Resource optimization method and system for router, gateway and camera

Abstract

The invention provides a resource optimization method and system for a router, a gateway and a camera in the technical field of edge computing. The method comprises the following steps: S1, training a network state prediction model and deploying the network state prediction model to edge computing equipment; s2, the edge computing device collects network index data based on a dynamic weight data collection mechanism and caches the network index data to a time sequence database; s3, inputting the network index data into the network state prediction model to obtain a network state prediction result; s4, generating a network resource optimization strategy based on the network state prediction result, verifying the network resource optimization strategy based on the digital twin technology, and executing the network resource optimization strategy; and S5, performing rolling updating on the network index data cached in the time sequence database based on the access times and the final access time. The method has the advantages that the flexibility and rationality of network resource configuration are greatly improved, and then the system performance and efficiency are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of edge computing, and particularly to a resource optimization method and system for routers, gateways, and cameras. Background Art

[0002] Although cloud computing can provide highly flexible computing and storage resources, due to the relatively high data transmission latency, especially in scenarios with large amounts of data and latency-sensitive requirements, the performance of cloud computing is not ideal. With the rapid development of emerging technologies such as the Internet of Things, artificial intelligence, and big data, edge computing, as an emerging computing model, is gradually becoming a development trend in various industries. With the rapid popularization and development of edge computing devices, more and more edge computing devices are being deployed, generating a large amount of real-time data, such as edge computing devices like routers, gateways, and cameras.

[0003] The real-time data generated by edge computing devices requires real-time response and processing. The centralized data center based on cloud computing cannot meet this real-time requirement, and the resources (network resources) of edge computing devices are relatively scarce. Therefore, there is a need to optimize the network resources of edge computing devices. However, traditional network resource optimization methods have the following problems:

[0004] 1. Defects in static resource allocation: Using a resource allocation algorithm based on a fixed threshold, for example, triggering an expansion operation when the bandwidth utilization rate exceeds 80%. This method cannot adapt to the dynamically changing network load characteristics and is prone to problems such as resource allocation lag or excessive redundancy in burst traffic scenarios, that is, the resource allocation is not flexible enough; a typical manifestation is that when a video surveillance device suddenly transmits high-resolution data at night, the static threshold strategy cannot quickly respond to bandwidth fluctuations. 2. Limitations of the data caching strategy: Only eliminating based on historical access time (such as the LRU algorithm), without considering the spatio-temporal characteristics of data popularity; for example, in a smart home scenario, the high-frequency access data of a security camera and the low-frequency data of a temperature and humidity sensor use the same elimination weight, resulting in a cache hit rate decrease of more than 30%. 3. Insufficient optimization of the network topology: Relying on manually preset rules and unable to autonomously identify the dynamic correlation between edge computing devices; for example, when a gateway and multiple cameras form a star network, it is difficult to detect link congestion in real time and reconstruct it into a distributed topology.

[0005] Therefore, how to provide a resource optimization method and system for routers, gateways, and cameras to improve the flexibility and rationality of network resource configuration, thereby improving system performance and efficiency, has become an urgent technical problem to be solved. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a resource optimization method and system for routers, gateways, and cameras, so as to improve the flexibility and rationality of network resource configuration, and further improve the system performance and efficiency.

[0007] In a first aspect, the present invention provides a resource optimization method for routers, gateways, and cameras, including the following steps:

[0008] Step S1: Create a network state prediction model based on an LSTM network, a GRU network, a feature fusion module, and an output module, train the network state prediction model, and deploy the trained network state prediction model to an edge computing device of a device type of router, gateway, or camera;

[0009] Step S2: The edge computing device collects network metric data based on a dynamic weight data collection mechanism, and caches each network metric data into a time series database;

[0010] Step S3: The edge computing device inputs the network metric data into the network state prediction model to obtain a network state prediction result;

[0011] Step S4: The edge computing device generates a network resource optimization strategy based on the network state prediction result, verifies the network resource optimization strategy based on digital twin technology, and then executes the network resource optimization strategy;

[0012] Step S5: The edge computing device performs rolling update on the network metric data cached in the time series database based on the access times and the last access time.

[0013] Further, the specific content of step S1 is as follows:

[0014] Create a network state prediction model based on an LSTM network, a GRU network, a feature fusion module, and an output module, and set the loss function of the network state prediction model as the mean square error function; the input end of the feature fusion module is respectively connected to the output end of the LSTM network and the output end of the GRU network, and the output end is connected to the input end of the output module;

[0015] The LSTM network is used to capture the periodic characteristics of the network state with a period of 72 hours; the GRU network is used to capture the instantaneous characteristics of the network state with a period of 30 minutes; the feature fusion module is used to fuse the periodic characteristics and the instantaneous characteristics to output fusion features, and the feature fusion module dynamically allocates the weights of the periodic characteristics and the instantaneous characteristics based on the attention mechanism; the output module is used to output the network state prediction result for the next 5 minutes, and the network state prediction result is that the network state is healthy or the network state is sub-healthy;

[0016] Obtain a large amount of historical network metric data including at least bandwidth, latency, packet loss rate, jitter, throughput, and utilization. After performing data cleaning operations on each piece of the historical network metric data, label the network status of each piece of the historical network metric data to construct a dataset;

[0017] Based on a preset splitting ratio, divide the dataset into a training set, a validation set, and a test set. Train the network status prediction model using the training set until the loss value of the loss function is less than a preset loss threshold. During the training process, compress the network status prediction model using knowledge distillation technology and dynamic pruning technology; Validate the trained network status prediction model using the validation set to determine whether the prediction accuracy is greater than a preset accuracy threshold. If not, the validation fails, expand the training set and continue training. If so, the validation succeeds; Test the network status prediction model that has passed the validation using the test set to determine whether the confidence level is greater than a preset confidence level threshold. If not, the test fails, expand the training set and continue training. If so, the test succeeds and end the training;

[0018] Deploy the trained network status prediction model to edge computing devices of device types such as routers, gateways, or cameras.

