Intelligent network communication link design method

Abstract

The application belongs to the technical field of network communication and relates to a smart network communication link design method, which designs a set of link quality maintenance mechanism: a joint feature vector containing a physical layer bit error rate and a logical layer queue depth is constructed, signal degradation acceleration and queue accumulation acceleration are captured based on second derivative calculation, adjacent link probability distribution differences are compared and layered sorted in a Pareto manner, and link quality fine evaluation and degradation judgment are realized; candidate backup links are screened based on physical disjoint constraints, real-time performance feedback of the candidate backup links is obtained through perturbation detection packets carrying traffic feature templates, service degradation risk probability is predicted in combination with quality scores of target links, resource margin and path cost information are interacted between network nodes based on a distributed negotiation mechanism, an optimal switching path is generated, and service flow is smoothly migrated, so that link maintenance is upgraded from passive response to active pre-inspection.

Description

Technical Field

[0001] This invention belongs to the field of network communication technology and relates to a method for designing intelligent network communication links. Background Technology

[0002] Communication links, as a fundamental element in building modern network topologies, bear the task of transmitting various information flows from source nodes to destination nodes. The quality of communication link design directly determines the degree to which upper-layer services meet the Quality of Service (QoS) protocols, thus affecting the reliability and practicality of network communication.

[0003] In existing technologies, link quality maintenance mainly employs a centralized network management system to periodically poll node status, or detects physical connectivity through hardware-level operation and management mechanisms. Path switching is triggered by link failure, and backup paths are configured based on link cost or static priority. Once a port is detected to be offline or physically interrupted, the controller issues flow table commands to redirect traffic to the backup link.

[0004] However, the above design concept is lagging behind in practical engineering applications. Specifically, the existing quality assessment design mechanism for communication links can only identify the two extreme states of the link: health and failure. The physical hardware status and the traffic queue status are independent of each other in terms of monitoring dimensions and are isolated from each other. Furthermore, the link quality assessment does not consider the spatial propagation law of the degradation status between adjacent links, making it difficult to identify the source of degradation and the frontier of diffusion, and intervention decisions lack spatial reference.

[0005] 2. The existing path switching design mechanism of communication links relies on passive response after a failure occurs for switching decisions. It cannot initiate switching in advance based on trend prediction of quality scores. At the same time, there is no micro-disturbance detection mechanism for candidate backup links before switching, so it is impossible to verify their actual carrying capacity. In addition, switching decisions are mostly controlled by a single node, making it difficult to obtain the real-time load and resource status of all nodes in the network, resulting in a lack of collaborative optimality in path selection. Summary of the Invention

[0006] In view of this, in order to solve the problems mentioned in the background technology, a design method for intelligent network communication links is proposed.

[0007] The objective of this invention can be achieved through the following technical solution: This invention provides an intelligent network communication link design method, comprising: using physical link fluctuation events as time anchors, performing windowed aggregation on packet forwarding process data, and generating a joint feature vector containing link bit error rate and queue depth.

[0008] The second derivative of the joint feature vector of a single link is calculated to capture the acceleration of signal degradation and queue accumulation. The probability distribution difference between the joint feature vectors of adjacent links is analyzed based on the network topology. The quality score of each link is evaluated based on the temporal and spatial performance, and it is determined whether the preset degradation conditions are met.

[0009] The link that meets the preset degradation conditions is marked as the target link, and a probe request is sent to at least one candidate backup link that is not physically intersecting with the target link according to the network topology.

[0010] Receive real-time performance feedback from candidate backup links, combine it with the quality score of the target link to predict the probability of service degradation risk of the target link, and broadcast it to network nodes on the candidate backup links to trigger path pre-switching negotiation.

