Multi-strategy based dynamic allocation method and system for distributed network resources

Abstract

The present invention relates to the technical field of network resource allocation and management, and discloses a multi-strategy-based distributed network resource dynamic allocation method and system. The method includes: collecting node demand-related data of each network node in the distributed network and device health status data of edge nodes and performing preprocessing; constructing an improved prediction network model and inputting the preprocessed data into the prediction network model for training; through the trained prediction network model, predicting the future resource requirements of each network node and the device alarm levels of each edge node; according to the predicted future resource requirements of each network node, adopting a dynamic allocation strategy to optimize the allocated resources; and updating the actual alarm levels of edge nodes according to the predicted alarm levels, and taking corresponding management measures according to different alarm levels. The present invention significantly improves the utilization efficiency of distributed network resources and the overall performance of the network, and enhances the stability and reliability of the network.

Description

Technical Field

[0001] The present invention relates to the technical field of network resource allocation and management, and particularly to a multi-strategy-based distributed network resource dynamic allocation method and system. Background Art

[0002] With the rapid development of information technology, distributed networks are increasingly widely used in all walks of life. The network scale is continuously expanding, and the application scenarios are becoming more and more complex. Most traditional network resource allocation methods are based on static rules or simple prediction models, and these methods are unable to cope when dealing with the dynamic changes of node requirements and the uncertain factors in complex network environments. When the node traffic suddenly increases or resource fluctuations are caused by network failures, traditional methods often lead to uneven resource allocation, key service response delays, and at the same time, there is a coexistence of idle and overloaded resources in some nodes. In addition, as an important part of the network, edge nodes lack an effective intelligent early warning mechanism for device management, usually relying on manual inspections or fixed threshold judgments, which makes the discovery of potential device failures lag behind and the maintenance is not timely, thus affecting the overall performance and reliability of the network, increasing the operation and maintenance costs, and raising the risk of service interruption. Therefore, the existing network resource allocation and management methods are difficult to meet the urgent needs of modern distributed networks for high efficiency, stability, and intelligent management. Summary of the Invention

[0003] In order to overcome the deficiencies of the prior art, the present invention aims to provide a multi-strategy-based distributed network resource dynamic allocation method and system, which can achieve optimal resource allocation and reliable operation of devices by accurately predicting node requirements and intelligently managing edge nodes, improve the overall network performance and operation and maintenance efficiency, and solve the technical problem that the existing network resource allocation and management methods are difficult to meet the requirements of modern distributed networks for high efficiency, stability, and intelligent management.

[0004] The object of the present invention can be achieved by the following technical solutions:

[0005] In a first aspect, the present invention provides a multi-strategy-based distributed network resource dynamic allocation method, including:

[0006] Collecting node demand-related data of each network node in the distributed network and device health status data of edge nodes and performing preprocessing;

[0007] Build an improved prediction network model and input the preprocessed data into the prediction network model for training. The improvement and training process of the prediction network model includes: capturing the long-range dependencies between nodes through an approximate attention mechanism, and processing the input features through several hidden layers to predict the future resource requirements of network nodes and the alarm levels of edge node devices; among them, a loss function composed of weighted data fitting loss, physical constraint loss, and alarm level constraint loss is constructed for training the model;

[0008] Through the trained prediction network model, predict the future resource requirements of each network node and the alarm levels of each edge node device;

[0009] According to the predicted future resource requirements of each network node, adopt a dynamic allocation strategy to optimize the allocated resources; and according to the predicted alarm levels of each edge node device and the preset alarm thresholds, update the actual alarm levels of the edge nodes, and take corresponding management measures according to different alarm levels.

[0010] Preferably, among them, the node demand-related data includes historical resource usage, network traffic characteristics, time characteristics, changes in the number of users or devices connected to the node, types of running application programs, and load conditions; the health status data of the edge node devices includes hardware parameters of the edge node devices, software operation logs, and network connection quality data; the preprocessing includes cleaning, standardizing, and time series alignment processing of the data.

[0011] Preferably, among them, the adopting a dynamic allocation strategy to optimize the allocated resources according to the predicted future resource requirements of each network node includes:

[0012] According to the predicted future resource requirements of each network node by the trained model, adopt a dynamic allocation strategy, aiming to minimize the difference between the predicted resource requirements and the actually allocated resources, and perform resource allocation by weighted summation according to the priorities of the nodes; consider the finiteness of the total resources and node operation constraints during the allocation process, and perform feedback adjustment according to the difference between the actual resource usage and the allocated resources to optimize the resource allocation strategy.