[0019] Furthermore, the specific steps of step S2 are as follows:

[0020] The edge computing device collects network metric data including at least bandwidth, latency, packet loss rate, jitter, throughput, and utilization through a lightweight distributed probe at a preset collection period. During the collection process, dynamically adjust the weights of the collection periods of each network metric data through a Kalman filter, and cache each piece of the collected network metric data into a time series database;

[0021] The lightweight distributed probe is SkyWalking or Vine; the time series database is Redis;

[0022] The specific steps of step S3 are as follows:

[0023] The edge computing device performs data cleaning operations on the network metric data and then inputs it into the network status prediction model in real time to obtain a network status prediction result.

[0024] Furthermore, the specific steps of step S4 are as follows:

[0025] The edge computing device analyzes the network status prediction result. When the network status prediction result is sub-healthy, generate a network resource optimization strategy including a network topology adjustment sub-strategy and a bandwidth adjustment sub-strategy;

[0026] The specific network topology adjustment sub-strategy is as follows: Calculate the link quality factor Q of each link in the network topology based on bandwidth and latency: Q = 0.7 * bandwidth + 0.3 * (1 - latency), and adjust the topology structure or transmission priority of the network topology based on the link quality factor Q; The specific bandwidth adjustment sub-strategy is as follows: Calculate the allocation ratio of bandwidth based on the DDPG algorithm;

[0027] The edge computing device synchronizes the network resource optimization strategy, network metric data, and its own hardware configuration to the cloud server through the virtual mapping system. The cloud server verifies the network resource optimization strategy based on digital twin technology, generates a verification result, and feeds it back to the edge computing device;

[0028] The edge computing device executes the network resource optimization strategy based on the verification result, calculates the QoE index after the execution of the network resource optimization strategy based on the QoE evaluation model, and performs a rollback operation of the network resource optimization strategy based on the QoE index.

[0029] Further, the specific step S5 is as follows:

[0030] The edge computing device calculates the data importance Score of each network metric data based on the access times and the last access time:

[0031] Score = 0.6 * log(access times) + 0.4 * (1 / (current time - last access time));

[0032] The edge computing device sorts each network metric data based on the data importance Score at preset time intervals, and deletes the network metric data at the end with a preset ratio to perform a rolling update on the network metric data cached in the time series database.

[0033] In a second aspect, the present invention provides a resource optimization system for routers, gateways, and cameras, including the following modules:

[0034] A network state prediction model deployment module, used to create a network state prediction model based on an LSTM network, a GRU network, a feature fusion module, and an output module, train the network state prediction model, and deploy the trained network state prediction model to an edge computing device of the device type of router, gateway, or camera;

[0035] A network metric data acquisition module, used for the edge computing device to acquire network metric data based on a dynamic weight data acquisition mechanism and cache each network metric data in a time series database;

[0036] A network state prediction module, used for the edge computing device to input the network metric data into the network state prediction model to obtain a network state prediction result;

[0037] A network resource optimization module, which is used for an edge computing device to generate a network resource optimization strategy based on the network state prediction result, verify the network resource optimization strategy based on digital twin technology, and then execute the network resource optimization strategy;

[0038] A cache rolling update module, which is used for an edge computing device to perform rolling update on the network metric data cached in the time series database based on the access times and the last access time.

[0039] Furthermore, the network state prediction model deployment module is specifically used for:

[0040] Create a network state prediction model based on an LSTM network, a GRU network, a feature fusion module, and an output module, and set the loss function of the network state prediction model as the mean square error function; the input end of the feature fusion module is respectively connected to the output end of the LSTM network and the output end of the GRU network, and the output end is connected to the input end of the output module;

[0041] The LSTM network is used to capture the periodic characteristics of the network state with a period of 72 hours; the GRU network is used to capture the instantaneous characteristics of the network state with a period of 30 minutes; the feature fusion module is used to fuse the periodic characteristics and the instantaneous characteristics to output fused features, and the feature fusion module dynamically allocates the weights of the periodic characteristics and the instantaneous characteristics based on the attention mechanism; the output module is used to output the network state prediction result for the next 5 minutes, and the network state prediction result is that the network state is healthy or the network state is sub-healthy;

[0042] Obtain a large amount of historical network metric data including at least bandwidth, latency, packet loss rate, jitter, throughput, and utilization rate, perform data cleaning operations on each piece of the historical network metric data, and then label the network state of each piece of the historical network metric data to construct a data set;

[0043] Divide the data set into a training set, a validation set, and a test set based on a preset splitting ratio, train the network state prediction model through the training set until the loss value of the loss function is less than a preset loss threshold, and compress the network state prediction model through knowledge distillation technology and dynamic pruning technology during the training process; verify the trained network state prediction model through the validation set, and judge whether the prediction accuracy is greater than a preset accuracy threshold. If not, the verification fails, and the training set is expanded and training continues. If so, the verification is successful; test the network state prediction model that has passed the verification through the test set, and judge whether the confidence level is greater than a preset confidence level threshold. If not, the test fails, and the training set is expanded and training continues. If so, the test is successful, and the training ends;

[0044] Deploy the trained network state prediction model to edge computing devices of router, gateway or camera device types.

[0045] Furthermore, the network metric data acquisition module is specifically used for:

[0046] The edge computing device collects network metric data including at least bandwidth, latency, packet loss rate, jitter, throughput, and utilization through lightweight distributed probes at a preset acquisition period, and dynamically adjusts the weights of the acquisition periods of each network metric data through a Kalman filter during the acquisition process, and caches the collected network metric data into a time series database.

[0047] The lightweight distributed probe is SkyWalking or Vine; the time series database is Redis.

[0048] The network state prediction module is specifically used for:

[0049] The edge computing device inputs the network metric data into the network state prediction model in real time after performing data cleaning operations, and obtains a network state prediction result.

[0050] Furthermore, the network resource optimization module is specifically used for:

[0051] The edge computing device analyzes the network state prediction result, and when the network state prediction result is sub-healthy, generates a network resource optimization strategy including a network topology adjustment sub-strategy and a bandwidth adjustment sub-strategy.

[0052] The network topology adjustment sub-strategy is specifically: calculating the link quality factor Q of each link in the network topology based on bandwidth and latency: Q = 0.7 * bandwidth + 0.3 * (1 - latency), and adjusting the topology structure or transmission priority of the network topology based on the link quality factor Q; the bandwidth adjustment sub-strategy is specifically: calculating the allocation ratio of bandwidth based on the DDPG algorithm.

