A method and system for dynamic management of virtual machine network addresses based on neural network

Abstract

The present invention provides a neural network-based dynamic management method and system for virtual machine network addresses in the field of cloud computing and artificial intelligence. The method includes: step S1, creating a network address management model; step S2, collecting a large amount of historical virtual machine data to build a data set; step S3, training the network address management model using the data set; step S4, deploying the tested network address management model on a server; step S5, the server creates an address pool, sets virtual machine priorities, collects real-time virtual machine data, inputs the real-time virtual machine data into the deployed network address management model to obtain a network address management policy, and dynamically manages network addresses based on the network address management policy, address pool, and virtual machine priority. The present invention has the advantage of greatly improving the stability and reliability of the virtual machine network.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud computing and artificial intelligence, and in particular to a method and system for dynamically managing virtual machine network addresses based on a neural network. Background Art

[0002] A virtual machine is a computer system simulated through software, creating one or more virtual computer environments (virtual machines) on a physical computer (host). Developers can create different development environments within virtual machines for software development and testing; students and learners can install different operating systems and software within virtual machines for learning and experimentation; businesses can use virtual machines to consolidate server resources, improve resource utilization, and reduce hardware costs; and users can install different operating systems within virtual machines for entertainment, office work, and other purposes. Virtual machines have a wide range of applications.

[0003] Commercial virtual machines are typically deployed on servers. Computing resources are purchased from service providers to create virtual machines, which then call upon the server's computing resources to perform related computing tasks. When creating a virtual machine, the server must assign a network address to the virtual machine so that it can communicate with other devices (including other virtual machines, the host machine, and external network devices). Network addresses include, but are not limited to, IP addresses, subnet masks, default gateways, DNS server addresses, MAC addresses, port numbers, and broadcast addresses.

[0004] Traditionally, static allocation is used to allocate network addresses to virtual machines, where administrators manually assign fixed network addresses to virtual machines. This method is suitable for virtual machines that require long-term stable network connections, such as servers and databases. However, manual management is inefficient and prone to address conflicts (multiple virtual machines are assigned the same IP address). Therefore, dynamic address allocation (DHCP) came into being.

[0005] However, dynamic address allocation relies on static strategies (such as fixed address pool ranges and uniform lease durations) and cannot optimize allocation (adjust strategies) based on real-time network load, service type, or historical behavior. This leads to frequent IP address conflicts, especially in scenarios where elastic scaling of virtual machines lacks predictive capabilities, which in turn affects the stability and reliability of the virtual machine network. For example, during a promotional period, an e-commerce platform needed to elastically scale out hundreds of virtual machines to cope with traffic peaks. Due to the fixed DHCP address pool capacity, expansion could result in new VMs being unable to connect to the network due to address exhaustion. Meanwhile, during off-peak periods, a large number of addresses remain idle, resulting in resource utilization rates of less than 40%. In a hybrid deployment environment, video stream processing VMs require a low-latency network, while batch computing VMs have lower bandwidth requirements. Traditional DHCP uses the same allocation policy for all VMs, potentially assigning high-priority business VMs to high-latency network segments. In containerized development and testing environments, test VMs are typically created daily at a fixed time and released in batches after their tasks are completed. Because DHCP hasn't learned this periodicity, new VMs may be forced to allocate temporary addresses when they start due to address reclaim delays (e.g., lease expiration). This increases the risk of address conflicts with the original VMs by over 50%. In VM migration scenarios, if the original host fails to release the address promptly but the DHCP server on the target host reallocates it, active-active conflicts can occur, requiring manual intervention and resulting in mean time to recovery (MTTR) exceeding 15 minutes.

[0006] Therefore, how to provide a method and system for dynamic management of virtual machine network addresses based on neural networks to improve the stability and reliability of virtual machine networks has become a technical problem that needs to be solved urgently. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method and system for dynamic management of virtual machine network addresses based on neural network, so as to improve the stability and reliability of the virtual machine network.

[0008] In a first aspect, the present invention provides a method for dynamically managing virtual machine network addresses based on a neural network, comprising the following steps:

[0009] Step S1: creating a network address management model for outputting a network address management policy based on the address demand prediction module, the conflict risk perception module, and the adaptive strategy optimization module, and setting a loss function of the network address management model;

[0010] Step S2: Collect a large amount of historical virtual machine data, pre-process and annotate each of the historical virtual machine data to construct a data set; the historical virtual machine data at least includes virtual machine metadata, virtual machine life cycle events, service type, network traffic time series data, network status, and address allocation records;

[0011] Step S3, dividing the data set into a training set, a validation set, and a test set using a ten-fold cross-validation method, training the network address management model using the training set until the loss value of the loss function is less than a preset loss threshold, and then verifying and testing the network address management model using the validation set and the test set in turn;

[0012] Step S4: deploying the network address management model that has passed the test on a server;

[0013] Step S5: The server creates an address pool, sets virtual machine priorities, collects real-time virtual machine data, inputs the real-time virtual machine data into a deployed network address management model to obtain a network address management policy, and dynamically manages network addresses based on the network address management policy, the address pool, and the virtual machine priorities.

[0014] Step S6: Record the address management log in real time, continuously optimize the network address management model based on the address management log, and encrypt and store the address management log.

[0015] Furthermore, in step S1, the address demand prediction module is constructed based on a first input unit, an LSTM unit, an attention unit, and a first output unit, wherein the first input unit is used to input virtual machine data; the LSTM unit is used to extract long-term dependency features of address demand from the virtual machine data; the attention unit is used to extract attention features from the long-term dependency features based on time-step attention weights; and the first output unit is used to output address demand probability distribution based on the attention features through a fully connected layer.

[0016] The conflict risk perception module is constructed based on a second input unit, a graph convolution unit, a one-dimensional convolution neural unit, and a second output unit. The second input unit is used to input virtual machine data; the graph convolution unit is used to aggregate the neighbor node features of the virtual machine based on the virtual machine data to generate a topological embedding vector; the one-dimensional convolution neural unit is used to convolve the time series data in the virtual machine data to extract abnormal features; the second output unit is used to output the conflict node and the corresponding conflict probability of the conflict node based on the topological embedding vector and the abnormal features through a fully connected layer;

[0017] The adaptive strategy optimization module is constructed based on a third input unit, a processing unit, and a third output unit; the third input unit is used to input address demand probability distribution, conflicting nodes, and conflict probability; the processing unit is constructed based on an Actor network and a Critic network; the Actor network is constructed by a three-layer full connection and is used to output an action space probability distribution based on address demand probability distribution, conflicting nodes, and conflict probability; the Critic network is constructed by a two-layer full connection and is used to output a state-value function based on address demand probability distribution, conflicting nodes, and conflict probability; the third output unit is used to output a network address management strategy based on the action space probability distribution and the state-value function;

[0018] The formula of the loss function is:

[0019] L total =λ1*L pred +λ2*L risk +λ3*L policy +λ4*L reg ;

[0020] Among them, L total Represents the loss value of the loss function; L pred represents the address prediction loss, using the mean square error function; L risk Represents the conflict risk loss, using the weighted cross entropy function; the L policy Represents the strategy optimization loss, using the A2C function; the L reg represents the regularization loss; λ1, λ2, λ3 and λ4 all represent weight coefficients.

[0021] Furthermore, the step S2 is specifically as follows:

[0022] Collecting a large amount of historical virtual machine data, performing preprocessing on each of the historical virtual machine data, including at least data cleaning, data standardization, and data normalization, annotating each of the preprocessed historical virtual machine data with address requirements, address conflicts, and network address management policies, and constructing a data set based on the annotated historical virtual machine data;

[0023] The historical virtual machine data includes at least virtual machine metadata, virtual machine life cycle events, service types, network traffic time series data, network status, and address allocation records;

[0024] The virtual machine metadata includes at least CPU configuration, memory configuration, storage capacity configuration, virtual machine type, virtual machine purpose and network topology; the virtual machine life cycle event is virtual machine creation, virtual machine migration or virtual machine destruction; the business type is server hosting, development and testing environment, desktop virtualization, cloud computing service, high-performance computing, security and isolation, education and training, backup and disaster recovery, Internet of Things or containerized environment; the network traffic time series data includes at least timestamp, traffic size, number of data packets, data packet size distribution, transmission rate, protocol type, source address, destination address, delay and jitter, packet loss rate and QoS indicator; the network status includes at least network connectivity, network bandwidth, network delay, network congestion, packet loss rate, network device status, network configuration status and network availability; the address allocation record includes at least the allocated network address, allocation time, address usage duration and address release time.

[0025] Furthermore, the step S3 is specifically as follows:

[0026] The data set is divided into 10 data subsets by a ten-fold cross-validation method, and 8 of the 10 data subsets are selected as training sets, 1 as a validation set, and 1 as a test set in turn;

[0027] The network address management model is trained using the training set, and the model parameters, hyperparameters, regularization parameters, loss function parameters, learning rate scheduler and initialization parameters of the network address management model are continuously optimized during the training process until the loss value of the loss function is less than a preset loss threshold; the model parameters include at least weights and biases; the hyperparameters include at least learning rate, batch size, training rounds, network structure and optimizer; the regularization parameters include at least L2 regularization parameters, random dropout rate and data augmentation parameters; the learning rate scheduler includes a learning rate decay value and a cyclic learning rate; the initialization parameters include at least initialization weights and initialization biases;

[0028] The prediction accuracy is calculated using the validation set to validate the trained network address management model. If the validation fails, the training set is expanded to continue training. If the validation passes, then:

[0029] The reliability index is calculated through the test set to test the verified network address management model. If the test fails, the training set is expanded to continue training; if the test passes, the training is terminated.

