A multi-network fusion communication link scheduling method and system based on deep learning

Abstract

The application discloses a kind of multi-network fusion communication link scheduling method and system based on deep learning, comprising the following steps: S1, link operation data in multi-network fusion communication environment is collected and preprocessed;S2, link data is Fourier multiscale decomposition and constructs space-time coupling feature map;S3, graph convolution and Transformer are jointly extracted link space-time characteristics;S4, multi-objective constraint link scheduling is carried out based on cloud network fusion resources and NFV virtualization network;S5, path search is carried out using reinforcement learning double policy iteration, and primary and backup link collaborative optimization scheduling is realized;S6, loss function is constructed based on link feedback, and the parameters of graph convolution and Transformer are adaptively updated and optimized.The application realizes the intelligent scheduling of multi-network fusion communication link and primary and backup collaborative optimization, improves link stability and resource utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of communication network and network resource scheduling technology, and in particular to a method and system for scheduling multi-network converged communication links based on deep learning. Background Technology

[0002] In multi-network converged communication environments, various heterogeneous networks such as short-range communication, narrowband IoT, and full-duplex communication coexist. These networks differ significantly in bandwidth capacity, latency characteristics, signal stability, and transmission protocols, resulting in highly dynamic communication links during actual operation. Existing technologies typically select and switch links using static rules or threshold-based methods, relying on current link quality indicators for decision-making. This lack of predictability regarding future link state trends makes it difficult to adapt to the continuously changing needs of complex and dynamic network environments.

[0003] Existing multi-network link scheduling methods often employ single-network metric evaluation mechanisms, such as link optimization based solely on latency or bandwidth. These methods fail to simultaneously consider multiple dimensions, including link stability, service load balancing, secure communication constraints, and access control. Furthermore, with the continuous development of cloud-network convergence architectures, the lack of an effective mapping mechanism between NFV virtualized network resource scheduling and the real physical network leads to discrepancies between the deployment of virtual link functional units and the actual communication link status. This results in low resource utilization efficiency and unreasonable link migration path planning.

[0004] Meanwhile, existing technologies typically employ traditional statistical methods or shallow machine learning models in link state analysis, making it difficult to extract deep features from multi-source heterogeneous link data and effectively uncover the temporal variation patterns and spatial topological relationships of links. In complex communication scenarios, link state is affected not only by network load but also by constraints such as access permission levels, network security communication protocol status, and service priorities. However, existing methods often neglect the role of security constraints and access control in the link scheduling process, resulting in deficiencies in the security and compliance aspects of the scheduling results.

[0005] Furthermore, existing reinforcement learning methods for communication link scheduling often employ single-policy decision structures, lacking the ability to collaboratively model the relationship between link risk factors and service benefits. In highly dynamic network environments, this leads to problems such as unstable policy convergence, oscillating decision paths, and frequent link switching, making it difficult to achieve stable collaborative scheduling between primary and backup links. Simultaneously, during link feedback optimization, existing methods lack a unified joint optimization mechanism for graph structure feature extraction parameters and temporal modeling parameters, resulting in weak overall model adaptability.

[0006] Therefore, how to provide a method and system for scheduling multi-network converged communication links based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a multi-network fusion communication link scheduling method and system based on deep learning. This invention integrates graph convolutional networks, Transformer temporal attention mechanism, digital twin communication topology modeling, cloud-network convergence and NFV virtualization resource scheduling, and reinforcement learning dual-strategy value iteration method to achieve spatiotemporal feature modeling of multi-network link states, link trend prediction, and adaptive scheduling optimization under security constraints. It has the advantages of high link prediction accuracy, strong real-time scheduling, strong cross-network collaborative capability, and high communication stability and security.

[0008] A multi-network converged communication link scheduling method based on deep learning according to an embodiment of the present invention includes the following steps: S1. Collect link operation data in a multi-network converged communication environment and perform preprocessing; S2. Perform multi-scale frequency domain decomposition on the preprocessed link operation data using discrete Fourier transform, and perform link mapping operation in combination with digital twin communication topology to generate spatiotemporal coupling feature map. S3. Perform graph convolution feature extraction on the spatiotemporal coupled feature map, construct a gated residual constraint structure at the cross-layer connection position, and perform link fluctuation trend learning processing in combination with the Transformer temporal attention structure. S4. Perform cloud-network converged resource association processing on the feature extraction results, construct a link migration sequence based on the NFV virtualized network, and perform multi-objective constraint scheduling operations in conjunction with network security communication protocols; S5. Perform reinforcement learning path search processing on the link migration sequence after constraint scheduling, calculate the link switching cost and service load balancing value using the dual-policy value iteration structure, establish a dynamic collaborative relationship between the main communication link and the backup communication link, and generate link scheduling results. S6. Perform link feedback analysis on the link scheduling results, construct the link feedback loss function, and perform adaptive update processing on the graph convolution parameters and Transformer temporal attention parameters.

[0009] Optionally, S1 specifically includes: S11. Deploy link monitoring terminals in short-range communication nodes, narrowband IoT nodes, and full-duplex communication nodes, and collect link delay values, received signal strength values, spectrum occupancy values, round-trip delay values, link jitter values, bit error rate values, and service queue length values ​​through high-speed sampling interfaces at a sampling period of 10ms. S12. The time acquisition of each communication node is aligned at the nanosecond level using a unified clock synchronization mechanism, and a link index identifier is generated according to the node number-service number-time stamp format. S13. Perform frame buffering processing on the collected link operation data and encapsulate the data within 20 consecutive sampling periods into a link status data block; S14. The missing data in the link status data block is compensated by cubic spline interpolation, and the data exceeding 1.5 times the interquartile range is removed by an anomaly detection method based on box distribution. S15. Perform minimum-maximum normalization processing on the data that has completed anomaly detection, and map the round-trip delay, bit error rate and spectrum occupancy values ​​of the link to the range of 0 to 1. S16. A sliding time window method is used to perform time-series splicing processing on 32 consecutive link status data blocks, specifically including: The continuous link status data blocks are queued according to the timestamp order. A data window of length 32 is established with 8 link status data blocks as the sliding step size. The link delay value, received signal strength value, spectrum occupancy value, round-trip delay value, link jitter value, bit error rate value and service queue length value in each data window are spliced ​​together to form a link sequence matrix with a dimension of 32×7. A time feature vector is added to each temporal position in the link sequence matrix using a position encoding method, and node embedding vectors are established for different short-range communication nodes, narrowband IoT nodes and full-duplex communication nodes based on the node number; The time feature vector and the node embedding vector are superimposed on the link sequence matrix to generate the link time-series feature matrix; A one-dimensional convolutional structure is used to perform local temporal correlation extraction on the link temporal feature matrix. The convolutional kernel length is set to 5 and the convolutional stride is set to 1. Convolutional mapping is performed on the link fluctuation trend between consecutive time slices. The state update process of the gated recurrent unit is performed on the convolution mapping result, and the hidden state vector corresponding to each time slice is calculated in time order. A self-attention mechanism is used to perform weight allocation on the hidden state vector and calculate the time dependency between the state data blocks of each link. The hidden state vector is weighted and aggregated based on the attention weight values, and then batch normalization and Dropout random deactivation are performed.

