Distributed large two-layer network intelligent routing method and system based on edge cloud nodes

Abstract

The invention relates to a distributed large two-layer network intelligent routing method and system based on edge cloud nodes, and belongs to the technical field of computer networks. Aiming at the problems of slow routing decision response and low fault recovery efficiency of the existing network in a dynamic environment, the invention provides a method for periodically collecting link bandwidth utilization rate, transmission delay and other state information through an edge cloud node, uploading the state information to a central cloud platform to generate a global routing strategy and issuing the global routing strategy to the edge node, and triggering fault switching in combination with local dynamic threshold monitoring. According to the technical scheme, bidirectional communication between edge nodes and a central platform, a neural network model of spatial-temporal feature fusion and a path priority generation mechanism of multi-objective optimization are included. The method and the system are suitable for a large-scale distributed network environment, and network resource utilization rate and service continuity can be improved.

Description

Technical Field

[0001] The present invention relates to the field of computer network technologies. More specifically, the present invention relates to a distributed large two-layer network intelligent routing method and system based on edge cloud nodes. Background Art

[0002] In the existing distributed network environment, traditional routing protocols (such as OSPF, BGP) rely on static configuration or centralized control and are difficult to adapt to the dynamically changing network state. Problems such as link bandwidth fluctuations between edge nodes, sudden increases in transmission delay, and uneven node loads often lead to lagging routing strategies, resulting in service interruptions or a decline in service quality. For example, in the industrial Internet of Things scenario, frequent device access and bursty data transmission are likely to cause local congestion, and centralized routing calculation cannot respond in a timely manner due to data transmission delays. In addition, existing failover mechanisms are mostly based on fixed thresholds and lack the ability to predict the trend of delay fluctuations, resulting in a relatively high failure rate of handover path verification.

[0003] The reasons for the above problems include: 1) Global network state awareness depends on periodic reporting, and the real-time nature of data is insufficient; 2) Routing strategy optimization does not incorporate spatio-temporal characteristics and it is difficult to balance multi-dimensional constraints (bandwidth, delay, load); 3) Local decision-making lacks an adaptive threshold adjustment mechanism and cannot cope with sudden anomalies. Existing solutions attempt to introduce SDN or AI models, but still face challenges such as limited edge computing resources, high model training overhead, and low cross-platform policy synchronization efficiency. Summary of the Invention

[0004] An object of the present invention is to solve at least the above problems and provide at least the advantages described later.

[0005] To achieve these and other advantages in accordance with the present invention, there is provided a distributed large two-layer network intelligent routing method based on edge cloud nodes, including the following steps: Multiple edge cloud nodes respectively establish two-way data connections with the central cloud platform through pre-configured communication protocols, and each edge cloud node periodically collects the network state information of the node where it is located. The network state information includes the link bandwidth utilization rate between adjacent nodes, the cross-node transmission delay, the network device port state, and the node local computing resource load rate; Each edge cloud node encapsulates the network state information into a structured data packet and uploads it to the central cloud platform through an encrypted tunnel; The central cloud platform generates a global routing strategy based on the network state information of all edge cloud nodes, using a neural network model based on spatio-temporal feature extraction. The global routing strategy includes a path priority list between each edge cloud node, a traffic allocation weight coefficient, and a failover path set; After each edge cloud node receives the global routing policy issued by the central cloud platform, it updates the forwarding table entries in the local routing table storage unit according to the path priority list and the traffic allocation weight coefficient, and synchronizes the updated forwarding table entries to adjacent nodes through a pre-configured communication protocol; Each edge cloud node monitors the change amount of the link transmission delay between adjacent nodes in real time. When it detects that the change amount of the link transmission delay exceeds the dynamic threshold pre-stored in the local routing decision module, it triggers the local routing decision module to reconfigure the local routing table based on the failover path set in the global routing policy, and marks the link status exception information and uploads it to the central cloud platform through an encrypted tunnel; The central cloud platform starts to recalculate the global routing policy according to the link status exception information, and distributes the updated global routing policy to the affected edge cloud nodes associated with the link status exception information.

[0006] Preferably, the neural network model based on spatio-temporal feature extraction includes a convolutional neural network branch and a long short-term memory network branch arranged in parallel, where, The convolutional neural network branch takes the physical topology connection relationship of the edge cloud nodes as the input, and extracts the spatial correlation features between adjacent nodes through three convolutional layers. The spatial correlation features include the link bandwidth fluctuation pattern between nodes and the port status change trend; The long short-term memory network branch takes the cross-node transmission delay sequence as the input, and extracts the delay fluctuation time series features within a preset time window through a bidirectional cyclic structure. The delay fluctuation time series features include the periodic congestion pattern and the bursty abnormal delay segment; The neural network model performs dynamic weight allocation on the spatial correlation features and the delay fluctuation time series features through the attention mechanism layer to generate a network state vector that fuses spatio-temporal features; The fully connected layer of the neural network model calculates and generates the global routing policy according to the network state vector. The generation process of the path priority list includes performing multi-objective optimization operations on the network state vector that fuses spatio-temporal features. The constraint conditions of the multi-objective optimization operations include the link bandwidth utilization threshold, the transmission delay upper limit, and the node load balancing coefficient.

[0007] Preferably, reconfiguring the local routing table storage unit based on the failover path set includes the following steps: When it detects that the change amount of the link transmission delay exceeds the dynamic threshold, extract the pre-sorted candidate path sequence from the failover path set. The candidate path sequence is sorted based on the historical transmission success rate and the path hop count; Perform link bandwidth availability detection and transmission delay verification on the candidate path with the highest priority. Among them, the link bandwidth availability detection is achieved by sending probe packets to the target node and calculating the packet loss rate, and the transmission delay verification is achieved by measuring the difference between the round-trip delay of the probe packet and the preset delay threshold; If the packet loss rate of the candidate path is lower than the preset packet loss threshold and the round-trip delay difference is within the preset tolerance range, then divert the current faulty link traffic to this candidate path according to a preset ratio, and generate a forwarding table entry update instruction including the diversion ratio and the path identifier; If the verification of the candidate path fails, trigger the dynamic weight adjustment of the candidate path sequence, recalculate the path priority according to the real-time load rate of adjacent nodes, and jump to the link bandwidth availability detection step to perform the verification of the sub-optimal candidate path; After completing the path switching, the local routing decision module associates and stores the updated forwarding table entry with the fault switching timestamp, and sends a routing table synchronization request to adjacent edge cloud nodes through a pre-configured communication protocol. The synchronization request includes the hash value and version number of the changed forwarding table entry; The network status detection unit continuously monitors the transmission quality indicators of the switched path. When the transmission delay fluctuation amount exceeds 50% of the dynamic threshold for three consecutive detection cycles, trigger the secondary path switching process and update the priority weight parameters of the fault switching path set.

[0008] Preferably, the execution process of the multi-objective optimization operation includes the following steps: Map the network state vector integrating spatio-temporal features into a set of decision variables including the bandwidth allocation ratio factor α i , the delay compensation factor β j and the node load balancing weight γ k , where i corresponds to the path number, j corresponds to the delay level identifier, and k corresponds to the node identifier; Construct a set of objective functions including the first objective function, the second objective function, and the third objective function. The first objective function is f1 = Σ(α i ×C i ), which is used to maximize the effective bandwidth utilization rate. The second objective function is f2 = Σ(β j ×D j ), which is used to minimize the transmission delay offset. The third objective function is f3 = Σ(γ k ×L k ), which is used to balance the node load difference degree. Among them, C i represents the available bandwidth capacity of path i, D j represents the delay compensation amount of path j, and L k represents the load deviation value of node k; Set a system of equations including the following constraint conditions: Σαi ≤ predefined link bandwidth utilization threshold, Σβ j ≤predefined upper limit of transmission delay, Σγ k ≤predefined node load balancing factor; A non-dominated sorting genetic algorithm with an elite retention strategy is used to solve the objective function set, which includes: Generate an initial solution set as the genetic algorithm population based on the path priority list, perform an adaptive crossover operation based on path similarity, dynamically adjust the crossover probability based on the path overlap rate of the individuals in the solution set, and perform an adjacent path replacement mutation operation to replace the current path with a candidate path that is directly adjacent in the physical topology; The following processing is performed on the solution set produced by each iteration: Delete the individuals that violate the constraint condition equation group, perform non-dominated sorting on the remaining individuals and calculate the crowding distance, and retain the first preset number of individuals with the highest sorting level and the largest crowding distance; From the Pareto frontier solution set of the final iterative result, select the individual with the largest crowding distance as the optimal solution, and extract the α corresponding to the individual i , β j and γ k The parameters are combined to generate a sorting weight value for the path priority list.

[0009] Preferably, the determination and update process of the dynamic threshold in the local routing decision module includes the following steps: The local routing decision module is initialized and loaded with the baseline threshold T0, which is calculated based on the normal distribution parameters of the historical transmission delay data. The calculation formula is T0=μ+3σ, where μ is the delay mean and σ is the delay standard deviation. The network status detection unit counts the sliding window average of the link transmission delay between adjacent nodes every five minutes. The window size is the most recent 30 sampling points. When the deviation between the sliding window average and the baseline threshold T0 exceeds 20%, the threshold adaptive adjustment is triggered. The threshold adaptive adjustment process performs the following operations: Collect the delay data set of each hour within 24 hours before the current time, remove the data points with more than three times the standard deviation, recalculate the mean μ' and standard deviation σ', generate the updated dynamic threshold T1=μ'+2.5σ' and overwrite and store it in the local routing decision module; Between two threshold adjustment cycles, if a single delay change is detected to exceed 150% of the current dynamic threshold, the emergency threshold correction process is immediately initiated, the current dynamic threshold is temporarily set to 80% of the abnormal delay value and marked as pending calibration; When the edge cloud node receives the updated global routing policy issued by the central cloud platform, it synchronously obtains the threshold calibration parameters included in the global routing policy, and uses the weighted average algorithm to fuse and calculate the local dynamic threshold with the calibration parameters to generate the finally effective dynamic threshold.