[0013] Preferably, among them, the updating the actual alarm levels of the edge nodes according to the predicted alarm levels of each edge node device and the preset alarm thresholds, and taking corresponding management measures according to different alarm levels includes:

[0014] According to the predicted alarm levels of each edge node device and the preset alarm thresholds, update the actual alarm levels of the edge nodes, and take corresponding management measures according to different alarm levels, including automatic inspection and remote diagnosis.

[0015] Preferably, the approximate attention mechanism uses the Nyström method to replace the traditional attention mechanism, forms new approximate queries and approximate keys through subsampling and mean aggregation to capture long-range dependencies between nodes, and then processes the allocation and alarm levels through the approximate attention mechanism.

[0016] Preferably, the data fitting loss in the loss function uses the mean square error to measure the error of resource demand prediction, and uses the absolute value error to measure the error of alarm level prediction; the physical constraint loss is constructed according to physical laws such as the law of conservation of resources; the alarm level constraint loss is constructed according to the correlation between the change of device alarm level and device status indicators.

[0017] Preferably, the dynamic allocation strategy is optimized using the following formula:

[0018] ;

[0019] where, is the allocated resource vector, representing the resource allocation result for each network node i; is the total number of nodes; is the resource allocated to network node i; is the predicted future resource demand of network node i; is the priority weight of network node i; The role of is to find a set of resource allocation schemes by optimizing , so that the weighted sum of squared errors between the predicted future resource demand of network node i and the corresponding actually allocated resource reaches the minimum.

[0020] Preferably, the feedback adjustment is performed according to the difference between the actual resource usage and the allocated resources, and optimizing the resource allocation strategy includes:

[0021] Calculate the difference between the actual resource usage of the node and the allocated resources. When the difference is greater than the set tolerance threshold, trigger the re-optimization of the resource allocation strategy.

[0022] Preferably, the steps of updating the device alarm level include: according to the predicted alarm level and the preset alarm threshold, classifying the actual alarm level of the edge node into normal, minor alarm and moderate alarm levels, and taking corresponding management measures respectively, including automatic inspection and remote diagnosis.

[0023] In a second aspect, the present invention also provides a multi-strategy-based distributed network resource dynamic allocation system, including:

[0024] A data acquisition module, which is used to collect data related to the node requirements of each network node in the distributed network and the health status data of edge devices and perform preprocessing;

[0025] A model construction module, which is used to construct an improved prediction network model and input the preprocessed data into the prediction network model for training. The improvement and training process of the prediction network model includes: capturing the long-range dependence relationship between nodes through an approximate attention mechanism, and processing the input features through several hidden layers to predict the future resource requirements of nodes and the alarm levels of edge node devices; among them, a loss function composed of weighted data fitting loss, physical constraint loss and alarm level constraint loss is constructed for training the model;

[0026] A prediction output module, which is used to predict the future resource requirements of each network node and the alarm levels of each edge node device through the trained prediction network model;

[0027] An optimization and update module, which is used to optimize the allocated resources by adopting a dynamic allocation strategy according to the predicted future resource requirements of each network node; and update the actual alarm levels of edge nodes according to the predicted alarm levels of each edge node device and a preset alarm threshold, and take corresponding management measures according to different alarm levels.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] (1) Precise prediction and dynamic allocation: By constructing an improved prediction network model, the present invention can accurately predict the future resource requirements of each node and the alarm levels of edge nodes. Based on these prediction results, a dynamic allocation strategy is adopted to optimize the allocation of resources, ensuring the efficient use of resources, avoiding resource waste and overload phenomena, and thus significantly improving the overall utilization rate of network resources.

[0030] (2) Improving network performance and stability: By adjusting the resource allocation strategy in real time, it can ensure that key services obtain sufficient support during peak resource demands, reduce response latency, improve the user experience, and enhance the adaptive ability of the network. At the same time, by intelligently managing edge nodes, potential faults can be discovered and processed in advance, enhancing the stability and reliability of the network and reducing the risk of service interruption.