[0011] Based on the path pre-switching negotiation results, the target link service flow will be migrated to the selected candidate backup link.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention constructs a joint feature vector containing the physical layer bit error rate and the logic layer queue depth, and calculates the capture signal degradation acceleration and queue accumulation acceleration based on the second derivative. It combines the probability distribution differences of adjacent links to perform Pareto comparison and hierarchical sorting, unifying the physical layer and logic layer states in the same feature space, eliminating the fragmentation defect, and realizing the refined evaluation of link quality and the judgment of degradation conditions.

[0013] (2) The present invention calculates KL divergence based on the probability distribution difference of adjacent links and sorts all links hierarchically through Pareto comparison, introduces spatial correlation analysis between links, makes up for the shortcomings of existing technologies that independently evaluate each link and ignore the law of degradation propagation, can identify the source of degradation and the frontier of diffusion, and make intervention decisions have spatial dimension reference information.

[0014] (3) Based on physical disjoint constraints, the present invention selects candidate backup links, obtains real-time performance feedback of backup links by carrying micro-disturbance probe messages with traffic feature templates, predicts the probability of service degradation risk by combining the quality score of the target link, and exchanges resource margin and path cost information between network nodes according to the distributed negotiation mechanism. It can start the switching preparation in advance based on the quality trend prediction before the failure occurs, and verify the actual carrying capacity of the candidate backup links by probe. Finally, the optimal switching path is generated and the smooth migration of the service flow is completed, realizing the distributed collaborative optimality of path selection. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the implementation steps of the method of the present invention;

[0017] Figure 2 This is a schematic diagram illustrating the logic for generating the joint feature vector of the present invention;

[0018] Figure 3 This is a schematic diagram of the judgment logic for the preset degradation conditions of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 As shown, this invention provides an intelligent network communication link design method, including: S1. Using physical link fluctuation events as time anchors, performing windowed aggregation on packet forwarding process data to generate a joint feature vector containing link bit error rate and queue depth. (Refer to...) Figure 2 As shown, in this embodiment, generating a joint feature vector containing the link bit error rate and queue depth includes: continuously monitoring the bit errors at the physical link receiver, specifically: reading the bit error counter of the physical layer interface at a fixed sampling period and recording the number of erroneous bits and the total number of transmitted bits in each period.

[0021] The link bit error rate is calculated at a fixed period. The link bit error rate is equal to the ratio of the number of erroneous bits in the period to the total number of transmitted bits. If the total number of transmitted bits in the period is zero, the link bit error rate is zero.

[0022] Cumulative sum detection is performed on the link bit error rate sequence consisting of continuous periods: After continuously sampling the link bit error rate for multiple periods during the initial link calibration phase, the average value is taken as the reference value of the link bit error rate under the link health state. The deviation between the link bit error rate of the current period and the reference value is calculated. The cumulative sum of the previous period is added to the deviation of the current period. If the sum is greater than zero, the cumulative sum is updated; otherwise, the cumulative sum is reset to zero.

[0023] Events that accumulate and exceed a preset threshold are considered link fluctuation events. Specifically, the preset threshold is set to 4 times the standard deviation of the link bit error rate of multiple consecutive sampling cycles during the initial link calibration phase. The setting is based on the control limit in the standard CUSUM control chart, which is set at a position 4 times the standard deviation from the center line by default. The threshold is reset to zero after the accumulated sum exceeds the preset threshold.

[0024] Using the end of the period in which the link fluctuation event occurs as the time anchor, a fixed-length time window is traced backward, and the link bit error rate and queue depth within the time window are collected at equal intervals to form a link bit error rate sequence and a queue depth sequence, respectively; the queue depth is directly read from the transmission buffer of the corresponding port of the link. Through sequence time alignment, the link bit error rate and queue depth at the same sampling time are combined into a two-dimensional vector, arranged in the order of collection time, to form a joint feature vector of a single link.

[0025] S2. Calculate the second derivative of the joint feature vector of a single link to capture the signal degradation acceleration and queue accumulation acceleration. Analyze the probability distribution differences between the joint feature vectors of adjacent links based on network topology. Evaluate the quality score of each link based on temporal and spatial performance and determine whether the preset degradation conditions are met.

