Data acquisition method and system based on frequency adaptive adjustment

Abstract

The invention discloses a data acquisition method and system based on frequency adaptive adjustment, and relates to the field of data acquisition, and the method comprises the steps: obtaining a monitoring index acquisition value set; extracting local time sequence data of the monitoring index acquisition value set under various sliding windows, and calculating a multi-scale difference entropy value; extracting a real-time multi-dimensional monitoring index matrix corresponding to an acquisition moment closest to the current time from the monitoring index acquisition value set, and calculating a mahalanobis distance between the real-time multi-dimensional monitoring index matrix and the historical multi-dimensional monitoring index mean matrix; constructing an input state of a reinforcement learning model according to the first acquisition frequency, the mahalanobis distance and the multi-scale difference entropy, so as to determine a target action in a reinforcement learning mode; and determining a second acquisition frequency according to the first acquisition frequency and acquisition frequency adjustment information indicated by the target action. Therefore, multi-scale time sequence characteristics and real-time deviation calculation are fused, and intelligent self-adaptive adjustment of the sampling frequency is realized by using a reinforcement learning model.

Description

Technical Field

[0001] The present application relates to the technical field of data acquisition, and in particular to a data acquisition method and system based on frequency adaptive regulation. Background Art

[0002] With the rapid development of data centers and cloud platforms, real-time collection and monitoring of key indicators in system operation (such as CPU, memory, disk I / O, network traffic, etc.) has become increasingly important.

[0003] At present, most data collection methods use a fixed sampling frequency, which cannot adjust the sampling accuracy in time according to real-time data fluctuations, resulting in the easy loss of abnormal event information in high-dynamic scenarios, and often causing waste of resources in stable stages. Specifically, for some monitoring indicators in a stable state, the computing, storage and transmission pressure brought by frequent data collection far exceeds the actual demand, adding unnecessary burdens. In some cases, the data values ​​at adjacent moments are very close, resulting in a large amount of redundancy between the data and a waste of storage space.

[0004] In addition, there are some related adaptive data collection mechanisms that rely on preset monitoring thresholds or abnormal trigger mechanisms. Once the monitored indicators exceed or fall below a certain set threshold, the system will actively switch to high-frequency sampling mode to collect more detailed data on abnormal situations. However, such methods are highly dependent on trigger conditions and can often only capture one-time or relatively sudden abnormalities, making it difficult to respond in a timely manner to some gradually accumulated hidden problems. Summary of the invention

[0005] The present application provides a data acquisition method, system, storage medium, computer program product and electronic device based on frequency adaptive regulation, which are used to at least solve the hidden problems of difficulty in balancing one-time bursts and gradual accumulation, data redundancy accumulation and system resource waste caused by relying on preset thresholds or single abnormal trigger mechanisms in the current related technologies.

[0006] In a first aspect, an embodiment of the present application provides a data collection method based on frequency adaptive adjustment, comprising: obtaining a monitoring indicator collection value set, wherein the monitoring indicator collection value set includes multiple collection moments and corresponding multidimensional monitoring indicator collection values ​​obtained based on a first collection frequency; extracting local time series data of the monitoring indicator collection value set under various types of sliding windows, and calculating the scale difference entropy values ​​corresponding to each of the local time series data, and generating multi-scale difference entropy values ​​through fusion; each type of sliding window has a corresponding time scale; extracting the real-time multi-dimensional monitoring indicator collection value corresponding to the collection moment closest to the current time from the monitoring indicator collection value set; A real-time multi-dimensional monitoring indicator matrix is ​​prepared, and the Mahalanobis distance between the real-time multi-dimensional monitoring indicator matrix and the historical multi-dimensional monitoring indicator mean matrix is ​​calculated; the historical multi-dimensional monitoring indicator mean matrix represents the mean of the multi-dimensional monitoring indicator matrix at each acquisition moment in a preset historical time period; an input state of a reinforcement learning model is constructed according to the first acquisition frequency, the Mahalanobis distance and the multi-scale difference entropy value, so as to determine a target action by reinforcement learning; the action space of the reinforcement learning model is defined according to a sampling frequency adjustment strategy; a second acquisition frequency is determined according to the first acquisition frequency and the acquisition frequency adjustment information indicated by the target action.

[0007] In a second aspect, an embodiment of the present application provides a data acquisition system based on frequency adaptive regulation, comprising: an original indicator acquisition unit, used to acquire a monitoring indicator acquisition value set, wherein the monitoring indicator acquisition value set includes multiple acquisition moments and corresponding multi-dimensional monitoring indicator acquisition values ​​obtained based on a first acquisition frequency; a multi-scale difference entropy extraction unit, used to extract local time series data of the monitoring indicator acquisition value set under various types of sliding windows, and calculate the scale difference entropy values ​​corresponding to each of the local time series data, and generate multi-scale difference entropy values ​​through fusion; each type of sliding window has a corresponding time scale; a Mahalanobis distance calculation unit, used to extract the acquisition moment closest to the current time from the monitoring indicator acquisition value set A corresponding real-time multidimensional monitoring indicator matrix is ​​obtained, and the Mahalanobis distance between the real-time multidimensional monitoring indicator matrix and the historical multidimensional monitoring indicator mean matrix is ​​calculated; the historical multidimensional monitoring indicator mean matrix represents the mean of the multidimensional monitoring indicator matrix at each acquisition moment in a preset historical time period; a reinforcement learning optimization unit is used to construct the input state of the reinforcement learning model according to the first acquisition frequency, the Mahalanobis distance and the multi-scale difference entropy value, so as to determine the target action by reinforcement learning; the action space of the reinforcement learning model is defined according to the sampling frequency adjustment strategy; an acquisition frequency adjustment unit is used to determine the second acquisition frequency according to the first acquisition frequency and the acquisition frequency adjustment information indicated by the target action.

[0008] According to a third aspect, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the steps of the data acquisition method based on frequency adaptive adjustment of any embodiment of the present application.