[0010] Preferably, the process of periodically collecting network status information includes the following steps: The network status detection unit configures multi-dimensional data collection tasks according to the preset collection policy. The collection policy includes that the collection period of link bandwidth utilization rate is 5 minutes, the collection period of cross-node transmission delay is 30 seconds, the collection period of network device port status is 1 minute, and the collection period of node local computing resource load rate is 2 minutes; When performing link bandwidth utilization rate collection, send ICMP probe packets to adjacent nodes and calculate the arithmetic mean of the bandwidth occupancy rates of the last 10 probe results. At the same time, obtain the real-time throughput counter value of the network device port through the SNMP protocol; When performing cross-node transmission delay collection, use UDP benchmark packets for two-way transmission testing, calculate the mean value of the time difference between the sending moment and the receiving confirmation moment as the effective delay value, and record the maximum delay fluctuation amount during the testing process; The raw data collected is processed by the preprocessing module to perform the following operations: Eliminate abnormal sampling points exceeding three times the standard deviation range, perform one-hot encoding conversion on the binary flag bits in the port status data, and convert the index data with different dimensions into eigenvalue of the same magnitude through the Z-score standardization method; When the network status detection unit monitors that the change rate of the adjacent two collection results of any index exceeds the preset sensitivity threshold, automatically shorten the collection period of this index to 50% of the original period, and start intensive sampling for three consecutive periods; All preprocessed network status information data is attached with a timestamp accurate to the millisecond level when stored, and the clock is synchronized and calibrated with the central cloud platform through the NTP protocol.

[0011] Preferably, the processing after the sub-optimal candidate path verification fails further includes the following steps: When the candidate path verification fails three times in a row, the local routing decision module performs the following operations: Roll back the traffic of the current faulty link to the path before switching, and generate a path rollback instruction to update the forwarding table entry; Send a path switching failure warning message to the central cloud platform. The warning message includes the failed path identifier, the verification failure reason code, and the current node load status; Start the local degradation mode, limit the forwarding rate of non-critical service traffic to the preset safety threshold, and maintain the minimum guaranteed bandwidth of critical service traffic; After receiving the path switching failure warning message, the central cloud platform marks this path as a high-risk path in the updated global routing policy and excludes it from the fault switching path set for at least 24 hours.

[0012] Provided is a distributed large-scale layer-2 network intelligent routing system based on edge cloud nodes, including: Multiple edge cloud nodes and a central cloud platform, where the edge cloud nodes establish a two-way data connection with the central cloud platform through a pre-configured communication protocol; Each edge cloud node includes a network status detection unit, a local routing decision module, and a routing table storage unit. The network status detection unit periodically collects the link bandwidth utilization rate, cross-node transmission delay, network device port status, and node local computing resource load rate between adjacent nodes; The local routing decision module encapsulates the data collected by the network status detection unit into a structured data packet and transmits it to the central cloud platform through an encrypted tunnel; The central cloud platform deploys a global routing analysis engine. The global routing analysis engine generates a global routing policy based on the network status information of all received edge cloud nodes, and the global routing policy includes a path priority list, a traffic allocation weight coefficient, and a fault switching path set; After receiving the global routing policy issued by the central cloud platform, the local routing decision module updates the forwarding table entries in the routing table storage unit according to the path priority list and the traffic allocation weight coefficient, and synchronizes the updated forwarding table entries to adjacent edge cloud nodes through a pre-configured communication protocol; The network status detection unit real-time monitors the change amount of the link transmission delay between adjacent nodes. When the detected change amount exceeds the dynamic threshold pre-stored in the routing table storage unit, it triggers the local routing decision module to reconfigure the routing table storage unit based on the fault switching path set, and uploads the marked link status abnormal information to the central cloud platform through an encrypted tunnel; The global routing analysis engine of the central cloud platform starts policy recalculation according to the link status abnormal information and distributes the updated global routing policy to the associated affected edge cloud nodes.

[0013] The present invention at least includes the following beneficial effects: First, by periodically collecting network status information (such as link bandwidth utilization rate, transmission delay) by edge cloud nodes and combining with dynamic threshold monitoring, the system can quickly perceive network status changes. The central cloud platform generates a global routing policy in real time based on a neural network model of spatio-temporal feature fusion, effectively coping with sudden traffic fluctuations and uneven node loads, and reducing the problem of policy failure caused by data lag in traditional routing protocols.

[0014] Second, adopt a distributed architecture. Edge nodes execute local decisions (such as failure - switch path verification) to reduce the computing pressure on the central platform. At the same time, the central platform provides a global path priority list through a multi - objective optimization algorithm. The two work together to ensure both the real - time nature of local responses and the optimal allocation of network - wide resources, avoiding the latency and resource bottlenecks of a single centralized control.

[0015] Third, by combining a dynamic threshold trigger mechanism (such as the time - delay change amount exceeding a preset threshold) and a multi - path verification process (packet - loss rate detection, time - delay tolerance verification), the system can quickly switch to the pre - sorted candidate paths. The local degradation mode restricts non - critical service traffic in case of continuous switching failures, giving priority to core services, significantly shortening the service interruption time and enhancing the network fault - tolerance ability.

[0016] Fourth, balance multi - dimensional constraints such as bandwidth utilization, transmission delay, and node load balancing through a multi - objective optimization algorithm to maximize the network resource utilization efficiency. Encrypted tunnel communication (such as the TLS protocol) and the policy signature mechanism ensure the security of data transmission and policy synchronization, preventing malicious attacks or data tampering and enhancing the overall stability of the system.

[0017] Fifth, the dynamic threshold adjustment mechanism (such as based on the fusion of sliding - window statistics and global calibration parameters) enables the system to adapt to changes in different network environments. The modular design of edge nodes and the central platform supports flexible expansion and is applicable to various large - scale distributed network scenarios such as industrial Internet of Things and cloud computing centers.

[0018] Other advantages, objectives, and features of the present invention will be partially reflected by the following description and partially understood by those skilled in the art through the research and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the network status collection and upload process of the present invention; Figure 2 It is a structural diagram of the spatio - temporal feature fusion neural network model of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0020] The following further elaborates on the present invention with reference to the accompanying drawings, enabling those skilled in the art to implement it according to the description in the specification.

[0021] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation schemes are all conventional methods, and the reagents and materials, unless otherwise specified, can all be obtained from commercial channels; in the description of the present invention, the orientation or positional relationship indicated by the terms is based on the orientation or positional relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be construed as a limitation to the present invention.

[0022] As Figures 1 - 2 shown, the present invention provides a distributed large two-layer network intelligent routing method based on edge cloud nodes, including the following steps: Multiple edge cloud nodes respectively establish bidirectional data connections with the central cloud platform through pre-configured communication protocols. Each edge cloud node periodically collects the network status information of its own node. The network status information includes the link bandwidth utilization rate between adjacent nodes, the cross-node transmission delay, the network device port status, and the node local computing resource load rate; Each edge cloud node encapsulates the network status information into a structured data packet and uploads it to the central cloud platform through an encrypted tunnel; The central cloud platform generates a global routing policy based on the network status information of all edge cloud nodes by using a neural network model based on spatio-temporal feature extraction. The global routing policy includes a path priority list between edge cloud nodes, a traffic allocation weight coefficient, and a failure switchover path set; After each edge cloud node receives the global routing policy issued by the central cloud platform, it updates the forwarding table entries in the local routing table storage unit according to the path priority list and the traffic allocation weight coefficient, and synchronizes the updated forwarding table entries to adjacent nodes through a pre-configured communication protocol; Each edge cloud node real-time monitors the change amount of the link transmission delay between adjacent nodes. When it detects that the change amount of the link transmission delay exceeds the dynamic threshold pre-stored in the local routing decision module, it triggers the local routing decision module to reconfigure the local routing table based on the failure switchover path set in the global routing policy, and marks the link status abnormal information and uploads it to the central cloud platform through an encrypted tunnel; The central cloud platform starts to recalculate the global routing policy according to the link status abnormal information and distributes the updated global routing policy to the affected edge cloud nodes associated with the link status abnormal information.

[0023] In the above technical solution, the network status information periodically collected by the edge cloud node includes link bandwidth utilization rate, cross-node transmission delay, network device port status, and node local computing resource load rate. The collection period can be set as follows: the link bandwidth utilization rate is once every 5 minutes, the cross-node transmission delay is once every 30 seconds, the port status is once every 1 minute, and the computing resource load rate is once every 2 minutes. The bandwidth utilization rate can be obtained by sending ICMP probe packets and calculating the arithmetic mean of the last 10 probe results, and at the same time obtaining the real-time throughput counter value of the port through the SNMP protocol. The cross-node transmission delay is tested by bidirectional transmission of UDP benchmark packets, and the mean value of the time difference is calculated as the effective delay value.