[0031] (3) Intelligent management and reduction of operation and maintenance costs: The intelligent early warning and management mechanism of the present invention can automatically inspect and remotely diagnose edge nodes, reduce the need for manual intervention, and reduce operation and maintenance costs. At the same time, by providing real-time feedback and adjusting the resource allocation strategy, the operation and maintenance efficiency is improved, making network management more intelligent and automated.

[0032] In summary, the present invention significantly improves the utilization efficiency of distributed network resources and the overall performance of the network, reduces the operation and maintenance costs, enhances the stability and reliability of the network, and provides strong support for the efficient, stable and intelligent management of modern distributed networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The above and / or additional aspects and advantages of the present invention will become apparent and be readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0034] Figure 1 FIG. is a flowchart of a method for dynamically allocating distributed network resources based on multiple policies provided by an embodiment of the present invention;

[0035] Figure 2 FIG. is a data processing flowchart of a method for dynamically allocating distributed network resources based on multiple policies provided by an embodiment of the present invention;

[0036] Figure 3 FIG. is a unit module diagram of a system for dynamically allocating distributed network resources based on multiple policies provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0038] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by terms such as up, down, front, back, left, right, etc. is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation of the present invention.

[0039] In the description of the present invention, the meaning of "several" is one or more, the meaning of "multiple" is two or more, and terms such as greater than, less than, exceeding, etc. are understood as not including the recited number, and terms such as above, below, within, etc. are understood as including the recited number. If there is a description of first and second, it is only for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features or implicitly indicating the sequence of the indicated technical features.

[0040] The present invention aims to construct a multi-strategy-based distributed network resource dynamic allocation method and system. The core objective is to achieve optimized resource allocation, reliable operation of devices, improved overall network performance, and operation and maintenance efficiency by accurately predicting node requirements and intelligently managing edge nodes.

[0041] Specifically, through data collection and preprocessing, the data characteristics and patterns of node requirements and the device status of edge nodes are mined. An improved prediction model is used to predict the future resource requirements of nodes and real-time evaluate the alarm levels of edge nodes, providing a key basis for resource allocation and device management. In the resource allocation link, based on the prediction results and priority strategies, combined with the total resource constraints and node operating conditions, the allocation strategy is dynamically optimized, and the actual resource usage deviation is responded to in real time through an adjustment mechanism to ensure that the resource allocation accurately adapts to business requirements, improve resource utilization rate, reduce energy consumption and costs, and ensure network service quality. In terms of edge node management, based on the predicted alarm levels and threshold systems, the alarm levels are updated in a timely and accurate manner, and corresponding management measures are automatically triggered, such as automatic inspection during minor alarms and remote diagnosis during moderate alarms, to achieve proactive prevention before failures and rapid response during failures, enhance the reliability and stability of edge nodes, reduce the operation and maintenance workload and service interruption duration, and ultimately improve the adaptability, operation efficiency, and service continuity of the distributed network in complex environments.

[0042] In the present invention, network nodes refer to general nodes in the network, which can be servers, client devices, or other network devices, and are responsible for handling data transmission, storage, and computing tasks in the network. The requirements and resource usage of network nodes reflect the basic operating status of the entire network. Edge nodes are specifically devices located at the edge of the network, which are usually closer to the user side and are used to provide low-latency and high-bandwidth services. The health status and management of edge nodes are crucial for ensuring the stability and reliability of the network.

[0043] Please refer to Figures 1-3 , an embodiment of the present invention provides a multi-strategy-based distributed network resource dynamic allocation method 100, including:

[0044] S110: Collect data related to node requirements of each network node in the distributed network and device health status data of edge nodes and perform preprocessing;

[0045] In this step S110, a data collection module is deployed on the devices of each network node and edge node in the distributed network to collect multi-dimensional data:

[0046] (1) Node demand - related data: including historical resource usage (such as CPU usage rate, memory occupancy, network bandwidth consumption, etc.), network traffic characteristics (such as packet size, latency, packet loss rate), time characteristics (such as timestamp, weekend / weekday), changes in the number of users or devices connected to the node, types of running application programs and load conditions, etc. Record the timestamp for analyzing time - series data.

[0047] (2) Device health status data of edge nodes: Collect hardware parameters (such as temperature, voltage, fan speed, load, etc.), software operation logs (such as system error messages, application crash records, etc.), and network connection quality data (such as latency, packet loss rate, etc.).