[0030] Furthermore, the step S4 is specifically as follows:

[0031] After dynamically pruning and knowledge distilling the network address management model that has passed the test, the network address management model is encapsulated as a high-performance service and deployed on a server;

[0032] The step S5 is specifically as follows:

[0033] The server creates an address pool with a preset number of network addresses, divides the network addresses in the address pool into regular addresses and backup addresses, and sets virtual machine priorities based on the service type;

[0034] The server collects real-time virtual machine data, inputs the real-time virtual machine data into a deployed network address management model, and obtains a network address management policy including at least a network address allocation plan, lease adjustment parameters, and conflict avoidance operations;

[0035] The server determines whether the regular addresses in the address pool can meet the allocation requirements based on the network address management policy. If so, the server performs a network address allocation operation based on each of the regular addresses. If not, the server:

[0036] Further determine whether the backup addresses in the address pool can meet the allocation requirements. If so, perform a network address allocation operation based on each of the backup addresses; if not, then:

[0037] Based on the virtual machine priority and the IP address drift mechanism, the network address of the low-priority virtual machine is allocated to the high-priority virtual machine for use;

[0038] The step S6 is specifically as follows:

[0039] The server records in real time an address management log including at least real-time virtual machine data, network address management policies, and network address adjustment records, and continuously optimizes the network address management model based on the address management log;

[0040] The server encrypts the address management log into first-level encrypted data using the RSA algorithm, converts the first-level encrypted data into binary data, performs XOR on the binary data using a preset key to obtain second-level encrypted data, encrypts the second-level encrypted data into an encrypted log using the 3DES algorithm, and stores and distributes the encrypted log.

[0041] In a second aspect, the present invention provides a virtual machine network address dynamic management system based on a neural network, comprising the following modules:

[0042] A network address management model creation module is used to create a network address management model for outputting a network address management policy based on the address demand prediction module, the conflict risk perception module, and the adaptive strategy optimization module, and to set a loss function for the network address management model;

[0043] A data set construction module is used to collect a large amount of historical virtual machine data, pre-process and annotate each of the historical virtual machine data to construct a data set; the historical virtual machine data includes at least virtual machine metadata, virtual machine life cycle events, service types, network traffic time series data, network status and address allocation records;

[0044] a network address management model training module, configured to divide the data set into a training set, a validation set, and a test set using a ten-fold cross-validation method, train the network address management model using the training set until the loss value of the loss function is less than a preset loss threshold, and then validate and test the network address management model using the validation set and the test set in turn;

[0045] A network address management model deployment module, configured to deploy the network address management model that has passed the test on a server;

[0046] A network address management module is used to create an address pool on the server, set virtual machine priorities, collect real-time virtual machine data, input the real-time virtual machine data into the deployed network address management model to obtain a network address management policy, and dynamically manage network addresses based on the network address management policy, address pool, and virtual machine priorities;

[0047] The address management log management module is used to record the address management log in real time, continuously optimize the network address management model based on the address management log, and encrypt and store the address management log.

[0048] Furthermore, in the network address management model creation module, the address demand prediction module is constructed based on a first input unit, an LSTM unit, an attention unit, and a first output unit. The first input unit is used to input virtual machine data; the LSTM unit is used to extract long-term dependency features of address demand from virtual machine data; the attention unit is used to extract attention features from long-term dependency features based on time-step attention weights; and the first output unit is used to output address demand probability distribution based on the attention features through a fully connected layer.

[0049] The conflict risk perception module is constructed based on a second input unit, a graph convolution unit, a one-dimensional convolution neural unit, and a second output unit. The second input unit is used to input virtual machine data; the graph convolution unit is used to aggregate the neighbor node features of the virtual machine based on the virtual machine data to generate a topological embedding vector; the one-dimensional convolution neural unit is used to convolve the time series data in the virtual machine data to extract abnormal features; the second output unit is used to output the conflict node and the corresponding conflict probability of the conflict node based on the topological embedding vector and the abnormal features through a fully connected layer;

[0050] The adaptive strategy optimization module is constructed based on a third input unit, a processing unit, and a third output unit; the third input unit is used to input address demand probability distribution, conflicting nodes, and conflict probability; the processing unit is constructed based on an Actor network and a Critic network; the Actor network is constructed by a three-layer full connection and is used to output an action space probability distribution based on address demand probability distribution, conflicting nodes, and conflict probability; the Critic network is constructed by a two-layer full connection and is used to output a state-value function based on address demand probability distribution, conflicting nodes, and conflict probability; the third output unit is used to output a network address management strategy based on the action space probability distribution and the state-value function;

[0051] The formula of the loss function is:

[0052] L total =λ1*L pred +λ2*L risk +λ3*L policy +λ4*L reg ;

[0053] Among them, L total Represents the loss value of the loss function; L pred represents the address prediction loss, using the mean square error function; L risk Represents the conflict risk loss, using the weighted cross entropy function; the L policy Represents the strategy optimization loss, using the A2C function; the L reg represents the regularization loss; λ1, λ2, λ3 and λ4 all represent weight coefficients.

[0054] Furthermore, the dataset construction module is specifically used to:

[0055] Collecting a large amount of historical virtual machine data, performing preprocessing on each of the historical virtual machine data, including at least data cleaning, data standardization, and data normalization, annotating each of the preprocessed historical virtual machine data with address requirements, address conflicts, and network address management policies, and constructing a data set based on the annotated historical virtual machine data;

[0056] The historical virtual machine data includes at least virtual machine metadata, virtual machine life cycle events, service types, network traffic time series data, network status, and address allocation records;

[0057] The virtual machine metadata includes at least CPU configuration, memory configuration, storage capacity configuration, virtual machine type, virtual machine purpose and network topology; the virtual machine life cycle event is virtual machine creation, virtual machine migration or virtual machine destruction; the business type is server hosting, development and testing environment, desktop virtualization, cloud computing service, high-performance computing, security and isolation, education and training, backup and disaster recovery, Internet of Things or containerized environment; the network traffic time series data includes at least timestamp, traffic size, number of data packets, data packet size distribution, transmission rate, protocol type, source address, destination address, delay and jitter, packet loss rate and QoS indicator; the network status includes at least network connectivity, network bandwidth, network delay, network congestion, packet loss rate, network device status, network configuration status and network availability; the address allocation record includes at least the allocated network address, allocation time, address usage duration and address release time.

[0058] Furthermore, the network address management model training module is specifically used to:

[0059] The data set is divided into 10 data subsets by a ten-fold cross-validation method, and 8 of the 10 data subsets are selected as training sets, 1 as a validation set, and 1 as a test set in turn;

[0060] The network address management model is trained using the training set, and the model parameters, hyperparameters, regularization parameters, loss function parameters, learning rate scheduler and initialization parameters of the network address management model are continuously optimized during the training process until the loss value of the loss function is less than a preset loss threshold; the model parameters include at least weights and biases; the hyperparameters include at least learning rate, batch size, training rounds, network structure and optimizer; the regularization parameters include at least L2 regularization parameters, random dropout rate and data augmentation parameters; the learning rate scheduler includes a learning rate decay value and a cyclic learning rate; the initialization parameters include at least initialization weights and initialization biases;

[0061] The prediction accuracy is calculated using the validation set to validate the trained network address management model. If the validation fails, the training set is expanded to continue training. If the validation passes, then:

[0062] The reliability index is calculated through the test set to test the verified network address management model. If the test fails, the training set is expanded to continue training; if the test passes, the training is terminated.

[0063] Furthermore, the network address management model deployment module is specifically used to:

[0064] After dynamically pruning and knowledge distilling the network address management model that has passed the test, the network address management model is encapsulated as a high-performance service and deployed on a server;

[0065] The network address management module is specifically used for:

[0066] The server creates an address pool with a preset number of network addresses, divides the network addresses in the address pool into regular addresses and backup addresses, and sets virtual machine priorities based on the service type;

[0067] The server collects real-time virtual machine data, inputs the real-time virtual machine data into a deployed network address management model, and obtains a network address management policy including at least a network address allocation plan, lease adjustment parameters, and conflict avoidance operations;

[0068] The server determines whether the regular addresses in the address pool can meet the allocation requirements based on the network address management policy. If so, the server performs a network address allocation operation based on each of the regular addresses. If not, the server:

[0069] Further determine whether the backup addresses in the address pool can meet the allocation requirements. If so, perform a network address allocation operation based on each of the backup addresses; if not, then:

[0070] Based on the virtual machine priority and the IP address drift mechanism, the network address of the low-priority virtual machine is allocated to the high-priority virtual machine for use;

[0071] The address management log management module is specifically used to:

[0072] The server records in real time an address management log including at least real-time virtual machine data, network address management policies, and network address adjustment records, and continuously optimizes the network address management model based on the address management log;

[0073] The server encrypts the address management log into first-level encrypted data using the RSA algorithm, converts the first-level encrypted data into binary data, performs XOR on the binary data using a preset key to obtain second-level encrypted data, encrypts the second-level encrypted data into an encrypted log using the 3DES algorithm, and stores and distributes the encrypted log.