[0024] For device selection, the edge cloud node can adopt a server with a multi-core processor, a network switch that supports the SNMP and ICMP protocols, and an embedded gateway that supports UDP testing. In terms of materials, the server chassis can be made of aluminum alloy, and the internal circuit board uses FR-4 epoxy resin substrate. The edge cloud node is usually deployed at the network edge position close to the terminal device, such as the base station side or inside the factory workshop.

[0025] During the working process, the network status detection unit executes the data collection task according to the preset period. After the preprocessing module eliminates the abnormal sampling points, it attaches a millisecond-level time stamp to the standardized data and uploads it to the central cloud platform through an encrypted tunnel.

[0026] Global routing policy generation and distribution: The central cloud platform uses a neural network model based on spatio-temporal feature extraction to generate the global routing policy. The convolutional neural network branch inputs the physical topology connection relationship and extracts spatial correlation features through three convolutional layers; the long short-term memory network branch inputs the delay sequence and extracts the time series features of delay fluctuations. The attention mechanism layer dynamically assigns weights to the two types of features, and the fully connected layer outputs the path priority list, traffic allocation weights, and the set of failover paths.

[0027] For device selection, the central cloud platform can adopt a high-performance server cluster that supports GPU acceleration, and the neural network model can be implemented based on the TensorFlow or PyTorch framework. In terms of materials, the server heat dissipation module can be a combination of copper heat pipes and aluminum alloy fins. The central cloud platform is deployed in the data center computer room and is connected to the edge cloud node through redundant fiber optic links.

[0028] During the working process, the global routing analysis engine receives the data of all edge cloud nodes, generates policies through a multi-objective optimization algorithm, and encapsulates the policies in JSON format for distribution. The sorting weights of the path priority list are solved by the non-dominated sorting genetic algorithm, and the constraint conditions include the link bandwidth utilization rate threshold (such as 85%), the upper limit of transmission delay (such as 100ms), and the node load balancing coefficient (such as 0.3).

[0029] Local Routing Table Update and Failover: The edge cloud node monitors the change in link transmission delay, and the dynamic threshold is initialized to T0 = μ + 3σ (for example, when μ = 50ms and σ = 10ms, T0 = 80ms). When the detected change in delay exceeds the threshold, the failover process is triggered. Candidate path verification includes packet loss rate detection (the preset packet loss threshold is 2%) and round-trip delay verification (the preset tolerance range is ±10ms). If the verification passes, the traffic of the faulty link is split to the candidate path according to a preset ratio (for example, 70%); if it fails, the path priority is dynamically adjusted and the sub-optimal path is tried.

[0030] For device selection, the local routing decision module can adopt a software-defined network (SDN) controller that supports the OpenFlow protocol, and the routing table storage unit can select a high-speed NVMe solid-state drive. In terms of materials, the device shell can be made of flame-retardant ABS plastic. When deploying the edge cloud node, it needs to be physically topologically directly connected to adjacent nodes, for example, through optical fibers or Cat6a network cables.

[0031] During operation, the local routing decision module updates the forwarding table entries after receiving the policy and synchronizes the change information to adjacent nodes through the BGP protocol. After the path is switched, the network status detection unit continuously monitors the transmission quality. If the delay fluctuation exceeds 50% of the dynamic threshold (for example, 40ms), a secondary switch is triggered and the failover path set is updated.

[0032] This method realizes real-time perception and global optimization of the network state through periodic acquisition and dynamic policy adjustment, reducing service interruptions caused by link congestion or node overload. The failover mechanism combines multi-path verification to improve the network fault tolerance. The neural network model that fuses spatio-temporal features enhances the adaptability and accuracy of routing decisions, ultimately improving the stability and transmission efficiency of large-scale layer-2 networks.

[0033] In another technical solution, the neural network model based on spatio-temporal feature extraction includes a convolutional neural network branch and a long short-term memory network branch arranged in parallel, where, The convolutional neural network branch takes the physical topology connection relationship of the edge cloud node as the input, and extracts the spatial correlation features between adjacent nodes through three convolutional layers. The spatial correlation features include the link bandwidth fluctuation pattern between nodes and the port state change trend; The long short-term memory network branch takes the cross-node transmission delay sequence as the input, and extracts the delay fluctuation time series features within a preset time window through a bidirectional cyclic structure. The delay fluctuation time series features include periodic congestion patterns and sudden abnormal delay segments; The neural network model dynamically assigns weights to the spatial correlation features and the delay fluctuation time series features through the attention mechanism layer to generate a network state vector that fuses spatio-temporal features; The fully connected layer of the neural network model calculates and generates a global routing policy based on the network state vector. The generation process of the path priority list includes performing multi-objective optimization operations on the network state vector that fuses spatio-temporal features. The constraint conditions of the multi-objective optimization operations include the link bandwidth utilization threshold, the upper limit of transmission delay, and the node load balancing coefficient.

[0034] In the above technical solution, the convolutional neural network branch inputs the physical topology connection relationship of the edge cloud nodes and adopts a three-layer convolutional layer structure. The number of convolutional kernels in each layer can be 32, 64, and 128 respectively, and the kernel size can be set to 3×3. The extraction of spatial correlation features includes the link bandwidth fluctuation pattern (such as the bandwidth change rate within 5 minutes) and the port status change trend (such as the port enable / disable frequency). The long short-term memory network branch inputs the cross-node transmission delay sequence, the time window can be set to 60 seconds, and the number of hidden layer units can be configured to 128. The extraction of temporal features includes the periodic congestion pattern (such as the peak delay per hour) and the bursty abnormal delay segment (such as the segment where the delay suddenly increases by more than 50ms).

[0035] In terms of device selection, the neural network model can be deployed on a GPU server that supports CUDA acceleration, and the implementation of the convolutional layer and the LSTM layer can be based on the TensorFlow framework. In terms of materials, the server motherboard can use a glass fiber-reinforced epoxy resin substrate, and the radiator can use an aluminum alloy material. The neural network module needs to be deployed in the high-performance computing nodes of the central cloud platform and is connected to the data storage unit through the PCIe bus.

[0036] During the working process, the physical topology data is input into the CNN branch in the form of an adjacency matrix, and the delay sequence data is input into the LSTM branch according to the time step. The CNN branch extracts the spatial correlation features between nodes through the convolutional layer, and the LSTM branch captures the delay fluctuation trend through the bidirectional cyclic structure. The output feature vectors are respectively transmitted to the attention mechanism layer.

[0037] The attention mechanism layer performs dynamic weight allocation on the spatial correlation features output by the CNN and the delay fluctuation temporal features output by the LSTM. The weight allocation can be calculated based on the cosine similarity of the feature vectors. For example, the spatial feature weight coefficient can be set to 0.6, and the temporal feature weight coefficient is 0.4. The dimension of the fused network state vector can be set to 256 dimensions and is mapped to a low-dimensional feature representation through the fully connected layer.

[0038] For device selection, the attention mechanism module can be implemented based on open-source deep learning libraries (such as the MultiheadAttention layer in PyTorch). In terms of materials, copper wires can be selected for the data bus, and RJ45 modules with gold-plated contacts can be selected for the interface connectors. This module needs to be integrated into the global routing analysis engine of the central cloud platform and interact with the neural network branches through high-speed memory channels.

[0039] During operation, the attention mechanism calculates the correlation scores between spatial features and temporal features, and weights and sums the feature vectors according to the score ratio. The fused network state vector undergoes normalization processing through a fully connected layer and is output to the multi-objective optimization module. During the feature fusion stage, it is necessary to ensure timestamp alignment to avoid feature misalignment caused by differences in data acquisition cycles.

[0040] The decision variables for multi-objective optimization operations include the bandwidth allocation ratio factor α (e.g., α1 = 0.7, α2 = 0.3), the delay compensation factor β (e.g., β1 = 1.2, β2 = 0.8), and the node load balancing weight γ (e.g., γ1 = 0.5, γ2 = 0.5). The constraint conditions include the link bandwidth utilization threshold (85%), the upper limit of transmission delay (100 ms), and the node load balancing coefficient (0.3). The optimization algorithm uses the non-dominated sorting genetic algorithm with elitist preservation strategy (NSGA-II), and the population size can be set to 200, with 100 generations of iterations.

[0041] For device selection, the multi-objective optimization module can run on a server equipped with a multi-core CPU, and the crossover and mutation operations of the genetic algorithm can be implemented based on the DEAP framework. In terms of materials, DRAM chips with DDR4 specifications can be selected for the server memory, and TLC NAND flash can be used for the storage unit. This module needs to be deployed in the policy calculation node of the central cloud platform and exchange data with the neural network module through shared memory.

[0042] During operation, the initial solution set is generated according to the historical routing strategy. The two-point crossover method is used for the crossover operation, and the crossover probability is dynamically adjusted according to the path overlap rate (e.g., when the overlap rate is higher than 80%, the probability drops to 0.3). The mutation operation replaces the current path with an adjacent candidate path in the physical topology (e.g., replacing path A with path B). The Pareto front solution set is screened by the crowding distance, and finally, the individual with the largest crowding distance is selected to generate the path priority list.

[0043] This method extracts spatial topology features and time-delay fluctuation rules through the CNN and LSTM branches respectively, and combines with the dynamic weight fusion mechanism to enhance the adaptability of the routing strategy to the network state. The multi-objective optimization algorithm generates the global optimal path under multiple constraints such as bandwidth, time-delay, and load balancing, reducing the risk of link congestion. The synergistic effect of the attention mechanism and the genetic algorithm improves the routing decision-making efficiency, which is suitable for the real-time traffic scheduling requirements in large-scale edge network environments.