[0053] The edge computing device synchronizes the network resource optimization strategy, network metric data, and its own hardware configuration to the cloud server through a virtual mapping system, and the cloud server verifies the network resource optimization strategy based on digital twin technology, generates a verification result and feeds it back to the edge computing device.

[0054] The edge computing device executes the network resource optimization strategy based on the verification result, calculates the QoE index after the execution of the network resource optimization strategy based on the QoE evaluation model, and performs a rollback operation of the network resource optimization strategy based on the QoE index.

[0055] Furthermore, the cache rolling update module is specifically used for:

[0056] The edge computing device calculates the data importance Score of each network metric data based on the access times and the last access time:

[0057] Score = 0.6 * log(access times) + 0.4 * (1 / (current time - last access time));

[0058] The edge computing device sorts each network metric data based on the data importance Score at preset time intervals, and deletes the network metric data at the end with a preset ratio, so as to perform rolling update on the network metric data cached in the time series database.

[0059] The advantages of the present invention are as follows:

[0060] 1. A network state prediction model is created through an LSTM network, a GRU network, a feature fusion module, and an output module, and after being trained, it is deployed to the edge computing device; the edge computing device collects network metric data based on a dynamic weight data acquisition mechanism and caches it in the time series database, inputs the network metric data into the network state prediction model to obtain a network state prediction result, generates a network resource optimization strategy based on the network state prediction result, verifies the network resource optimization strategy based on digital twin technology, executes the network resource optimization strategy, and performs rolling update on the network metric data cached in the time series database based on the access times and the last access time; that is, network metric data is collected based on a dynamic weight data acquisition mechanism to adapt to the dynamically changing network load. For example, when sudden traffic is detected, the collection period of bandwidth and packet loss rate is shortened to increase the corresponding data volume for subsequent analysis; the network resource optimization strategy is also dynamically updated based on the network state prediction result to make the strategy fit the actual network state and dynamically optimize the network topology and bandwidth; moreover, the caching of network metric data combines the access times and the last access time (heat - timeliness two - dimensional elimination algorithm), effectively improving the cache hit rate, and ultimately greatly improving the flexibility and rationality of network resource allocation, thereby greatly improving the system performance and efficiency.

[0061] 2. Create a network state prediction model through an LSTM network, a GRU network, a feature fusion module, and an output module. Set the loss function of the network state prediction model as the mean square error function. The LSTM network is used to capture the periodic features of the network state with a period of 72 hours. The GRU network is used to capture the instantaneous features of the network state with a period of 30 minutes. The feature fusion module is used to fuse the periodic features and the instantaneous features and output the fused features. The output module is used to output the network state prediction results for the next 5 minutes. That is, the network state prediction model synchronously captures the periodic features and the instantaneous features through a dual-channel neural network and fuses them, effectively improving the feature extraction ability. The mean square error function punishes the square of the error and is sensitive to small errors, which is suitable for scenarios with high prediction accuracy requirements, thereby greatly improving the accuracy of network state prediction.

[0062] 3. By dividing the data set into a training set, a validation set, and a test set, train the network state prediction model with the training set until the loss value of the loss function is less than the preset loss threshold. During the training process, compress the network state prediction model through knowledge distillation technology and dynamic pruning technology. Calculate the prediction accuracy through the validation set to verify the trained network state prediction model, and calculate the confidence through the test set to test the network state prediction model that has passed the verification. If the test is successful, deploy the network state prediction model to edge computing devices of router, gateway, or camera device types. That is, during the training process of the network state prediction model, continuous compression, verification, and testing are carried out to effectively balance the model volume and prediction accuracy of the network state prediction model, so as to better deploy it on edge computing devices.

[0063] 4. During the network metric data collection process, dynamically adjust the weights of the collection periods of each network metric data through a Kalman filter. When facing sudden changes in network state or network load, more network metric data can be collected for analysis, so as to better and more timely optimize network resources, greatly improving the reliability and timeliness of network resource optimization.

[0064] 5. Calculate the link quality factor Q of each link in the network topology through bandwidth and delay. Adjust the topology structure or transmission priority of the network topology based on the link quality factor Q. Calculate the allocation ratio of bandwidth through the DDPG algorithm. Verify the network resource optimization strategy through digital twin technology. Perform a fallback operation on the network resource optimization strategy based on the QoE metric. That is, the edge computing device dynamically adjusts the network topology and bandwidth allocation. Before the adjustment, perform a stress test through digital twin technology. If it is determined that the optimization effect is not good based on the QoE metric, an automatic fallback operation can be performed, greatly improving the flexibility and rationality of network resource configuration. Description of the Drawings

[0065] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0066] Figure 1 It is a flowchart of a resource optimization method for routers, gateways, and cameras according to the present invention.

[0067] Figure 2 It is a schematic structural diagram of a resource optimization system for routers, gateways, and cameras according to the present invention. Specific embodiments

[0068] The technical solution in the embodiments of the present application has the following general idea: collecting network metric data based on a dynamic weight data collection mechanism to adapt to the dynamically changing network load; the network resource optimization strategy is also dynamically updated based on the network state prediction result to make the strategy fit the actual network state and dynamically optimize the network topology and bandwidth; and the caching of network metric data combines the access times and the last access time, effectively improving the cache hit rate, thereby enhancing the flexibility and rationality of network resource allocation.

[0069] Please refer to Figures 1 to 2 As shown, a preferred embodiment of a resource optimization method for routers, gateways, and cameras according to the present invention includes the following steps:

[0070] Step S1: Create a network state prediction model based on an LSTM network, a GRU network, a feature fusion module, and an output module, train the network state prediction model, and deploy the trained network state prediction model to an edge computing device of a device type of router, gateway, or camera;

[0071] Step S2: The edge computing device collects network metric data based on a dynamic weight data collection mechanism and caches each network metric data into a time series database;

[0072] Step S3: The edge computing device inputs the network metric data into the network state prediction model to obtain a network state prediction result;

[0073] Step S4: The edge computing device generates a network resource optimization strategy based on the network state prediction result, verifies the network resource optimization strategy based on digital twin technology, and then executes the network resource optimization strategy;

[0074] Step S5: The edge computing device performs rolling updates on the network metric data cached in the time series database based on the access times and the last access time.