[0048] Among them, pre - processing includes data cleaning, standardization, and time - series alignment processing, etc. Data cleaning is used to remove significantly incorrect or abnormal data points, such as extremely large or small values caused by sensor failures. Data standardization is used to convert different types of data into a unified numerical range, for example, converting data into values between 0 and 1 to facilitate subsequent model processing. Time - series alignment is used to ensure the time synchronization of node demand data and edge node device status data for joint analysis.

[0049] S120: Construct an improved prediction network model and input the pre - processed data into the prediction network model for training. The improvement and training process of the prediction network model include: capturing long - range dependencies between nodes through an approximate attention mechanism, and processing input features through several hidden layers to predict future resource requirements of nodes and alarm levels of edge node devices; among them, construct a loss function composed of weighted data fitting loss, physical constraint loss, and alarm level constraint loss for training the model;

[0050] This step S120 is used to implement the construction and improvement of the network model, and input the data obtained through step S110 into the improved prediction network model for training. Taking the following process as an example, this step S120 includes the following steps:

[0051] S121: Define each variable;

[0052] (1) Resource demand vector : The resource demand vector of network node at time , where can represent the demand for the - th resource (such as CPU, memory, bandwidth, etc.) of network node at time , being the number of resource types.

[0053] (2)Device status vector : Edge node At time The device status vector of , where Can represent the edge node At time The th device status indicator (such as temperature, voltage, packet loss rate, etc.), Is the number of device status indicators.

[0054] (3)Input variable : Take the time And the network node Or edge node 's identifier (such as node number, edge node number, etc., which can be converted into a numerical form through a suitable coding method) as the input variable of the neural network, denoted as

[0055] (4)Output variable: Includes two main output variables: One is the predicted future resource demand vector of the network node, denoted as ; The other is the predicted device alarm level of the edge node (which can represent different levels through a suitable numerical coding, such as 0 for normal, 1 for minor alarm, 2 for moderate alarm, etc.), denoted as .

[0056] S122: Construct and improve the prediction network model;

[0057] Step S1221: Approximate attention mechanism;

[0058] Traditional attention mechanisms have high computational complexity when dealing with a large amount of node data and are difficult to apply in real time. Therefore, the present invention adopts an approximate attention mechanism based on the Nyström method to replace the traditional attention mechanism to capture long-range dependencies between nodes and reduce computational complexity. This Nyström method is a prior art proposed in 1977 and will not be elaborated here.

[0059] Among them, the model adopts an approximate attention mechanism to form new approximate queries and approximate keys through subsampling and mean aggregation to capture long-range dependencies between nodes.

[0060] Specifically, it includes:

[0061] (1)Input feature quantization;

[0062] Regard the node demand and device status data in the input feature vector As the input query And key( ) to capture long-range dependencies in this data.

[0063] ;

[0064] Among them, is the number of nodes, is the dimension of the feature, represents the real number field.

[0065] (2) Subsampling and mean aggregation;

[0066] Subsample the query and the key ([[]] ) to obtain:

[0067] ;

[0068] Among them, is the dimension after subsampling (smaller than the dimension of the original matrix ), and a new approximate query and an approximate key ([[]] ) are formed through mean aggregation.

[0069] (3) Approximate attention calculation:

[0070] A. Calculate the attention weights using the approximate query and the approximate key ([[]] ), and the formula is:

[0071] ;

[0072] Here, the softmax function is used to normalize the attention weights so that their sum is 1.

[0073] B. Adjust the calculation of the dependencies between nodes through the pseudo-inverse matrix, and the formula is:

[0074] ;

[0075] The pseudo-inverse matrix is used to handle possible singular matrix situations and ensure the stability of the calculation.

[0076] Finally, the processing results of resource allocation and alarm level through the approximate attention mechanism are:

[0077] ;

[0078] In the above formula, represents the device status or other decision variables; is the output of the approximate attention mechanism, which contains the long-range dependencies between the input features; softmax normalizes each element of the matrix through the exponential function, making the output values within the interval, and the sum of each row is 1; represents the dot product of the approximate query matrix and the approximate key matrix, indicating the relationship matrix between the subsampled nodes; Scale and normalize the relationship matrix to prevent the problem of gradient vanishing or explosion in the attention mechanism.