[0009] In a fourth aspect, an embodiment of the present application provides a storage medium on which a computer program is stored, characterized in that when the program is executed by a processor, the steps of the data acquisition method based on frequency adaptive adjustment of any embodiment of the present application are implemented.

[0010] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the data acquisition method based on frequency adaptive adjustment of any embodiment of the present application.

[0011] The data acquisition method and system based on frequency adaptive adjustment provided by the present application can produce at least the following technical effects: (1) The multi-scale difference entropy and Mahalanobis distance are used to capture the multi-scale temporal complexity and real-time change trend of the current monitoring indicators of the system, and the system state is dynamically constructed. The reward relationship between the historical state and the behavior is learned through the reinforcement learning model, which can adaptively output the optimal sampling frequency adjustment strategy, so that the system can improve the sampling accuracy when the data fluctuates violently, and reduce the sampling frequency when the system is stable, thereby effectively avoiding the coexistence of abnormal information loss and resource waste.

[0012] (2) Compared with the traditional method that relies on a single abnormal threshold trigger mechanism, the introduction of the Mahalanobis distance, a statistical distance between multi-dimensional indicators, to judge the degree of deviation between real-time data and historical patterns has a stronger overall perception ability and can discover potential anomalies with nonlinear correlations in the data. In addition, the introduction of multi-scale difference entropy enables the system to evaluate the complexity and volatility of data at multiple time granularities, and has the ability to detect abnormal features that evolve slowly and accumulate gradually, thereby significantly improving the system's recognition effect on non-sudden anomalies.

[0013] (3) Through the continuous updating of the relationship between system state-action-reward by the reinforcement learning model, it has the ability of online learning and adaptive adjustment, and can continuously optimize the indicator collection strategy under different business load modes, operating environments and abnormal behavior modes, forming a closed-loop optimized data collection mechanism, which is significantly better than the traditional method that relies on manually set strategies or static rules.

[0014] Through this technical solution, multi-scale time series features are integrated with real-time deviation calculation, and the reinforcement learning model is used to realize intelligent adaptive adjustment of the sampling frequency, thereby optimizing resource utilization, improving the sensitivity of anomaly detection, and enhancing the self-learning ability of the system, providing an efficient and accurate intelligent solution for the monitoring system of data centers and cloud platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0016] Figure 1 A flowchart of an example of a data acquisition method based on frequency adaptive adjustment according to an embodiment of the present application is shown; Figure 2 A schematic diagram showing an example of a state transition action in a reinforcement learning model; Figure 3 An operation flow chart of an example of calculating multi-scale difference entropy values ​​according to an embodiment of the present application is shown; Figure 4 An operation flow chart of an example of calculating Mahalanobis distance according to an embodiment of the present application is shown; Figure 5 A structural block diagram of an example of a data acquisition system based on frequency adaptive adjustment according to an embodiment of the present application is shown; Figure 6 It is a schematic structural diagram of an embodiment of an electronic device of the present application. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0018] It should be noted that in network telemetry of data centers and cloud platforms, the acquisition frequency directly affects the measurement cost and measurement accuracy. If the acquisition frequency is too low, key information (such as microbursts) may be missed when the indicator changes significantly, resulting in inaccurate measurement results; while if the sampling frequency is too high, it will increase the measurement overhead. Therefore, how to study adaptive variable frequency measurement algorithms according to the dynamic changes of the network has become the key to reducing the measurement cost.

[0019] Among the current related technologies, Moneo is an intuitive adaptive frequency control method that adjusts the measurement collection frequency according to the change of traffic peak value to reduce the measurement overhead. However, it cannot accurately observe the sudden rise and fall of performance. Moreover, there is a lack of theoretical basis for frequency conversion, and there is no way to ensure the inference accuracy of fine-grained data after frequency conversion to collect coarse-grained data. In addition, some experts and scholars have proposed an adaptive sampling method that uses artificial neural networks to reduce the traffic measurement overhead, but this method is only applicable to smooth traffic. Cuckoo sampling is an adaptive sampling method designed based on Cuckoo hashing, which can adjust the sampling rate according to the memory usage of the hash table. Event-driven adaptive sampling methods usually trigger some specific abnormal events, and the monitoring system will actively schedule the sensor nodes to increase the sampling frequency. In order to improve the efficiency of event monitoring and resource utilization, e-Sampling adjusts the sampling rate according to the frequency characteristics of the signal. Based on the adaptive sampling method of compressed sensing, under the premise of given perceptual quality, the minimum required sampling rate is estimated in each sampling window, and the sampling frequency of the sensor is adjusted accordingly. However, perceptual quality estimation will bring high computational costs. On the other hand, the two-stage pattern sampling method collects basic information of historical data in the first stage and uses the basic information of historical data to determine the sampling frequency in the second stage. Although the above methods have made some discussions on variable frequency measurement, the existing variable frequency measurement research lacks theoretical basis, making it difficult to achieve a balance between measurement accuracy and cost, and cannot capture microbursts.

[0020] It should be understood that the purpose of the above description of the current related art is only to facilitate the public to better understand the inventive spirit and motivation of the present application, and is not to be regarded as a limitation of the present application. In addition, the technical solutions described in the above-mentioned current related art are not prior art, and may also be undisclosed technical solutions, such as solutions under research or in the laboratory stage.

[0021] In the technical solution of this application, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved shall comply with the provisions of relevant laws and regulations and shall not violate public order and good morals.

[0022] Figure 1 A flowchart of an example of a data acquisition method based on frequency adaptive adjustment according to an embodiment of the present application is shown.

[0023] Regarding the execution subject of the method of the embodiment of the present application, it can be any controller or processor with computing or processing capabilities. Specifically, it can be implemented by a data acquisition management platform or a data acquisition management system, which constructs input features based on multidimensional statistical difference measurement and time series complexity analysis, and combines reinforcement learning to achieve intelligent adjustment of sampling frequency, effectively improving the robustness, agility and resource efficiency of the data acquisition system in a complex operating environment.

[0024] In some examples, it may be integrated and configured in an electronic device or terminal by means of software, hardware, or a combination of software and hardware, and the type of the terminal or electronic device may be diverse, such as a mobile phone, a tablet computer, or a desktop computer, etc.