[0026] In this embodiment, the quality score of each link based on temporal and spatial performance includes: taking non-negative values ​​for the signal degradation acceleration and queue accumulation acceleration of each link respectively, setting them to zero if they are negative, adding the two non-negative values ​​to obtain a sum, dividing each non-negative value by the sum to obtain two probability values, forming a binary probability distribution, and calculating its Shannon entropy to obtain the link's own entropy value.

[0027] The network topology map is used to retrieve each adjacent link of each link. The adjacent link refers to the link that shares the same network node with the current link. The probability distribution of the current link and the probability distribution of the adjacent links are calculated using KL divergence. The arithmetic mean of the KL divergences of all adjacent links is taken as the average relative entropy.

[0028] Perform a Pareto comparison on any two links in the network topology to determine the quality order between the two links.

[0029] Specifically, the Pareto comparison is performed as follows: for any two links in the network topology graph... and Each of them obtains its own entropy value. , and average relative entropy , .

[0030] Determine the link Quality is better than link The conditions are: ( )or or .

[0031] If link Quality is better than link Then it is called Dominate If neither of the following conditions are met Dominate Not satisfied Dominate ,but and They have a non-dominant relationship.

[0032] Based on the Pareto comparison results, all links are hierarchically sorted. Specifically, first, all links in the network topology are grouped into a set. All links in the current set that are not dominated by any other link are retrieved and grouped into the same level, and assigned a level number of 1. Then, the links that have been assigned to a level are removed from the current set. For the remaining links, the search for links that are not dominated by any other link is repeated. The level number is incremented by 1 in each round until all links are assigned a level number. Finally, the level number of each link is obtained, where a smaller level number indicates a better link quality.

[0033] The difference between the tier number and the total number of tiers is normalized and used as the link quality score. The specific normalization process is to add 1 to the difference between the total number of tiers and the current tier number and then divide it by the total number of tiers to obtain the quality score for each link.

[0034] Reference Figure 3 As shown, in this embodiment, determining whether the preset degradation condition is met includes: obtaining the signal degradation acceleration and queue accumulation acceleration of all links in the entire network, generating a set of coordinate points on a two-dimensional plane with the signal degradation acceleration as the horizontal axis and the queue accumulation acceleration as the vertical axis, and each link corresponding to one coordinate point.

[0035] The set of all coordinate points is taken as input, and the convex hull boundary of the coordinate point set is calculated using the Andrew algorithm or Graham scan method. The links corresponding to the coordinate points located on the convex hull boundary are extracted as boundary links.

[0036] Taking Andrew's algorithm as an example, the process of obtaining the convex hull boundary of the coordinate point set is as follows: Sort all coordinate points in ascending order of x-coordinate, and if the x-coordinates are the same, sort them in ascending order of y-coordinate; construct the lower convex hull by scanning sequentially, and use the cross product to determine the direction, where the top of the stack is popped when the cross product is less than or equal to 0; construct the upper convex hull by scanning in descending order; merge the upper and lower convex hulls to obtain the complete convex hull boundary. This method is a well-known technique, and this invention will not elaborate on it in detail.

[0037] For each boundary link, count the number of adjacent links that are also located on the convex hull boundary, and use this as the diffusion clustering degree.

[0038] The boundary link with the highest diffusion aggregation degree is identified as the degradation center. If multiple boundary links have the same highest diffusion aggregation degree, the one with the largest Euclidean distance between the signal degradation acceleration and the queue accumulation acceleration is taken as the degradation center.

[0039] Links in the boundary links that have at least one adjacent link path to the degradation center and whose diffusion aggregation degree is less than that of the degradation center are identified as target links that meet the preset degradation conditions.

[0040] S3. Mark the link that meets the preset degradation conditions as the target link, and send a probe request to at least one candidate backup link that is not physically intersecting with the target link according to the network topology.