[0010] Optionally, S2 specifically includes: S21. Construct discrete time series signals for the preprocessed link operation data according to the 7-dimensional link attribute dimensions, perform fast Fourier transform processing on each dimension signal, and calculate the corresponding frequency domain amplitude spectrum and phase spectrum. S22. Divide the frequency range into low-frequency sub-band, mid-frequency sub-band and high-frequency sub-band, and perform bandpass filtering and energy normalization on different sub-bands to form a multi-scale frequency domain feature matrix. S23. Perform tensor fusion operation on the multi-scale frequency domain feature matrix and the link time-series feature matrix to generate a time-frequency joint representation tensor; S24. Construct a digital twin communication topology model, mapping near-range communication nodes, narrowband IoT nodes, and full-duplex communication nodes to topology graph structure nodes according to physical and logical connection relationships, and attaching access permission level vectors and link security state vectors to each topology node. S25. Based on the link index identifier, map each link state unit in the time-frequency joint representation tensor to the corresponding topology node and edge connection relationship in the digital twin communication topology, and perform topology alignment mapping processing; S26. Perform graph structure encoding processing on the topologically aligned node features, and use a propagation update mechanism based on the adjacency matrix to calculate the node embedding representation; S27. Perform element-wise fusion operation between the node embedding representation and the multi-scale frequency domain feature matrix to generate a spatiotemporal coupled feature map containing access permission constraint weights.

[0011] Optionally, S24 specifically includes: S241. Based on the node registry, uniquely identify and encode short-range communication nodes, narrowband IoT nodes and full-duplex communication nodes, and construct a topology node set according to the node's geographical location coordinates, link connection port number and communication frequency band parameters. S242. Construct an adjacency matrix based on the node interaction relationships recorded in the link operation data, and assign numerical values ​​to different link connection strengths using a weighted method to form an initial topology edge weight matrix. S243. Map the initial topology edge weight matrix to the graph structure representation space to generate a digital twin communication topology graph; S244. Extract access permission level data for each topology node, divide it into four access permission vectors according to the permission level, and perform one-hot encoding on the access permission vectors to form permission feature vectors. S245. Parse the link security communication protocol status and convert the encryption protocol type, authentication status and intrusion detection identifier into a security status feature vector. S246. Concatenate and splice the permission feature vector and the security status feature vector to generate a node attribute vector, and attach the node attribute vector to the corresponding topology node in the digital twin communication topology diagram to form a set of topology nodes with constraint attributes. S247. Perform normalization and vector standardization on the set of topology nodes to generate a digital twin communication topology model that includes access permission level constraints and link security state constraints.

[0012] Optionally, S3 specifically includes: S31. Expand the spatiotemporal coupling feature map according to the node index, construct the node feature matrix and the topological adjacency matrix, and divide the node feature matrix into continuous sequence input units according to time slices. S32. After normalizing the node features corresponding to each time slice, input them into a three-layer graph convolutional structure, where: In the first graph convolutional layer, the topological adjacency matrix is ​​enhanced by self-connection. In the second convolutional layer, an edge weight decay coefficient is introduced to weighted dilution of the adjacency relationship; In the third graph convolutional layer, neighborhood feature aggregation is performed to generate higher-order node representation vectors; S33. Set up a cross-layer connection channel between the second and third convolutional layers, and perform alignment and rearrangement processing on the output features of the first layer and the input features of the third layer. Use node index sorting to ensure that the cross-layer features correspond one-to-one. S34. Construct a gated residual constraint structure in the cross-layer connection channel, and generate a gated coefficient matrix and a residual coefficient matrix for the cross-layer features respectively, where: The gating coefficient matrix is ​​constrained by the Sigmoid mapping function, which controls the feature flow ratio. The residual coefficient matrix is ​​used to preserve the original low-level structure information; S35. Weighted fusion of gated features and residual features in a node-by-node, dimension-by-dimensional manner to generate constrained fusion feature vector; S36. Reconstruct the constrained fusion feature vectors in chronological order, concatenate the node representations of each time slice into a fixed-length sequence matrix, and insert the time position encoding vector and the node type encoding vector at the beginning of the sequence matrix. S37. Input the encoded sequence matrix into a multi-layer Transformer encoding structure, perform multi-head attention calculation on the sequence in each layer, model the correlation of link state changes between different time slices, and perform normalization to form an attention distribution matrix. S38. Based on the attention distribution matrix, perform weighted convergence processing on the sequence features to generate a global temporal representation vector; S39. Perform feedforward fully connected mapping and nonlinear activation processing on the global time series representation vector in sequence. Superimpose a gated residual constraint structure on the output end. Perform dimension-wise fusion calculation between the constraint fusion feature vector and the global time series representation vector to obtain the feature extraction result. Continue to perform link fluctuation trend learning processing. Optionally, S4 specifically includes: S41. The feature extraction results are split into business dimensions to extract the link stability score vector, latency fluctuation prediction vector and bandwidth availability vector, and then the node attribute vectors in the digital twin communication topology are mapped and matched node by node to construct a cloud-network converged resource mapping matrix. S42. Based on the resource mapping matrix, the virtual network functional units in the NFV virtualization network are instantiated and deployed, generating a set of virtual link functional nodes. The virtual link functional nodes are sorted and encoded according to the computing resource utilization rate, forwarding capability parameters and scheduling response delay parameters. S43. Perform link migration path construction processing on the sorted virtual link functional nodes, and perform path constraint search according to time priority and link stability score to generate a set of candidate link migration paths. S44. Perform path compression processing on the candidate link migration path set, merge and reconstruct consecutive redundant jump nodes, and generate a standardized link migration sequence. S45. Introduce network security communication protocol constraints into the link migration sequence, parse the encryption authentication status, key negotiation status and data integrity verification identifier, and map the security status into a constraint weight matrix; S46. The constraint weight matrix and the link migration sequence are fused to generate a multi-objective scheduling evaluation matrix. The link-by-link score is calculated according to the three objective functions of minimizing latency, maximizing bandwidth utilization and maximizing security level. S47. Perform sorting and filtering processing on the multi-objective scheduling evaluation matrix to select the optimal link migration sequence, that is, the link migration sequence after constrained scheduling.