[0079] Step S1222: The input features are passed to the hidden layer;

[0080] Next, the model will process through several hidden layers , and the model consists of hidden layers, where the forward propagation process of the -th layer is expressed by the formula:

[0081] ;

[0082] Among them, the initial input feature is , that is, the approximate attention output obtained through step S1221. is the output vector of the -th hidden layer, is the activation function (such as ReLU, etc.), and are the weight matrix and bias vector of the -th layer respectively, is the total number of hidden layers; is the predicted future resource demand vector of the node; is the predicted alarm level of the edge node device (different levels can be represented by appropriate numerical encodings, such as 0 for normal, 1 for minor alarm, 2 for moderate alarm, etc.).

[0083] Step S1223: Construct the loss function;

[0084] The loss function is composed of three weighted parts: the data fitting loss , the physical constraint loss (defined according to the relevant physical laws of the distributed network) and the alarm level constraint loss :

[0085] ;

[0086] Among them, are the corresponding weight coefficients, used to balance the importance of different loss parts in training.

[0087] Preferably, the data fitting loss in the loss function The mean square error is used to measure the error of resource demand prediction, and the absolute value error is used to measure the error of alarm level prediction; the physical constraint loss is constructed according to physical laws such as the law of conservation of resources; the alarm level constraint loss is constructed according to the correlation between the change of the device alarm level and the device status index. Specifically,

[0088] (1) Data fitting loss :

[0089] ;

[0090] Among them, is the number of training data points, and are respectively the th and the th input data points, and are respectively the corresponding true resource demand vector and true alarm level. Here, the mean square error is used to measure the error of resource demand prediction, and the absolute value error is used to measure the error of alarm level prediction. are respectively the corresponding future resource demand vector of the node predicted by the model and the alarm level of the edge node device; This term is the mean square error, which is used to measure the difference between the resource demand predicted by the model and the true resource demand. By minimizing this term, the model will adjust the parameters to reduce the error of resource demand prediction. This term is the absolute error, which is used to measure the difference between the alarm level predicted by the model and the true alarm level. Since the alarm level is usually a discrete level classification, it is more appropriate to use the absolute value error.

[0091] (2)Physical constraint loss :

[0092] In a distributed network, physical constraints are constructed according to physical laws such as the law of conservation of resources. Assume that the total amount of resources in the entire distributed network is fixed, denoted by , and the resource demand predictions of all nodes are transformed into the form of a loss function to obtain the resource conservation constraint loss for time . Then, during the training process, the physical constraint losses at different times are accumulated and averaged to obtain the total physical constraint loss :

[0093] ;

[0094] Law of conservation of resources (constraint condition): ;

[0095] Among them, is the resource demand change rate; is the predicted demand of network node i for the th type of resource (such as CPU, memory, bandwidth, etc.) at time represents the total amount of the th type of resource (in the entire distributed network); represents the number of resource types and is used to index different types of resources. is the total number of nodes; is the length of the time range considered.

[0096] (3) Alarm level constraint loss :

[0097] Build constraints based on the correlation between the change of device alarm level and device status indicators. If a certain device status indicator exceeds the normal range, it will cause the alarm level to rise, which can be expressed as:

[0098] ;

[0099] Among them, is the number of data points related to the alarm level, and the device status vector , where represents the th device status indicator (such as temperature, voltage, packet loss rate, etc.) of the edge node at time .

[0100] Among them, is a function that describes the correlation between the change of device alarm level and device status indicators. First, judge whether each device status indicator exceeds the normal range. Here, represents the threshold corresponding to the th device status indicator of the edge node . It is set separately according to different edge nodes and the characteristics of different device status indicators, and is used to accurately judge whether the indicator is in an abnormal state at the current node, thereby affecting the alarm level. Among them, is used to indicate whether the indicator is in an abnormal state at the current node:

[0101] ;

[0102] Calculate the device alarm level based on the triggering conditions of the above-mentioned status indicators and their weights. The formula is as follows:

[0103] ;

[0104] In the above formula, represents the weight corresponding to the th device status indicator of the edge node and reflects the differential importance of different device status indicators on different edge nodes k to the alarm level. The floor function is used to map the calculation result to discrete alarm level values (for example, we can set that when the calculation result is in the interval , the alarm level is 0, indicating normal; when in the interval , the alarm level is 1, indicating a minor alarm; when in the interval

[0105] S130: Through the trained prediction network model, predict the future resource requirements of each network node and the alarm levels of each edge node device;

[0106] In this step S130: Two main output variables are output through the trained prediction network model, namely the predicted future resource requirements of each node and the predicted alarm levels of the edge node devices .