[0075] The specific content of step S1 is as follows:

[0076] Create a network state prediction model based on the LSTM network, GRU network, feature fusion module, and output module, and set the loss function of the network state prediction model to the mean squared error function; the input end of the feature fusion module is respectively connected to the output end of the LSTM network and the output end of the GRU network, and the output end is connected to the input end of the output module; the mean squared error function is applicable to regression problems, that is, predicting continuous values (such as network latency, bandwidth utilization, etc.);

[0077] The LSTM network is used to capture the periodic features of the network state with a period of 72 hours; the GRU network is used to capture the instantaneous features of the network state with a period of 30 minutes; the feature fusion module is used to fuse the periodic features and instantaneous features to output fused features, and the feature fusion module dynamically allocates the weights of the periodic features and instantaneous features based on the attention mechanism; the output module is used to output the network state prediction result for the next 5 minutes, and the network state prediction result is that the network state is healthy or the network state is sub-healthy;

[0078] Create a network state prediction model through the LSTM network, GRU network, feature fusion module, and output module, and set the loss function of the network state prediction model to the mean squared error function; the LSTM network is used to capture the periodic features of the network state with a period of 72 hours; the GRU network is used to capture the instantaneous features of the network state with a period of 30 minutes; the feature fusion module is used to fuse the periodic features and instantaneous features to output fused features; the output module is used to output the network state prediction result for the next 5 minutes; that is, the network state prediction model synchronously captures the periodic features and instantaneous features through a dual-channel neural network and fuses them, effectively improving the feature extraction ability, and the mean squared error function punishes the square of the error and is sensitive to small errors, suitable for scenarios with high requirements for prediction accuracy, thus greatly improving the accuracy of network state prediction.

[0079] Obtain a large amount of historical network metric data at least including bandwidth, latency, packet loss rate, jitter, throughput, and utilization. After performing data cleaning operations on each piece of the historical network metric data, then label the network state of each piece of the historical network metric data to construct a dataset;

[0080] Divide the dataset into a training set, a validation set, and a test set based on a preset splitting ratio. Train the network state prediction model using the training set until the loss value of the loss function is less than a preset loss threshold. During the training process, compress the network state prediction model using knowledge distillation technology and dynamic pruning technology. Validate the trained network state prediction model using the validation set and determine whether the prediction accuracy is greater than a preset accuracy threshold. If not, the validation fails, and the training set is expanded for continued training. If so, the validation succeeds. Test the network state prediction model that has passed the validation using the test set and determine whether the confidence level is greater than a preset confidence threshold. If not, the test fails, and the training set is expanded for continued training. If so, the test succeeds, and the training ends;

[0081] Deploy the trained network state prediction model to an edge computing device with a device type of router, gateway, or camera.

[0082] By dividing the dataset into a training set, a validation set, and a test set, training the network state prediction model using the training set until the loss value of the loss function is less than a preset loss threshold, and compressing the network state prediction model using knowledge distillation technology and dynamic pruning technology during the training process; validating the trained network state prediction model by calculating the prediction accuracy using the validation set, testing the network state prediction model that has passed the validation by calculating the confidence level using the test set, and deploying the network state prediction model to an edge computing device with a device type of router, gateway, or camera if the test is successful; that is, continuously compressing, validating, and testing during the training process of the network state prediction model to effectively balance the model volume and prediction accuracy of the network state prediction model for better deployment on edge computing devices.

[0083] Step S2 is specifically as follows:

[0084] The edge computing device collects network metric data including at least bandwidth, latency, packet loss rate, jitter, throughput, and utilization rate through lightweight distributed probes at a preset collection period. During the collection process, dynamically adjust the weights of the collection periods of each network metric data through a Kalman filter (i.e., dynamically adjust the collection period, and shorten the collection period when abnormal conditions occur to collect more data), and cache the collected network metric data in a time series database. The core idea of the Kalman filter is to use a dynamic model and an observation model, combine prior knowledge and observation data, and recursively estimate the state of the system;

[0085] The lightweight distributed probe is SkyWalking or Vine; the time series database is Redis; the lightweight distributed probe is a tool for collecting service call chain information in real time in a distributed system. It is usually embedded in microservices in the form of a lightweight proxy program to monitor and analyze the service call process; the main function of the lightweight distributed probe is to collect call chain information, performance metrics, and log data of the application in a non-invasive manner through bytecode injection or proxy. These probes are usually designed to have less impact on application performance while being able to efficiently transmit data to the backend system for analysis;

[0086] During the process of collecting network metric data, by dynamically adjusting the weights of the collection periods of each network metric data through a Kalman filter, more network metric data can be collected for analysis when facing sudden changes in network status or network load, so as to optimize network resources better and more timely, greatly improving the reliability and timeliness of network resource optimization.

[0087] The specific content of step S3 is as follows:

[0088] The edge computing device performs data cleaning operations on the network metric data and then inputs it into the network state prediction model in real time to obtain the network state prediction result.

[0089] The specific content of step S4 is as follows:

[0090] The edge computing device analyzes the network state prediction result. When the network state prediction result is sub-healthy, it generates a network resource optimization strategy including a network topology adjustment sub-strategy and a bandwidth adjustment sub-strategy;

[0091] The specific content of the network topology adjustment sub-strategy is as follows: Calculate the link quality factor Q of each link in the network topology based on bandwidth and delay: Q = 0.7 * bandwidth + 0.3 * (1 - delay), and adjust the topology structure or transmission priority of the network topology based on the link quality factor Q, that is, reconstruct the device connection based on the improved Prim algorithm to dynamically generate a minimum spanning tree; The Prim algorithm is a classic algorithm for finding the minimum spanning tree in a weighted connected graph. It is based on the idea of a greedy algorithm and constructs the spanning tree by gradually selecting the edge with the smallest weight;

[0092] The bandwidth adjustment sub-strategy is specifically: calculating the bandwidth allocation ratio based on the DDPG algorithm; the DDPG (Deep Deterministic Policy Gradient) algorithm is an algorithm based on deep reinforcement learning, which is suitable for solving problems in continuous action spaces and combines deterministic strategies with a model-independent reinforcement learning algorithm of deep neural networks; the calculation of the bandwidth allocation ratio is realized by the DDPG algorithm, which can effectively deal with the resource allocation problem in a dynamic network environment and show good performance in a variety of scenarios;

[0093] The edge computing device synchronizes the network resource optimization strategy, network indicator data, and the hardware configuration of the local device to the cloud server through the virtual mapping system. The cloud server verifies the network resource optimization strategy based on the digital twin technology, generates the verification result, and feeds it back to the edge computing device. In real time, the cloud server can use the Monte Carlo method to simulate extreme load scenarios (such as 200% burst traffic).