[0025] like Figure 1 As shown, in step S110, a monitoring indicator collection value set is obtained, and the monitoring indicator collection value set includes multiple collection moments obtained based on a first collection frequency and corresponding multi-dimensional monitoring indicator collection values.

[0026] Specifically, the system collects multi-dimensional system monitoring indicators in real time from multiple running entities (such as physical servers, virtual machines, containers, database instances, etc.) distributed in a data center or cloud platform according to a preset first collection frequency (for example, 10 seconds or 30 seconds, etc.), and generates a multi-dimensional vector at each sampling point (or sampling moment). It should be understood that the indicator dimensions indicated by the multi-dimensional vector of each sampling point can be diverse, such as CPU frequency, application performance indicators, etc., which are not limited here.

[0027] In some examples of the embodiments of the present application, the monitoring indicator dimensions indicated by the multi-dimensional monitoring indicator collection values ​​include at least one of the following: computing resource indicators, memory resource indicators, storage resource indicators, or network resource indicators.

[0028] More specifically, computing resource indicators include CPU utilization, processing power of computing tasks, thread occupancy, etc., which reflect the use of computing resources and can help determine whether the system has a processing power bottleneck. Memory resource indicators include memory usage, memory allocation and recovery, swap space usage, etc., which help understand system performance and avoid memory leaks. Storage resource indicators include disk read and write speed, disk space usage, I / O wait time, etc., which can help identify disk I / O bottlenecks and insufficient storage problems. Network resource indicators include network bandwidth, latency, packet loss rate, number of connections, etc., which can reflect latency or bandwidth bottlenecks in the data transmission process. As a result, the collected data covers different dimensions of computing, memory, storage, and network resources, ensuring comprehensive monitoring of system status.

[0029] During the data collection phase, it is necessary to ensure that the system can efficiently obtain the above multi-dimensional monitoring indicators and store the collected data according to the timestamp to accurately record the time series characteristics. In addition, the collected data is also verified during each collection, such as by comparing the timestamps of two consecutive data collections, the fluctuation of the collected values, etc., to ensure that no data is lost during the data collection process.

[0030] In step S120, local time series data of the monitoring indicator collection value set under each type of sliding window are extracted, and the scale difference entropy values ​​corresponding to each local time series data are calculated, and multi-scale difference entropy values ​​are generated by fusion, and each type of sliding window has a corresponding time scale.

[0031] In some embodiments, for the acquired monitoring indicator collection value set, the system constructs sliding windows of multiple time scales, and each type of sliding window corresponds to a different time span (e.g., short-term window, medium-term window, and long-term window), such as 1 minute, 5 minutes, and 10 minutes. Then, in each type of sliding window, the corresponding local time series data sequence is extracted, and then the scale difference entropy value of the time series data is calculated using a specific entropy algorithm (e.g., sample entropy, permutation entropy, fuzzy entropy algorithm, etc.), which reflects the complexity and uncertainty of data fluctuations at the corresponding time scale. Then, the difference entropy values ​​calculated at different time scales are fused to obtain a multi-scale difference entropy value, so that the multi-scale entropy value can effectively capture the subtle changes and mutations of the data within different time ranges, and can effectively produce the comprehensive characteristics of the system's short-term shocks and long-term trends, while also avoiding the risk of "overfitting" a single window size to a specific behavior pattern, thereby improving the generalization ability of the system.

[0032] In step S130, a real-time multidimensional monitoring indicator matrix corresponding to the collection time closest to the current time is extracted from the monitoring indicator collection value set, and the Mahalanobis distance between the real-time multidimensional monitoring indicator matrix and the historical multidimensional monitoring indicator mean matrix is ​​calculated.

[0033] Here, the historical multidimensional monitoring indicator mean matrix represents the mean of the multidimensional monitoring indicator matrix at each collection time within a preset historical time period. Specifically, the sampling time closest to the current time is selected from the monitoring indicator collection value set to construct a real-time multidimensional monitoring indicator matrix, and each column in the matrix corresponds to a monitoring indicator. At the same time, the multidimensional indicator mean at each collection time is calculated from the data collected within the preset historical time period to form a historical multidimensional monitoring indicator mean matrix to represent the normal operation status of the system. Then, the real-time matrix is ​​compared with the historical mean matrix using the Mahalanobis distance calculation method to quantify the degree of deviation between the current state and the historical normal state.

[0034] It should be noted that Mahalanobis distance is a distance measurement method used to measure the similarity between multidimensional data points. Unlike Euclidean distance, Mahalanobis distance takes into account the covariance structure of the data, so it can better handle the correlation and scale differences in high-dimensional data. In addition, because Mahalanobis distance takes into account the covariance between each indicator, when facing multidimensional data, the system can not only comprehensively consider the correlation between each indicator, but also quickly discover abnormal conditions. Its results are more suitable for reflecting the actual distance between multidimensional data.

[0035] In step S140, an input state of a reinforcement learning model is constructed according to the first acquisition frequency, the Mahalanobis distance and the multi-scale difference entropy value to determine a target action by reinforcement learning, wherein the action space of the reinforcement learning model is defined according to a sampling frequency adjustment strategy.

[0036] In some embodiments, the first acquisition frequency, Mahalanobis distance, and multi-scale difference entropy value currently used for sampling of the system are fused to form a multi-dimensional state vector, which fully reflects the real-time changes and historical deviations of the current monitoring data. The Mahalanobis distance is used to reflect the difference between real-time multi-dimensional monitoring data and historical data. If the Mahalanobis distance is large, it means that the current data has a large deviation from the historical state, which may indicate that an abnormality or load change has occurred in the system, and the reinforcement learning model may need to consider increasing the sampling frequency. The volatility and complexity of the system data are measured by multi-scale difference entropy values. A higher difference entropy value indicates that the data changes greatly, and more frequent sampling may be required to capture the changes; a lower difference entropy value indicates that the system is stable and the sampling frequency can be appropriately reduced.