[0074] The advantages of the present invention are:

[0075] 1. Create a network address management model for outputting network address management strategies through the address demand prediction module, the conflict risk perception module, and the adaptive strategy optimization module, and set the loss function of the network address management model; then collect a large amount of historical virtual machine data, pre-process and annotate each historical virtual machine data to construct a data set; the historical virtual machine data at least includes virtual machine metadata, virtual machine life cycle events, business types, network traffic time series data, network status, and address allocation records; then divide the data set into a training set, a validation set, and a test set through the ten-fold cross-validation method, train the network address management model through the training set until the loss value of the loss function is less than the preset loss threshold, and then verify and test the network address management model through the validation set and the test set in turn; then deploy the network address management model that has passed the test on the server; the server creates an address pool , set virtual machine priority, collect real-time virtual machine data, input real-time virtual machine data into the deployed network address management model to obtain the network address management strategy, and dynamically manage the network address based on the network address management strategy, address pool and virtual machine priority; record the address management log in real time, continuously optimize the network address management model based on the address management log, and encrypt and store the address management log; that is, manage the network address through the pre-trained network address management model, which is constructed by the address demand prediction module, the conflict risk perception module and the adaptive strategy optimization module. It can dynamically adjust the network address management strategy based on the address demand and conflict risk, and fully consider the real-time network load, business type and historical behavior in combination with the virtual machine priority, so as to avoid network problems caused by network address allocation as much as possible, and ultimately greatly improve the stability and reliability of the virtual machine network.

[0076] 2. The address demand prediction module (LSTM + attention mechanism) can capture long-term temporal dependencies and focus on key time points, improving demand forecast accuracy. The conflict risk perception module (graph convolution + one-dimensional convolution) combines topological structure and temporal anomaly detection to enhance the dynamic perception of network conflicts. The adaptive policy optimization module (Actor-Critic reinforcement learning) achieves dynamic policy optimization, balances demand and risk, and improves the global optimality of network address management strategies. The clear division of labor and efficient collaboration between modules greatly improve the accuracy and stability of network address allocation, thereby greatly improving the stability and reliability of virtual machine networks.

[0077] 3. Loss function fusion address prediction (L pred ), conflict risk (L risk ), Strategy Optimization (L policy ) and regularization (L reg) four losses, and balance the optimization objectives of different tasks through weight coefficients (λ1, λ2, λ3, λ4), avoid overfitting of a single task, enhance the generalization ability of the network address management model, and meet the multi-objective requirements of dynamic management.

[0078] 4. By setting historical virtual machine data to cover virtual machine metadata, virtual machine life cycle events, business types, network traffic time series data, network status and address allocation records, the full-scene characteristics of network address management are covered; by setting the preprocessing of historical virtual machine data including data cleaning, data standardization and data normalization, the data quality is effectively improved; by marking address requirements, address conflicts and network address management strategies, the strong correlation labels required for supervised learning are constructed, thereby greatly improving the quality of the data set, supporting the network address management model to learn dynamic laws in complex scenarios, thereby greatly improving the rationality of the generation of network address management strategies.

[0079] 5. By using the ten-fold cross-validation method to divide the data set, the utilization of the data set is maximized and the risk of overfitting is reduced; through the training-validation-testing phase, the model performance of the network address management model is double-verified by accuracy and reliability indicators, thereby greatly improving the robustness of the network address management model and ensuring the reliability of the network address management model in the real environment.

[0080] 6. By dividing the network addresses in the address pool into regular addresses and backup addresses, regular addresses are allocated first, and backup addresses are enabled when resources are insufficient; virtual machine priorities are set based on business types, and address allocation is dynamically adjusted in combination with the IP drift mechanism, effectively improving the utilization of network addresses, ensuring the resource requirements of high-priority businesses, and reducing the probability of conflicts.

[0081] 7. After dynamic pruning and knowledge distillation of the tested network address management model, the network address management model is encapsulated as a high-performance service and deployed on the server, that is, the network address management model is deployed in a lightweight manner to meet low latency requirements. Combined with the collection of real-time virtual machine data input into the network address management model, a network address management strategy including network address allocation schemes, lease adjustment parameters, and conflict avoidance operations is obtained, realizing real-time adaptive management of network addresses with fast response speed, which is suitable for high-concurrency scenarios.

[0082] 8. Encrypt the address management log into first-level encrypted data using the RSA algorithm, convert the first-level encrypted data into binary data, perform XOR on the binary data using a preset key to obtain second-level encrypted data, encrypt the second-level encrypted data into an encrypted log using the 3DES algorithm, and store and distribute the encrypted log; that is, use the RSA algorithm, XOR operation, and 3DES triple encryption to ensure the security of address management log storage; and improve data disaster recovery capabilities through distributed backup.

[0083] 9. Through innovative composite neural network models, multi-dimensional data fusion, dynamic priority scheduling and secure encryption mechanisms, efficient, secure and adaptive management of virtual machine network addresses is achieved, with high-precision prediction, low conflict risk, strong scalability and continuous optimization capabilities.

[0084] 10. By dividing the address pool into regular addresses and backup addresses, regular addresses are used first and backup addresses are enabled only when resources are insufficient to avoid address waste; combined with the IP drift mechanism, the network addresses of low-priority virtual machines are dynamically recycled to improve the reuse rate of network addresses, thereby significantly reducing the idle rate of network address resources, which is suitable for large-scale virtual machine scenarios.

[0085] 11. The size of the network address management model is optimized through dynamic pruning (removing redundant neurons) and knowledge distillation (compressing the number of model parameters), effectively reducing server computing and storage overhead; after being deployed as a high-performance service, the response latency is low and high-concurrency requests are supported, thereby greatly reducing hardware resource consumption and greatly reducing operation and maintenance costs.

[0086] 12. Graph convolution units aggregate topological features, one-dimensional convolution units detect timing anomalies, and dual mechanisms identify potential address conflicts. The conflict probability output assists the model in avoiding high-risk allocation strategies in advance, thereby effectively reducing network failure rates and improving system stability.

[0087] 13. By recording and storing address management logs, if a new allocation strategy causes a conflict or performance degradation, it can be quickly rolled back to the last stable strategy based on the address management logs. By monitoring the network status in real time, trigger threshold alarms can be set and emergency plans can be automatically triggered (such as temporarily expanding the backup address pool), thereby effectively enhancing system fault tolerance and ensuring business continuity.

[0088] 14. Address management log encryption uses RSA (asymmetric encryption) + 3DES (symmetric encryption) combined algorithm, taking into account both efficiency and security.

[0089] 15. Continuously optimize model parameters through address management logs to form a closed loop of "data collection → model training → strategy deployment → effect feedback", effectively reducing the workload of manual parameter adjustment and realizing the long-term self-evolution of the network address management model. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0091] Figure 1 This is a flow chart of a method for dynamic management of virtual machine network addresses based on neural network of the present invention.

[0092] Figure 2It is a structural diagram of a neural network-based dynamic management system for virtual machine network addresses of the present invention. DETAILED DESCRIPTION

[0093] The technical solution in the embodiments of the present application has the following overall idea: network addresses are managed through a pre-trained network address management model. The network address management model is constructed by an address demand prediction module, a conflict risk perception module, and an adaptive strategy optimization module. The network address management strategy can be dynamically adjusted based on address demand and conflict risk. Combined with the virtual machine priority, it fully considers the real-time network load, business type, and historical behavior to avoid network problems caused by network address allocation as much as possible, thereby improving the stability and reliability of the virtual machine network.

[0094] Please refer to Figures 1 to 2 As shown, a preferred embodiment of a method for dynamic management of virtual machine network addresses based on a neural network of the present invention includes the following steps:

[0095] Step S1: creating a network address management model for outputting a network address management policy based on the address demand prediction module, the conflict risk perception module, and the adaptive strategy optimization module, and setting a loss function of the network address management model;

[0096] Step S2: Collect a large amount of historical virtual machine data, pre-process and annotate each of the historical virtual machine data to construct a data set; the historical virtual machine data at least includes virtual machine metadata, virtual machine life cycle events, service type, network traffic time series data, network status, and address allocation records;

[0097] Step S3, dividing the data set into a training set, a validation set, and a test set using a ten-fold cross-validation method, training the network address management model using the training set until the loss value of the loss function is less than a preset loss threshold, and then verifying and testing the network address management model using the validation set and the test set in turn;

[0098] Step S4: deploying the network address management model that has passed the test on a server;

[0099] Step S5: The server creates an address pool, sets virtual machine priorities, collects real-time virtual machine data, inputs the real-time virtual machine data into a deployed network address management model to obtain a network address management policy, and dynamically manages network addresses based on the network address management policy, the address pool, and the virtual machine priorities.