[0013] Optionally, S5 specifically includes: S51. Expand the link migration sequence after constraint scheduling by time step, construct a set of state transition sequences, and concatenate the link resource occupancy status, link stability score, bandwidth utilization and security constraint weight at each time step to form a state vector, and map the state vector to the state space of the reinforcement learning environment. S52. Define the action space as a combination of short-range communication link switching actions, narrowband IoT link switching actions and full-duplex link switching actions, and assign a corresponding link switching control identifier to each action. S53. Construct a dual-strategy value iteration structure, in which the main strategy network is used to generate the probability distribution of link selection actions, and the auxiliary strategy network is used to generate the probability distribution of link risk suppression actions. Perform element-wise fusion operation on the two sets of probability distributions to form a joint strategy output. S54. Perform value function evaluation calculation on the joint strategy output, take the link switching delay change, packet loss rate change and access permission constraint conflict degree as cost factors, construct the link switching cost vector, and construct the service load balancing value vector based on the service traffic distribution offset and node load balancing deviation. S55. Perform weighted aggregation calculation on the link switching cost vector and the service load balancing value vector to generate an action evaluation value, and perform policy gradient update processing based on the action evaluation value to synchronously update the parameters of the main policy network and the auxiliary policy network. S56. Perform dynamic path selection operation at each state transition node, take the action with the best evaluation value as the current link switching decision result, and string together the continuous decision results to form a link scheduling path. S57. Perform primary and backup link binding processing on the link scheduling path, bind the primary communication link and the backup communication link according to time slice redundancy mapping, form a dynamic collaborative scheduling relationship structure, and output the link scheduling result.

[0014] A deep learning-based multi-network converged communication link scheduling system according to an embodiment of the present invention includes: The data processing module is used to collect link operation data in a multi-network converged communication environment and perform preprocessing. The decomposition and mapping module is used to perform multi-scale frequency domain decomposition processing on the preprocessed link running data using discrete Fourier transform, and to perform link mapping operation in combination with digital twin communication topology to generate a spatiotemporal coupling feature map. The feature extraction module is used to perform graph convolution feature extraction processing on the spatiotemporally coupled feature map, construct gated residual constraint structures at cross-layer connection positions, and perform link fluctuation trend learning processing in combination with the Transformer temporal attention structure. The constraint scheduling module is used to perform cloud-network converged resource association processing on the feature extraction results, construct link migration sequences based on NFV virtualized networks, and perform multi-objective constraint scheduling operations in conjunction with network security communication protocols; The scheduling decision module is used to perform reinforcement learning path search processing on the link migration sequence after constraint scheduling, calculate the link switching cost and service load balancing value using a dual-policy value iteration structure, establish a dynamic collaborative relationship between the primary communication link and the backup communication link, and generate link scheduling results. The feedback update module is used to perform link feedback analysis on the link scheduling results, construct the link feedback loss function, and perform adaptive update processing on the graph convolution parameters and Transformer temporal attention parameters.

[0015] The beneficial effects of this invention are: First, this invention constructs a link operation data acquisition and preprocessing mechanism in a multi-network converged communication environment, and combines digital twin communication topology and multi-scale frequency domain decomposition method to achieve unified modeling and spatiotemporal coupling expression of the link status of heterogeneous networks such as short-range communication, narrowband IoT and full-duplex communication. This improves the link status information from a single indicator description to a multi-dimensional joint feature expression, thereby significantly enhancing the integrity and continuity of link status perception.

[0016] Secondly, this invention introduces graph convolutional networks and Transformer temporal attention structures to perform deep joint modeling of the spatiotemporal features of links. It also combines cloud-network convergence and NFV virtualization resource scheduling mechanisms to constrain and optimize link migration paths and dynamically generate them. At the same time, it incorporates access permission levels and network security communication protocol status into the scheduling constraint system, enabling the link scheduling process to have both resource optimization capabilities and security compliance control capabilities, thereby improving the accuracy and stability of link selection in multi-network environments.

[0017] Finally, this invention constructs a reinforcement learning dual-strategy value iteration structure to achieve coordinated optimization of link switching cost and service load balancing, and performs dynamic coordinated binding and adaptive switching control of the main communication link and backup communication link, enabling the link scheduling strategy to have continuous learning and adaptive evolution capabilities, thereby effectively reducing link switching latency and communication interruption probability, and improving the robustness and resource utilization efficiency of the overall communication system. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based multi-network fusion communication link scheduling method proposed in this invention; Figure 2 This is a flowchart of the multi-network heterogeneous communication data acquisition and timing preprocessing of a multi-network fusion communication link scheduling method based on deep learning proposed in this invention; Figure 3 This is a schematic diagram of the spatiotemporal feature modeling and prediction structure based on graph convolution and Transformer fusion for a multi-network fusion communication link scheduling method based on deep learning proposed in this invention. Figure 4 This invention presents a flowchart of a multi-strategy link scheduling optimization method based on cloud network resources and reinforcement learning, which is proposed as a deep learning-based multi-network fusion communication link scheduling method. Figure 5 This is a module structure diagram of a deep learning-based multi-network converged communication link scheduling system proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-4 A deep learning-based method for scheduling multi-network converged communication links includes the following steps: S1. Collect link operation data in a multi-network converged communication environment and perform preprocessing; S2. Perform multi-scale frequency domain decomposition on the preprocessed link operation data using discrete Fourier transform, and perform link mapping operation in combination with digital twin communication topology to generate a spatiotemporal coupling feature map. S3. Perform graph convolution feature extraction on the spatiotemporal coupled feature map, construct a gated residual constraint structure at the cross-layer connection position, and perform link fluctuation trend learning processing in combination with the Transformer temporal attention structure. S4. Perform cloud-network converged resource association processing on the feature extraction results, construct a link migration sequence based on the NFV virtualized network, and perform multi-objective constraint scheduling operations in conjunction with network security communication protocols; S5. Perform reinforcement learning path search processing on the link migration sequence after constraint scheduling, calculate the link switching cost and service load balancing value using the dual-policy value iteration structure, establish a dynamic collaborative relationship between the main communication link and the backup communication link, and generate link scheduling results. S6. Perform link feedback analysis on the link scheduling results, construct the link feedback loss function, and perform adaptive update processing on the graph convolution parameters and Transformer temporal attention parameters.