[0037] Then, based on the state vector, a decision model is established using a reinforcement learning model algorithm (such as deep Q network, policy gradient, etc.). During the training process, the model continuously optimizes the strategy according to the feedback of the action in the environment, so that the optimal sampling frequency adjustment action can be output in different environments. In the inference stage, the target action output by the training optimized action space is used to determine the corresponding sampling frequency adjustment information, such as increasing the sampling frequency by a certain amount, reducing the sampling frequency by a certain amount, or maintaining the sampling frequency unchanged.

[0038] Figure 2 A schematic diagram showing an example of state transition actions in a reinforcement learning model.

[0039] like Figure 2 As shown, it involves multiple states S The action corresponding to the state transition in the state space of 1~Sn, for example a 1 means from S 1 to S 2. The action of state transition,a 2 means from S 2 to S 1 action, a 3 means from S 1 to S 3 state transition, and so on. Here, the corresponding state transition can occur based on the strategy, and each state transition strategy can be used to cause different transitions. S 1 transfer strategy, action can occur a 2 or a 3.

[0040] It should be noted that the range of states that a state can transfer to (also called transferable states) may be restricted or conditional, for example S 1~ S None of the 3 S 4~ S n occurs between states, and for states S 1 can be transferred to the state S 2 and S 3, etc.

[0041] In some embodiments, each action has a corresponding action reward, and each action reward can be determined based on a preset reward function. Generally, if the transfer reward is larger, the transfer action is considered more valuable, and the system will give priority to executing this action. a The reward corresponding to 1 is greater than a The reward corresponding to 3 indicates the transfer action. a 1 is more valuable.

[0042] In an embodiment of the present application, the action of the reinforcement learning model is used to express the adjustment strategy for the acquisition frequency, which includes increasing the sampling frequency by a certain amount, reducing the sampling frequency by a certain amount, or maintaining the sampling frequency unchanged. In this way, the sampling frequency adjustment decision process is regarded as an "action selection" problem, and the reinforcement learning model is used to learn the best sampling frequency adjustment strategy through continuous trial and error. The input state of the reinforcement learning model consists of the first acquisition frequency currently used by the system and the real-time calculated Mahalanobis distance and multi-scale difference entropy value, which together affect the sampling frequency adjustment decision.

[0043] In step S150, a second acquisition frequency is determined according to the first acquisition frequency and acquisition frequency adjustment information indicated by the target action.

[0044] More specifically, according to the target action output by the reinforcement learning model, the system interprets the corresponding sampling frequency adjustment information, dynamically transforms the first acquisition frequency, and obtains a new second acquisition frequency. For example, the new acquisition frequency can be increased significantly to capture the data details of sudden abnormalities, it can also be appropriately decreased to reduce invalid data collection when the system is stable, and it can also remain unchanged to maintain the current monitoring indicator collection status. Thus, efficient resource management is achieved through adaptive adjustment of the sampling frequency, forming a good data acquisition feedback loop, which can ensure that key data is not lost in a highly dynamic environment, while reducing the computing, storage and transmission burdens in the stable stage.

[0045] Through the embodiments of the present application, real-time multi-scale data feature extraction is integrated with reinforcement learning decision-making. By combining multi-scale difference entropy and Mahalanobis distance to comprehensively reflect the system status and potential anomalies, the sampling frequency can be quickly increased when an abnormality breaks out to ensure the capture of key data. The reinforcement learning model is called to continuously correct the strategy through actual sampling feedback, thereby realizing a complete closed loop from data collection, feature extraction to decision feedback, and realizing intelligent adaptive adjustment of the data collection frequency.

[0046] It should be noted that, in the embodiments of the present application, the action space of the reinforcement learning model is relatively small, and it only involves three types of actions (i.e., increasing, decreasing or maintaining the acquisition frequency), and the Deep Q-Network (DQN) can effectively deal with the problem of discrete action space. Therefore, in some examples of the embodiments of the present application, the reinforcement learning model adopts a deep Q network. In addition, the deep Q network can also use the experience replay technology, using the experience replay buffer to store samples of historical state-action-reward-next state, and randomly sample training to break the sample correlation and thus reduce the correlation between samples, accelerating training convergence. Therefore, by using the architecture of the deep Q network, the system can efficiently learn strategies with the help of a large number of training samples, avoid the limitations brought about by manually setting the sampling frequency, and improve the efficiency of online decision-making.

[0047] As described above, the policy network of the reinforcement learning model is trained through training samples generated by the interaction between the agent and the environment, and its optimization direction is determined by the goal defined by the reward function.

[0048] In some examples of the embodiments of the present application, a reward function is used to reflect the reward that can be obtained by the “state transition” caused by taking an action.

[0049] Specifically, the reward function is expressed as follows: , Formula (1) In the formula, is the reward function, indicating that The reward generated by the state transition action; represents the Mahalanobis distance term, represents the multi-scale difference entropy term, Represents the sampling frequency deviation term.

[0050] , Formula (2) In the formula, Indicates the current status The corresponding Mahalanobis distance is Indicates the next state The corresponding Mahalanobis distance is Represents the Mahalanobis distance reward correction factor.

[0051] It should be noted that in formula (2), the Mahalanobis distance term mainly measures the distance between the state before and after the state transition (i.e., state arrive )The difference between real-time monitoring data and historical data benchmarks changes. When the gap between the current state data and the historical mean decreases after a certain action, it means that the action helps to return the system state to normal; otherwise, it means that the current action has not achieved the expected effect. In this way, the degree of deviation between the abnormal and normal states of the system is effectively captured, so that the positive feedback generated by the action can guide the intelligent agent to minimize the distance from the historical normal state in the state transition, and improve the response speed of anomaly detection and system stability recovery.

[0052] , Formula (3) In the formula, Indicates the current status The corresponding multi-scale difference entropy value, Indicates the next state The corresponding multi-scale difference entropy value, Represents the multi-scale difference entropy reward correction coefficient.