[0100] Step S6: Record the address management log in real time, continuously optimize the network address management model based on the address management log, and encrypt and store the address management log.

[0101] Through innovative composite neural network models, multi-dimensional data fusion, dynamic priority scheduling and secure encryption mechanisms, efficient, secure and adaptive management of virtual machine network addresses is achieved, with high-precision prediction, low conflict risk, strong scalability and continuous optimization capabilities.

[0102] In step S1, the address demand prediction module is constructed based on a first input unit, an LSTM unit, an attention unit, and a first output unit. The first input unit is used to input virtual machine data; the LSTM unit is used to extract long-term dependency features of address demand from the virtual machine data; the attention unit is used to extract attention features from the long-term dependency features based on time-step attention weights; and the first output unit is used to output address demand probability distribution based on the attention features through a fully connected layer.

[0103] The conflict risk perception module is constructed based on a second input unit, a graph convolution unit, a one-dimensional convolution neural unit, and a second output unit. The second input unit is used to input virtual machine data; the graph convolution unit is used to aggregate the neighbor node features of the virtual machine based on the virtual machine data to generate a topological embedding vector; the one-dimensional convolution neural unit is used to convolve the time series data in the virtual machine data to extract abnormal features; the second output unit is used to output the conflict node and the corresponding conflict probability of the conflict node based on the topological embedding vector and the abnormal features through a fully connected layer;

[0104] The graph convolution unit aggregates topological features, the one-dimensional convolution unit detects timing anomalies, and the dual mechanism identifies potential address conflicts; the conflict probability output assists the model in avoiding high-risk allocation strategies in advance, thereby effectively reducing network failure rates and improving system stability.

[0105] The adaptive strategy optimization module is constructed based on a third input unit, a processing unit, and a third output unit; the third input unit is used to input address demand probability distribution, conflicting nodes, and conflict probability; the processing unit is constructed based on an Actor network and a Critic network; the Actor network is constructed by a three-layer full connection and is used to output an action space probability distribution based on address demand probability distribution, conflicting nodes, and conflict probability; the Critic network is constructed by a two-layer full connection and is used to output a state-value function based on address demand probability distribution, conflicting nodes, and conflict probability; the third output unit is used to output a network address management strategy based on the action space probability distribution and the state-value function;

[0106] The address demand prediction module (LSTM + attention mechanism) can capture long-term temporal dependencies and focus on key time points, improving the accuracy of demand forecasting. The conflict risk perception module (graph convolution + one-dimensional convolution) combines topological structure and temporal anomaly detection to enhance the dynamic perception of network conflicts. The adaptive policy optimization module (Actor-Critic reinforcement learning) achieves dynamic optimization of policies, balances demand and risk, and improves the global optimality of network address management policies. The clear division of labor and efficient collaboration between modules greatly improve the accuracy and stability of network address allocation, thereby greatly improving the stability and reliability of virtual machine networks.

[0107] The formula of the loss function is:

[0108] L total =λ1*L pred +λ2*L risk +λ3*L policy +λ4*L reg ;

[0109] Among them, L total Represents the loss value of the loss function; L pred represents the address prediction loss, using the mean square error function; L risk Represents the conflict risk loss, using the weighted cross entropy function; the L policy Represents the strategy optimization loss, using the A2C function; the L reg represents the regularization loss; λ1, λ2, λ3 and λ4 all represent weight coefficients.

[0110] Loss function fusion address prediction (L pred ), conflict risk (L risk ), Strategy Optimization (L policy ) and regularization (L reg ) four losses, and balance the optimization objectives of different tasks through weight coefficients (λ1, λ2, λ3, λ4), avoid overfitting of a single task, enhance the generalization ability of the network address management model, and meet the multi-objective requirements of dynamic management.

[0111] The step S2 is specifically as follows:

[0112] Collecting a large amount of historical virtual machine data, performing preprocessing on each of the historical virtual machine data, including at least data cleaning, data standardization, and data normalization, annotating each of the preprocessed historical virtual machine data with address requirements, address conflicts, and network address management policies, and constructing a data set based on the annotated historical virtual machine data;

[0113] The historical virtual machine data includes at least virtual machine metadata, virtual machine life cycle events, service types, network traffic time series data, network status, and address allocation records;

[0114] By setting historical virtual machine data to cover virtual machine metadata, virtual machine lifecycle events, business types, network traffic time series data, network status and address allocation records, the full-scenario characteristics of network address management are covered; by setting the preprocessing of historical virtual machine data to include data cleaning, data standardization and data normalization, the data quality is effectively improved; by marking address requirements, address conflicts and network address management strategies, the strong correlation labels required for supervised learning are constructed, thereby greatly improving the quality of the data set, supporting the network address management model to learn dynamic laws in complex scenarios, thereby greatly improving the rationality of the generation of network address management strategies.

[0115] The virtual machine metadata includes at least CPU configuration, memory configuration, storage capacity configuration, virtual machine type, virtual machine purpose and network topology; the virtual machine life cycle event is virtual machine creation, virtual machine migration or virtual machine destruction; the business type is server hosting, development and testing environment, desktop virtualization, cloud computing service, high-performance computing, security and isolation, education and training, backup and disaster recovery, Internet of Things or containerized environment; the network traffic time series data includes at least timestamp, traffic size, number of data packets, data packet size distribution, transmission rate, protocol type, source address, destination address, delay and jitter, packet loss rate and QoS indicator; the network status includes at least network connectivity, network bandwidth, network delay, network congestion, packet loss rate, network device status, network configuration status and network availability; the address allocation record includes at least the allocated network address, allocation time, address usage duration and address release time.

[0116] The step S3 is specifically as follows:

[0117] The data set is divided into 10 data subsets by a ten-fold cross-validation method, and 8 of the 10 data subsets are selected as training sets, 1 as a validation set, and 1 as a test set in turn;

[0118] The network address management model is trained using the training set, and the model parameters, hyperparameters, regularization parameters, loss function parameters, learning rate scheduler and initialization parameters of the network address management model are continuously optimized during the training process until the loss value of the loss function is less than a preset loss threshold; the model parameters include at least weights and biases; the hyperparameters include at least learning rate, batch size, training rounds, network structure and optimizer; the regularization parameters include at least L2 regularization parameters, random dropout rate and data augmentation parameters; the learning rate scheduler includes a learning rate decay value and a cyclic learning rate; the initialization parameters include at least initialization weights and initialization biases;

[0119] The prediction accuracy is calculated using the validation set to validate the trained network address management model. If the validation fails, the training set is expanded to continue training. If the validation passes, then:

[0120] The reliability index is calculated through the test set to test the verified network address management model. If the test fails, the training set is expanded to continue training; if the test passes, the training is terminated.

[0121] By using the ten-fold cross-validation method to divide the data set, the utilization of the data set is maximized and the risk of overfitting is reduced; through the training-validation-testing phase, the model performance of the network address management model is doubly verified by accuracy and reliability indicators, thereby greatly improving the robustness of the network address management model and ensuring the reliability of the network address management model in the real environment.

[0122] The step S4 is specifically as follows:

[0123] After dynamically pruning and knowledge distilling the network address management model that has passed the test, the network address management model is encapsulated as a high-performance service and deployed on a server;

[0124] After dynamic pruning and knowledge distillation of the tested network address management model, the network address management model is encapsulated as a high-performance service and deployed on the server. That is, the network address management model is deployed in a lightweight manner to meet low latency requirements. Combined with the collection of real-time virtual machine data input into the network address management model, a network address management strategy including network address allocation schemes, lease adjustment parameters, and conflict avoidance operations is obtained, realizing real-time adaptive management of network addresses with fast response speed, which is suitable for high-concurrency scenarios.

[0125] By optimizing the size of the network address management model through dynamic pruning (removing redundant neurons) and knowledge distillation (compressing the number of model parameters), server computing and storage overhead are effectively reduced. By deploying it as a high-performance service, the response latency is low and high-concurrency requests are supported, which greatly reduces hardware resource consumption and greatly reduces operation and maintenance costs.

[0126] The step S5 is specifically as follows:

[0127] The server creates an address pool with a preset number of network addresses, divides the network addresses in the address pool into regular addresses and backup addresses, and sets virtual machine priorities based on the service type;

[0128] The server collects real-time virtual machine data, inputs the real-time virtual machine data into a deployed network address management model, and obtains a network address management policy including at least a network address allocation plan, lease adjustment parameters, and conflict avoidance operations;

[0129] The server determines whether the regular addresses in the address pool can meet the allocation requirements based on the network address management policy. If so, the server performs a network address allocation operation based on each of the regular addresses. If not, the server:

[0130] Further determine whether the backup addresses in the address pool can meet the allocation requirements. If so, perform a network address allocation operation based on each of the backup addresses; if not, then:

[0131] Based on the virtual machine priority and the IP address drift mechanism, the network address of the low-priority virtual machine is allocated to the high-priority virtual machine for use;

[0132] By dividing the network addresses in the address pool into regular addresses and backup addresses, regular addresses are allocated first, and backup addresses are enabled when resources are insufficient. By setting virtual machine priorities based on business types and dynamically adjusting address allocation in combination with the IP drift mechanism, the utilization rate of network addresses is effectively improved, the resource requirements of high-priority businesses are guaranteed, and the probability of conflicts is reduced.