[0021] In this embodiment, S1 specifically includes: S11. Deploy link monitoring terminals in short-range communication nodes, narrowband IoT nodes, and full-duplex communication nodes, and collect link delay values, received signal strength values, spectrum occupancy values, round-trip delay values, link jitter values, bit error rate values, and service queue length values ​​through high-speed sampling interfaces at a sampling period of 10ms. S12. The time acquisition of each communication node is aligned at the nanosecond level using a unified clock synchronization mechanism, and a link index identifier is generated according to the node number-service number-time stamp format. S13. Perform frame buffering processing on the collected link operation data and encapsulate the data within 20 consecutive sampling periods into a link status data block; S14. The missing data in the link status data block is compensated by cubic spline interpolation, and the data exceeding 1.5 times the interquartile range is removed by an anomaly detection method based on box distribution. S15. Perform minimum-maximum normalization processing on the data that has completed anomaly detection, and map the round-trip delay, bit error rate and spectrum occupancy values ​​of the link to the range of 0 to 1. S16. A sliding time window method is used to perform time-series splicing processing on 32 consecutive link status data blocks, specifically including: The continuous link status data blocks are queued according to the timestamp order. A data window of length 32 is established with 8 link status data blocks as the sliding step size. The link delay value, received signal strength value, spectrum occupancy value, round-trip delay value, link jitter value, bit error rate value and service queue length value in each data window are spliced ​​together to form a link sequence matrix with a dimension of 32×7. A time feature vector is added to each temporal position in the link sequence matrix using a position encoding method, and node embedding vectors are established for different short-range communication nodes, narrowband IoT nodes and full-duplex communication nodes based on the node number; The time feature vector and the node embedding vector are superimposed on the link sequence matrix to generate the link time-series feature matrix; A one-dimensional convolutional structure is used to perform local temporal correlation extraction on the link temporal feature matrix. The convolutional kernel length is set to 5 and the convolutional stride is set to 1. Convolutional mapping is performed on the link fluctuation trend between consecutive time slices. The state update process of the gated recurrent unit is performed on the convolution mapping result, and the hidden state vector corresponding to each time slice is calculated in time order. A self-attention mechanism is used to perform weight allocation on the hidden state vector and calculate the time dependency between the state data blocks of each link. The hidden state vector is weighted and aggregated based on the attention weight values, and then batch normalization and Dropout random deactivation are performed.

[0022] In this embodiment, S2 specifically includes: S21. Construct discrete time series signals for the preprocessed link operation data according to the 7-dimensional link attribute dimensions, perform fast Fourier transform processing on each dimension signal, and calculate the corresponding frequency domain amplitude spectrum and phase spectrum. S22. Divide the frequency range into low-frequency sub-band, mid-frequency sub-band and high-frequency sub-band, and perform bandpass filtering and energy normalization on different sub-bands to form a multi-scale frequency domain feature matrix. S23. Perform tensor fusion operation on the multi-scale frequency domain feature matrix and the link time-series feature matrix to generate a time-frequency joint representation tensor; S24. Construct a digital twin communication topology model, mapping near-range communication nodes, narrowband IoT nodes, and full-duplex communication nodes to topology graph structure nodes according to physical and logical connection relationships, and attaching access permission level vectors and link security state vectors to each topology node. S25. Based on the link index identifier, map each link state unit in the time-frequency joint representation tensor to the corresponding topology node and edge connection relationship in the digital twin communication topology, and perform topology alignment mapping processing; S26. Perform graph structure encoding processing on the topologically aligned node features, and use a propagation update mechanism based on the adjacency matrix to calculate the node embedding representation; S27. Perform element-wise fusion operation between the node embedding representation and the multi-scale frequency domain feature matrix to generate a spatiotemporal coupled feature map containing access permission constraint weights.

[0023] In this embodiment, S24 specifically includes: S241. Based on the node registry, uniquely identify and encode short-range communication nodes, narrowband IoT nodes and full-duplex communication nodes, and construct a topology node set according to the node's geographical location coordinates, link connection port number and communication frequency band parameters. S242. Construct an adjacency matrix based on the node interaction relationships recorded in the link operation data, and assign numerical values ​​to different link connection strengths using a weighted method to form an initial topology edge weight matrix. S243. Map the initial topology edge weight matrix to the graph structure representation space to generate a digital twin communication topology graph; S244. Extract access permission level data for each topology node, divide it into four access permission vectors according to the permission level, and perform one-hot encoding on the access permission vectors to form permission feature vectors. S245. Parse the link security communication protocol status and convert the encryption protocol type, authentication status and intrusion detection identifier into a security status feature vector. S246. Concatenate and splice the permission feature vector and the security status feature vector to generate a node attribute vector, and attach the node attribute vector to the corresponding topology node in the digital twin communication topology diagram to form a set of topology nodes with constraint attributes. S247. Perform normalization and vector standardization on the set of topology nodes to generate a digital twin communication topology model that includes access permission level constraints and link security state constraints.