[0041] In this embodiment, the step of sending a probe request to at least one candidate backup link that is physically disjoint from the target link based on the network topology includes: obtaining the starting node and ending node of the target link; calling the network topology map, which contains each node, each link and its physical attributes, including port identifier, node identifier and transmission medium type; and running a path search algorithm to find all acyclic paths from the starting node to the ending node except for the target link itself. The path search algorithm can be the K-shortest path algorithm, where K is preset to a value of 3 to 5.

[0042] From all the loop-free paths found, paths that do not share any physical ports, intermediate nodes, or the same transmission medium as the target link are selected as candidate backup links.

[0043] As an example of screening candidate backup links, a shared risk group label can be pre-assigned to each node and each link in the network topology based on physical attributes. Components in the same physical cabinet, the same fiber optic duct, or the same power domain are assigned the same shared risk group label, excluding loop-free paths that share any of the same shared risk group labels with the target link, thus achieving physical non-intersection.

[0044] In addition, if no candidate backup link that meets the physical disjointness constraint is found, the pre-switching attempt is abandoned, the target link is marked as high risk and reported as an alarm, and manual intervention or traditional failover mechanism is waited for.

[0045] Send a probe request message to the starting node of the candidate backup link. The probe request message carries at least the traffic feature template of the target link service flow, the probe request duration, and the performance parameter tolerance limit.

[0046] It should be noted that the traffic feature template includes at least the average packet size, packet interval duration, and burst size, which are used to instruct the receiving end to simulate the same traffic behavior.

[0047] The duration of the request detection is a preset fixed value, which is set based on the following: it is greater than the upper limit of the link fault detection time specified by the network protocol of the target link, so as to ensure that the detection and negotiation process is completed before the protocol layer determines the fault; as an example, for PROFINET MRP ring network, the value is 300 milliseconds; for EtherCAT network, the value is 150 milliseconds; for data center Ethernet, the value is 1 second.

[0048] The performance parameter tolerance limits are pre-configured according to the service quality protocol requirements of the target link service flow, and include at least the maximum end-to-end latency, the maximum latency jitter, and the minimum available bandwidth.

[0049] S4. Receive real-time performance feedback from candidate backup links, combine the quality score of the target link to predict the probability of service degradation risk of the target link, and broadcast it to network nodes on the candidate backup links to trigger path pre-switching negotiation.

[0050] In this embodiment, the prediction of the service degradation risk probability of the target link includes: receiving the probe response message returned by the candidate backup link, and extracting the end-to-end transmission delay, available bandwidth and delay jitter measured by the simulated traffic feature template of the candidate backup link.

[0051] By comparing the performance parameter tolerance limits carried in the probe request message, the over-limit values ​​of end-to-end transmission delay, available bandwidth, and delay jitter are calculated respectively, and combined with the quality score of the target link to form a four-dimensional feature vector.

[0052] By performing gradient descent regression training on historical link switching event samples, the weight vector and bias constant are calibrated.

[0053] Calculate the dot product of the four-dimensional feature vector and the weight vector, and substitute the sum of the dot product and the bias constant into a preset function to predict the service degradation risk probability of the target link. The preset function is S.

[0054] It should be noted that the specific calculation process of the over-limit amplitude is as follows: Mark the parameter over-limit status, where the end-to-end transmission delay and delay jitter are greater than their respective tolerance limits (maximum values) and the available bandwidth is less than the minimum available bandwidth. For the parameter marked as over-limit, divide the absolute difference between the current parameter measurement value and its tolerance limit by the tolerance limit to obtain the over-limit amplitude; for the parameter not marked as over-limit, directly set the over-limit amplitude to 0.

[0055] The gradient descent regression training process is as follows: Step 1, sample collection: collect at least 500 sets of historical link switching event samples. Each set of samples contains a four-dimensional feature vector and a corresponding service degradation label. The service degradation label takes the value of 1 or 0, where 1 indicates that the link where the sample is located has experienced service degradation within a preset time period after detection, and 0 indicates that it has not occurred.