[0094] The edge computing device executes the network resource optimization strategy based on the verification result, calculates the QoE index after the execution of the network resource optimization strategy based on the QoE evaluation model, and performs a rollback operation of the network resource optimization strategy based on the QoE index. The QoE evaluation model is specifically:

[0095] QoE = 0.5*(1-freezing duration / total duration)+0.3*resolution level+0.2*operation response speed.

[0096] The link quality factor Q of each link in the network topology is calculated through bandwidth and delay, and the topological structure or transmission priority of the network topology is adjusted based on the link quality factor Q. The bandwidth allocation ratio is calculated through the DDPG algorithm, and the network resource optimization strategy is verified through digital twin technology. The fallback operation of the network resource optimization strategy is performed based on the QoE indicator. That is, the edge computing device dynamically adjusts the network topology and bandwidth allocation, and performs stress testing through digital twin technology before adjustment. If the optimization effect is not good based on the QoE indicator, the fallback operation can be automatically performed, which greatly improves the flexibility and rationality of network resource configuration.

[0097] The step S5 is specifically as follows:

[0098] The edge computing device calculates the data importance score of each network indicator data based on the number of accesses and the last access time:

[0099] Score = 0.6*log(number of visits)+0.4*(1 / (current time-last visit time));

[0100] At every preset time interval, the edge computing device sorts the network metric data based on the data importance Score, and deletes the network metric data at the end with a preset ratio, so as to perform rolling update on the network metric data cached in the time series database.

[0101] A preferred embodiment of a resource optimization system for routers, gateways, and cameras according to the present invention includes the following modules:

[0102] A network status prediction model deployment module, configured to create a network status prediction model based on an LSTM network, a GRU network, a feature fusion module, and an output module, train the network status prediction model, and deploy the trained network status prediction model to an edge computing device of a device type of a router, a gateway, or a camera;

[0103] A network metric data acquisition module, configured to enable an edge computing device to acquire network metric data based on a dynamic weight data acquisition mechanism, and cache each network metric data into a time series database;

[0104] A network status prediction module, configured to enable an edge computing device to input the network metric data into the network status prediction model to obtain a network status prediction result;

[0105] A network resource optimization module, configured to enable an edge computing device to generate a network resource optimization strategy based on the network status prediction result, verify the network resource optimization strategy based on digital twin technology, and execute the network resource optimization strategy;

[0106] A cache rolling update module, configured to enable an edge computing device to perform rolling update on the network metric data cached in the time series database based on the access times and the last access time.

[0107] The network status prediction model deployment module is specifically configured to:

[0108] Create a network status prediction model based on an LSTM network, a GRU network, a feature fusion module, and an output module, and set the loss function of the network status prediction model as a mean square error function; the input end of the feature fusion module is respectively connected to the output end of the LSTM network and the output end of the GRU network, and the output end is connected to the input end of the output module; the mean square error function is applicable to regression problems, that is, predicting continuous values (such as network latency, bandwidth utilization, etc.);

[0109] The LSTM network is used to capture the periodic features of the network state with a period of 72 hours; the GRU network is used to capture the instantaneous features of the network state with a period of 30 minutes; the feature fusion module is used to fuse the periodic features and the instantaneous features to output fused features, and the feature fusion module dynamically allocates the weights of the periodic features and the instantaneous features based on the attention mechanism; the output module is used to output the prediction result of the network state in the next 5 minutes, and the prediction result of the network state is that the network state is healthy or the network state is sub-healthy.

[0110] A network state prediction model is created through the LSTM network, the GRU network, the feature fusion module and the output module. The loss function of the network state prediction model is set as the mean square error function; the LSTM network is used to capture the periodic features of the network state with a period of 72 hours; the GRU network is used to capture the instantaneous features of the network state with a period of 30 minutes; the feature fusion module is used to fuse the periodic features and the instantaneous features to output fused features; the output module is used to output the prediction result of the network state in the next 5 minutes; that is, the network state prediction model synchronously captures the periodic features and the instantaneous features through a dual-channel neural network and fuses them, effectively improving the feature extraction ability. The mean square error function penalizes the square of the error and is sensitive to small errors, which is suitable for scenarios with high requirements for prediction accuracy, thus greatly improving the accuracy of network state prediction.

[0111] A large amount of historical network metric data including at least bandwidth, latency, packet loss rate, jitter, throughput, and utilization is obtained. After performing data cleaning operations on each piece of the historical network metric data, each piece of the historical network metric data is labeled with the network state to construct a data set.

[0112] The data set is divided into a training set, a validation set, and a test set based on a preset splitting ratio. The network state prediction model is trained through the training set until the loss value of the loss function is less than a preset loss threshold. During the training process, the network state prediction model is compressed through knowledge distillation technology and dynamic pruning technology; the trained network state prediction model is verified through the validation set to determine whether the prediction accuracy rate is greater than a preset accuracy threshold. If not, the verification fails, and the training set is expanded and training continues. If so, the verification succeeds; the network state prediction model that has passed the verification is tested through the test set to determine whether the confidence level is greater than a preset confidence level threshold. If not, the test fails, and the training set is expanded and training continues. If so, the test succeeds and the training ends.

[0113] The trained network state prediction model is deployed to edge computing devices of device types such as routers, gateways, or cameras.

[0114] By dividing the data set into a training set, a validation set, and a test set, training the network state prediction model with the training set until the loss value of the loss function is less than a preset loss threshold, and compressing the network state prediction model during the training process through knowledge distillation technology and dynamic pruning technology; calculating the prediction accuracy through the validation set to verify the trained network state prediction model, calculating the confidence through the test set to test the network state prediction model that has passed the verification, and if the test is successful, deploying the network state prediction model to an edge computing device with a device type of router, gateway, or camera; that is, continuously compressing, validating, and testing during the training process of the network state prediction model to effectively balance the model volume and prediction accuracy of the network state prediction model, so as to be better deployed on the edge computing device.