[0044] In another technical solution, reconfiguring the local routing table storage unit based on the set of failover paths includes the following steps: When it is detected that the change amount of the link transmission time-delay exceeds the dynamic threshold, extract the pre-sorted candidate path sequence from the set of failover paths, and the candidate path sequence is sorted by priority based on the historical transmission success rate and the path hop count; Perform link bandwidth availability detection and transmission time-delay verification on the candidate path with the highest priority, where the link bandwidth availability detection is achieved by sending probe packets to the target node and calculating the packet loss rate, and the transmission time-delay verification is achieved by measuring the difference between the round-trip time-delay of the probe packets and the preset time-delay threshold; If the packet loss rate of the candidate path is lower than the preset packet loss threshold and the round-trip time-delay difference is within the preset tolerance range, then split the current faulty link traffic to the candidate path according to the preset ratio, and generate a forwarding table entry update instruction including the split ratio and the path identifier; If the candidate path verification fails, trigger the dynamic weight adjustment of the candidate path sequence, recalculate the path priority according to the real-time load rate of adjacent nodes, and jump to the link bandwidth availability detection step to perform the verification of the sub-optimal candidate path; After the path switching is completed, the local routing decision module associates and stores the updated forwarding table entry with the failover timestamp, and sends a routing table synchronization request to the adjacent edge cloud node through the pre-configured communication protocol, and the synchronization request includes the hash value and version number of the changed forwarding table entry; The network state detection unit continuously monitors the transmission quality indicators of the switched path. When the transmission time-delay fluctuation amount exceeds 50% of the dynamic threshold for three consecutive detection cycles, trigger the secondary path switching process and update the priority weight parameters of the set of failover paths.

[0045] In the above technical solution, when it is detected that the change amount of the link transmission time-delay exceeds the dynamic threshold (for example, 80 ms), extract the pre-sorted candidate path sequence from the set of failover paths. The priority sorting of the candidate paths is based on the historical transmission success rate (for example, paths with a success rate ≥ 95% are prioritized) and the path hop count (for example, paths with a hop count ≤ 3 are prioritized). The historical transmission success rate is calculated by statistically analyzing the packet loss rate of the path in the past 24 hours, and the path hop count is obtained in real time from the network topology database.

[0046] For device selection, the failure switchover path set can be stored in a memory database that supports the Redis protocol, and the path priority calculation module can be installed on an embedded processor with multi-threaded processing capabilities. In terms of materials, the storage chips of the memory database can be selected with the LPDDR4 specification, and the processor packaging substrate can be selected with FR-4 epoxy resin material. This module needs to be deployed in the local routing decision module of the edge cloud node and connected to the network status detection unit through an internal bus.

[0047] During operation, the local routing decision module triggers the extraction of candidate paths based on dynamic thresholds, and preferentially selects paths with high historical transmission success rates and few hops. The candidate path sequence is stored in the routing table storage unit in JSON format, and the data integrity is ensured through a hash algorithm.

[0048] Perform link bandwidth availability detection on the candidate path with the highest priority. Calculate the packet loss rate by sending probe packets (the preset packet loss threshold is 2%), and measure the round-trip delay at the same time (the preset tolerance range is ±10 ms). If the verification passes (for example, the packet loss rate is 1.5% and the delay difference is 8 ms), split the current failed link traffic to this path according to a preset ratio (for example, 70%), and generate a forwarding table entry update instruction that includes the splitting ratio and the path ID.

[0049] For device selection, the sending of probe packets can be based on a hardware probe that supports the IP SLA protocol, and a high-precision clock synchronization chip can be selected for delay measurement. In terms of materials, the probe housing can be selected with flame-retardant ABS plastic, and the clock chip packaging can be selected with a ceramic substrate. The detection module needs to be integrated into the network interface card of the edge cloud node and communicate with the routing decision module through a PCIe interface.

[0050] During operation, the verification module sends UDP probe packets to the target node, counts the packet loss rate, and calculates the average delay. If the result meets the threshold, the routing table storage unit updates the forwarding table entry and synchronizes the change information to adjacent nodes through the BGP protocol. The splitting ratio is dynamically adjusted according to the remaining bandwidth of the path. For example, when the remaining bandwidth ≥ 50 Mbps, 70% of the traffic is split.

[0051] If the verification of the candidate path fails (for example, the packet loss rate is 3% or the delay difference is 15 ms), trigger the dynamic adjustment of the path priority. Recalculate the path priority according to the real-time load rate of adjacent nodes (for example, the weight of nodes with a load rate ≥ 80% is reduced), and try the sub-optimal candidate path. If the verification fails three times in a row, roll back the failed link traffic to the original path and limit the forwarding rate of non-critical service traffic to a preset safety threshold (for example, 10 Mbps).

[0052] For device selection, the load rate collection can be based on a network traffic analyzer that supports the NetFlow protocol, and the rate limit can be implemented on a switch chip that supports QoS policies. In terms of materials, the analyzer housing can be made of aluminum alloy, and the heat sink can be made of copper heat pipes. This module needs to be deployed in the traffic control unit of the edge cloud node and interact with the routing decision module through shared memory.

[0053] During operation, the dynamic adjustment module reorders the candidate paths according to the node load data and triggers the verification of sub-optimal paths. During the fallback operation, the forwarding table entries are restored to the state before the switch, and an alarm message is sent to the central cloud platform through the SNMP protocol. The transmission quality of the switched path is continuously monitored by the network status detection unit. If the delay fluctuation amount exceeds 50% of the dynamic threshold (e.g., 40ms) within three consecutive cycles, a secondary path switch is triggered and the priority weights of the failed switch path set are updated.

[0054] This method ensures the reliability and timeliness of the failover through pre-ordering candidate paths and a multi-dimensional verification mechanism. The dynamic priority adjustment avoids secondary congestion caused by node overload, and the fallback mechanism and traffic limit reduce the impact of service interruption. The continuous monitoring and secondary switching functions further optimize the path stability, which is suitable for the fast fault recovery requirements in a highly dynamic network environment.

[0055] In another technical solution, the execution process of the multi-objective optimization operation includes the following steps: Map the network state vector that fuses spatio-temporal features into a decision variable set including the bandwidth allocation ratio factor α i , the delay compensation factor β j and the node load balancing weight γ k , where i corresponds to the path number, j corresponds to the delay level identifier, and k corresponds to the node identifier; Construct an objective function set including the first objective function, the second objective function, and the third objective function. The first objective function is f1 = Σ(α i ×C i ), which is used to maximize the effective bandwidth utilization rate. The second objective function is f2 = Σ(β j ×D j ), which is used to minimize the transmission delay offset. The third objective function is f3 = Σ(γ k ×L k ), which is used to balance the node load difference degree, where C i represents the available bandwidth capacity of path i, D j represents the delay compensation amount of path j, and L k represents the load deviation value of node k; Set up a system of equations including the following constraint conditions: Σα i≤ Pre-defined link bandwidth utilization threshold, Σβ j ≤ Pre-defined upper limit of transmission delay, Σγ k ≤ Pre-defined node load balancing coefficient; Solve the objective function set using a non-dominated sorting genetic algorithm with an elitist retention strategy, which includes: Generate an initial solution set as the genetic algorithm population according to the path priority list, perform an adaptive crossover operation based on path similarity, the crossover probability is dynamically adjusted according to the path overlap rate of individuals in the solution set, perform an adjacent path replacement mutation operation, and replace the current path with a candidate path directly adjacent in the physical topology; Perform the following processing on the solution set generated in each iteration: Delete individuals that violate the system of constraint equations, perform non-dominated sorting on the remaining individuals and calculate the crowding distance, and retain the top pre-set number of individuals with the highest sorting level and the largest crowding distance; Select the individual with the largest crowding distance from the Pareto front solution set of the final iteration result as the optimal solution, and extract the corresponding α i , β j and γ k The parameter combination generates the sorting weight value of the path priority list.

[0056] In the above technical solution, the network state vector integrating spatio-temporal features is mapped to a bandwidth allocation ratio factor α (for example, α1 = 0.6, α2 = 0.4), a delay compensation factor β (for example, β1 = 1.1, β2 = 0.9), and a node load balancing weight γ (for example, γ1 = 0.5, γ2 = 0.5). The objective functions include maximizing the effective bandwidth utilization (f1 = Σα i ×C i , C i is the available bandwidth capacity of the path), minimizing the transmission delay offset (f2 = Σ(β j ×D j ), D j is the path delay compensation amount), and balancing the node load difference degree (f3 = Σ(γ k ×L k ), L k is the node load deviation value).

[0057] In terms of device selection, the mapping calculation can run on a multi-core CPU server supporting floating-point operation acceleration, and the objective function calculation module can be implemented based on the NumPy library. In terms of materials, the server motherboard can use an FR-4 epoxy resin substrate, and the memory module can be equipped with DRAM chips of the DDR4 specification. This module needs to be deployed in the policy calculation node of the central cloud platform and connected to the neural network module through a high-speed data bus.

[0058] During the working process, the network state vector is input into the variable mapping module after being normalized by the fully connected layer, and a set of decision variables is generated according to the predefined weight rules. The objective function calculates the bandwidth, delay, and load metrics of each path in real time, and realizes multi-objective correlation analysis through matrix operations.