[0056] The second step is parameter initialization: set all four components of the weight vector to 0, set the bias constant to 0, set the learning rate to 0.01, and set the maximum number of iterations to 1000.

[0057] Special note: If there are insufficient historical link switching event samples, all four components in the weight vector will be set to the default value of 0.25 to avoid cold start.

[0058] The third step is iterative training: For each iteration, the order of all samples is randomly shuffled, the service degradation risk probability of each sample is calculated, and the difference between the calculated service degradation risk probability and the sample service degradation label is used to obtain the prediction error.

[0059] The bias update amount is obtained by multiplying the prediction error by the learning rate, and the updated bias constant is obtained by subtracting the bias update amount from the current bias constant.

[0060] Subtract the product of the bias update amount and the corresponding feature value in the four-dimensional feature vector from each component of the current weight vector to obtain the updated weight vector.

[0061] The updated bias constant and weight vector are used as input for the next iteration.

[0062] After each full sample traversal, the average loss of the current iteration is calculated using the logarithmic loss function. If the loss change over 5 consecutive iterations is less than [a certain value], then [the outcome is considered successful]. If the loss converges, training is stopped, the converged parameters are marked, and the converged parameters are stored in the database for direct use.

[0063] The formula for the logarithmic loss function is as follows: .

[0064] In the formula, This represents the average loss of the current iteration. Indicates the first The service degradation label corresponding to each sample Indicates the first The probability of service degradation risk calculated for each sample in the current iteration. Indicates the sample number. ,in This indicates the total number of samples.

[0065] This item only takes effect when the service degradation label is 1, and is used to penalize situations where degradation actually occurs but the predicted risk is very low. The natural logarithm of the probability of service degradation risk is used when the probability of service degradation risk is close to 1 (the prediction is correct). When the probability approaches zero, the loss is small, but when the probability of service degradation is close to zero (prediction error), the loss is small. A very large negative number becomes a positive loss after taking the negative sign.

[0066] This item only takes effect when the service degradation label is 0, and is used to penalize situations where no actual degradation has occurred but the predicted risk is high. This indicates the probability that a service degradation will not occur. If the probability of service degradation is close to 0 (the prediction is correct), then... Close to 1 If the probability is close to 0, the loss is minimal; if the probability of service degradation is close to 1 (prediction error), then... Approaching 0 It is a very large negative number, which becomes a positive loss after being negative.

[0067] The entire formula means that for all samples, the loss caused by a lower predicted probability when a service degradation actually occurred is added to the loss caused by a higher predicted probability when a service degradation did not actually occur. The average of these two losses is then taken. The smaller the average loss, the closer the predicted service degradation risk probability is to the actual situation. During training, the weights and biases are continuously adjusted using gradient descent, causing the average loss to gradually decrease and eventually converge to its minimum.

[0068] It should be further explained that the above-mentioned historical link switching event sample collection process is as follows: In actual network operation, each time a probe process (steps S3-S4) is executed, the four-dimensional feature vector at that time is recorded, and the link is continuously monitored for service degradation in the future within a preset time period. Service degradation can be defined as a packet loss rate exceeding 1% or the over-limit value of any performance feedback parameter reaching half of its original tolerance limit. If it occurs, the label is 1; otherwise, it is 0. Each probe is stored as a sample in the historical database. When the number of samples reaches 500 sets, offline training is started.

[0069] Furthermore, the specific values ​​involved in the above gradient training process, such as the number of samples, packet loss rate threshold, learning rate, maximum number of iterations, and loss change convergence threshold, are all exemplary parameters in this embodiment. Those skilled in the art can make adaptive adjustments according to the actual network scale and service quality requirements, and such adjustments do not depart from the technical essence of this invention.

[0070] In this embodiment, the trigger path pre-switching negotiation includes: encapsulating the service degradation risk probability into a broadcast message and sending it to network nodes on each candidate backup link.

[0071] After receiving the broadcast message, each network node performs a local evaluation of its own candidate backup link and calculates the resource margin and path cost required to take over the target link's service flow.