[0053] It should be noted that the multi-scale differential entropy term reflects the volatility and uncertainty of data at different time scales, and thus reflects the complexity of the system state. Large system fluctuations often indicate potential anomalies. By looking at the difference in differential entropy before and after the state transition, it is possible to assess whether the data volatility has been improved after taking a certain action. Specifically, during the training process, the model compares the state and The entropy value of the data can capture the improvement or deterioration of data stability. In this way, the fluctuations of the system in the short and long term can be more comprehensively captured, so that the monitoring strategy can better adapt to different dynamic situations, further reduce the risk of false alarms caused by data fluctuations, and more accurately guide the intelligent agent to adjust the sampling frequency.

[0054] , Formula (4) In the formula, represents the ideal sampling frequency indicated by the labels of the training samples, Indicates the current status The corresponding acquisition frequency, Indicates the next state The corresponding acquisition frequency, Indicates the sampling frequency deviation correction factor.

[0055] In formula (4), It is used to measure the deviation between the current system sampling frequency and the preset target sampling frequency. The core is to quantify the impact of the action on the sampling frequency adjustment effect. By comparing the sampling frequency deviation before and after the state transition, the square form of the normalized difference is used to punish excessive deviation, so that the intelligent agent can alleviate overly frequent or insufficient frequency adjustments and make the sampling frequency as close to the ideal target as possible. In this way, it ensures that the system tends to the target frequency when adjusting the sampling frequency, optimizes the balance of data collection, and helps reduce the number of collection frequency adjustments, thereby reducing resource waste.

[0056] Explanation of the various correction coefficients involved in the above reward function: and Mainly controls the impact of Mahalanobis distance and multi-scale difference entropy on rewards. Generally, data with larger fluctuations require higher sampling frequencies, so these two coefficients are larger. The penalty for adjusting sampling frequency deviation ensures that the system avoids unnecessary frequency adjustments. More specifically, these correction coefficients can be optimized through the sample training process so that the model can adjust the sampling frequency according to the actual situation of the system.

[0057] By using the reward function provided in the embodiment of the present application, the state is calculated arrive The changes in various indicators of the system not only focus on the state evaluation at a single moment, but also truly capture the transfer effect of the system state after the action is executed, so that the intelligent agent can learn to take the best action in different situations through training, which can not only improve the abnormal state, but also ensure the balance between sampling frequency and resource consumption. In addition, the three key indicators of Mahalanobis distance, multi-scale difference entropy and sampling frequency deviation system load are introduced into the reward function at the same time, so that a joint feedback mechanism is formed in anomaly detection, data fluctuation balance and resource management. The intelligent agent can automatically increase the sampling frequency when the data fluctuation is large, and automatically reduce the sampling frequency when the load is high, so as to realize system indicator monitoring based on adaptive frequency adjustment.

[0058] Figure 3The following is a flowchart showing an example of calculating the multi-scale difference entropy value according to an embodiment of the present application. In the embodiment of the present application, a non-parametric entropy calculation method is adopted.

[0059] like Figure 3 As shown, in step S310, each local time series data is processed based on the kernel density estimation method to respectively determine the scale probability density function at the corresponding time scale.

[0060] It should be noted that Kernel Density Estimation (KDE) is a non-parametric probability density estimation method, which is used to estimate continuous probability density functions from discrete data. In the context of multi-dimensional monitoring data, KDE can help estimate the distribution of local time series data. In addition, the KDE method is particularly suitable for processing monitoring data with high noise, which can effectively remove meaningless fluctuations in the data and extract the real change pattern of the data.

[0061] Specifically, by smoothing the data in each time scale window, the probability density function of the data in the window can be obtained using the kernel density estimation method: , which reveals the volatility and randomness of the data.

[0062] , Formula (5) In the formula, For the The probability density function of the data in the time scale window is similar to that of the local time series data at the point The probability density at ; represents the bandwidth parameter of the kernel density estimate, which determines the degree of data smoothing; Indicates The total number of sampling points within the class time scale window, is the Gaussian kernel function, Indicates the corresponding The first The value of the sampling point.

[0063] Therefore, by applying KDE, the calculation of differential entropy does not rely on traditional discretization or binning methods, avoids the inaccuracy caused by data cutting, and can handle smoother continuous data distribution.

[0064] In step S320, based on the scale probability density function of each time scale, the sliding window difference entropy of the corresponding time scale is calculated respectively.

[0065] Here, differential entropy is used to measure the "complexity" or "randomness" of a data set. A higher differential entropy means greater variability in the data. In the calculation of multi-scale differential entropy, the data distribution in each time scale window is evaluated, and the differential entropy of the data in the window is calculated, which can effectively quantify the volatility at different time scales and reveal the stability of the system at each time scale.

[0066] Specifically, for the data in each time scale window, the difference entropy is calculated according to the probability density obtained by KDE, and then the distribution of each window is processed using the formula of difference entropy to ensure that the volatility and change pattern of the data can be reflected.

[0067] , Formula (6) In the formula, Indicates Sliding window difference entropy corresponding to the class time scale window.

[0068] Therefore, by calculating the difference entropy, the complexity of local time series data can be captured. Especially when facing complex multi-dimensional data, the difference entropy provides a method to quantify and compare volatility and uncertainty, which helps the system detect changes at different levels such as short-term fluctuations, medium-term trends and long-term stability. In addition, in the embodiment of the present application, the difference entropy calculation does not only rely on data of a single dimension, but combines data of multiple monitoring indicator dimensions (such as computing resources, memory resources, storage resources and network resources), so that the system can comprehensively evaluate the interaction of multi-dimensional monitoring data.

[0069] In step S330, weighted summation is performed on the difference entropies of the sliding windows to obtain corresponding multi-scale difference entropy values.

[0070] Here, weighted summation is to weight the difference entropies of different time scales according to their importance to obtain a comprehensive multi-scale difference entropy value. By assigning different weights to each time scale, the contribution of each scale to the final result can be more flexibly adjusted in different dynamic environments.