[0115] The network metric data collection module is specifically used for:

[0116] The edge computing device collects network metric data including at least bandwidth, latency, packet loss rate, jitter, throughput, and utilization rate through a lightweight distributed probe at a preset collection period. During the collection process, the weights of the collection periods of each network metric data are dynamically adjusted through a Kalman filter (i.e., dynamically adjusting the collection period, shortening the collection period when abnormal conditions occur to collect more data), and caching the collected network metric data into a time series database; the core idea of the Kalman filter is to use a dynamic model and an observation model, combine prior knowledge and observation data, and recursively estimate the state of the system.

[0117] The lightweight distributed probe is SkyWalking or Vine; the time series database is Redis; the lightweight distributed probe is a tool for collecting service call link information in a distributed system in real time, usually embedded in microservices in the form of a lightweight agent program to monitor and analyze the service call process; the main function of the lightweight distributed probe is to non-intrusively collect the call link information, performance metrics, and log data of the application through bytecode injection or proxy. These probes are usually designed to have less impact on application performance while being able to efficiently transmit data to the backend system for analysis.

[0118] By dynamically adjusting the weights of the collection periods of each network metric data through a Kalman filter during the network metric data collection process, more network metric data can be collected for analysis in the face of sudden changes in network state or network load, so as to better and more timely optimize network resources, greatly improving the reliability and timeliness of network resource optimization.

[0119] The network state prediction module is specifically used for:

[0120] The edge computing device performs data cleaning operations on the network indicator data and inputs the data into the network status prediction model in real time to obtain a network status prediction result.

[0121] The network resource optimization module is specifically used for:

[0122] The edge computing device analyzes the network status prediction result, and when the network status prediction result is sub-healthy, generates a network resource optimization strategy including a network topology adjustment sub-strategy and a bandwidth adjustment sub-strategy;

[0123] The network topology adjustment sub-strategy is specifically as follows: based on bandwidth and delay, the link quality factor Q of each link in the network topology is calculated: Q=0.7*bandwidth+0.3*(1-delay); based on the link quality factor Q, the topology structure or transmission priority of the network topology is adjusted, that is, the device connection is reconstructed based on the improved Prim algorithm, and the minimum spanning tree is dynamically generated; the Prim algorithm is a classic algorithm for finding the minimum spanning tree in a weighted connected graph, which is based on the idea of ​​a greedy algorithm and constructs a spanning tree by gradually selecting the edge with the minimum weight;

[0124] The bandwidth adjustment sub-strategy is specifically as follows: calculating the bandwidth allocation ratio based on the DDPG algorithm; the DDPG (Deep Deterministic Policy Gradient) algorithm is an algorithm based on deep reinforcement learning, which is suitable for solving problems in continuous action spaces and combines deterministic strategies with a model-independent reinforcement learning algorithm of deep neural networks; the calculation of the bandwidth allocation ratio is realized by the DDPG algorithm, which can effectively deal with the resource allocation problem in a dynamic network environment and show good performance in a variety of scenarios;

[0125] The edge computing device synchronizes the network resource optimization strategy, network indicator data, and the hardware configuration of the local device to the cloud server through the virtual mapping system. The cloud server verifies the network resource optimization strategy based on the digital twin technology, generates the verification result, and feeds it back to the edge computing device. In real time, the cloud server can use the Monte Carlo method to simulate extreme load scenarios (such as 200% burst traffic).

[0126] The edge computing device executes the network resource optimization strategy based on the verification result, calculates the QoE index after the execution of the network resource optimization strategy based on the QoE evaluation model, and performs a rollback operation of the network resource optimization strategy based on the QoE index. The QoE evaluation model is specifically:

[0127] QoE = 0.5*(1-freezing duration / total duration)+0.3*resolution level+0.2*operation response speed.

[0128] Calculate the link quality factor Q of each link in the network topology through bandwidth and latency, adjust the topology structure or transmission priority of the network topology based on the link quality factor Q, calculate the allocation ratio of bandwidth through the DDPG algorithm, verify the network resource optimization strategy through digital twin technology, and perform a fallback operation on the network resource optimization strategy based on the QoE index. That is, the edge computing device dynamically adjusts the network topology and bandwidth allocation. Before the adjustment, a stress test is carried out through digital twin technology. If it is determined that the optimization effect is not good based on the QoE index, a fallback operation can be automatically performed, greatly improving the flexibility and rationality of network resource allocation.

[0129] The cache rolling update module is specifically used for:

[0130] The edge computing device calculates the data importance Score of each network metric data based on the access count and the last access time:

[0131] Score = 0.6 * log(access count) + 0.4 * (1 / (current time - last access time));

[0132] The edge computing device sorts each network metric data based on the data importance Score at preset intervals and deletes the network metric data at the end with a preset ratio to perform a rolling update on the network metric data cached in the time series database.

[0133] In summary, the advantages of the present invention are as follows:

[0134] 1. Create a network state prediction model through the LSTM network, GRU network, feature fusion module, and output module, train the network state prediction model and deploy it to the edge computing device; the edge computing device collects network metric data based on the dynamic weight data collection mechanism and caches it in the time series database, inputs the network metric data into the network state prediction model to obtain the network state prediction result, generates a network resource optimization strategy based on the network state prediction result, verifies the network resource optimization strategy through digital twin technology, executes the network resource optimization strategy, and performs a rolling update on the network metric data cached in the time series database based on the access count and the last access time; that is, collect network metric data based on the dynamic weight data collection mechanism to adapt to the dynamically changing network load. For example, when sudden traffic is detected, shorten the collection period of bandwidth and packet loss rate to increase the corresponding data volume for subsequent analysis; the network resource optimization strategy is also dynamically updated based on the network state prediction result to make the strategy fit the actual network state and dynamically optimize the network topology and bandwidth; and the caching of network metric data combines the access count and the last access time (heat - timeliness two - dimensional elimination algorithm), effectively improving the cache hit rate, and ultimately greatly improving the flexibility and rationality of network resource allocation, thereby greatly improving the system performance and efficiency.