[0072] Specifically, the calculation process for the resource margin required to take over the target link service flow is as follows: obtain the current available bandwidth of the candidate backup link and the priority of each service flow it carries. The priority of the service flow is pre-configured according to the service type, with a value range of 1 to 7. The smaller the value, the higher the priority. As an example, real-time control services are configured with priority 1, ordinary data acquisition services are configured with priority 4, and background services are configured with priority 7.

[0073] The priority of the target link service flow is compared with the priority of each service flow already carried on the candidate backup link. The set of low-priority service flows that need to be preempted or downgraded is determined according to the following rules: i. Service flows whose priority value is greater than the priority value of the target link among the service flows already carried on the candidate backup link are marked as preemptible candidates.

[0074] ii. If the currently available bandwidth already meets the bandwidth requirements of the target link's service flow, then the set of low-priority service flows is empty.

[0075] iii. If the current available bandwidth does not meet the demand, preemptible candidate service flows are selected in descending order of priority value and merged into the low-priority service flow set. The bandwidth occupied by the preemptible candidate service flows in the set is accumulated synchronously until the accumulated bandwidth is greater than or equal to the difference between the bandwidth requirement of the target link service flow and the current available bandwidth of the candidate backup link. The bandwidth occupied by a single preemptible candidate service flow is obtained by calculating the average transmission rate of the service flow in the most recent sliding time window.

[0076] Calculate the total bandwidth occupied by the set of low-priority service flows, and use that as the free bandwidth.

[0077] The total available bandwidth is obtained by adding the currently available bandwidth of the candidate backup link to the bandwidth that can be released.

[0078] Obtain the bandwidth requirements of the target link's service flow, subtract the bandwidth requirements from the total available bandwidth, and obtain the resource margin for the candidate backup link to take over the target link's service flow.

[0079] Specifically, the calculation process for the path cost required to take over the target link's service flow is as follows: obtain the path hop increment, transmission delay increment, and delay jitter increment of the candidate backup link relative to the target link.

[0080] The ratio of each parameter increment to the current parameter measurement value of the target link is analyzed to obtain the parameter increment ratio.

[0081] By summing up the incremental ratios of each parameter, the path cost required to take over the target link's service flow is obtained.

[0082] Each network node interacts and compares its resource adequacy and path cost, obtained from local assessments, along with the probability of service degradation risk. The network node with the largest resource adequacy is selected as the temporary coordination node. If multiple network nodes have the same largest resource adequacy, the path costs of these nodes are compared, and the node with the lowest path cost is selected as the temporary coordination node. If both resource adequacy and path cost are the same, the probability of service degradation risk is compared, and the node with the lowest probability of risk is selected as the temporary coordination node.

[0083] The temporary coordination node collects the resource margin and path cost information of all candidate backup links, sorts all candidate backup links in ascending order of path cost, selects the candidate backup link with the lowest path cost and the resource margin greater than the bandwidth requirement of the target link's service flow, and adds an optimal path identifier.

[0084] The negotiation result, which includes the optimal path identifier, is encapsulated into an acknowledgment message and broadcast to all participating network nodes. Each network node receives and acknowledges the message, thus completing the path pre-switching negotiation.

[0085] S5. Based on the path pre-switching negotiation results, migrate the target link service flow to the selected candidate backup link.

[0086] In this embodiment, migrating the target link service flow to the selected candidate backup link includes: configuring forwarding table entries for network nodes on the selected candidate backup link. The forwarding table entries contain matching characteristics of the target service flow (such as source IP, destination IP, VLAN tag, port number) and next-hop pointing information. Each network node updates its local forwarding table in sequence to establish a backup forwarding path.

[0087] Within a preset time period, the target link service flow is simultaneously copied and forwarded to the target link and the backup forwarding path, and the average number of successfully received data packets within the same time length of the target link is counted.