[0024] In this embodiment, S3 specifically includes: S31. Expand the spatiotemporal coupling feature map according to the node index, construct the node feature matrix and the topological adjacency matrix, and divide the node feature matrix into continuous sequence input units according to time slices. S32. After normalizing the node features corresponding to each time slice, input them into a three-layer graph convolutional structure, where: In the first graph convolutional layer, the topological adjacency matrix is ​​enhanced by self-connection. In the second convolutional layer, an edge weight decay coefficient is introduced to weighted dilution of the adjacency relationship; In the third graph convolutional layer, neighborhood feature aggregation is performed to generate higher-order node representation vectors; S33. Set up a cross-layer connection channel between the second and third convolutional layers, and perform alignment and rearrangement processing on the output features of the first layer and the input features of the third layer. Use node index sorting to ensure that the cross-layer features correspond one-to-one. S34. Construct a gated residual constraint structure in the cross-layer connection channel, and generate a gated coefficient matrix and a residual coefficient matrix for the cross-layer features respectively, where: The gating coefficient matrix is ​​constrained by the Sigmoid mapping function, which controls the feature flow ratio. The residual coefficient matrix is ​​used to preserve the original low-level structure information; S35. Weighted fusion of gated features and residual features in a node-by-node, dimension-by-dimensional manner to generate constrained fusion feature vector; S36. Reconstruct the constrained fusion feature vectors in chronological order, concatenate the node representations of each time slice into a fixed-length sequence matrix, and insert the time position encoding vector and the node type encoding vector at the beginning of the sequence matrix. S37. Input the encoded sequence matrix into a multi-layer Transformer encoding structure, perform multi-head attention calculation on the sequence in each layer, model the correlation of link state changes between different time slices, and perform normalization to form an attention distribution matrix. S38. Based on the attention distribution matrix, perform weighted convergence processing on the sequence features to generate a global temporal representation vector; S39. Perform feedforward fully connected mapping and nonlinear activation processing on the global time series representation vector in sequence. Superimpose a gated residual constraint structure on the output end. Perform dimension-wise fusion calculation between the constraint fusion feature vector and the global time series representation vector to obtain the feature extraction result. Continue to perform link fluctuation trend learning processing. In this embodiment, S4 specifically includes: S41. The feature extraction results are split into business dimensions to extract the link stability score vector, latency fluctuation prediction vector and bandwidth availability vector, and then the node attribute vectors in the digital twin communication topology are mapped and matched node by node to construct a cloud-network converged resource mapping matrix. S42. Based on the resource mapping matrix, the virtual network functional units in the NFV virtualization network are instantiated and deployed, generating a set of virtual link functional nodes. The virtual link functional nodes are sorted and encoded according to the computing resource utilization rate, forwarding capability parameters and scheduling response delay parameters. S43. Perform link migration path construction processing on the sorted virtual link functional nodes, and perform path constraint search according to time priority and link stability score to generate a set of candidate link migration paths, specifically including: An ordered node queue is constructed from the set of virtual link functional nodes according to the node sorting and encoding results; The computational resource utilization, forwarding capability parameters, and scheduling response latency parameters of each virtual link functional node are vectorized to form a node feature vector. The node feature vector is then concatenated with the corresponding link stability score vector to form a path weight initialization matrix. Based on time priority constraints, the node queue is divided into time windows, the link migration process is divided into fixed-length scheduling cycles, and a set of reachable nodes is built within each scheduling cycle. Perform path expansion processing on the set of reachable nodes, generate multi-branch path candidate nodes according to the topological connection relationship starting from the current node, and accumulate and calculate the link stability score attenuation value and time delay penalty value for each branch path; A hybrid path search strategy combining depth-first search and breadth-first search is adopted. Constraint filtering is performed on each node selection in the path expansion process to eliminate node paths whose computing resource utilization exceeds a preset threshold. For the reserved paths, the path cost is cumulatively calculated by gradually accumulating the inverse value of the forwarding capability parameter, the scheduling response delay, and the time priority offset to form a path cost vector. All candidate paths are sorted according to their path cost vectors, and the top N paths with the lowest costs are retained as the set of candidate link migration paths. S44. Perform path compression processing on the candidate link migration path set, merge and reconstruct consecutive redundant jump nodes, and generate a standardized link migration sequence. S45. Introduce network security communication protocol constraints into the link migration sequence, parse the encryption authentication status, key negotiation status and data integrity verification identifier, and map the security status into a constraint weight matrix; S46. The constraint weight matrix and the link migration sequence are fused to generate a multi-objective scheduling evaluation matrix. The link-by-link score is calculated according to the three objective functions of minimizing latency, maximizing bandwidth utilization and maximizing security level. S47. Perform sorting and filtering processing on the multi-objective scheduling evaluation matrix to select the optimal link migration sequence, that is, the link migration sequence after constrained scheduling.

[0025] In this embodiment, S5 specifically includes: S51. Expand the link migration sequence after constraint scheduling by time step, construct a set of state transition sequences, and concatenate the link resource occupancy status, link stability score, bandwidth utilization and security constraint weight at each time step to form a state vector, and map the state vector to the state space of the reinforcement learning environment. S52. Define the action space as a combination of short-range communication link switching actions, narrowband IoT link switching actions and full-duplex link switching actions, and assign a corresponding link switching control identifier to each action. S53. Construct a dual-strategy value iteration structure, where the main strategy network generates the probability distribution of link selection actions, and the auxiliary strategy network generates the probability distribution of link risk suppression actions. Element-by-element fusion operations are performed on the two probability distributions to form a joint strategy output, specifically including: The motion space control identifier output by S52 is processed by vector embedding encoding to construct a set of motion feature vectors, and then matched with the state vector generated by S51 one by one according to time step to form a state-action pair input sequence. The state-action pair input sequence is fed into the main policy network and the auxiliary policy network respectively. The main policy network adopts a multi-layer fully connected neural network and an attention weight allocation structure. The output link selects the action probability vector and performs softmax normalization on the probability vector to form a standardized action distribution. The auxiliary policy network employs a risk assessment coding layer and a constraint reinforcement layer to jointly encode the link stability score, security constraint weight, and access permission conflict factor in the state vector, outputting a link risk suppression probability vector, and then applying sigmoid gating constraint processing to it to form a risk control distribution. The link selection action probability vector output by the main policy network and the link risk suppression probability vector output by the auxiliary policy network are fused element-wise to generate a joint policy probability matrix. Temperature coefficient adjustment is applied to the joint strategy probability matrix to reduce the probability of high-risk actions and increase the probability of low-risk, high-return actions. Based on the adjusted joint policy probability matrix, sampling selection processing is performed to generate a set of candidate link switching actions. The set of candidate link switching actions is then mapped to the action execution interface of the reinforcement learning environment to form the joint policy output result. S54. Perform value function evaluation calculation on the joint strategy output, take the link switching delay change, packet loss rate change and access permission constraint conflict degree as cost factors, construct the link switching cost vector, and construct the service load balancing value vector based on the service traffic distribution offset and node load balancing deviation. S55. Perform weighted aggregation calculation on the link switching cost vector and the service load balancing value vector to generate an action evaluation value, and perform policy gradient update processing based on the action evaluation value to synchronously update the parameters of the main policy network and the auxiliary policy network. S56. Perform dynamic path selection operation at each state transition node, take the action with the best evaluation value as the current link switching decision result, and string together the continuous decision results to form a link scheduling path. S57. Perform primary and backup link binding processing on the link scheduling path, bind the primary communication link and the backup communication link according to time slice redundancy mapping, form a dynamic collaborative scheduling relationship structure, and output the link scheduling result.