[0133] By dividing the address pool into regular addresses and backup addresses, regular addresses are used first and backup addresses are enabled only when resources are insufficient to avoid address waste. The IP drift mechanism is combined to dynamically recycle the network addresses of low-priority virtual machines, thereby improving the reuse rate of network addresses and significantly reducing the idle rate of network address resources. It is suitable for large-scale virtual machine scenarios.

[0134] The step S6 is specifically as follows:

[0135] The server records in real time an address management log including at least real-time virtual machine data, network address management policies, and network address adjustment records, and continuously optimizes the network address management model based on the address management log;

[0136] By recording and storing address management logs, if a new allocation strategy causes a conflict or performance degradation, the system can quickly roll back to the previous stable strategy based on the address management logs. By monitoring the network status in real time, trigger threshold alarms can be set and emergency plans can be automatically triggered (such as temporarily expanding the backup address pool), thereby effectively enhancing system fault tolerance and ensuring business continuity.

[0137] The server encrypts the address management log into first-level encrypted data using the RSA algorithm, converts the first-level encrypted data into binary data, performs XOR on the binary data using a preset key to obtain second-level encrypted data, encrypts the second-level encrypted data into an encrypted log using the 3DES algorithm, and stores and distributes the encrypted log.

[0138] The address management log is encrypted into first-level encrypted data using the RSA algorithm, the first-level encrypted data is converted into binary data, the binary data is XORed using a preset key to obtain second-level encrypted data, the second-level encrypted data is encrypted into an encrypted log using the 3DES algorithm, and the encrypted log is stored and distributedly backed up; that is, the RSA algorithm, XOR operation, and 3DES triple encryption are used to ensure the security of address management log storage; and distributed backup is used to improve data disaster recovery capabilities.

[0139] The address management log encryption adopts the RSA (asymmetric encryption) + 3DES (symmetric encryption) combined algorithm, taking into account both efficiency and security.

[0140] By continuously optimizing model parameters through address management logs, a closed loop of "data collection → model training → strategy deployment → effect feedback" is formed, effectively reducing the workload of manual parameter adjustment and realizing the long-term self-evolution of the network address management model.

[0141] A preferred embodiment of a virtual machine network address dynamic management system based on a neural network of the present invention includes the following modules:

[0142] A network address management model creation module is used to create a network address management model for outputting a network address management policy based on the address demand prediction module, the conflict risk perception module, and the adaptive strategy optimization module, and to set a loss function for the network address management model;

[0143] A data set construction module is used to collect a large amount of historical virtual machine data, pre-process and annotate each of the historical virtual machine data to construct a data set; the historical virtual machine data includes at least virtual machine metadata, virtual machine life cycle events, service types, network traffic time series data, network status and address allocation records;

[0144] a network address management model training module, configured to divide the data set into a training set, a validation set, and a test set using a ten-fold cross-validation method, train the network address management model using the training set until the loss value of the loss function is less than a preset loss threshold, and then validate and test the network address management model using the validation set and the test set in turn;

[0145] A network address management model deployment module, configured to deploy the network address management model that has passed the test on a server;

[0146] A network address management module is used to create an address pool on the server, set virtual machine priorities, collect real-time virtual machine data, input the real-time virtual machine data into the deployed network address management model to obtain a network address management policy, and dynamically manage network addresses based on the network address management policy, address pool, and virtual machine priorities;

[0147] The address management log management module is used to record the address management log in real time, continuously optimize the network address management model based on the address management log, and encrypt and store the address management log.

[0148] Through innovative composite neural network models, multi-dimensional data fusion, dynamic priority scheduling and secure encryption mechanisms, efficient, secure and adaptive management of virtual machine network addresses is achieved, with high-precision prediction, low conflict risk, strong scalability and continuous optimization capabilities.

[0149] In the network address management model creation module, the address demand prediction module is constructed based on a first input unit, an LSTM unit, an attention unit, and a first output unit. The first input unit is used to input virtual machine data; the LSTM unit is used to extract long-term dependency features of address demand from virtual machine data; the attention unit is used to extract attention features from long-term dependency features based on time-step attention weights; and the first output unit is used to output address demand probability distribution based on the attention features through a fully connected layer.

[0150] The conflict risk perception module is constructed based on a second input unit, a graph convolution unit, a one-dimensional convolution neural unit, and a second output unit. The second input unit is used to input virtual machine data; the graph convolution unit is used to aggregate the neighbor node features of the virtual machine based on the virtual machine data to generate a topological embedding vector; the one-dimensional convolution neural unit is used to convolve the time series data in the virtual machine data to extract abnormal features; the second output unit is used to output the conflict node and the corresponding conflict probability of the conflict node based on the topological embedding vector and the abnormal features through a fully connected layer;

[0151] The graph convolution unit aggregates topological features, the one-dimensional convolution unit detects timing anomalies, and the dual mechanism identifies potential address conflicts; the conflict probability output assists the model in avoiding high-risk allocation strategies in advance, thereby effectively reducing network failure rates and improving system stability.

[0152] The adaptive strategy optimization module is constructed based on a third input unit, a processing unit, and a third output unit; the third input unit is used to input address demand probability distribution, conflicting nodes, and conflict probability; the processing unit is constructed based on an Actor network and a Critic network; the Actor network is constructed by a three-layer full connection and is used to output an action space probability distribution based on address demand probability distribution, conflicting nodes, and conflict probability; the Critic network is constructed by a two-layer full connection and is used to output a state-value function based on address demand probability distribution, conflicting nodes, and conflict probability; the third output unit is used to output a network address management strategy based on the action space probability distribution and the state-value function;

[0153] The address demand prediction module (LSTM + attention mechanism) can capture long-term temporal dependencies and focus on key time points, improving the accuracy of demand forecasting. The conflict risk perception module (graph convolution + one-dimensional convolution) combines topological structure and temporal anomaly detection to enhance the dynamic perception of network conflicts. The adaptive policy optimization module (Actor-Critic reinforcement learning) achieves dynamic optimization of policies, balances demand and risk, and improves the global optimality of network address management policies. The clear division of labor and efficient collaboration between modules greatly improve the accuracy and stability of network address allocation, thereby greatly improving the stability and reliability of virtual machine networks.

[0154] The formula of the loss function is:

[0155] L total =λ1*L pred +λ2*L risk +λ3*L policy +λ4*L reg ;

[0156] Among them, L total Represents the loss value of the loss function; L pred represents the address prediction loss, using the mean square error function; L risk Represents the conflict risk loss, using the weighted cross entropy function; the L policy Represents the strategy optimization loss, using the A2C function; the L reg represents the regularization loss; λ1, λ2, λ3 and λ4 all represent weight coefficients.

[0157] Loss function fusion address prediction (L pred ), conflict risk (Lrisk ), Strategy Optimization (L policy ) and regularization (L reg ) four losses, and balance the optimization objectives of different tasks through weight coefficients (λ1, λ2, λ3, λ4), avoid overfitting of a single task, enhance the generalization ability of the network address management model, and meet the multi-objective requirements of dynamic management.

[0158] The dataset construction module is specifically used for:

[0159] Collecting a large amount of historical virtual machine data, performing preprocessing on each of the historical virtual machine data, including at least data cleaning, data standardization, and data normalization, annotating each of the preprocessed historical virtual machine data with address requirements, address conflicts, and network address management policies, and constructing a data set based on the annotated historical virtual machine data;

[0160] The historical virtual machine data includes at least virtual machine metadata, virtual machine life cycle events, service types, network traffic time series data, network status, and address allocation records;

[0161] By setting historical virtual machine data to cover virtual machine metadata, virtual machine lifecycle events, business types, network traffic time series data, network status and address allocation records, the full-scenario characteristics of network address management are covered; by setting the preprocessing of historical virtual machine data to include data cleaning, data standardization and data normalization, the data quality is effectively improved; by marking address requirements, address conflicts and network address management strategies, the strong correlation labels required for supervised learning are constructed, thereby greatly improving the quality of the data set, supporting the network address management model to learn dynamic laws in complex scenarios, thereby greatly improving the rationality of the generation of network address management strategies.

[0162] The virtual machine metadata includes at least CPU configuration, memory configuration, storage capacity configuration, virtual machine type, virtual machine purpose and network topology; the virtual machine life cycle event is virtual machine creation, virtual machine migration or virtual machine destruction; the business type is server hosting, development and testing environment, desktop virtualization, cloud computing service, high-performance computing, security and isolation, education and training, backup and disaster recovery, Internet of Things or containerized environment; the network traffic time series data includes at least timestamp, traffic size, number of data packets, data packet size distribution, transmission rate, protocol type, source address, destination address, delay and jitter, packet loss rate and QoS indicator; the network status includes at least network connectivity, network bandwidth, network delay, network congestion, packet loss rate, network device status, network configuration status and network availability; the address allocation record includes at least the allocated network address, allocation time, address usage duration and address release time.