[0088] Continuously receive data packet confirmation information returned by the backup forwarding path. When the number of consecutive successfully received data packets of the backup forwarding path reaches or exceeds the average number of successfully received data packets, and the qualified state continues for at least two consecutive cycles, stop forwarding the service flow to the target link and dismantle the forwarding path of the target link.

[0089] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A method for designing intelligent network communication links, characterized in that, include: Using physical link fluctuation events as time anchors, the packet forwarding process data is windowed and aggregated to generate a joint feature vector containing link bit error rate and queue depth; The second derivative of the joint feature vector of a single link is calculated to capture the signal degradation acceleration and queue accumulation acceleration. The probability distribution difference between the joint feature vectors of adjacent links is analyzed based on the network topology. The quality score of each link is evaluated based on the temporal and spatial performance, and it is determined whether the preset degradation conditions are met. The link that meets the preset degradation conditions is marked as the target link, and a probe request is sent to at least one candidate backup link that is not physically intersected with the target link according to the network topology. Receive real-time performance feedback from candidate backup links, combine it with the quality score of the target link to predict the probability of service degradation risk of the target link, and broadcast it to network nodes on the candidate backup links to trigger path pre-switching negotiation. Based on the path pre-switching negotiation results, the target link service flow will be migrated to the selected candidate backup link.

2. The intelligent network communication link design method according to claim 1, characterized in that, The generation of a joint feature vector containing link bit error rate and queue depth includes: Continuously monitor bit errors at the physical link receiver, calculate the link bit error rate at fixed intervals, perform cumulative sum detection on the link bit error rate sequence composed of continuous periods, and treat events where the cumulative sum exceeds a preset threshold as link fluctuation events. Using the end of the cycle in which the link fluctuation event occurs as the time anchor point, a fixed-length time window is traced back, and the link bit error rate and queue depth within the time window are collected at equal intervals to form a link bit error rate sequence and a queue depth sequence, respectively. By aligning the sequence time, the link bit error rate and queue depth at the same sampling time are combined into a two-dimensional vector, which is then arranged in the order of sampling time to form a joint feature vector for a single link.

3. The intelligent network communication link design method according to claim 1, characterized in that, The quality score for each link is evaluated based on temporal and spatial performance, including: The signal degradation acceleration and queue accumulation acceleration of each link are normalized into a probability distribution, and the Shannon entropy of the probability distribution is calculated to obtain the entropy value of the link itself. Call the network topology map to retrieve each link's adjacent links, calculate the KL divergence between the probability distribution of the current link and the probability distribution of the adjacent links, and take the arithmetic mean of the KL divergences of all adjacent links as the average relative entropy. Perform a Pareto comparison on any two links in the network topology to determine the quality order between the two links; Based on the Pareto comparison results, all links are sorted hierarchically to obtain the hierarchical number of each link; The difference between the tier number and the total number of tiers is normalized and used as the link quality score, thus obtaining the quality score for each link.

4. The intelligent network communication link design method according to claim 1, characterized in that, The determination of whether the preset degradation condition is met includes: Obtain the signal degradation acceleration and queue accumulation acceleration of all links in the entire network, and generate a set of coordinate points on a two-dimensional plane with the signal degradation acceleration as the x-axis and the queue accumulation acceleration as the y-axis; Calculate the convex hull boundary of the coordinate point set, and extract the links corresponding to the coordinate points located on the convex hull boundary as boundary links; For each boundary link, count the number of its adjacent links that are also located on the convex hull boundary, and use this as the diffusion clustering degree; The boundary link with the highest diffusion aggregation degree is identified as the deterioration center; Links in the boundary links that have at least one adjacent link path to the degradation center and whose diffusion aggregation degree is less than that of the degradation center are identified as target links that meet the preset degradation conditions.

