource deformation sequence, extract time-series features and fit trends to generate a deformation trend description that represents the long-term direction of change. The feature clustering module is used to generate stable clusters based on deformation trend description using the K-means clustering algorithm, extract center point features, and construct multi-dimensional contour feature parameters; The trend prediction module is used to train an LSTM network and generate future deformation trend prediction values ​​when there are non-linear fluctuations in the contour feature parameters. The deviation calculation module is used to compare the predicted value of future deformation trend with the current state data, calculate the deviation value, and adjust the aggregation timing parameters accordingly. The uniformity calculation module is used to statistically analyze the current capacity distribution and calculate the uniformity index based on aggregation timing parameters and speed parameters. The instruction output module is used to generate and output aggregation control instructions by iteratively optimizing and updating the speed parameters when the uniformity index is lower than the preset uniformity threshold, combined with the aggregation timing parameters.

Images