[0107] S140: According to the predicted future resource requirements of each node, optimize the allocated resources using a dynamic allocation strategy; and according to the predicted alarm levels of each edge node device and the preset alarm threshold, update the actual alarm levels of the edge nodes, and take corresponding management measures according to different alarm levels.

[0108] In the embodiments of the present invention, the network nodes optimize their resource usage through a dynamic resource allocation strategy to ensure the efficient utilization of network resources. In addition to resource allocation, the edge nodes need to perform intelligent early warning and management based on the health status data of the devices, such as automatic inspection, remote diagnosis, etc., to detect and solve potential faults in advance.

[0109] This step S140 includes the following two steps, namely distributed network resource dynamic allocation and updating the device alarm level, specifically including the following steps:

[0110] S141: According to the predicted future resource requirements of each node, optimize the allocated resources using a dynamic allocation strategy, including:

[0111] According to the future resource requirements of each network node predicted by the trained model, a dynamic allocation strategy is adopted. With the goal of minimizing the difference between the predicted resource requirements and the actual allocated resources, resource allocation is carried out by weighted summation according to the priorities of the nodes; during the allocation process, the finiteness of the total resources and the node operation constraints are considered, and feedback adjustment is performed according to the difference between the actual resource usage and the allocated resources to optimize the resource allocation strategy. More specifically, it includes:

[0112] S1411: Resource allocation;

[0113] Through the trained model, obtain the predicted future resource requirement vectors of each network node , adopt a dynamic allocation strategy, with the goal of minimizing the difference between the predicted resource requirements and the actual allocated resources, and perform weighted summation according to the priorities of the nodes to optimize the allocated resources:

[0114] ;

[0115] Among them, is the allocated resource vector, representing the resource allocation results for each network node i (a total of nodes); represents the resource allocated to network node i; is the predicted resource requirement of network node i, which is the predicted value obtained through the model in step S130; is the priority weight of network node i, used to measure the importance of this node in resource allocation. Nodes with larger weights will receive more resource allocations; The role of (the resource allocation amount for each node) is to find a set of resource allocation schemes through optimization such that the weighted sum of squares of the errors between the predicted demand and the actual allocation

[0116] S1412: Add constraint conditions;

[0117] During the resource allocation process, some constraint conditions need to be considered to ensure the rationality and feasibility of resource allocation.

[0118] 1) Finiteness of total resources: That is, the sum of the resources allocated to all nodes cannot exceed the total resource amount.

[0119] ;

[0120] 2) Node operation constraints: That is, the resource allocation for each node must be within its minimum and maximum demand ranges.

[0121] ;

[0122] Among them, is the minimum resource requirement of network node i, representing the minimum resource allocation that the node can accept; is the maximum resource requirement of network node i, representing the maximum resource allocation that the node can accept.

[0123] S1413: Feedback adjustment;

[0124] To perform dynamic optimization according to the actual resource utilization, a feedback adjustment mechanism is set up: including calculating the difference between the actual resource usage of the node and the allocated resources. When the difference is large, specifically when the difference is greater than the set tolerance threshold, trigger the re-optimization of the resource allocation strategy and update the resource allocation strategy. Specifically:

[0125] ;

[0126] Among them, is the actual resource usage of network node i at time (obtained through monitoring data); is the resource adjustment amount of network node i, representing the difference between the allocated resources and the actual used resources.

[0127] Update condition:

[0128] If , then the resource allocation strategy needs to be updated, that is, re-optimize the resource allocation. Among them, is the set tolerance threshold. If the resource difference is greater than this threshold, trigger the re-optimization.

[0129] Through this feedback adjustment mechanism, the model can optimize the allocation strategy in real time according to the actual resource usage, reducing resource waste or shortage.

[0130] S142: According to the predicted alarm levels of each edge node device and the preset alarm threshold, update the actual alarm level of the edge node, and take corresponding management measures according to different alarm levels, including:

[0131] In this step S142, according to the predicted alarm levels of each edge node device and the preset alarm threshold, update the actual alarm level of the edge node, and take corresponding management measures according to different alarm levels, including automatic inspection and remote diagnosis. The steps to update the device alarm level include: according to the predicted alarm levels of each edge node device and the preset alarm threshold, divide the actual alarm level of the edge node into levels such as normal, minor alarm, and moderate alarm, and take corresponding management measures respectively, including automatic inspection and remote diagnosis.