[0135] 2. Create a network state prediction model through an LSTM network, a GRU network, a feature fusion module, and an output module. Set the loss function of the network state prediction model as the mean squared error function. The LSTM network is used to capture the periodic features of the network state with a 72-hour period. The GRU network is used to capture the instantaneous features of the network state with a 30-minute period. The feature fusion module is used to fuse the periodic features and the instantaneous features and output the fused features. The output module is used to output the network state prediction results for the next 5 minutes. That is, the network state prediction model synchronously captures the periodic features and the instantaneous features through a dual-channel neural network and performs fusion, effectively improving the feature extraction ability. The mean squared error function punishes the square of the error and is sensitive to small errors, which is suitable for scenarios with high prediction accuracy requirements, thus greatly improving the accuracy of network state prediction.

[0136] 3. Divide the dataset into a training set, a validation set, and a test set. Train the network state prediction model with the training set until the loss value of the loss function is less than the preset loss threshold. During the training process, compress the network state prediction model through knowledge distillation technology and dynamic pruning technology. Calculate the prediction accuracy through the validation set to verify the trained network state prediction model. Calculate the confidence through the test set to test the network state prediction model that has passed the verification. If the test is successful, deploy the network state prediction model to edge computing devices of router, gateway, or camera types. That is, during the training process of the network state prediction model, continuous compression, verification, and testing are performed to effectively balance the model volume and prediction accuracy of the network state prediction model, so as to better deploy it on edge computing devices.

[0137] 4. During the network metric data collection process, dynamically adjust the weights of the collection periods of each network metric data through a Kalman filter. When facing sudden changes in network state or network load, more network metric data can be collected for analysis, so as to perform network resource optimization better and more timely, greatly improving the reliability and timeliness of network resource optimization.

[0138] 5. Calculate the link quality factor Q of each link in the network topology through bandwidth and delay. Adjust the topology structure or transmission priority of the network topology based on the link quality factor Q. Calculate the allocation ratio of bandwidth through the DDPG algorithm. Verify the network resource optimization strategy through digital twin technology. Perform a fallback operation on the network resource optimization strategy based on the QoE index. That is, the edge computing device dynamically adjusts the network topology and bandwidth allocation. Before the adjustment, perform a stress test through digital twin technology. If it is judged that the optimization effect is not good based on the QoE index, an automatic fallback operation can also be performed, greatly improving the flexibility and rationality of network resource allocation.

[0139] Although the specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments we described are illustrative only and not used to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered by the scope protected by the claims of the present invention.

Claims

1. A resource optimization method for routers, gateways, and cameras, characterized in that: The steps include: Step S1, creating a network state prediction model based on the LSTM network, the GRU network, the feature fusion module and the output module, training the network state prediction model, and deploying the trained network state prediction model to an edge computing device whose device type is a router, a gateway or a camera; Step S2: The edge computing device collects network indicator data based on a dynamic weight data collection mechanism, and caches each of the network indicator data into a time series database; Step S3: The edge computing device inputs the network indicator data into a network status prediction model to obtain a network status prediction result; Step S4: The edge computing device generates a network resource optimization strategy based on the network status prediction result, verifies the network resource optimization strategy based on the digital twin technology, and then executes the network resource optimization strategy; Step S5: The edge computing device performs rolling updates on the network indicator data cached in the time series database based on the number of accesses and the last access time.

2. A resource optimization method for routers, gateways, and cameras as claimed in claim 1, characterized in that: The step S1 is specifically as follows: A network state prediction model is created based on the LSTM network, the GRU network, the feature fusion module and the output module, and the loss function of the network state prediction model is set to be a mean square error function; the input end of the feature fusion module is respectively connected to the output end of the LSTM network and the output end of the GRU network, and the output end is connected to the input end of the output module; The LSTM network is used to capture the periodic features of the network state with a period of 72 hours; the GRU network is used to capture the instantaneous features of the network state with a period of 30 minutes; the feature fusion module is used to fuse the periodic features and the instantaneous features to output fusion features, and the feature fusion module dynamically allocates the weights of the periodic features and the instantaneous features based on the attention mechanism; the output module is used to output the network state prediction result for the next 5 minutes, and the network state prediction result is a healthy network state or a sub-healthy network state; Obtain a large amount of historical network indicator data including at least bandwidth, delay, packet loss rate, jitter, throughput, and utilization rate, perform data cleaning operations on each of the historical network indicator data, and then annotate the network status of each of the historical network indicator data to construct a data set; Based on a preset segmentation ratio, the data set is divided into a training set, a validation set and a test set. The network state prediction model is trained by the training set until the loss value of the loss function is less than a preset loss threshold. During the training process, the network state prediction model is compressed by the knowledge distillation technology and the dynamic pruning technology; the trained network state prediction model is verified by the validation set to determine whether the prediction accuracy is greater than the preset accuracy threshold. If not, the verification fails, and the training set is expanded to continue the training. If so, the verification succeeds; the successfully verified network state prediction model is tested by the test set to determine whether the confidence is greater than the preset confidence threshold. If not, the test fails, and the training set is expanded to continue the training. If so, the test succeeds and the training ends; The trained network status prediction model is deployed to an edge computing device whose device type is a router, gateway or camera.

3. A resource optimization method for routers, gateways, and cameras as claimed in claim 1, characterized in that: The step S2 is specifically as follows: The edge computing device uses lightweight distributed probes to collect network indicator data including at least bandwidth, delay, packet loss rate, jitter, throughput, and utilization according to a preset collection period. During the collection process, the Kalman filter is used to dynamically adjust the weight of the collection period of each network indicator data, and the collected network indicator data is cached in a time series database; The lightweight distributed probe is SkyWalking or Vine; the time series database is Redis; The step S3 is specifically as follows: The edge computing device performs data cleaning operations on the network indicator data and inputs the data into the network status prediction model in real time to obtain a network status prediction result.

4. A resource optimization method for routers, gateways, and cameras as claimed in claim 1, characterized in that: The step S4 is specifically as follows: The edge computing device analyzes the network status prediction result, and when the network status prediction result is sub-healthy, generates a network resource optimization strategy including a network topology adjustment sub-strategy and a bandwidth adjustment sub-strategy; The network topology adjustment sub-strategy is specifically: based on bandwidth and delay, the link quality factor Q of each link in the network topology is calculated: Q = 0.7*bandwidth+0.3*(1-delay), and the topology structure or transmission priority of the network topology is adjusted based on the link quality factor Q; the bandwidth adjustment sub-strategy is specifically: based on the DDPG algorithm, the bandwidth allocation ratio is calculated; The edge computing device synchronizes the network resource optimization strategy, network indicator data, and the hardware configuration of the local device to the cloud server through the virtual mapping system. The cloud server verifies the network resource optimization strategy based on the digital twin technology, generates the verification result, and feeds it back to the edge computing device. The edge computing device executes the network resource optimization strategy based on the verification result, calculates the QoE indicator after the execution of the network resource optimization strategy based on the QoE evaluation model, and performs a rollback operation of the network resource optimization strategy based on the QoE indicator.