[0163] The network address management model training module is specifically used to:

[0164] The data set is divided into 10 data subsets by a ten-fold cross-validation method, and 8 of the 10 data subsets are selected as training sets, 1 as a validation set, and 1 as a test set in turn;

[0165] The network address management model is trained using the training set, and the model parameters, hyperparameters, regularization parameters, loss function parameters, learning rate scheduler and initialization parameters of the network address management model are continuously optimized during the training process until the loss value of the loss function is less than a preset loss threshold; the model parameters include at least weights and biases; the hyperparameters include at least learning rate, batch size, training rounds, network structure and optimizer; the regularization parameters include at least L2 regularization parameters, random dropout rate and data augmentation parameters; the learning rate scheduler includes a learning rate decay value and a cyclic learning rate; the initialization parameters include at least initialization weights and initialization biases;

[0166] The prediction accuracy is calculated using the validation set to validate the trained network address management model. If the validation fails, the training set is expanded to continue training. If the validation passes, then:

[0167] The reliability index is calculated through the test set to test the verified network address management model. If the test fails, the training set is expanded to continue training; if the test passes, the training is terminated.

[0168] By using the ten-fold cross-validation method to divide the data set, the utilization of the data set is maximized and the risk of overfitting is reduced; through the training-validation-testing phase, the model performance of the network address management model is doubly verified by accuracy and reliability indicators, thereby greatly improving the robustness of the network address management model and ensuring the reliability of the network address management model in the real environment.

[0169] The network address management model deployment module is specifically used to:

[0170] After dynamically pruning and knowledge distilling the network address management model that has passed the test, the network address management model is encapsulated as a high-performance service and deployed on a server;

[0171] After dynamic pruning and knowledge distillation of the tested network address management model, the network address management model is encapsulated as a high-performance service and deployed on the server. That is, the network address management model is deployed in a lightweight manner to meet low latency requirements. Combined with the collection of real-time virtual machine data input into the network address management model, a network address management strategy including network address allocation schemes, lease adjustment parameters, and conflict avoidance operations is obtained, realizing real-time adaptive management of network addresses with fast response speed, which is suitable for high-concurrency scenarios.

[0172] By optimizing the size of the network address management model through dynamic pruning (removing redundant neurons) and knowledge distillation (compressing the number of model parameters), server computing and storage overhead are effectively reduced. By deploying it as a high-performance service, the response latency is low and high-concurrency requests are supported, which greatly reduces hardware resource consumption and greatly reduces operation and maintenance costs.

[0173] The network address management module is specifically used for:

[0174] The server creates an address pool with a preset number of network addresses, divides the network addresses in the address pool into regular addresses and backup addresses, and sets virtual machine priorities based on the service type;

[0175] The server collects real-time virtual machine data, inputs the real-time virtual machine data into a deployed network address management model, and obtains a network address management policy including at least a network address allocation plan, lease adjustment parameters, and conflict avoidance operations;

[0176] The server determines whether the regular addresses in the address pool can meet the allocation requirements based on the network address management policy. If so, the server performs a network address allocation operation based on each of the regular addresses. If not, the server:

[0177] Further determine whether the backup addresses in the address pool can meet the allocation requirements. If so, perform a network address allocation operation based on each of the backup addresses; if not, then:

[0178] Based on the virtual machine priority and the IP address drift mechanism, the network address of the low-priority virtual machine is allocated to the high-priority virtual machine for use;

[0179] By dividing the network addresses in the address pool into regular addresses and backup addresses, regular addresses are allocated first, and backup addresses are enabled when resources are insufficient. By setting virtual machine priorities based on business types and dynamically adjusting address allocation in combination with the IP drift mechanism, the utilization rate of network addresses is effectively improved, the resource requirements of high-priority businesses are guaranteed, and the probability of conflicts is reduced.

[0180] By dividing the address pool into regular addresses and backup addresses, regular addresses are used first and backup addresses are enabled only when resources are insufficient to avoid address waste. The IP drift mechanism is combined to dynamically recycle the network addresses of low-priority virtual machines, thereby improving the reuse rate of network addresses and significantly reducing the idle rate of network address resources. It is suitable for large-scale virtual machine scenarios.

[0181] The address management log management module is specifically used to:

[0182] The server records in real time an address management log including at least real-time virtual machine data, network address management policies, and network address adjustment records, and continuously optimizes the network address management model based on the address management log;

[0183] By recording and storing address management logs, if a new allocation strategy causes a conflict or performance degradation, the system can quickly roll back to the previous stable strategy based on the address management logs. By monitoring the network status in real time, trigger threshold alarms can be set and emergency plans can be automatically triggered (such as temporarily expanding the backup address pool), thereby effectively enhancing system fault tolerance and ensuring business continuity.

[0184] The server encrypts the address management log into first-level encrypted data using the RSA algorithm, converts the first-level encrypted data into binary data, performs XOR on the binary data using a preset key to obtain second-level encrypted data, encrypts the second-level encrypted data into an encrypted log using the 3DES algorithm, and stores and distributes the encrypted log.

[0185] The address management log is encrypted into first-level encrypted data using the RSA algorithm, the first-level encrypted data is converted into binary data, the binary data is XORed using a preset key to obtain second-level encrypted data, the second-level encrypted data is encrypted into an encrypted log using the 3DES algorithm, and the encrypted log is stored and distributedly backed up; that is, the RSA algorithm, XOR operation, and 3DES triple encryption are used to ensure the security of address management log storage; and distributed backup is used to improve data disaster recovery capabilities.

[0186] The address management log encryption adopts the RSA (asymmetric encryption) + 3DES (symmetric encryption) combined algorithm, taking into account both efficiency and security.

[0187] By continuously optimizing model parameters through address management logs, a closed loop of "data collection → model training → strategy deployment → effect feedback" is formed, effectively reducing the workload of manual parameter adjustment and realizing the long-term self-evolution of the network address management model.

[0188] In summary, the advantages of the present invention are:

[0189] 1. Create a network address management model for outputting network address management strategies through the address demand prediction module, the conflict risk perception module, and the adaptive strategy optimization module, and set the loss function of the network address management model; then collect a large amount of historical virtual machine data, pre-process and annotate each historical virtual machine data to construct a data set; the historical virtual machine data at least includes virtual machine metadata, virtual machine life cycle events, business types, network traffic time series data, network status, and address allocation records; then divide the data set into a training set, a validation set, and a test set through the ten-fold cross-validation method, train the network address management model through the training set until the loss value of the loss function is less than the preset loss threshold, and then verify and test the network address management model through the validation set and the test set in turn; then deploy the network address management model that has passed the test on the server; the server creates an address pool , set virtual machine priority, collect real-time virtual machine data, input real-time virtual machine data into the deployed network address management model to obtain the network address management strategy, and dynamically manage the network address based on the network address management strategy, address pool and virtual machine priority; record the address management log in real time, continuously optimize the network address management model based on the address management log, and encrypt and store the address management log; that is, manage the network address through the pre-trained network address management model, which is constructed by the address demand prediction module, the conflict risk perception module and the adaptive strategy optimization module. It can dynamically adjust the network address management strategy based on the address demand and conflict risk, and fully consider the real-time network load, business type and historical behavior in combination with the virtual machine priority, so as to avoid network problems caused by network address allocation as much as possible, and ultimately greatly improve the stability and reliability of the virtual machine network.

[0190] 2. The address demand prediction module (LSTM + attention mechanism) can capture long-term temporal dependencies and focus on key time points, improving demand forecast accuracy. The conflict risk perception module (graph convolution + one-dimensional convolution) combines topological structure and temporal anomaly detection to enhance the dynamic perception of network conflicts. The adaptive policy optimization module (Actor-Critic reinforcement learning) achieves dynamic policy optimization, balances demand and risk, and improves the global optimality of network address management strategies. The clear division of labor and efficient collaboration between modules greatly improve the accuracy and stability of network address allocation, thereby greatly improving the stability and reliability of virtual machine networks.

[0191] 3. Loss function fusion address prediction (L pred ), conflict risk (L risk ), Strategy Optimization (L policy ) and regularization (L reg) four losses, and balance the optimization objectives of different tasks through weight coefficients (λ1, λ2, λ3, λ4), avoid overfitting of a single task, enhance the generalization ability of the network address management model, and meet the multi-objective requirements of dynamic management.

[0192] 4. By setting historical virtual machine data to cover virtual machine metadata, virtual machine life cycle events, business types, network traffic time series data, network status and address allocation records, the full-scene characteristics of network address management are covered; by setting the preprocessing of historical virtual machine data including data cleaning, data standardization and data normalization, the data quality is effectively improved; by marking address requirements, address conflicts and network address management strategies, the strong correlation labels required for supervised learning are constructed, thereby greatly improving the quality of the data set, supporting the network address management model to learn dynamic laws in complex scenarios, thereby greatly improving the rationality of the generation of network address management strategies.