[0059] The constraint conditions include the link bandwidth utilization threshold (85%), the upper limit of transmission delay (100 ms), and the node load balancing coefficient (0.3). The optimization algorithm uses the non-dominated sorting genetic algorithm with elitist retention strategy (NSGA-II). The population size can be set to 200, and the number of iterations is 100 generations. The crossover operation uses the two-point crossover method, and the crossover probability is dynamically adjusted according to the path overlap rate (for example, when the overlap rate ≥ 80%, the probability drops to 0.3). The mutation operation is the replacement of adjacent paths (for example, replacing path A with topologically adjacent path B).

[0060] In terms of device selection, the genetic algorithm solution can run on a computing cluster that supports multi-threaded parallelism, and the crossover and mutation operations can be implemented based on the DEAP framework. In terms of materials, copper heat sinks and aluminum alloy chassis can be selected for the cluster nodes, and TLC NAND flash memory can be used for the storage unit. This module needs to be integrated into the global routing analysis engine of the central cloud platform and interact with the objective function module through shared memory.

[0061] During the working process, the initial solution set is generated according to the historical routing strategy. Each generation of the population generates new individuals through crossover and mutation, and the solutions that violate the constraint conditions are deleted. The remaining individuals are sorted according to the non-dominated level, and the crowding distance is used for priority screening within the same level. Finally, the top 50 optimal individuals are retained for the next generation of iteration.

[0062] The crowding distance of the Pareto front solution set of the final iteration result is calculated, and the individual with the largest distance is selected as the optimal solution (for example, individuals with a crowding distance ≥ 0.8). The α, β, γ parameter combinations of the optimal solution are extracted and converted into the sorting weight values of the path priority list (for example, α1 = 0.6 corresponds to a weight of 60% for path 1). The path priority list is sent to the edge cloud nodes in JSON format.

[0063] In terms of device selection, the crowding distance calculation can be based on a mathematical acceleration library that supports vectorized operations, and the parameter extraction module can be installed on a processor with a high-precision floating-point arithmetic unit. In terms of materials, ceramic substrates can be selected for the processor package, and PCIe slots with gold-plated contacts can be selected for the data interface. This module needs to be deployed in the policy distribution node of the central cloud platform and connected to the genetic algorithm module through a high-speed network.

[0064] During the working process, the Pareto solution set is sorted by non-dominated sorting, and the congestion distance of each solution is calculated. The optimal solution parameters are converted into path weights through linear interpolation, and executable routing policy instructions are generated. The policy instructions are encrypted and distributed to the associated edge cloud nodes through optical fiber links.

[0065] This method uses a multi-objective optimization algorithm to balance multi-dimensional constraints such as bandwidth utilization, transmission delay, and node load balancing to improve the global rationality of the routing strategy. The NSGA-II algorithm combines the elite retention strategy to ensure the diversity and convergence efficiency of the solution set. The generation process of the path priority list takes into account both the real-time network status and the rules of historical data, and is suitable for dynamic traffic scheduling needs in large-scale distributed network environments.

[0066] In another technical solution, the process of determining and updating the dynamic threshold in the local routing decision module includes the following steps: The local routing decision module is initialized and loaded with the baseline threshold T0, which is calculated based on the normal distribution parameters of the historical transmission delay data. The calculation formula is T0=μ+3σ, where μ is the delay mean and σ is the delay standard deviation. The network status detection unit counts the sliding window average of the link transmission delay between adjacent nodes every five minutes. The window size is the most recent 30 sampling points. When the deviation between the sliding window average and the baseline threshold T0 exceeds 20%, the threshold adaptive adjustment is triggered. The threshold adaptive adjustment process performs the following operations: Collect the delay data set of each hour within 24 hours before the current time, remove the data points with more than three times the standard deviation, recalculate the mean μ' and standard deviation σ', generate the updated dynamic threshold T1=μ'+2.5σ' and overwrite and store it in the local routing decision module; If, between two threshold adjustment cycles, a single delay change is detected that exceeds 150% of the current dynamic threshold, the emergency threshold correction process is immediately initiated, the current dynamic threshold is temporarily set to 80% of the abnormal delay value and marked as pending calibration; When the edge cloud node receives the updated global routing policy issued by the central cloud platform, it synchronously obtains the threshold calibration parameters contained in the global routing policy, and fuses the local dynamic threshold with the calibration parameters through the weighted average algorithm to generate the final effective dynamic threshold.

[0067] In the above technical solution, when the local routing decision module is initialized, the baseline threshold T0 = μ + 3σ is loaded (for example, when μ = 50ms and σ = 10ms, T0 = 80ms). The historical delay data is used to generate the normal distribution parameters by analyzing the transmission records of the past 7 days. The network status detection unit statistically calculates the sliding window average of the link delay between adjacent nodes every 5 minutes, and the window size is the last 30 sampling points (for example, sampling once every 30 seconds, and the window covers 15 minutes of data). When the deviation of the sliding window average from T0 exceeds 20% (for example, triggering an adjustment when the average value becomes 96ms).

[0068] For device selection, an embedded database supporting the SQLite protocol can be selected for delay data storage, and a microcontroller with a floating-point arithmetic unit can be used for sliding window calculation. In terms of materials, an eMMC flash memory can be selected for the database storage chip, and an FR-4 epoxy resin material can be selected for the microcontroller substrate. This module needs to be deployed in the local routing decision module of the edge cloud node and is connected to the network interface unit through the SPI bus.

[0069] During the working process, in the initialization stage, the delay data is loaded from the historical database, μ and σ are calculated, and after generating T0, it is written into the non-volatile memory. When the sliding window statistical period is triggered, the latest 30 delay sampling points are collected to calculate the arithmetic average, and the deviation is compared with T0.

[0070] When the deviation of the sliding window average from T0 exceeds 20%, the threshold adaptive adjustment is triggered. The delay data at the whole hour of the previous 24 hours (abnormal values exceeding 3σ are excluded) is collected, μ' and σ' are recalculated, and the updated dynamic threshold T1 = μ' + 2.5σ' is generated (for example, when μ' = 55ms and σ' = 8ms, T1 = 75ms). If the single delay change amount exceeds 150% of the current threshold (for example, when the current threshold T1 = 75ms, a delay of 112.5ms is detected), the emergency correction is immediately started, the threshold is temporarily set to 80% of the abnormal value (for example, 90ms) and marked as the to-be-calibrated state.

[0071] For device selection, the outlier removal algorithm can be based on a statistical processor supporting Z-score calculation, and the emergency correction module can be equipped with a programmable logic controller (PLC). In terms of materials, an aluminum alloy material can be selected for the heat sink of the statistical processor, and a flame-retardant PC plastic can be selected for the PLC housing. This module needs to be integrated into the threshold management unit of the edge cloud node and interacts with the sliding window statistical module through the internal data bus.

[0072] During operation, in the adaptive adjustment phase, 24-hour data is extracted from the latency database, and the threshold parameters are recalculated after abnormal filtering. During emergency correction, the temporary threshold takes effect and overwrites the original value, and at the same time, a threshold anomaly warning is sent to the central cloud platform. The temporary threshold is reset in the next adjustment cycle or when the global policy is issued.

[0073] When the edge cloud node receives the global routing policy issued by the central cloud platform, it synchronously obtains the threshold calibration parameters included in the policy (for example, the calibration coefficient k = 0.8). The local dynamic threshold is fused with the calibration parameters through a weighted average algorithm to generate the final threshold T final = w1 × T local + w2 × T center (for example, weights w1 = 0.6, w2 = 0.4). The weights of the calibration parameters are dynamically allocated according to the success rate of the node's historical policy execution (for example, when the success rate ≥ 90%, w1 is increased to 0.7). Among them, T final represents the finally effective dynamic threshold, which is generated by fusing the local threshold and the central platform calibration parameters, T local represents the dynamic threshold calculated locally by the edge cloud node, usually based on the statistical analysis of historical latency data, T center represents the global calibration parameters issued by the central cloud platform, which are optimized based on the network-wide status data. w1 and w2 respectively represent the weight coefficients of the local threshold and the central platform parameters, used to balance the influence of both on the final threshold.

[0074] In terms of device selection, a coprocessor supporting matrix operations can be installed for weighted average calculation, and an EEPROM chip can be selected for storing the calibration parameters. In terms of materials, the package of the coprocessor can use a ceramic substrate, and the pins of the storage chip can use gold-plated copper alloy. This module needs to be deployed in the policy execution unit of the edge cloud node and communicate with the central cloud platform through a TLS encrypted link.

[0075] During operation, the global policy parsing module extracts the calibration parameters and calculates the weighted average through the floating-point arithmetic unit. The final threshold T final is written into the routing table storage unit and synchronized to the threshold management modules of adjacent nodes. If the calibration parameters conflict with the local threshold, the central platform parameters are given priority to perform the overwrite.

[0076] This method realizes the accurate perception of the network latency status through baseline threshold initialization and sliding window dynamic statistics. The adaptive adjustment mechanism reduces the problem of threshold failure caused by network environment changes, and the emergency correction function improves the processing ability of sudden anomalies. The global policy synchronization ensures the consistency of the threshold with the central platform policy, enhancing the coordination and reliability of threshold management in a large-scale network environment.