[0132] Specifically, according to the predicted alarm levels of each edge node device and a preset alarm threshold (set threshold to distinguish normal and minor alarms, the threshold to distinguish minor and moderate alarms, etc.), update the actual alarm level of edge node k, and take corresponding management measures. According to different alarm levels take different measures. For , start the automatic inspection program; for , arrange for operation and maintenance personnel to perform remote diagnostic operations:

[0133] ;

[0134] is the actual alarm level of edge node , which is determined according to the predicted alarm level and the alarm threshold ( , ), and is used to determine what management measures to take for edge node k.

[0135] According to the above embodiments of the present invention, through accurate prediction, dynamic allocation, intelligent management, and adaptive adjustment, the utilization efficiency of distributed network resources and the overall performance of the network are significantly improved, the operation and maintenance costs are reduced, the stability and reliability of the network are enhanced, and strong support is provided for the efficient, stable, and intelligent management of modern distributed networks.

[0136] The above is the introduction of the method embodiments. The following further illustrates the solution of the present invention through device embodiments.

[0137] Figure 3 shows a module schematic diagram of a multi-strategy-based distributed network resource dynamic allocation system 200 according to an embodiment of the present invention. As Figure 3 shown, a multi-strategy-based distributed network resource dynamic allocation system 200 of the present invention includes:

[0138] A data acquisition module 201, configured to collect data related to node requirements of each network node in the distributed network and device health status data of edge nodes and perform preprocessing;

[0139] A model construction module 202, configured to construct an improved prediction network model and input the preprocessed data into the prediction network model for training. The improvement and training process of the prediction network model includes: capturing long-range dependencies between nodes through an approximate attention mechanism, and processing input features through several hidden layers to predict future resource requirements of network nodes and alarm levels of edge node devices; wherein, constructing a loss function composed of weighted data fitting loss, physical constraint loss, and alarm level constraint loss for training the model;

[0140] The prediction output module 203 is used to predict the future resource requirements of each network node and the alarm levels of each edge node device through a trained prediction network model;

[0141] The optimization and update module 204 is used to optimize the allocated resources according to the predicted future resource requirements of each node by adopting a dynamic allocation strategy; and update the actual alarm levels of the edge nodes according to the predicted alarm levels of each edge node device and a preset alarm threshold, and take corresponding management measures according to different alarm levels.

[0142] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the described modules can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0143] The same contents of the method of this embodiment and the above reaction device will not be elaborated herein.

[0144] The above has described in detail an embodiment of the present invention, but the described content is only a preferred embodiment of the present invention and cannot be considered as limiting the scope of implementation of the present invention. All equivalent changes and improvements made according to the scope of the present invention application shall still fall within the scope covered by the claims of the present invention.

Claims

1. A multi-strategy-based dynamic allocation method for distributed network resources, characterized in that Including: Collecting node demand-related data of each network node in the distributed network and device health status data of the edge nodes and performing preprocessing; Constructing an improved prediction network model and inputting the preprocessed data into the prediction network model for training. The improvement and training process of the prediction network model includes: capturing long-range dependencies between nodes through an approximate attention mechanism, and processing the input features through several hidden layers to predict the future resource demands of network nodes and the alarm levels of edge node devices; wherein, constructing a loss function composed of weighted data fitting loss, physical constraint loss, and alarm level constraint loss for training the model; wherein, the data fitting loss in the loss function uses the mean square error to measure the error of resource demand prediction and the absolute value error to measure the error of alarm level prediction; the physical constraint loss is constructed according to the physical law of resource conservation; the alarm level constraint loss is constructed according to the correlation between the change of device alarm level and device status indicators; Through the trained prediction network model, predicting the future resource demands of each network node and the alarm levels of each edge node device; According to the predicted future resource demands of each network node, using a dynamic allocation strategy to optimize the allocated resources; and according to the predicted alarm levels of each edge node device and a preset alarm threshold, updating the actual alarm levels of the edge nodes, and taking corresponding management measures according to different alarm levels.