[0193] 5. By using the ten-fold cross-validation method to divide the data set, the utilization of the data set is maximized and the risk of overfitting is reduced; through the training-validation-testing phase, the model performance of the network address management model is double-verified by accuracy and reliability indicators, thereby greatly improving the robustness of the network address management model and ensuring the reliability of the network address management model in the real environment.

[0194] 6. By dividing the network addresses in the address pool into regular addresses and backup addresses, regular addresses are allocated first, and backup addresses are enabled when resources are insufficient; virtual machine priorities are set based on business types, and address allocation is dynamically adjusted in combination with the IP drift mechanism, effectively improving the utilization of network addresses, ensuring the resource requirements of high-priority businesses, and reducing the probability of conflicts.

[0195] 7. After dynamic pruning and knowledge distillation of the tested network address management model, the network address management model is encapsulated as a high-performance service and deployed on the server, that is, the network address management model is deployed in a lightweight manner to meet low latency requirements. Combined with the collection of real-time virtual machine data input into the network address management model, a network address management strategy including network address allocation schemes, lease adjustment parameters, and conflict avoidance operations is obtained, realizing real-time adaptive management of network addresses with fast response speed, which is suitable for high-concurrency scenarios.

[0196] 8. Encrypt the address management log into first-level encrypted data using the RSA algorithm, convert the first-level encrypted data into binary data, perform XOR on the binary data using a preset key to obtain second-level encrypted data, encrypt the second-level encrypted data into an encrypted log using the 3DES algorithm, and store and distribute the encrypted log; that is, use the RSA algorithm, XOR operation, and 3DES triple encryption to ensure the security of address management log storage; and improve data disaster recovery capabilities through distributed backup.

[0197] 9. Through innovative composite neural network models, multi-dimensional data fusion, dynamic priority scheduling and secure encryption mechanisms, efficient, secure and adaptive management of virtual machine network addresses is achieved, with high-precision prediction, low conflict risk, strong scalability and continuous optimization capabilities.

[0198] 10. By dividing the address pool into regular addresses and backup addresses, regular addresses are used first and backup addresses are enabled only when resources are insufficient to avoid address waste; combined with the IP drift mechanism, the network addresses of low-priority virtual machines are dynamically recycled to improve the reuse rate of network addresses, thereby significantly reducing the idle rate of network address resources, which is suitable for large-scale virtual machine scenarios.

[0199] 11. The size of the network address management model is optimized through dynamic pruning (removing redundant neurons) and knowledge distillation (compressing the number of model parameters), effectively reducing server computing and storage overhead; after being deployed as a high-performance service, the response latency is low and high-concurrency requests are supported, thereby greatly reducing hardware resource consumption and greatly reducing operation and maintenance costs.

[0200] 12. Graph convolution units aggregate topological features, one-dimensional convolution units detect timing anomalies, and dual mechanisms identify potential address conflicts. The conflict probability output assists the model in avoiding high-risk allocation strategies in advance, thereby effectively reducing network failure rates and improving system stability.

[0201] 13. By recording and storing address management logs, if a new allocation strategy causes a conflict or performance degradation, it can be quickly rolled back to the last stable strategy based on the address management logs. By monitoring the network status in real time, trigger threshold alarms can be set and emergency plans can be automatically triggered (such as temporarily expanding the backup address pool), thereby effectively enhancing system fault tolerance and ensuring business continuity.

[0202] 14. Address management log encryption uses RSA (asymmetric encryption) + 3DES (symmetric encryption) combined algorithm, taking into account both efficiency and security.

[0203] 15. Continuously optimize model parameters through address management logs to form a closed loop of "data collection → model training → strategy deployment → effect feedback", effectively reducing the workload of manual parameter adjustment and realizing the long-term self-evolution of the network address management model.

[0204] Although the specific embodiments of the present invention are described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and are not intended to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for dynamic management of virtual machine network addresses based on a neural network, characterized by: The steps include: Step S1: Create a network address management model for outputting a network address management policy based on the address demand prediction module, the conflict risk perception module, and the adaptive strategy optimization module, and set a loss function for the network address management model; the formula of the loss function is: L total =λ1*L pred +λ2*L risk +λ3*L policy +λ4*L reg ; Among them, L total Represents the loss value of the loss function; L pred represents the address prediction loss, using the mean square error function; L risk Represents the conflict risk loss, using the weighted cross entropy function; the L policy Represents the strategy optimization loss, using the A2C function; the L reg represents the regularization loss; λ1, λ2, λ3 and λ4 all represent weight coefficients; Step S2: Collect a large amount of historical virtual machine data, pre-process and annotate each of the historical virtual machine data to construct a data set; the historical virtual machine data at least includes virtual machine metadata, virtual machine life cycle events, service type, network traffic time series data, network status, and address allocation records; Step S3, dividing the data set into a training set, a validation set, and a test set using a ten-fold cross-validation method, training the network address management model using the training set until the loss value of the loss function is less than a preset loss threshold, and then verifying and testing the network address management model using the validation set and the test set in turn; Step S4: deploying the network address management model that has passed the test on a server; Step S5: The server creates an address pool, sets virtual machine priorities, collects real-time virtual machine data, inputs the real-time virtual machine data into a deployed network address management model to obtain a network address management policy, and dynamically manages network addresses based on the network address management policy, the address pool, and the virtual machine priorities. Step S6: Record the address management log in real time, continuously optimize the network address management model based on the address management log, and encrypt and store the address management log.

2. The method for dynamic management of virtual machine network addresses based on a neural network according to claim 1, wherein: In step S1, the address demand prediction module is constructed based on a first input unit, an LSTM unit, an attention unit, and a first output unit. The first input unit is used to input virtual machine data; the LSTM unit is used to extract long-term dependency features of address demand from the virtual machine data; the attention unit is used to extract attention features from the long-term dependency features based on time-step attention weights; and the first output unit is used to output address demand probability distribution based on the attention features through a fully connected layer. The conflict risk perception module is constructed based on a second input unit, a graph convolution unit, a one-dimensional convolution neural unit, and a second output unit. The second input unit is used to input virtual machine data; the graph convolution unit is used to aggregate the neighbor node features of the virtual machine based on the virtual machine data to generate a topological embedding vector; the one-dimensional convolution neural unit is used to convolve the time series data in the virtual machine data to extract abnormal features; the second output unit is used to output the conflict node and the corresponding conflict probability of the conflict node based on the topological embedding vector and the abnormal features through a fully connected layer; The adaptive strategy optimization module is constructed based on a third input unit, a processing unit and a third output unit; the third input unit is used to input address demand probability distribution, conflict nodes and conflict probability; the processing unit is constructed based on the Actor network and the Critic network; the Actor network is constructed by 3 layers of full connection, and is used to output the action space probability distribution based on the address demand probability distribution, conflict nodes and conflict probability; the Critic network is constructed by 2 layers of full connection, and is used to output the state value function based on the address demand probability distribution, conflict nodes and conflict probability; the third output unit is used to output the network address management strategy based on the action space probability distribution and the state value function.

3. The method for dynamic management of virtual machine network addresses based on a neural network according to claim 1, wherein: The step S2 is specifically as follows: Collecting a large amount of historical virtual machine data, performing preprocessing on each of the historical virtual machine data, including at least data cleaning, data standardization, and data normalization, annotating each of the preprocessed historical virtual machine data with address requirements, address conflicts, and network address management policies, and constructing a data set based on the annotated historical virtual machine data; The historical virtual machine data includes at least virtual machine metadata, virtual machine life cycle events, service types, network traffic time series data, network status, and address allocation records; The virtual machine metadata includes at least CPU configuration, memory configuration, storage capacity configuration, virtual machine type, virtual machine purpose and network topology; the virtual machine life cycle event is virtual machine creation, virtual machine migration or virtual machine destruction; the business type is server hosting, development and testing environment, desktop virtualization, cloud computing service, high-performance computing, security and isolation, education and training, backup and disaster recovery, Internet of Things or containerized environment; the network traffic time series data includes at least timestamp, traffic size, number of data packets, data packet size distribution, transmission rate, protocol type, source address, destination address, delay and jitter, packet loss rate and QoS indicator; the network status includes at least network connectivity, network bandwidth, network delay, network congestion, packet loss rate, network device status, network configuration status and network availability; the address allocation record includes at least the allocated network address, allocation time, address usage duration and address release time.

4. The method for dynamic management of virtual machine network addresses based on a neural network according to claim 1, wherein: The step S3 is specifically as follows: The data set is divided into 10 data subsets by a ten-fold cross-validation method, and 8 of the 10 data subsets are selected as training sets, 1 as a validation set, and 1 as a test set in turn; The network address management model is trained using the training set, and the model parameters, hyperparameters, regularization parameters, loss function parameters, learning rate scheduler and initialization parameters of the network address management model are continuously optimized during the training process until the loss value of the loss function is less than a preset loss threshold; the model parameters include at least weights and biases; the hyperparameters include at least learning rate, batch size, training rounds, network structure and optimizer; the regularization parameters include at least L2 regularization parameters, random dropout rate and data augmentation parameters; the learning rate scheduler includes a learning rate decay value and a cyclic learning rate; the initialization parameters include at least initialization weights and initialization biases; The prediction accuracy is calculated using the validation set to validate the trained network address management model. If the validation fails, the training set is expanded to continue training. If the validation passes, then: The reliability index is calculated through the test set to test the verified network address management model. If the test fails, the training set is expanded to continue training; if the test passes, the training is terminated.