[0026] refer to Figure 5 A deep learning-based multi-network converged communication link scheduling system includes: The data processing module is used to collect link operation data in a multi-network converged communication environment and perform preprocessing. The decomposition and mapping module is used to perform multi-scale frequency domain decomposition processing on the preprocessed link running data using discrete Fourier transform, and to perform link mapping operation in combination with digital twin communication topology to generate a spatiotemporal coupling feature map. The feature extraction module is used to perform graph convolution feature extraction processing on the spatiotemporally coupled feature map, construct gated residual constraint structures at cross-layer connection positions, and perform link fluctuation trend learning processing in combination with the Transformer temporal attention structure. The constraint scheduling module is used to perform cloud-network converged resource association processing on the feature extraction results, construct link migration sequences based on NFV virtualized networks, and perform multi-objective constraint scheduling operations in conjunction with network security communication protocols; The scheduling decision module is used to perform reinforcement learning path search processing on the link migration sequence after constraint scheduling, calculate the link switching cost and service load balancing value using a dual-policy value iteration structure, establish a dynamic collaborative relationship between the primary communication link and the backup communication link, and generate link scheduling results. The feedback update module is used to perform link feedback analysis on the link scheduling results, construct the link feedback loss function, and perform adaptive update processing on the graph convolution parameters and Transformer temporal attention parameters.

[0027] Example 1: To verify the feasibility of the present invention in practice, the present invention was applied to... To verify the feasibility of this invention in practice, it was applied to a high-dynamic service scheduling scenario of multi-network converged communication. This scenario includes three types of heterogeneous networks: short-range communication links, narrowband IoT links, and full-duplex communication links. The service types include high-frequency control command streams, continuous video streams, and low-speed sensor data streams. The three types of services are transmitted concurrently within the same time window and the load fluctuates significantly. Among them, control services account for about 18%, video services account for about 47%, and sensor services account for about 35%. The link status exhibits obvious sudden fluctuation characteristics at a 10ms sampling period.

[0028] During the data acquisition phase, link monitoring units were deployed simultaneously on three types of network nodes. Data on link latency, bandwidth utilization, packet loss rate, signal strength, link jitter, and queue backlog length were acquired within a continuous sampling period. Each link generated 32 frames of continuous status data within a statistical window. In the raw data, the average latency fluctuation range for short-range communication links was 28ms to 95ms, the average packet loss rate for narrowband IoT links was 3.2% to 7.6%, and the bandwidth utilization fluctuation range for full-duplex links was 41% to 88%.

[0029] During the preprocessing stage, timestamp alignment and outlier removal were performed on the collected data. The proportion of outlier samples was approximately 4.7%, mainly concentrated in the range of sudden congestion and signal attenuation. After processing, the data was normalized and mapped to the [0,1] interval, where the standard deviation of delay decreased from the original 0.42 to 0.19, significantly improving data stability.

[0030] In the feature modeling stage, the processed data underwent Discrete Fourier Transform. After multi-scale frequency domain decomposition, the energy proportions of low-frequency components were approximately 62.3%, mid-frequency components approximately 27.5%, and high-frequency components approximately 10.2%. High-frequency abrupt changes mainly corresponded to link switching and sudden congestion events. Node mapping relationships were constructed based on the digital twin communication topology, generating a total of 128 topology nodes and 412 edge connections. The access permission levels in the node attributes were distributed as follows: Level 3 services accounted for 41%, Level 2 services for 37%, and Level 1 services for 22%.

[0031] In the graph convolution and temporal modeling stages, the output node feature dimension of the graph convolutional network is expanded from the original 7 dimensions to a 64-dimensional representation vector. The cross-layer gated residual structure reduces the feature gradient decay rate from 0.38 to 0.11. After the Transformer temporal attention structure models the link state at 32 time steps, the concentration of key temporal dependency weights increases to 0.86, indicating that the model can effectively capture the trend changes before the link abruptly changes.

[0032] During the cloud-network convergence scheduling phase, the number of NFV virtualization link instances increased from an average of 6 under the traditional fixed allocation method to 11 resource mapping units under dynamic scheduling. The average link migration path length decreased from 9.3 hops to 5.1 hops, and the path redundancy rate decreased by approximately 43.7%. After the introduction of security constraints, the access permission conflict rate decreased from the original 8.4% to 0%.

[0033] During the reinforcement learning scheduling phase, the dual-policy value iteration structure converged after approximately 3000 training iterations, with the policy loss function decreasing from an initial 2.31 to 0.27, representing a convergence speed improvement of approximately 38% compared to the single-policy model. The average cost of link switching decreased by 41.6%, and the business load balancing index increased from 0.58 to 0.93.

[0034] The final scheduling results are shown in Table 1. Under the same service pressure, the method of this invention performed 42 link switching operations in a continuous 30-minute simulation cycle, while the traditional method performed 78. The average link switching latency was reduced from 74ms to 33ms. The number of service interruptions was reduced from 12 to 2. The success rate of high-priority service transmission was increased from 89.1% to 99.4%.

[0035] The corresponding performance comparison data is shown in the table below: Table 1. Quantitative Comparison of Multi-Network Converged Communication Link Scheduling Performance

[0036] As can be seen from Table 1, this invention significantly improves the stability, real-time performance, security, and resource utilization of multi-network converged communication links through spatiotemporal feature modeling, digital twin mapping, joint learning of graph convolution and Transformer, and dual-policy reinforcement learning scheduling mechanism. It effectively solves the problems of link scheduling lag, insufficient cross-network collaboration, and lack of security constraints in the prior art.

