Deterministic network redundancy protection methods

Abstract

This invention provides a deterministic network redundancy protection method that dynamically decides whether to enable PREOF, where to enable it, and how to select redundant paths based on the real-time network status, thereby meeting the SLA requirements of deterministic transmission, such as extremely low packet loss rate, low latency, and low jitter. The method employs a three-level state assessment mechanism, dividing the network into normal, warning, and protection states based on the current and future packet loss rate. In the warning state, only the PREOF policy is issued, while the PREOF mechanism is triggered only when the network transitions to the protection state. This ensures that PREOF is only activated when network transmission quality is abnormal or shows a deteriorating trend, avoiding the waste of network resources caused by constantly activating redundancy protection.

Description

Technical Field

[0001] This invention belongs to the field of deterministic network technology, specifically relating to a method for redundancy protection in deterministic networks. Background Technology

[0002] Deterministic Networking (DetNet) aims to provide network layer transport services with deterministic Quality of Service (QoS) for mission-critical applications. Its core objective is to ensure that data streams have guaranteed performance boundaries during network transmission. According to industry consensus, DetNet must meet requirements for extremely low end-to-end latency, microsecond-level jitter, and reliability exceeding 99.999%. These metrics are crucial for high-security applications such as autonomous driving, vehicle-to-everything (V2X) communication, and industrial control.

[0003] The Internet Engineering Task Force (IETF) standard RFC 8655 specifies that the deterministic transmission performance supported by the DetNet network includes extremely low upper bounds on latency, jitter, and packet loss rate, as well as limited out-of-order packet delivery. It also proposes three key technologies in the network architecture: explicit routing, resource reservation, and packet replication, elimination, and ordering functions (PREOF). Explicit routing specifies the transmission path for DetNet data streams, avoiding temporary route switching due to network topology changes and mitigating transmission interruptions, packet loss, and out-of-order issues caused by route convergence. Resource reservation ensures that nodes along the route reserve bandwidth, buffers, and other data plane resources for DetNet data streams, and also shapes and limits the output rate of nodes, completely eliminating packet loss caused by traffic contention. PREOF, on the other hand, overcomes packet loss caused by device failures and link errors by copying DetNet packets to multiple paths at designated PRF nodes and eliminating duplicate packets at downstream PEF nodes. At the same time, POF nodes reorder packets to avoid out-of-order transmission caused by multi-path transmission, ensuring orderly transmission.

[0004] The DetNet network architecture is divided into a forwarding sub-layer and a service sub-layer from bottom to top. Nodes in the DetNet network include transit nodes and relay nodes. Transit nodes operate only in the DetNet forwarding sub-layer, performing routing and forwarding of packets. Relay nodes have both service and forwarding sub-layer functions; PREOF explicitly operates in the DetNet service sub-layer of the relay node. In engineering implementation, to protect the DetNet service flow, PRF, PEF, and POF functional entities are activated in the service sub-layer of the relay node. Transit nodes provide explicit forwarding paths with resource reservations to the relay nodes, collaboratively ensuring the deterministic transmission of DetNet flows.

[0005] However, the PREOF strategy currently used in deterministic networks has clear shortcomings, mainly:

[0006] First, the configuration method is static. Existing solutions typically require network administrators to pre-configure the location of replication nodes and redundant paths, which cannot be dynamically adjusted according to real-time network conditions. When network topology changes or link quality fluctuates, statically configured redundant paths may no longer be effective or optimal.

[0007] Second, resource utilization efficiency is low. Existing solutions always enable PREOF for DetNet flows that require packet loss protection, lacking continuous assessment of whether the current network state truly requires redundant protection, resulting in a continuous waste of network resources such as bandwidth and queues.

[0008] Third, there is a lack of a systematic decision-making framework. Existing solutions lack systematic decision-making methods for aspects such as the selection of replication node locations, redundant path search, and sorting node settings. The parameter relationships between each aspect (such as the relationship between latency difference threshold and sorting cache depth) are not fully considered. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a deterministic network redundancy protection method to solve the technical problems of how to enable and disable redundancy protection on demand, and how to automatically optimize and select PRF, PEF node locations and redundant paths based on real-time network status.

[0010] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0011] A deterministic network redundancy protection method:

[0012] S1.1, Identify the target flow status as follows:

[0013] Normal state: along the original path, both currently and in the future. Throughout the period, the packet loss rate met the target requirements;

[0014] Warning status: Along the original path, the current packet loss rate meets the target requirements, but it is predicted that the loss rate will increase in the future. The target will be exceeded within the specified timeframe;

[0015] Protected state: Along the original path, the current packet loss rate does not meet the target requirements;

[0016] S1.2, Execute protection actions based on the target flow status:

[0017] When the target flow changes from a normal state to an alert state, the PREOF strategy is issued;

[0018] When the target flow changes from the warning state to the protection state, the PREOF mechanism is executed based on the issued policy.