5. A resource optimization method for routers, gateways, and cameras as claimed in claim 1, characterized in that: The step S5 is specifically as follows: The edge computing device calculates the data importance score of each network indicator data based on the number of accesses and the last access time: Score = 0.6*log(number of visits)+0.4*(1 / (current time-last visit time)); The edge computing device sorts each network indicator data based on the data importance Score at a preset time interval, and deletes the network indicator data of the last preset proportion, so as to perform a rolling update on the network indicator data cached in the time series database.

6. A resource optimization system for routers, gateways, and cameras, characterized in that: Includes the following modules: A network status prediction model deployment module is used to create a network status prediction model based on an LSTM network, a GRU network, a feature fusion module, and an output module, train the network status prediction model, and deploy the trained network status prediction model to an edge computing device whose device type is a router, a gateway, or a camera; A network indicator data collection module is used for edge computing devices to collect network indicator data based on a dynamic weight data collection mechanism, and cache each of the network indicator data into a time series database; A network status prediction module, used for the edge computing device to input the network indicator data into a network status prediction model to obtain a network status prediction result; A network resource optimization module, used for the edge computing device to generate a network resource optimization strategy based on the network status prediction result, and to execute the network resource optimization strategy after verifying the network resource optimization strategy based on the digital twin technology; The cache rolling update module is used for the edge computing device to roll over the network indicator data cached in the time series database based on the number of accesses and the last access time.

7. A resource optimization system for routers, gateways, and cameras as claimed in claim 6, characterized in that: The network status prediction model deployment module is specifically used for: A network state prediction model is created based on the LSTM network, the GRU network, the feature fusion module and the output module, and the loss function of the network state prediction model is set to be a mean square error function; the input end of the feature fusion module is respectively connected to the output end of the LSTM network and the output end of the GRU network, and the output end is connected to the input end of the output module; The LSTM network is used to capture the periodic features of the network state with a period of 72 hours; the GRU network is used to capture the instantaneous features of the network state with a period of 30 minutes; the feature fusion module is used to fuse the periodic features and the instantaneous features to output fusion features, and the feature fusion module dynamically allocates the weights of the periodic features and the instantaneous features based on the attention mechanism; the output module is used to output the network state prediction result for the next 5 minutes, and the network state prediction result is a healthy network state or a sub-healthy network state; Obtain a large amount of historical network indicator data including at least bandwidth, delay, packet loss rate, jitter, throughput, and utilization rate, perform data cleaning operations on each of the historical network indicator data, and then annotate the network status of each of the historical network indicator data to construct a data set; Based on a preset segmentation ratio, the data set is divided into a training set, a validation set and a test set. The network state prediction model is trained by the training set until the loss value of the loss function is less than a preset loss threshold. During the training process, the network state prediction model is compressed by the knowledge distillation technology and the dynamic pruning technology; the trained network state prediction model is verified by the validation set to determine whether the prediction accuracy is greater than the preset accuracy threshold. If not, the verification fails, and the training set is expanded to continue the training. If so, the verification succeeds; the successfully verified network state prediction model is tested by the test set to determine whether the confidence is greater than the preset confidence threshold. If not, the test fails, and the training set is expanded to continue the training. If so, the test succeeds and the training ends; The trained network status prediction model is deployed to an edge computing device whose device type is a router, gateway or camera.

8. A resource optimization system for routers, gateways, and cameras as claimed in claim 6, characterized in that: The network indicator data collection module is specifically used for: The edge computing device uses lightweight distributed probes to collect network indicator data including at least bandwidth, delay, packet loss rate, jitter, throughput, and utilization according to a preset collection period. During the collection process, the Kalman filter is used to dynamically adjust the weight of the collection period of each network indicator data, and the collected network indicator data is cached in a time series database; The lightweight distributed probe is SkyWalking or Vine; the time series database is Redis; The network status prediction module is specifically used for: The edge computing device performs data cleaning operations on the network indicator data and inputs the data into the network status prediction model in real time to obtain a network status prediction result.

9. A resource optimization system for routers, gateways, and cameras as claimed in claim 6, characterized in that: The network resource optimization module is specifically used for: The edge computing device analyzes the network status prediction result, and when the network status prediction result is sub-healthy, generates a network resource optimization strategy including a network topology adjustment sub-strategy and a bandwidth adjustment sub-strategy; The network topology adjustment sub-strategy is specifically: based on bandwidth and delay, the link quality factor Q of each link in the network topology is calculated: Q = 0.7*bandwidth+0.3*(1-delay), and the topology structure or transmission priority of the network topology is adjusted based on the link quality factor Q; the bandwidth adjustment sub-strategy is specifically: based on the DDPG algorithm, the bandwidth allocation ratio is calculated; The edge computing device synchronizes the network resource optimization strategy, network indicator data, and the hardware configuration of the local device to the cloud server through the virtual mapping system. The cloud server verifies the network resource optimization strategy based on the digital twin technology, generates the verification result, and feeds it back to the edge computing device. The edge computing device executes the network resource optimization strategy based on the verification result, calculates the QoE indicator after the execution of the network resource optimization strategy based on the QoE evaluation model, and performs a rollback operation of the network resource optimization strategy based on the QoE indicator.

10. A resource optimization system for routers, gateways, and cameras as claimed in claim 6, characterized in that: The cache rolling update module is specifically used for: The edge computing device calculates the data importance score of each network indicator data based on the number of accesses and the last access time: Score = 0.6*log(number of visits)+0.4*(1 / (current time-last visit time)); The edge computing device sorts each network indicator data based on the data importance Score at a preset time interval, and deletes the network indicator data of the last preset proportion, so as to perform a rolling update on the network indicator data cached in the time series database.

Images