[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-network converged communication link scheduling method based on deep learning, characterized in that, Includes the following steps: S1. Collect link operation data in a multi-network converged communication environment and perform preprocessing; S2. Perform multi-scale frequency domain decomposition on the preprocessed link operation data using discrete Fourier transform, and perform link mapping operation in combination with digital twin communication topology to generate a spatiotemporal coupling feature map. S3. Perform graph convolution feature extraction on the spatiotemporal coupled feature map, construct a gated residual constraint structure at the cross-layer connection position, and perform link fluctuation trend learning processing in combination with the Transformer temporal attention structure. S4. Perform cloud-network convergence resource association processing on the feature extraction results, construct a link migration sequence based on the NFV virtualization network, and perform multi-objective constraint scheduling operations in conjunction with network security communication protocols; S5. Perform reinforcement learning path search processing on the link migration sequence after constraint scheduling, calculate the link switching cost and service load balancing value using the dual-policy value iteration structure, establish a dynamic collaborative relationship between the main communication link and the backup communication link, and generate link scheduling results. S6. Perform link feedback analysis on the link scheduling results, construct the link feedback loss function, and perform adaptive update processing on the graph convolution parameters and Transformer temporal attention parameters.

2. The method for scheduling multi-network fusion communication links based on deep learning according to claim 1, characterized in that, S1 specifically includes: S11. Deploy link monitoring terminals in short-range communication nodes, narrowband IoT nodes, and full-duplex communication nodes, and collect link delay values, received signal strength values, spectrum occupancy values, round-trip delay values, link jitter values, bit error rate values, and service queue length values ​​through high-speed sampling interfaces at a sampling period of 10ms. S12. The time acquisition of each communication node is aligned at the nanosecond level using a unified clock synchronization mechanism, and a link index identifier is generated according to the node number-service number-time stamp format. S13. Perform frame buffering processing on the collected link operation data and encapsulate the data within 20 consecutive sampling periods into a link status data block; S14. The missing data in the link status data block is compensated by cubic spline interpolation, and the data exceeding 1.5 times the interquartile range is removed by an anomaly detection method based on box distribution. S15. Perform minimum-maximum normalization processing on the data that has completed anomaly detection, and map the round-trip delay, bit error rate and spectrum occupancy values ​​of the link to the range of 0 to 1. S16. A sliding time window method is used to perform time-series splicing processing on 32 consecutive link status data blocks, specifically including: The continuous link status data blocks are queued according to the timestamp order. A data window of length 32 is established with 8 link status data blocks as the sliding step size. The link delay value, received signal strength value, spectrum occupancy value, round-trip delay value, link jitter value, bit error rate value and service queue length value in each data window are spliced ​​together to form a link sequence matrix with a dimension of 32×7. A time feature vector is added to each temporal position in the link sequence matrix using a position encoding method, and node embedding vectors are established for different short-range communication nodes, narrowband IoT nodes and full-duplex communication nodes based on the node number; The time feature vector and the node embedding vector are superimposed on the link sequence matrix to generate the link time series feature matrix; A one-dimensional convolutional structure is used to perform local temporal correlation extraction on the link temporal feature matrix, where the convolutional kernel length is set to 5 and the convolutional stride is set to 1, and the link fluctuation trend between consecutive time slices is convolved and mapped. The state update process of the gated recurrent unit is performed on the convolution mapping result, and the hidden state vector corresponding to each time slice is calculated in time order. A self-attention mechanism is used to perform weight allocation on the hidden state vector and calculate the time dependency between the state data blocks of each link. The hidden state vector is weighted and aggregated based on the attention weight values, and then batch normalization and Dropout random deactivation are performed.

3. The method for scheduling multi-network converged communication links based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Construct discrete time series signals for the preprocessed link operation data according to the 7-dimensional link attribute dimensions, perform fast Fourier transform processing on each dimension signal, and calculate the corresponding frequency domain amplitude spectrum and phase spectrum. S22. Divide the frequency range into low-frequency sub-band, mid-frequency sub-band and high-frequency sub-band, and perform bandpass filtering and energy normalization on different sub-bands to form a multi-scale frequency domain feature matrix. S23. Perform tensor fusion operation on the multi-scale frequency domain feature matrix and the link time-series feature matrix to generate a time-frequency joint representation tensor; S24. Construct a digital twin communication topology model, mapping near-range communication nodes, narrowband IoT nodes, and full-duplex communication nodes to topology graph structure nodes according to physical and logical connection relationships, and attaching access permission level vectors and link security state vectors to each topology node. S25. Based on the link index identifier, map each link state unit in the time-frequency joint representation tensor to the corresponding topology node and edge connection relationship in the digital twin communication topology, and perform topology alignment mapping processing; S26. Perform graph structure encoding processing on the topologically aligned node features, and use a propagation update mechanism based on the adjacency matrix to calculate the node embedding representation; S27. Perform element-wise fusion operation between the node embedding representation and the multi-scale frequency domain feature matrix to generate a spatiotemporal coupled feature map containing access permission constraint weights.

4. The multi-network fusion communication link scheduling method based on deep learning according to claim 3, characterized in that, S24 specifically includes: S241. Based on the node registry, uniquely identify and encode short-range communication nodes, narrowband IoT nodes and full-duplex communication nodes, and construct a topology node set according to the node's geographical location coordinates, link connection port number and communication frequency band parameters. S242. Construct an adjacency matrix based on the node interaction relationships recorded in the link operation data, and assign numerical values ​​to different link connection strengths using a weighted method to form an initial topology edge weight matrix. S243. Map the initial topology edge weight matrix to the graph structure representation space to generate a digital twin communication topology graph; S244. Extract access permission level data for each topology node, divide it into four access permission vectors according to the permission level, and perform one-hot encoding on the access permission vectors to form permission feature vectors. S245. Parse the link security communication protocol status and convert the encryption protocol type, authentication status and intrusion detection identifier into a security status feature vector. S246. Concatenate and splice the permission feature vector and the security status feature vector to generate a node attribute vector, and attach the node attribute vector to the corresponding topology node in the digital twin communication topology diagram to form a set of topology nodes with constraint attributes. S247. Perform normalization and vector standardization on the set of topology nodes to generate a digital twin communication topology model that includes access permission level constraints and link security state constraints.