[0071] Specifically, the weight of the time scale is calculated according to the inverse ratio of the window size or the exponential decay method, ensuring that the window with a longer time scale has a greater impact on the final difference entropy value. By weighted summing the difference entropies of all time scales, the final multi-scale difference entropy value can fully reflect the dynamic fluctuations of the system state.

[0072] , Formula (7) , Formula (8) In the formula, Indicates The weight of the class time scale window indicates the contribution of this time scale to the final multi-scale difference entropy value; Represents the attenuation factor, which is used to control the influence of time scale on weight; represents the multi-scale difference entropy value, represents the total number of classes in the time scale, Indicates The time scale size of the class time scale window; represents the category index of the time scale window, is the normalization factor for all time scale weights, which is used to ensure that the sum of the weights of time scale windows under all categories is 1.

[0073] Therefore, through the weighted summation method, the fusion of multi-scale information is achieved, allowing the system to more comprehensively perceive the volatility of data in different time ranges. The longer time scale window can capture long-term trend changes, while the short time scale window can accurately capture instantaneous fluctuations, avoiding the shortcomings of a single time scale and improving the system's monitoring and anomaly detection capabilities.

[0074] Compared with the traditional fixed sampling frequency and simple threshold judgment method, the multi-scale difference entropy calculation used in the embodiment of the present application can comprehensively analyze the dynamic changes of the system. By using the kernel density estimation method to process the data distribution, complex data patterns can be more accurately reflected, reducing the errors caused by data segmentation and discretization in traditional methods. In addition, the advantage of difference entropy in abnormal data processing enables the system to maintain high accuracy and robustness when facing complex and noisy environments.

[0075] Figure 4 An operational flow chart of an example of calculating the Mahalanobis distance according to an embodiment of the present application is shown.

[0076] It should be noted that in the traditional Mahalanobis distance calculation, the distance calculation is based on the difference of data without considering the correlation between data. In the actual monitoring system, various resources (such as CPU, memory, storage, network, etc.) may be highly correlated. Therefore, the covariance matrix is ​​incorporated into the calculation of the Mahalanobis distance in the embodiment of the present application. The covariance matrix reflects the joint change pattern between various monitoring indicators, so that the calculated Mahalanobis distance not only reflects the difference between the mean of current data and historical data, but also takes into account the coordinated changes between different monitoring indicators.

[0077] Specifically, first select the monitoring data at the nearest time. Assuming the system sampling frequency is (for example, 1Hz), then in the system time stream, the monitoring data closest to the current time can be selected by timestamp. The nearest time refers to the time from the current timestamp. The shortest data point at a sampling moment.

[0078] Formula (8) In the formula, Indicates the timestamp of the current moment. The timestamp representing each data point at the moment of collection.

[0079] In this way, the system can ensure that each time anomaly detection is performed, the monitoring data closest to the current moment is used, ensuring real-time and accuracy.

[0080] like Figure 4 As shown, in step S410, the mean values ​​of the monitoring indicators within a preset historical time period are calculated to obtain a historical multi-dimensional monitoring indicator mean matrix.

[0081] Here, by calculating the mean of historical data, a normal state benchmark is obtained for comparison with real-time data. The historical mean represents the normal operation state of the system in the past period of time and is the basis for current data analysis.

[0082] Specifically, the historical time period can be calculated The mean of all sampling points in the matrix It contains the average value of each monitoring indicator in the historical time period. The calculation can flexibly adjust the time period length (for example, the past 5 minutes, the past 24 hours, etc.) to meet different analysis needs.

[0083] , Formula (9) In the formula, Represents the mean matrix of historical multi-dimensional monitoring indicators within the historical time period, Indicates the historical period The multi-dimensional monitoring indicator matrix at each collection time, Represents the total number of data points in the historical time period.

[0084] Therefore, by constructing a historical mean matrix, a benchmark for the normal state of the system is provided.

[0085] In step S420, the covariance matrix is ​​calculated.

[0086] Here, the covariance matrix reflects the correlation between various monitoring indicators. For multi-dimensional monitoring data, if there is a correlation between different resource indicators, directly calculating their Euclidean distance will not accurately capture the overall changes of the system. In contrast, the Mahalanobis distance can consider the relationship between different dimensions by introducing the covariance matrix.

[0087] Specifically, the covariance matrix of historical data is calculated, which measures the joint changes between various monitoring indicators. For example, CPU and memory usage may show a certain positive correlation, and the covariance matrix can accurately capture this relationship.

[0088] , Formula (10) In the formula, is the covariance matrix, which represents the covariance between various monitoring indicators; Represents a vector transpose operation.

[0089] Therefore, the covariance matrix reveals the mutual influence between different monitoring dimensions, and the Mahalanobis distance can eliminate the deviations between different monitoring indicators due to different units and scales, ensuring the accuracy of the calculation.

[0090] In step S430, the Mahalanobis distance between the real-time multi-dimensional monitoring indicator matrix and the historical multi-dimensional monitoring indicator mean matrix is ​​calculated.

[0091] The Mahalanobis distance is used to measure the difference between current data and historical data. The Mahalanobis distance takes into account the covariance matrix of the data, so it can comprehensively evaluate the changes in multidimensional data and take into account the correlation between different dimensions.

[0092] Specifically, the Mahalanobis distance is calculated using the difference between real-time data and historical data. The inverse matrix of the covariance matrix can be introduced in the calculation to eliminate the influence of inconsistent scales between dimensions. The larger the Mahalanobis distance, the more the current data deviates from the historical data, and there may be anomalies.

[0093] , Formula (11) In the formula, represents the real-time multi-dimensional monitoring indicator matrix, express and The Mahalanobis distance between Represents the covariance matrix The inverse matrix of .

[0094] Therefore, by calculating the Mahalanobis distance, the system can accurately evaluate the difference between the current state and the normal historical state, thereby improving the accuracy of anomaly detection. In particular, the Mahalanobis distance calculated in combination with the covariance matrix can comprehensively consider the changes in multi-dimensional monitoring indicators and provide more accurate monitoring feedback.