[0077] In another technical solution, the process of periodically collecting network status information includes the following steps: The network status detection unit configures multi-dimensional data collection tasks according to a preset collection strategy. The collection strategy includes a link bandwidth utilization collection period of 5 minutes, an inter-node transmission delay collection period of 30 seconds, a network device port status collection period of 1 minute, and a node local computing resource load rate collection period of 2 minutes; When performing link bandwidth utilization collection, send ICMP probe packets to adjacent nodes and calculate the arithmetic mean of the bandwidth occupancy rates of the last 10 probe results. At the same time, obtain the real-time throughput counter value of the network device port through the SNMP protocol; When performing inter-node transmission delay collection, use UDP benchmark packets for two-way transmission testing, calculate the mean of the time difference between the sending time and the receiving confirmation time as the effective delay value, and record the maximum delay fluctuation during the testing process; The collected raw data is processed by the preprocessing module to perform the following operations: Eliminate abnormal sampling points exceeding three times the standard deviation range, perform one-hot encoding conversion on the binary flag bits in the port status data, and convert index data with different dimensions into eigenvalue of the same magnitude through the Z-score normalization method; When the network status detection unit monitors that the change rate of two adjacent collection results of any index exceeds the preset sensitivity threshold, automatically shorten the collection period of this index to 50% of the original period, and start intensive sampling for three consecutive periods; When storing all preprocessed network status information data, append a timestamp accurate to the millisecond level, and perform clock synchronization calibration with the central cloud platform through the NTP protocol.

[0078] In the above technical solution, the network status detection unit configures a link bandwidth utilization collection period of 5 minutes, an inter-node transmission delay collection period of 30 seconds, a port status collection period of 1 minute, and a node load rate collection period of 2 minutes. The link bandwidth utilization is obtained by sending ICMP probe packets (such as 10 packets per second) to adjacent nodes and calculating the arithmetic mean of the last 10 probe results. At the same time, obtain the real-time throughput counter value of the switch port (such as a 64-bit counter) through the SNMP protocol.

[0079] In terms of device selection, the network status detection unit can be equipped with an embedded processor supporting multi-threaded scheduling, and the ICMP detection function can be based on a network interface card supporting IPv4 / v6 protocols. In terms of materials, the processor substrate can be made of FR-4 epoxy resin substrate, and the interface card housing can be made of aluminum alloy. This module needs to be deployed in the data collection unit of the edge cloud node and connected to the network interface through a PCIe slot.

[0080] During the operation, the acquisition task scheduler triggers each metric acquisition thread according to a preset period. ICMP probe packets are sent to the target node at fixed intervals, and the SNMP protocol reads port throughput data through OID identifiers. The acquisition results are temporarily stored in a circular buffer, waiting for processing by the preprocessing module.

[0081] For cross-node transmission delay acquisition, UDP benchmark test packets are used for two-way transmission tests. The average time difference (e.g., the average of 10 tests) between the sending and receiving confirmation times is calculated, and the maximum delay fluctuation amount (e.g., a single fluctuation exceeding 50 ms) is recorded. The preprocessing module performs outlier removal (sampling points exceeding 3 times the standard deviation) on the original data, one-hot encoding conversion of the port status binary flag bits (e.g., "port enabled" is encoded as [1,0]), and converts metrics with different dimensions (such as bandwidth in Mbps and delay in ms) to a unified magnitude through Z-score standardization.

[0082] In terms of device selection, UDP tests can be based on network cards supporting hardware timestamps, and the preprocessing module can run on a coprocessor with a floating-point operation accelerator. In terms of materials, the PCB board of the network card can be made of fiberglass-reinforced substrate, and the heat sink of the coprocessor can be made of copper. This module needs to be integrated into the data processing unit of the edge cloud node and directly interact with the acquisition unit through the DMA channel.

[0083] During the operation, the UDP test packets carry nanosecond-level timestamps, and the receiving end calculates the round-trip delay and filters out outliers caused by network jitter (e.g., data points with a delay > 200 ms). After the preprocessing module encodes the port status, the normalization process maps the bandwidth utilization rate to the [-1,1] interval and the delay value to the [0,1] interval for unified processing by subsequent analysis modules.

[0084] When the change rate of the acquisition results of any metric exceeds a preset sensitivity threshold (e.g., the change rate of bandwidth utilization rate ≥ 15%), the acquisition period of this metric is automatically shortened to 50% of the original period (e.g., adjusted from 5 minutes to 2.5 minutes), and intensive sampling for three consecutive periods is started. All preprocessed data is appended with a timestamp accurate to the millisecond level (e.g., 2023-10-01T12:34:56.789), and clock synchronization and calibration are performed with the central cloud platform through the NTP protocol (error ≤ 1 ms).

[0085] In terms of device selection, clock synchronization can be based on a hardware clock module supporting the PTP protocol, and dynamic period adjustment can be carried by a programmable timer chip. In terms of materials, the crystal oscillator of the clock module can be made of quartz, and the pins of the timer can be made of gold-plated copper alloy. This module needs to be deployed in the clock management unit of the edge cloud node and connected to the data storage unit through the I2C bus.

[0086] During operation, the change rate detection module calculates the metric difference in real time (e.g., (current value - previous value) / previous value). If it exceeds the threshold, a periodic adjustment instruction is triggered. During intensive sampling, the sampling frequency is increased to 50% of the original period and restored to the default setting after three periods. The timestamp generator synchronizes time through an NTP server to ensure the time consistency of cross-node data.

[0087] This method improves the timeliness and accuracy of network status data through multi-dimensional periodic acquisition and dynamic adjustment mechanisms. Standardization and anomaly filtering processing reduce noise interference and enhance the reliability of subsequent analysis. Clock synchronization and timestamp calibration ensure the temporal consistency of cross-node data, and are applicable to large-scale distributed network monitoring scenarios sensitive to time.

[0088] In another technical solution, the processing after the failure of the sub-optimal candidate path verification further includes the following steps: When the candidate path verification fails three consecutive times, the local routing decision module performs the following operations: Roll back the traffic of the current faulty link to the path before switching, and generate a path rollback instruction to update the forwarding table entry; Send a path switching failure warning message to the central cloud platform. The warning message includes the failed path identifier, the verification failure reason code, and the current node load status; Start the local degradation mode, limit the forwarding rate of non-critical service traffic to the preset security threshold, and at the same time maintain the minimum guaranteed bandwidth for critical service traffic; After receiving the path switching failure warning message, the central cloud platform marks the path as a high-risk path in the updated global routing policy and excludes it from the fault switching path set for at least 24 hours.

[0089] In the above technical solution, when the candidate path verification fails three consecutive times (e.g., packet loss rate ≥ 3% or delay difference ≥ 15 ms), the local routing decision module rolls back the traffic of the faulty link to the path before switching and generates a path rollback instruction to update the forwarding table entry. The warning message includes the failed path identifier (e.g., path ID = 0x5A), the verification failure reason code (e.g., code 0x01 indicates delay overrun), and the current node load status (e.g., CPU load rate 75%). The warning message is uploaded to the central cloud platform through a TLS encrypted tunnel, and the upload frequency can be set to once per second until a platform confirmation response is received.

[0090] For device selection, the path rollback operation can be based on an SDN controller that supports the OpenFlow protocol, and the alarm information encapsulation can be carried by an embedded processor that supports JSON serialization. In terms of materials, the circuit board of the controller can use FR-4 epoxy resin substrate, and the processor package can use ceramic substrate. This module needs to be deployed in the local routing decision module of the edge cloud node and connected to the routing table storage unit through a high-speed bus.

[0091] During operation, the rollback instruction triggers the forwarding table entries to be restored to the historical version and synchronizes the changes to adjacent nodes through the BGP protocol. After the alarm information is generated, a timestamp and a digital signature are added to ensure data integrity and source credibility.

[0092] When starting the local downgrade mode, the forwarding rate of non-critical service traffic (such as video streaming) is limited to a preset security threshold (such as 10Mbps), and critical service traffic (such as VoIP) maintains a minimum guaranteed bandwidth (such as 2Mbps). The rate limit is achieved through QoS policies. For example, the token bucket algorithm is used to control traffic shaping, and the token generation rate can be set to 1Mbps / ms.

[0093] For device selection, traffic control can be based on a switch chip that supports the DiffServ protocol, and the token bucket algorithm can be carried by a network processor with a hardware counter. In terms of materials, the heat sink of the switch chip can use aluminum alloy material, and the processor pins can use gold-plated copper alloy. This module needs to be integrated into the traffic management unit of the edge cloud node and communicate with the routing decision module through the PCIe interface.

[0094] During operation, the traffic classification module distinguishes service types according to the DSCP field. Non-critical service traffic is marked as low priority and enters the rate-limiting queue. Critical service traffic directly enters the high-priority queue to ensure that the minimum bandwidth is not affected by the downgrade. The rate limit parameters are dynamically configured through the CLI command line or the SNMP protocol.

[0095] After receiving the alarm information, the central cloud platform marks the failed path as a high-risk path (such as marked in red) in the updated global routing policy and excludes it from the failure switchover path set for at least 24 hours. During the exclusion period, this path does not participate in the candidate path priority sorting until it is manually unmarked or automatically expires.

[0096] For device selection, high-risk path management can be based on a network topology management system that supports a graph database, and the path exclusion policy can be carried by a rule engine module. In terms of materials, the database storage medium can use SSD solid-state drives, and the radiator of the rule engine can use copper heat pipes. This module needs to be deployed in the policy management node of the central cloud platform and connected to the global routing analysis engine through Gigabit Ethernet.