5. The multi-network fusion communication link scheduling method based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31. Expand the spatiotemporal coupling feature map according to the node index, construct the node feature matrix and the topological adjacency matrix, and divide the node feature matrix into continuous sequence input units according to time slices. S32. After normalizing the node features corresponding to each time slice, input them into a three-layer graph convolutional structure, where: In the first graph convolutional layer, the topological adjacency matrix is ​​enhanced by self-connection. In the second convolutional layer, an edge weight decay coefficient is introduced to weighted dilution of the adjacency relationship; In the third graph convolutional layer, neighborhood feature aggregation is performed to generate higher-order node representation vectors; S33. Set up a cross-layer connection channel between the second and third convolutional layers, and perform alignment and rearrangement processing on the output features of the first layer and the input features of the third layer. Use node index sorting to ensure that the cross-layer features correspond one-to-one. S34. Construct a gated residual constraint structure in the cross-layer connection channel, and generate a gated coefficient matrix and a residual coefficient matrix for the cross-layer features respectively, where: The gating coefficient matrix is ​​constrained by the Sigmoid mapping function, which controls the feature flow ratio. The residual coefficient matrix is ​​used to preserve the original low-level structure information; S35. Weighted fusion of gated features and residual features in a node-by-node, dimension-by-dimensional manner to generate constrained fusion feature vector; S36. Reconstruct the constrained fusion feature vectors in chronological order, concatenate the node representations of each time slice into a fixed-length sequence matrix, and insert the time position encoding vector and the node type encoding vector at the beginning of the sequence matrix. S37. Input the encoded sequence matrix into a multi-layer Transformer encoding structure, perform multi-head attention calculation on the sequence in each layer, model the correlation of link state changes between different time slices, and perform normalization to form an attention distribution matrix. S38. Based on the attention distribution matrix, perform weighted convergence processing on the sequence features to generate a global temporal representation vector; S39. Perform feedforward fully connected mapping and nonlinear activation processing on the global time series representation vector in sequence. Superimpose a gated residual constraint structure on the output end. Perform dimension-wise fusion calculation between the constraint fusion feature vector and the global time series representation vector to obtain the feature extraction result. Continue to perform link fluctuation trend learning processing.

6. The multi-network converged communication link scheduling method based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. The feature extraction results are split into business dimensions to extract the link stability score vector, latency fluctuation prediction vector and bandwidth availability vector, and then the node attribute vectors in the digital twin communication topology are mapped and matched node by node to construct a cloud-network converged resource mapping matrix. S42. Based on the resource mapping matrix, the virtual network functional units in the NFV virtualization network are instantiated and deployed, generating a set of virtual link functional nodes. The virtual link functional nodes are sorted and encoded according to the computing resource utilization rate, forwarding capability parameters and scheduling response delay parameters. S43. Perform link migration path construction processing on the sorted virtual link functional nodes, and perform path constraint search according to time priority and link stability score to generate a set of candidate link migration paths. S44. Perform path compression processing on the candidate link migration path set, merge and reconstruct consecutive redundant jump nodes, and generate a standardized link migration sequence. S45. Introduce network security communication protocol constraints into the link migration sequence, parse the encryption authentication status, key negotiation status and data integrity verification identifier, and map the security status into a constraint weight matrix; S46. The constraint weight matrix and the link migration sequence are fused to generate a multi-objective scheduling evaluation matrix. The link-by-link score is calculated according to the three objective functions of minimizing latency, maximizing bandwidth utilization and maximizing security level. S47. Perform sorting and filtering processing on the multi-objective scheduling evaluation matrix to select the optimal link migration sequence, that is, the link migration sequence after constrained scheduling.

7. The multi-network fusion communication link scheduling method based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Expand the link migration sequence after constraint scheduling by time step, construct a set of state transition sequences, and concatenate the link resource occupancy status, link stability score, bandwidth utilization and security constraint weight at each time step to form a state vector, and map the state vector to the state space of the reinforcement learning environment. S52. Define the action space as a combination of short-range communication link switching actions, narrowband IoT link switching actions and full-duplex link switching actions, and assign a corresponding link switching control identifier to each action. S53. Construct a dual-strategy value iteration structure, in which the main strategy network is used to generate the probability distribution of link selection actions, and the auxiliary strategy network is used to generate the probability distribution of link risk suppression actions. Perform element-wise fusion operation on the two sets of probability distributions to form a joint strategy output. S54. Perform value function evaluation calculation on the joint strategy output, take the link switching delay change, packet loss rate change and access permission constraint conflict degree as cost factors, construct the link switching cost vector, and construct the service load balancing value vector based on the service traffic distribution offset and node load balancing deviation. S55. Perform weighted aggregation calculation on the link switching cost vector and the service load balancing value vector to generate an action evaluation value, and perform policy gradient update processing based on the action evaluation value to synchronously update the parameters of the main policy network and the auxiliary policy network. S56. Perform dynamic path selection operation at each state transition node, take the action with the best evaluation value as the current link switching decision result, and string together the continuous decision results to form a link scheduling path. S57. Perform primary and backup link binding processing on the link scheduling path, bind the primary communication link and the backup communication link according to time slice redundancy mapping, form a dynamic collaborative scheduling relationship structure, and output the link scheduling result.

8. A deep learning-based multi-network converged communication link scheduling system, executing the deep learning-based multi-network converged communication link scheduling method according to any one of claims 1 to 7, characterized in that, include: The data processing module is used to collect link operation data in a multi-network converged communication environment and perform preprocessing. The decomposition and mapping module is used to perform multi-scale frequency domain decomposition processing on the preprocessed link running data using discrete Fourier transform, and to perform link mapping operation in combination with digital twin communication topology to generate a spatiotemporal coupling feature map. The feature extraction module is used to perform graph convolution feature extraction processing on the spatiotemporally coupled feature map, construct gated residual constraint structures at cross-layer connection positions, and perform link fluctuation trend learning processing in combination with the Transformer temporal attention structure. The constraint scheduling module is used to perform cloud-network converged resource association processing on the feature extraction results, construct link migration sequences based on NFV virtualized networks, and perform multi-objective constraint scheduling operations in conjunction with network security communication protocols; The scheduling decision module is used to perform reinforcement learning path search processing on the link migration sequence after constraint scheduling, calculate the link switching cost and service load balancing value using a dual-policy value iteration structure, establish a dynamic collaborative relationship between the primary communication link and the backup communication link, and generate link scheduling results. The feedback update module is used to perform link feedback analysis on the link scheduling results, construct the link feedback loss function, and perform adaptive update processing on the graph convolution parameters and Transformer temporal attention parameters.

Images