[0095] Through the embodiment of the present application, the Mahalanobis distance and covariance matrix are introduced. By comparing the real-time data with the historical mean and combining the covariance matrix, the global state of the system can be comprehensively evaluated, which can better reflect the overall health of the system, rather than just the fluctuations in a single dimension. Based on the covariance matrix, the system can capture the correlation between the various monitoring indicators and take it into account in the Mahalanobis distance calculation, so as to achieve a comprehensive consideration of the coordinated changes of multi-dimensional resources, especially in the case of highly correlated monitoring data (such as CPU and memory, disk I / O and network bandwidth, etc.), so that the system can more accurately identify anomalies and avoid the synergy that may be ignored by the single-dimensional monitoring method.

[0096] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of actions combined, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application. In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0097] Figure 5 A structural block diagram of an example of a data acquisition system based on frequency adaptive adjustment according to an embodiment of the present application is shown.

[0098] like Figure 5 As shown, the data acquisition system 500 based on frequency adaptive adjustment includes an original indicator acquisition unit 510, a multi-scale difference entropy extraction unit 520, a Mahalanobis distance calculation unit 530, a reinforcement learning optimization unit 540 and an acquisition frequency adjustment unit 550.

[0099] The original indicator acquisition unit 510 is used to acquire a monitoring indicator collection value set, where the monitoring indicator collection value set includes multiple collection moments obtained based on a first collection frequency and corresponding multi-dimensional monitoring indicator collection values.

[0100] The multi-scale difference entropy extraction unit 520 is used to extract the local time series data of the monitoring indicator collection value set under each type of sliding window, and calculate the scale difference entropy value corresponding to each of the local time series data, and generate a multi-scale difference entropy value by fusion; each type of sliding window has a corresponding time scale.

[0101] The Mahalanobis distance calculation unit 530 is used to extract the real-time multidimensional monitoring indicator matrix corresponding to the collection time closest to the current time from the monitoring indicator collection value set, and calculate the Mahalanobis distance between the real-time multidimensional monitoring indicator matrix and the historical multidimensional monitoring indicator mean matrix; the historical multidimensional monitoring indicator mean matrix represents the mean of the multidimensional monitoring indicator matrix at each collection time within a preset historical time period.

[0102] The reinforcement learning optimization unit 540 is used to construct the input state of the reinforcement learning model according to the first acquisition frequency, the Mahalanobis distance and the multi-scale difference entropy value, so as to determine the target action through reinforcement learning; the action space of the reinforcement learning model is defined according to the sampling frequency adjustment strategy.

[0103] The acquisition frequency adjustment unit 550 is used to determine a second acquisition frequency according to the first acquisition frequency and the acquisition frequency adjustment information indicated by the target action.

[0104] In some embodiments, an embodiment of the present application provides a non-volatile computer-readable storage medium, in which one or more programs including execution instructions are stored. The execution instructions can be read and executed by an electronic device (including but not limited to a computer, a server, or a network device, etc.) to execute the steps of any of the above-mentioned data acquisition methods based on frequency adaptive adjustment in the present application.

[0105] In some embodiments, the embodiments of the present application also provide a computer program product, which includes a computer program stored on a non-volatile computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer performs any step of the above-mentioned data acquisition method based on frequency adaptive adjustment.

[0106] In some embodiments, the embodiments of the present application also provide an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the steps of the data acquisition method based on frequency adaptive adjustment.

[0107] Figure 6 is a schematic diagram of the hardware structure of an electronic device for executing a data acquisition method based on frequency adaptive adjustment provided by another embodiment of the present application, such as Figure 6 As shown, the device includes: One or more processors 610 and memory 620, Figure 6 A processor 610 is taken as an example.

[0108] The device for executing the data collection method based on frequency adaptive regulation may further include: an input device 630 and an output device 640 .

[0109] The processor 610, the memory 620, the input device 630 and the output device 640 may be connected via a bus or other means. Figure 6 The example of connecting through bus is taken in the following.

[0110] The memory 620, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer executable programs and modules, such as program instructions / modules corresponding to the data acquisition method based on frequency adaptive adjustment in the embodiment of the present application. The processor 610 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions and modules stored in the memory 620, that is, the data acquisition method based on frequency adaptive adjustment in the above method embodiment is implemented.

[0111] The memory 620 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 620 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 620 may optionally include a memory remotely arranged relative to the processor 610, and these remote memories may be connected to the electronic device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0112] The input device 630 can receive input digital or character information and generate signals related to user settings and function control of the electronic device. The output device 640 can include a display device such as a display screen.

[0113] The one or more modules are stored in the memory 620, and when executed by the one or more processors 610, the data acquisition method based on frequency adaptive adjustment in any of the above method embodiments is executed.

[0114] The above-mentioned product can execute the method provided in the embodiment of the present application, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the method provided in the embodiment of the present application.

[0115] The electronic devices of the embodiments of the present application exist in various forms, including but not limited to: (1) Mobile communication equipment: This type of equipment is characterized by having mobile communication functions and its main purpose is to provide voice and data communications. This type of terminal includes: smart phones, multimedia phones, functional phones, and low-end phones.

[0116] (2) Ultra-mobile personal computer devices: These devices belong to the category of personal computers, have computing and processing functions, and generally also have mobile Internet access features. These terminals include: PDA, MID and UMPC devices, etc.

[0117] (3) Portable entertainment devices: These devices can display and play multimedia content. They include audio and video players, handheld game consoles, e-books, smart toys, and portable car navigation devices.

[0118] (4) Other onboard electronic devices with data interaction functions, such as on-board devices installed in vehicles.