[0097] During the working process, the policy management node parses the path ID in the alarm information and updates the path status attributes in the graph database. The dynamic update task of the failover path set is executed every 5 minutes to ensure that high-risk paths are excluded in time. After the exclusion period expires, the path automatically rejoins the candidate pool and participates in subsequent verification.

[0098] This method reduces the impact of business interruption through path fallback and alarm mechanism to ensure the continuity of core services. The local degradation mode prioritizes key services when resources are limited, improving the hierarchical guarantee capability of network services. The dynamic exclusion of high-risk paths reduces the risk of repeated switching failures and enhances the robustness of the fault recovery process, which is suitable for edge network scenarios with high reliability requirements.

[0099] The present invention provides a distributed large layer 2 network intelligent routing system based on edge cloud nodes, comprising: Multiple edge cloud nodes and central cloud platform, the edge cloud nodes establish a two-way data connection with the central cloud platform through a pre-configured communication protocol; Each edge cloud node includes a network status detection unit, a local routing decision module, and a routing table storage unit. The network status detection unit periodically collects link bandwidth utilization, cross-node transmission delay, network device port status, and node local computing resource load rate between adjacent nodes; The local routing decision module encapsulates the data collected by the network status detection unit into structured data packets and transmits them to the central cloud platform through an encrypted tunnel; The central cloud platform is deployed with a global routing analysis engine. Based on the network status information of all edge cloud nodes received, the global routing analysis engine uses a spatiotemporal feature extraction neural network model to generate a global routing strategy. The global routing strategy includes a path priority list, a traffic distribution weight coefficient, and a set of failover paths. After receiving the global routing policy issued by the central cloud platform, the local routing decision module updates the forwarding table items in the routing table storage unit according to the path priority list and the traffic distribution weight coefficient, and synchronizes the updated forwarding table items to the adjacent edge cloud nodes through the pre-configured communication protocol; The network status detection unit monitors the variation of link transmission delay between adjacent nodes in real time. When it is detected that the variation exceeds the dynamic threshold pre-stored in the routing table storage unit, the local routing decision module is triggered to reconfigure the routing table storage unit based on the set of fault switching paths, and upload the marked link status abnormality information to the central cloud platform via an encrypted tunnel; The global routing analysis engine of the central cloud platform initiates policy recalculation based on link status anomaly information and distributes the updated global routing policy to the associated affected edge cloud nodes.

[0100] In the above technical solution, the edge cloud node establishes a two-way connection with the central cloud platform through an IPsec VPN tunnel. The communication protocol can support MQTT or gRPC, and the heartbeat detection period can be set to 10 seconds. Each edge cloud node contains a network status detection unit (periodically collecting link bandwidth utilization, latency, etc.), a local routing decision module (executing path switching decisions) and a routing table storage unit (storing forwarding table entries).

[0101] In terms of equipment selection, edge cloud nodes can use industrial-grade servers equipped with dual network ports, and the network status detection unit can be equipped with an embedded controller that supports SNMP and ICMP protocols. In terms of materials, the server chassis can be made of aluminum alloy, and the internal circuit board uses FR-4 epoxy resin substrate. Edge cloud nodes are deployed at the edge of the network close to terminal devices, such as in the workshop switch cabinet of a smart factory, and are directly connected to adjacent nodes via Cat6a network cables.

[0102] During the working process, the edge cloud node automatically establishes an encrypted tunnel with the central cloud platform when it starts, and regularly sends heartbeat packets to maintain the connection. The network status detection unit collects data at a preset period (for example, bandwidth is collected every 5 minutes), and the local routing decision module updates the routing table after receiving the policy issued by the central platform, and synchronizes the changes to adjacent nodes through the BGP protocol.

[0103] The global routing analysis engine of the central cloud platform adopts a model that connects convolutional neural networks (CNN) and long short-term memory networks (LSTM) in parallel. The input includes node topology and delay sequence, and the output is a path priority list (for example, priority 1 is path A, with a weight of 70%). After the strategy is generated, it is encapsulated in JSON format and sent to the edge cloud node through a TLS encrypted link. The sending delay can be controlled within 200ms.

[0104] In terms of equipment selection, the global routing analysis engine can run on a server cluster that supports GPU acceleration, and the neural network model can be implemented based on the TensorFlow framework. In terms of materials, the server cooling module can use a combination of copper heat pipes and aluminum alloy fins, and the data storage unit can be equipped with an NVMe solid-state drive. The central cloud platform is deployed in the data center computer room and connected to the edge nodes through redundant optical fiber links.

[0105] During the working process, after the central cloud platform receives the status data of all edge nodes, the neural network model extracts the spatiotemporal features and generates a routing strategy. The policy file is compressed and signed and distributed to the target node. The local routing decision module of the edge node parses the JSON instruction and updates the forwarding table, and sends a synchronization request to the adjacent node.

[0106] The edge cloud node monitors the change in the link transmission delay between adjacent nodes in real time, and the dynamic threshold is initialized to T0 = μ + 3σ (for example, when μ = 50ms and σ = 10ms, T0 = 80ms). When the detected change in delay exceeds the threshold (for example, reaches 100ms), the local routing decision module is triggered to select a candidate path (for example, path B) from the set of fault switching paths, and perform packet loss rate detection (threshold 2%) and delay verification (tolerance ±10ms). When the verification fails, an exception message (including the delay value, path ID, and node load status) is sent to the central cloud platform.

[0107] For device selection, delay monitoring can be based on network cards that support hardware timestamps, and fault switching verification can be equipped with a probe module that supports the IP SLA protocol. In terms of materials, the PCB substrate of the network card can be made of fiberglass, and the probe housing can be made of flame-retardant ABS plastic. This module needs to be integrated into the network interface unit of the edge cloud node and communicate with the routing decision module through the PCIe interface.

[0108] During operation, an abnormal link state triggers local path switching. After the candidate path verification passes, the traffic is split (for example, 70% of the traffic is switched to path B). If the verification fails three times in a row, it will fallback to the original path and limit the bandwidth of non-critical services to 10Mbps. The exception message is marked with a timestamp and uploaded through an encrypted tunnel. After receiving it, the central platform starts a global policy recalculation.

[0109] This system realizes the real-time perception and dynamic optimization of the network state through a distributed architecture and a central coordination mechanism. The local decision of the edge node reduces the computing pressure on the central platform, and the fault switching process ensures business continuity. Encrypted communication and policy signature enhance the security of data transmission, meeting the requirements of high-reliability network operation and maintenance in large-scale edge computing scenarios.

[0110] Although the embodiments of the present invention have been disclosed as above, it is not limited to the applications listed in the specification and embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the field, additional modifications can be easily made. Therefore, without departing from the general concept defined by the claims and the equivalent scope, the present invention is not limited to the specific details and the illustrated and described examples here.

Claims

1. A distributed large two-layer network intelligent routing method based on edge cloud nodes, characterized in that, It includes the following steps: Multiple edge cloud nodes respectively establish bidirectional data connections with the central cloud platform through pre-configured communication protocols. Each edge cloud node periodically collects the network status information of its own node. The network status information includes the link bandwidth utilization rate between adjacent nodes, the cross-node transmission delay, the network device port status, and the node local computing resource load rate; Each edge cloud node encapsulates the network status information into a structured data packet and uploads it to the central cloud platform through an encrypted tunnel; Based on the network status information of all edge cloud nodes, the central cloud platform generates a global routing policy using a neural network model based on spatio-temporal feature extraction. The global routing policy includes a path priority list between edge cloud nodes, a traffic allocation weight coefficient, and a set of failover paths; After receiving the global routing policy issued by the central cloud platform, each edge cloud node updates the forwarding table entries in the local routing table storage unit according to the path priority list and the traffic allocation weight coefficient, and synchronizes the updated forwarding table entries to adjacent nodes through the pre-configured communication protocol; Each edge cloud node monitors the change amount of the link transmission delay between adjacent nodes in real time. When it detects that the change amount of the link transmission delay exceeds the dynamic threshold pre-stored in the local routing decision module, it triggers the local routing decision module to reconfigure the local routing table based on the set of failover paths in the global routing policy, and marks the link status exception information and uploads it to the central cloud platform through an encrypted tunnel; The central cloud platform starts to recalculate the global routing policy according to the link status exception information and distributes the updated global routing policy to the affected edge cloud nodes associated with the link status exception information.

2. The distributed large two-tier network intelligent routing method based on an edge cloud node according to claim 1, characterized in that The neural network model based on spatio-temporal feature extraction includes a convolutional neural network branch and a long short-term memory network branch arranged in parallel. Among them, The convolutional neural network branch takes the physical topology connection relationship of the edge cloud nodes as the input, and extracts the spatial association features between adjacent nodes through three convolutional layers. The spatial association features include the link bandwidth fluctuation pattern between nodes and the port status change trend; The long short-term memory network branch takes the cross-node transmission delay sequence as the input, and extracts the delay fluctuation time series features within a preset time window through a bidirectional cyclic structure. The delay fluctuation time series features include a periodic congestion pattern and a sudden abnormal delay segment; The neural network model dynamically assigns weights to the spatial association features and the delay fluctuation time series features through an attention mechanism layer to generate a network state vector that fuses spatio-temporal features; The fully connected layer of the neural network model calculates and generates a global routing policy according to the network state vector. The generation process of the path priority list includes performing a multi-objective optimization operation on the network state vector that fuses spatio-temporal features. The constraint conditions of the multi-objective optimization operation include the link bandwidth utilization rate threshold, the transmission delay upper limit, and the node load balancing coefficient.