5. The method for dynamic management of virtual machine network addresses based on a neural network according to claim 1, wherein: The step S4 is specifically as follows: After dynamically pruning and knowledge distilling the network address management model that has passed the test, the network address management model is encapsulated as a high-performance service and deployed on a server; The step S5 is specifically as follows: The server creates an address pool with a preset number of network addresses, divides the network addresses in the address pool into regular addresses and backup addresses, and sets virtual machine priorities based on the service type; The server collects real-time virtual machine data, inputs the real-time virtual machine data into a deployed network address management model, and obtains a network address management policy including at least a network address allocation plan, lease adjustment parameters, and conflict avoidance operations; The server determines whether the regular addresses in the address pool can meet the allocation requirements based on the network address management policy. If so, the server performs a network address allocation operation based on each of the regular addresses. If not, the server: Further determine whether the backup addresses in the address pool can meet the allocation requirements. If so, perform a network address allocation operation based on each of the backup addresses; if not, then: Based on the virtual machine priority and the IP address drift mechanism, the network address of the low-priority virtual machine is allocated to the high-priority virtual machine for use; The step S6 is specifically as follows: The server records in real time an address management log including at least real-time virtual machine data, network address management policies, and network address adjustment records, and continuously optimizes the network address management model based on the address management log; The server encrypts the address management log into first-level encrypted data using the RSA algorithm, converts the first-level encrypted data into binary data, performs XOR on the binary data using a preset key to obtain second-level encrypted data, encrypts the second-level encrypted data into an encrypted log using the 3DES algorithm, and stores and distributes the encrypted log.

6. A virtual machine network address dynamic management system based on neural network, characterized by: Includes the following modules: The network address management model creation module is used to create a network address management model for outputting a network address management policy based on the address demand prediction module, the conflict risk perception module, and the adaptive strategy optimization module, and set a loss function for the network address management model; the formula of the loss function is: L total =λ1*L pred +λ2*L risk +λ3*L policy +λ4*L reg ; Among them, L total Represents the loss value of the loss function; L pred represents the address prediction loss, using the mean square error function; L risk Represents the conflict risk loss, using the weighted cross entropy function; the L policy Represents the strategy optimization loss, using the A2C function; the L reg represents the regularization loss; λ1, λ2, λ3 and λ4 all represent weight coefficients; A data set construction module is used to collect a large amount of historical virtual machine data, pre-process and annotate each of the historical virtual machine data to construct a data set; the historical virtual machine data includes at least virtual machine metadata, virtual machine life cycle events, service types, network traffic time series data, network status and address allocation records; a network address management model training module, configured to divide the data set into a training set, a validation set, and a test set using a ten-fold cross-validation method, train the network address management model using the training set until the loss value of the loss function is less than a preset loss threshold, and then validate and test the network address management model using the validation set and the test set in turn; A network address management model deployment module, configured to deploy the network address management model that has passed the test on a server; A network address management module is used to create an address pool on the server, set virtual machine priorities, collect real-time virtual machine data, input the real-time virtual machine data into the deployed network address management model to obtain a network address management policy, and dynamically manage network addresses based on the network address management policy, address pool, and virtual machine priorities; The address management log management module is used to record the address management log in real time, continuously optimize the network address management model based on the address management log, and encrypt and store the address management log.

7. The neural network-based dynamic management system for virtual machine network addresses according to claim 6, characterized in that: In the network address management model creation module, the address demand prediction module is constructed based on a first input unit, an LSTM unit, an attention unit, and a first output unit. The first input unit is used to input virtual machine data; the LSTM unit is used to extract long-term dependency features of address demand from virtual machine data; the attention unit is used to extract attention features from long-term dependency features based on time-step attention weights; and the first output unit is used to output address demand probability distribution based on the attention features through a fully connected layer. The conflict risk perception module is constructed based on a second input unit, a graph convolution unit, a one-dimensional convolution neural unit, and a second output unit. The second input unit is used to input virtual machine data; the graph convolution unit is used to aggregate the neighbor node features of the virtual machine based on the virtual machine data to generate a topological embedding vector; the one-dimensional convolution neural unit is used to convolve the time series data in the virtual machine data to extract abnormal features; the second output unit is used to output the conflict node and the corresponding conflict probability of the conflict node based on the topological embedding vector and the abnormal features through a fully connected layer; The adaptive strategy optimization module is constructed based on a third input unit, a processing unit and a third output unit; the third input unit is used to input address demand probability distribution, conflict nodes and conflict probability; the processing unit is constructed based on the Actor network and the Critic network; the Actor network is constructed by 3 layers of full connection, and is used to output the action space probability distribution based on the address demand probability distribution, conflict nodes and conflict probability; the Critic network is constructed by 2 layers of full connection, and is used to output the state value function based on the address demand probability distribution, conflict nodes and conflict probability; the third output unit is used to output the network address management strategy based on the action space probability distribution and the state value function.

8. The neural network-based dynamic management system for virtual machine network addresses according to claim 6, characterized in that: The dataset construction module is specifically used for: Collecting a large amount of historical virtual machine data, performing preprocessing on each of the historical virtual machine data, including at least data cleaning, data standardization, and data normalization, annotating each of the preprocessed historical virtual machine data with address requirements, address conflicts, and network address management policies, and constructing a data set based on the annotated historical virtual machine data; The historical virtual machine data includes at least virtual machine metadata, virtual machine life cycle events, service types, network traffic time series data, network status, and address allocation records; The virtual machine metadata includes at least CPU configuration, memory configuration, storage capacity configuration, virtual machine type, virtual machine purpose and network topology; the virtual machine life cycle event is virtual machine creation, virtual machine migration or virtual machine destruction; the business type is server hosting, development and testing environment, desktop virtualization, cloud computing service, high-performance computing, security and isolation, education and training, backup and disaster recovery, Internet of Things or containerized environment; the network traffic time series data includes at least timestamp, traffic size, number of data packets, data packet size distribution, transmission rate, protocol type, source address, destination address, delay and jitter, packet loss rate and QoS indicator; the network status includes at least network connectivity, network bandwidth, network delay, network congestion, packet loss rate, network device status, network configuration status and network availability; the address allocation record includes at least the allocated network address, allocation time, address usage duration and address release time.

9. The neural network-based dynamic management system for virtual machine network addresses according to claim 6, characterized in that: The network address management model training module is specifically used to: The data set is divided into 10 data subsets by a ten-fold cross-validation method, and 8 of the 10 data subsets are selected as training sets, 1 as a validation set, and 1 as a test set in turn; The network address management model is trained using the training set, and the model parameters, hyperparameters, regularization parameters, loss function parameters, learning rate scheduler and initialization parameters of the network address management model are continuously optimized during the training process until the loss value of the loss function is less than a preset loss threshold; the model parameters include at least weights and biases; the hyperparameters include at least learning rate, batch size, training rounds, network structure and optimizer; the regularization parameters include at least L2 regularization parameters, random dropout rate and data augmentation parameters; the learning rate scheduler includes a learning rate decay value and a cyclic learning rate; the initialization parameters include at least initialization weights and initialization biases; The prediction accuracy is calculated using the validation set to validate the trained network address management model. If the validation fails, the training set is expanded to continue training. If the validation passes, then: The reliability index is calculated through the test set to test the verified network address management model. If the test fails, the training set is expanded to continue training; if the test passes, the training is terminated.

10. The neural network-based dynamic management system for virtual machine network addresses according to claim 6, characterized in that: The network address management model deployment module is specifically used to: After dynamically pruning and knowledge distilling the network address management model that has passed the test, the network address management model is encapsulated as a high-performance service and deployed on a server; The network address management module is specifically used for: The server creates an address pool with a preset number of network addresses, divides the network addresses in the address pool into regular addresses and backup addresses, and sets virtual machine priorities based on the service type; The server collects real-time virtual machine data, inputs the real-time virtual machine data into a deployed network address management model, and obtains a network address management policy including at least a network address allocation plan, lease adjustment parameters, and conflict avoidance operations; The server determines whether the regular addresses in the address pool can meet the allocation requirements based on the network address management policy. If so, the server performs a network address allocation operation based on each of the regular addresses. If not, the server: Further determine whether the backup addresses in the address pool can meet the allocation requirements. If so, perform a network address allocation operation based on each of the backup addresses; if not, then: Based on the virtual machine priority and the IP address drift mechanism, the network address of the low-priority virtual machine is allocated to the high-priority virtual machine for use; The address management log management module is specifically used to: The server records in real time an address management log including at least real-time virtual machine data, network address management policies, and network address adjustment records, and continuously optimizes the network address management model based on the address management log; The server encrypts the address management log into first-level encrypted data using the RSA algorithm, converts the first-level encrypted data into binary data, performs XOR on the binary data using a preset key to obtain second-level encrypted data, encrypts the second-level encrypted data into an encrypted log using the 3DES algorithm, and stores and distributes the encrypted log.

Images