2. The dynamic distribution method of distributed network resources based on multiple strategies according to claim 1, wherein, Wherein, The node demand-related data includes historical resource usage, network traffic characteristics, time characteristics, changes in the number of users or devices connected to the node, types of running application programs, and load conditions; the device health status data of the edge nodes includes hardware parameters of the edge node devices, software operation logs, and network connection quality data; the preprocessing includes cleaning, standardizing, and time series alignment processing of the data.

3. The dynamic allocation method of distributed network resources based on multiple strategies according to claim 2, wherein Wherein, The adopting a dynamic allocation strategy to optimize the allocated resources according to the predicted future resource demands of each node includes: According to the predicted future resource demands of each network node by the trained model, using a dynamic allocation strategy, aiming to minimize the difference between the predicted resource demands and the actually allocated resources, and performing resource allocation by weighted summation according to the priorities of the network nodes; considering the finiteness of total resources and node operation constraints during the allocation process, and performing feedback adjustment according to the difference between the actual resource usage and the allocated resources to optimize the resource allocation strategy.

4. The dynamic allocation method of distributed network resources based on multiple strategies according to claim 3, characterized in that, Wherein, The updating the actual alarm levels of the edge nodes according to the predicted alarm levels and a preset alarm threshold, and taking corresponding management measures according to different alarm levels includes: According to the predicted alarm levels of each edge node device and a preset alarm threshold, updating the actual alarm levels of the edge nodes, and taking corresponding management measures according to different alarm levels, including automatic inspection and remote diagnosis.

5. The dynamic allocation method of distributed network resources based on multiple strategies according to claim 1, characterized in that Wherein, The approximate attention mechanism uses the Nyström method to replace the traditional attention mechanism. It forms new approximate queries and approximate keys through subsampling and mean aggregation to capture long-range dependencies between nodes, and then processes resource allocation and alarm levels through the approximate attention mechanism.

6. The dynamic allocation method of distributed network resources based on multiple strategies according to claim 3, characterized in that The dynamic allocation strategy is optimized using the following formula: ， Among them, is the allocated resource vector, representing the resource allocation result for each network node i; N total is the total number of network nodes; is the resource allocated to network node i; is the predicted future resource demand of network node i; Wi is the priority weight of network node i; argmin x is to find a set of resource allocation schemes by optimizing X, so that the predicted future resource demand of network node i and the corresponding actually allocated resources reach the minimum of the weighted sum of squared errors between them.

7. The dynamic allocation method of distributed network resources based on multiple strategies according to claim 4, characterized in that Feedback adjustment is performed based on the difference between the actual resource usage and the allocated resources. The optimized resource allocation strategy includes: Calculating the difference between the actual resource usage of the node and the allocated resources. When the difference is greater than the set tolerance threshold, trigger the re-optimization of the resource allocation strategy.

8. The dynamic allocation method of distributed network resources based on multiple strategies according to claim 1, wherein The steps for updating the device alarm level include: based on the predicted alarm levels of each edge node device and the preset alarm threshold, classifying the actual alarm levels of the edge nodes into normal, minor alarm, and moderate alarm, and taking corresponding management measures respectively, including automatic inspection and remote diagnosis.

9. A distributed network resource dynamic allocation system based on multiple strategies, characterized in that, Including: A data acquisition module, which is used to collect data related to node requirements of each network node in the distributed network and health status data of edge devices and perform preprocessing; A model construction module, which is used to construct an improved prediction network model and input the preprocessed data into the prediction network model for training. The improvement and training process of the prediction network model includes: capturing long-range dependencies between nodes through an approximate attention mechanism, and processing input features through several hidden layers to predict the future resource requirements of network nodes and the alarm levels of edge node devices; among them, a loss function composed of weighted data fitting loss, physical constraint loss, and alarm level constraint loss is constructed for training the model; among them, the data fitting loss in the loss function uses the mean square error to measure the error of resource requirement prediction, and uses the absolute value error to measure the error of alarm level prediction; the physical constraint loss is constructed according to the physical law of conservation of resources; the alarm level constraint loss is constructed according to the correlation between the change of device alarm level and device status indicators; A prediction output module, which is used to predict the future resource requirements of each network node and the alarm levels of each edge node device through the trained prediction network model; An optimization and update module, which is used to optimize the allocated resources using a dynamic allocation strategy according to the predicted future resource requirements of each network node; and update the actual alarm levels of edge nodes according to the predicted alarm levels of each edge node device and the preset alarm threshold, and take corresponding management measures according to different alarm levels.

Images