3. The distributed large two-layer network intelligent routing method based on edge cloud nodes according to claim 1, characterized in that, Reconfiguring the local routing table storage unit based on the set of failover paths includes the following steps: When it detects that the change amount of the link transmission delay exceeds the dynamic threshold, extract a pre-sorted candidate path sequence from the set of failover paths. The candidate path sequence is sorted according to the historical transmission success rate and the path hop count; Perform link bandwidth availability detection and transmission delay verification on the candidate path with the highest priority. Link bandwidth availability detection is achieved by sending a probe data packet to the target node and calculating the packet loss rate. Transmission delay verification is achieved by measuring the difference between the round-trip delay of the probe data packet and the preset delay threshold. If the packet loss rate of the candidate path is lower than the preset packet loss threshold and the round-trip delay difference is within the preset tolerance range, the current fault link traffic is diverted to the candidate path according to the preset ratio, and a forwarding table entry update instruction including the diversion ratio and the path identifier is generated; If the candidate path verification fails, the dynamic weight adjustment of the candidate path sequence is triggered, the path priority is recalculated according to the real-time load rate of the adjacent nodes, and the link bandwidth availability detection step is jumped to perform the suboptimal candidate path verification; After the path switching is completed, the local routing decision module stores the updated forwarding table entry in association with the failover timestamp, and sends a routing table synchronization request to the adjacent edge cloud node through a preconfigured communication protocol, wherein the synchronization request includes the changed forwarding table entry hash value and version number; The network status detection unit continuously monitors the transmission quality indicators of the switched path. When the transmission delay fluctuation exceeds 50% of the dynamic threshold within three consecutive detection cycles, the secondary path switching process is triggered and the priority weight parameters of the failover path set are updated.

4. The distributed large two-tier network intelligent routing method based on edge cloud nodes according to claim 2, wherein, The execution process of multi-objective optimization operation includes the following steps: Map the network state vector that fuses spatio-temporal features into a set of decision variables including the bandwidth allocation ratio factor α i , the delay compensation factor β j and the node load balancing weight γ k , where i corresponds to the path number, j corresponds to the delay level identifier, and k corresponds to the node identifier; Construct a set of objective functions including the first objective function, the second objective function, and the third objective function. The first objective function is f1 = Σ(α i ×C i ), which is used to maximize the effective bandwidth utilization. The second objective function is f2 = Σ(β j ×D j ), which is used to minimize the transmission delay offset. The third objective function is f3 = Σ(γ k ×L k ), which is used to balance the node load difference. Among them, C i represents the available bandwidth capacity of path i, D j represents the delay compensation of path j, and L k represents the load deviation value of node k; Set up a system of equations with the following constraints: Σα i ≤ Predefined link bandwidth utilization threshold, Σβ j ≤ Predefined upper limit of transmission delay, Σγ k ≤ Predefined node load balancing coefficient; A non-dominated sorting genetic algorithm with an elite retention strategy is used to solve the objective function set, which includes: Generate an initial solution set as the genetic algorithm population based on the path priority list, perform an adaptive crossover operation based on path similarity, dynamically adjust the crossover probability based on the path overlap rate of the individuals in the solution set, and perform an adjacent path replacement mutation operation to replace the current path with a candidate path that is directly adjacent in the physical topology; The following processing is performed on the solution set produced by each iteration: Delete the individuals in the equation group that violate the constraint conditions, perform non-dominated sorting on the remaining individuals and calculate the crowding distance, and retain the first preset number of individuals with the highest sorting level and the largest crowding distance; Select the individual with the largest crowding distance from the Pareto front solution set of the final iteration result as the optimal solution, and extract the corresponding α, β, and γ parameter combinations of this individual to generate the sorting weight value of the path priority list. i , β j and γ k parameter combinations to generate the sorting weight value of the path priority list.

5. The distributed large two-layer network intelligent routing method based on an edge cloud node according to claim 1, characterized in that, The process of determining and updating the dynamic threshold in the local routing decision module includes the following steps: The local routing decision module is initialized and loaded with the baseline threshold T0, which is calculated based on the normal distribution parameters of the historical transmission delay data. The calculation formula is T0=μ+3σ, where μ is the delay mean and σ is the delay standard deviation. The network status detection unit counts the sliding window average of the link transmission delay between adjacent nodes every five minutes. The window size is the most recent 30 sampling points. When the deviation between the sliding window average and the baseline threshold T0 exceeds 20%, the threshold adaptive adjustment is triggered. The threshold adaptive adjustment process performs the following operations: Collect the delay data set of each hour within 24 hours before the current time, remove the data points with more than three times the standard deviation, recalculate the mean μ' and standard deviation σ', generate the updated dynamic threshold T1=μ'+2.5σ' and overwrite and store it in the local routing decision module; Between two threshold adjustment cycles, if it is detected that the single delay change amount exceeds 150% of the current dynamic threshold, an emergency threshold correction process is immediately started, and the current dynamic threshold is temporarily set to 80% of the abnormal delay value and marked as the to-be-calibrated state; When the edge cloud node receives the updated global routing policy issued by the central cloud platform, it synchronously obtains the threshold calibration parameters included in the global routing policy, and fuses and calculates the local dynamic threshold and the calibration parameters through the weighted average algorithm to generate the finally effective dynamic threshold.

6. The distributed large two-layer network intelligent routing method based on an edge cloud node according to claim 1, characterized in that The process of periodically collecting network status information includes the following steps: The network status detection unit configures multi-dimensional data collection tasks according to the preset collection strategy. The collection strategy includes that the collection period of the link bandwidth utilization rate is 5 minutes, the collection period of the cross-node transmission delay is 30 seconds, the collection period of the network device port status is 1 minute, and the collection period of the node local computing resource load rate is 2 minutes; When performing the collection of the link bandwidth utilization rate, send ICMP probe packets to adjacent nodes and calculate the arithmetic mean of the bandwidth occupancy rates of the last 10 detection results. At the same time, obtain the real-time throughput counter value of the network device port through the SNMP protocol; When performing the cross-node transmission delay collection, perform a two-way transmission test using UDP benchmark packets, calculate the mean value of the time difference between the sending moment and the receiving confirmation moment as the effective delay value, and record the maximum delay fluctuation amount during the test process; The following operations are performed on the collected raw data by the preprocessing module: Abnormal sampling points exceeding three times the standard deviation range are removed, the binary flag bits in the port status data are converted by one-hot encoding, and the index data with different dimensions are converted into feature values of the same magnitude through the Z-score normalization method; When the network status detection unit monitors that the change rate of the adjacent two collection results of any index exceeds the preset sensitivity threshold, the collection period of this index is automatically shortened to 50% of the original period, and intensive sampling for three consecutive periods is started; All preprocessed network status information data are appended with a timestamp accurate to the millisecond level during storage, and clock synchronization and calibration are performed with the central cloud platform through the NTP protocol.

7. The distributed large two-tier network intelligent routing method based on an edge cloud node according to claim 3, wherein The processing after the failure of the sub-optimal candidate path verification further includes the following steps: When the candidate path verification fails three consecutive times, the local routing decision module performs the following operations: Roll back the traffic of the current faulty link to the path before switching, and generate a path rollback instruction to update the forwarding table entry; Send a path switching failure alarm message to the central cloud platform. The alarm message includes the failed path identifier, the verification failure reason code, and the current node load status; Start the local downgrade mode, limit the forwarding rate of non-critical service traffic to the preset safety threshold, and maintain the minimum guaranteed bandwidth of critical service traffic at the same time; After receiving the path switching failure alarm message, the central cloud platform marks this path as a high-risk path in the updated global routing policy and excludes it from the fault switching path set for at least 24 hours.

8. A distributed large-scale two-tier network intelligent routing system based on edge cloud nodes, characterized in that, For implementing the distributed large two-layer network intelligent routing method based on edge cloud nodes described in any one of claims 1 to 7, it includes: Multiple edge cloud nodes and a central cloud platform, and the edge cloud nodes establish a two-way data connection with the central cloud platform through a pre-configured communication protocol; Each edge cloud node includes a network status detection unit, a local routing decision module, and a routing table storage unit. The network status detection unit periodically collects the link bandwidth utilization rate between adjacent nodes, the cross-node transmission delay, the network device port status, and the local computing resource load rate of the node; The local routing decision module encapsulates the data collected by the network status detection unit into structured data packets and transmits them to the central cloud platform through an encrypted tunnel; The central cloud platform deploys a global routing analysis engine. Based on the network status information of all received edge cloud nodes, the global routing analysis engine uses a spatio-temporal feature extraction neural network model to generate a global routing policy. The global routing policy includes a path priority list, a traffic allocation weight coefficient, and a set of failover paths; After receiving the global routing policy issued by the central cloud platform, the local routing decision module updates the forwarding table entries in the routing table storage unit according to the path priority list and the traffic allocation weight coefficient, and synchronizes the updated forwarding table entries to adjacent edge cloud nodes through a pre-configured communication protocol; The network status detection unit real-time monitors the change amount of the link transmission delay between adjacent nodes. When it detects that the change amount exceeds the dynamic threshold pre-stored in the routing table storage unit, it triggers the local routing decision module to reconfigure the routing table storage unit based on the set of failover paths, and uploads the marked link status exception information to the central cloud platform through an encrypted tunnel; The global routing analysis engine of the central cloud platform starts policy recalculation according to the link status exception information and distributes the updated global routing policy to the associated affected edge cloud nodes.

Images