5. The intelligent network communication link design method according to claim 1, characterized in that, The step of sending a probe request to at least one candidate backup link that is physically disjoint from the target link based on network topology includes: Obtain the start and end nodes of the target link, call the network topology map, and run the path search algorithm to find all acyclic paths from the start node to the end node except for the target link itself; From all the loop-free paths found, select those that do not share any physical ports, intermediate nodes, or the same transmission medium as the target link, and use them as candidate backup links. Send a probe request message to the starting node of the candidate backup link. The probe request message carries at least the traffic feature template of the target link service flow, the probe request duration, and the performance parameter tolerance limit.

6. The intelligent network communication link design method according to claim 5, characterized in that, The predicted probability of service degradation risk for the target link includes: Receive probe response messages returned by candidate backup links, and extract the end-to-end transmission delay, available bandwidth, and delay jitter measured by the simulated traffic feature template of the candidate backup links; By comparing the performance parameter tolerance limits carried in the probe request message, the over-limit values ​​of end-to-end transmission delay, available bandwidth, and delay jitter are calculated respectively, and combined with the quality score of the target link to form a four-dimensional feature vector; By performing gradient descent regression training on historical link switching event samples, the weight vector and bias constant are calibrated. Calculate the dot product of the four-dimensional feature vector and the weight vector, and substitute the sum of the dot product and the bias constant into a preset function to predict the probability of service degradation risk of the target link.

7. The intelligent network communication link design method according to claim 1, characterized in that, The trigger path pre-switching negotiation includes: The probability of service degradation risk is encapsulated into a broadcast message and sent to network nodes on each candidate backup link; After receiving the broadcast message, each network node performs a local evaluation of its own candidate backup link and calculates the resource margin and path cost required to take over the target link's service flow. Each network node uses the resource sufficiency and path cost obtained from local assessment, along with the probability of service degradation risk, to perform interactive comparisons and select temporary coordination nodes. The temporary coordination node collects the resource margin and path cost information of all candidate backup links, sorts all candidate backup links in ascending order of path cost, selects the candidate backup link with the lowest path cost and the resource margin greater than the bandwidth requirement of the target link's service flow, and adds an optimal path identifier. The negotiation result, which includes the optimal path identifier, is encapsulated into an acknowledgment message and broadcast to all participating network nodes. Each network node receives and acknowledges the message, thus completing the path pre-switching negotiation.

8. The intelligent network communication link design method according to claim 7, characterized in that, The resource margin required for the calculation to take over the target link service flow includes: Obtain the currently available bandwidth of the candidate backup links and the priority of each service flow they carry; The priority of the target link service flow is compared with the priority of each service flow already carried on the candidate backup link to determine the set of low-priority service flows that need to be preempted or downgraded. Calculate the total bandwidth occupied by the set of low-priority service flows, and use it as the free bandwidth; The total available bandwidth is obtained by adding the currently available bandwidth of the candidate backup link to the bandwidth that can be released. Obtain the bandwidth requirements of the target link's service flow, subtract the bandwidth requirements from the total available bandwidth, and obtain the resource margin for the candidate backup link to take over the target link's service flow.

9. The intelligent network communication link design method according to claim 7, characterized in that, Calculate the path cost required to take over the target link's service flow, including: Obtain the path hop count increment, transmission delay increment, and delay jitter increment of the candidate backup link relative to the target link; The ratio of each parameter increment to the current parameter measurement value of the target link is analyzed to obtain the parameter increment ratio; By summing up the incremental ratios of each parameter, the path cost required to take over the target link's service flow is obtained.

10. The intelligent network communication link design method according to claim 1, characterized in that, The migration of the target link service flow to the selected candidate backup link includes: Configure forwarding table entries on the selected candidate backup links to establish backup forwarding paths; Within a preset time period, the target link service flow is simultaneously copied and forwarded to the target link and the backup forwarding path, and the average number of successfully received data packets within the same time length of the target link is counted. Continuously receive data packet confirmation information returned by the backup forwarding path. When the number of consecutive successfully received data packets of the backup forwarding path reaches or exceeds the average number of successfully received data packets, and the qualified state continues for at least two consecutive cycles, stop forwarding the service flow to the target link and dismantle the forwarding path of the target link.

Images