[0119] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0120] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a general hardware platform, and of course, by hardware. Based on this understanding, the above technical solution, in essence, or the part that contributes to the relevant technology, can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiment.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A data acquisition method based on frequency adaptive regulation, comprising: Acquire a monitoring indicator collection value set, wherein the monitoring indicator collection value set includes a plurality of collection moments obtained based on a first collection frequency and corresponding multi-dimensional monitoring indicator collection values; Extract the local time series data of the monitoring indicator collection value set under each type of sliding window, and calculate the scale difference entropy value corresponding to each of the local time series data, and generate multi-scale difference entropy value by fusion; each type of sliding window has a corresponding time scale; Extracting the real-time multidimensional monitoring indicator matrix corresponding to the collection time closest to the current time from the monitoring indicator collection value set, and calculating the Mahalanobis distance between the real-time multidimensional monitoring indicator matrix and the historical multidimensional monitoring indicator mean matrix; the historical multidimensional monitoring indicator mean matrix represents the mean of the multidimensional monitoring indicator matrix at each collection time within a preset historical time period; An input state of a reinforcement learning model is constructed according to the first acquisition frequency, the Mahalanobis distance and the multi-scale difference entropy value, so as to determine a target action by reinforcement learning; the action space of the reinforcement learning model is defined according to a sampling frequency adjustment strategy; A second acquisition frequency is determined according to the first acquisition frequency and acquisition frequency adjustment information indicated by the target action.

2. The method according to claim 1, wherein: The monitoring indicator dimension indicated by the multi-dimensional monitoring indicator collection value includes at least one of the following: a computing resource indicator, a memory resource indicator, a storage resource indicator or a network resource indicator.

3. The method according to claim 1, wherein: The calculating the scale difference entropy values ​​respectively corresponding to the local time series data and fusing them to generate multi-scale difference entropy values ​​includes: The local time series data are processed based on the kernel density estimation method to respectively determine the scale probability density function at the corresponding time scale: ， In the formula, For the The probability density function of the data in the time scale window is similar to that of the local time series data at the point The probability density at ; represents the bandwidth parameter of the kernel density estimate, which determines the degree of data smoothing; Indicates The total number of sampling points within the class time scale window, is the Gaussian kernel function, Indicates the corresponding The first The value of the sampling point; Based on the scale probability density function of each time scale, the sliding window difference entropy of the corresponding time scale is calculated respectively: ， In the formula, Indicates Sliding window difference entropy corresponding to the class time scale window; Perform weighted summation of the difference entropy of each sliding window to obtain the corresponding multi-scale difference entropy value: ， ， In the formula, Indicates The weight of the class time scale window indicates the contribution of this time scale to the final multi-scale difference entropy value; Represents the attenuation factor, which is used to control the influence of time scale on weight; represents the multi-scale difference entropy value, represents the total number of classes in the time scale, Indicates The time scale size of the class time scale window; represents the category index of the time scale window, is the normalization factor for all time scale weights, which is used to ensure that the sum of the weights of time scale windows under all categories is 1.

4. The method according to claim 3, wherein: The calculating of the Mahalanobis distance between the real-time multi-dimensional monitoring indicator matrix and the historical multi-dimensional monitoring indicator mean matrix includes: Calculate the mean of each monitoring indicator within the preset historical time period to obtain the historical multi-dimensional monitoring indicator mean matrix: ， In the formula, Represents the mean matrix of historical multi-dimensional monitoring indicators within the historical time period, Indicates the historical period The multi-dimensional monitoring indicator matrix at each collection time, Indicates the total number of data points in the historical time period; Compute the covariance matrix: ， In the formula, is the covariance matrix, which represents the covariance between various monitoring indicators; Represents a vector transpose operation; Calculate the Mahalanobis distance between the real-time multidimensional monitoring indicator matrix and the historical multidimensional monitoring indicator mean matrix: ， In the formula, represents the real-time multi-dimensional monitoring indicator matrix, express and The Mahalanobis distance between Represents the covariance matrix The inverse matrix of .

5. The method according to claim 4, wherein: The policy network of the reinforcement learning model is trained through training samples generated by the interaction between the agent and the environment, and its optimization direction is determined by the goal defined by the reward function, where the reward function is expressed by the following formula: ， In the formula, is the reward function, indicating that The reward generated by the state transition action; represents the Mahalanobis distance term, represents the multi-scale difference entropy term, represents the sampling frequency deviation term; ， In the formula, Indicates the current status The corresponding Mahalanobis distance is Indicates the next state The corresponding Mahalanobis distance is Represents the Mahalanobis distance reward correction coefficient; ， In the formula, Indicates the current status The corresponding multi-scale difference entropy value, Indicates the next state The corresponding multi-scale difference entropy value, Represents the multi-scale difference entropy reward correction coefficient; ， In the formula, represents the ideal sampling frequency indicated by the labels of the training samples, Indicates the current status The corresponding acquisition frequency, Indicates the next state The corresponding acquisition frequency, Indicates the sampling frequency deviation correction factor.

6. The method according to claim 1 or 5, wherein: The reinforcement learning model adopts a deep Q network.

7. A data acquisition system based on frequency adaptive regulation, comprising: An original indicator acquisition unit, used to acquire a monitoring indicator collection value set, wherein the monitoring indicator collection value set includes a plurality of collection moments obtained based on a first collection frequency and corresponding multi-dimensional monitoring indicator collection values; A multi-scale difference entropy extraction unit is used to extract the local time series data of the monitoring indicator collection value set under each type of sliding window, and calculate the scale difference entropy value corresponding to each of the local time series data, and generate a multi-scale difference entropy value by fusion; each type of sliding window has a corresponding time scale; A Mahalanobis distance calculation unit is used to extract the real-time multidimensional monitoring indicator matrix corresponding to the collection time closest to the current time from the monitoring indicator collection value set, and calculate the Mahalanobis distance between the real-time multidimensional monitoring indicator matrix and the historical multidimensional monitoring indicator mean matrix; the historical multidimensional monitoring indicator mean matrix represents the mean of the multidimensional monitoring indicator matrix at each collection time in a preset historical time period; A reinforcement learning optimization unit, configured to construct an input state of a reinforcement learning model according to the first acquisition frequency, the Mahalanobis distance and the multi-scale difference entropy value, so as to determine a target action by reinforcement learning; the action space of the reinforcement learning model is defined according to a sampling frequency adjustment strategy; The acquisition frequency adjustment unit is used to determine a second acquisition frequency according to the first acquisition frequency and the acquisition frequency adjustment information indicated by the target action.

Images