[0019] Furthermore, in the PREOF strategy, the PRF and PEF node locations and redundant paths are determined as follows:

[0020] Phase 1, Minimum Coverage Attempt:

[0021] The nearest upstream node of the target subnet is designated as the PRF node and the nearest downstream node as the PEF node. If there is a redundant path between the two nodes that meets the requirements, the PRF and PEF node positions and the redundant path are determined. If there is no redundant path between the two nodes that meets the requirements, then proceed to stage 2.

[0022] Phase 2, Orderly Extended Search:

[0023] By extension distance The search proceeds step by step in ascending order, for any of the extended distances... :

[0024] P1, enumerate all satisfying... The node combinations (PRF, PEF), where This represents the number of relay hops from the PRF node to the nearest upstream node. This represents the number of relay hops from the nearest downstream node to the PEF node.

[0025] P2, for each combination of nodes, search for redundant paths that meet the requirements;

[0026] P3 selects the redundant path with the lowest overall cost and its corresponding combination as the final PRF, PEF nodes and redundant paths.

[0027] Furthermore, the search requirements for the redundant path include: ① isolation from the original path, and ② time delay difference from the original path. Less than the delay difference threshold ③ It does not form a loop with the upstream or downstream segments of the original path;

[0028] Search according to the following graded degradation mechanism:

[0029] Level 1: Redundant paths have no common intermediate nodes with the original paths;

[0030] Level 2: Redundant paths have no common link with the original path, but are allowed to pass through common nodes;

[0031] Level 3: The redundant path has the fewest common links and nodes with the original path;

[0032] Starting from level one, if no redundant path satisfying conditions ② and ③ is found at the current level, the process degenerates to the next level; if multiple redundant paths satisfying the conditions exist at the current level, the delay difference is selected. The shortest path.

[0033] Furthermore, the comprehensive cost is calculated as follows:

[0034]

[0035] In the formula, For the overall cost, As a cost of time delay, For the cost of resources, For the cost of risk, , , These are the weighting coefficients;

[0036] The costs are calculated as follows:

[0037]

[0038]

[0039]

[0040] In the formula, This represents the number of hops for redundant path nodes between the PRF and PEF. This represents the hop count for the corresponding original path nodes. and This represents the number of nodes and links on the original path from PRF to PEF. and To determine the number of common nodes and common links between redundant paths and the original paths after removing the target subnet, and These are the weighting coefficients;

[0041] Weighting coefficient , , Calculate as follows:

[0042]

[0043]

[0044] In the formula, For the Sigmoid function, The time delay difference sensitivity coefficient, The threshold for focusing on latency difference. For resource sensitivity coefficient, For resource attention threshold, Take a constant value.

[0045] Furthermore, the weighting coefficients and The sensitivity coefficients for time delay difference are 0.3 and 0.7 respectively. The threshold for latency difference is 6. The resource sensitivity coefficient is 0.5. The resource attention threshold is 5. It is 0.5. It is 0.2.

[0046] Furthermore, the time delay difference threshold Determine as follows:

[0047]

[0048]

[0049]

[0050] In the formula, To address the end-to-end delay upper bound of the target flow based on the SLA agreement, This is the upper bound of the actual delay of the target stream being transmitted along the original path. This refers to the maximum time window that a POF node can cache for a target flow, where periodic traffic has... , The maximum buffer depth that the device provides for the target stream packets. The message sending period; For safety margin, .

[0051] Further, set as follows :

[0052]

[0053] In the formula, This refers to the convergence time required for the strategy to be calculated and deployed to the data plane devices to take effect. The handover time required from the detection of an anomaly to the complete takeover by the redundant path. This is a preset additional safety margin.

[0054] Furthermore, in the PREOF strategy, for the determined PRF, PEF nodes and redundant paths, the POF nodes are configured as follows:

[0055] like If the target flow requires ordered message transmission, then a POF node is set;

[0056] like If the target flow can accept out-of-order packets, then it is permissible not to set a POF node;

[0057] The POF node is located either within the PEF node itself or as the first relay node downstream of the PEF node; the buffer size reserved for the target flow is at least [size missing]. ×MTU bytes;

[0058] MTU is the maximum number of bytes that can be sent in a single message in the network. The buffer depth reserved by the device for the target stream packets. Indicates rounding up. This represents the time delay difference between the redundant path and the original path. For message sending period, This represents the maximum jitter.

[0059] Furthermore, the strategy is compiled and distributed:

[0060] PREOF policies are compiled into data plane forwarding entries and distributed to the corresponding network devices through the southbound interface. The forwarding entries include: target flow identifier; replication entries of PRF nodes, containing the outgoing interface list and explicit routing information of redundant paths; elimination entries of PEF nodes; and sorting entries of POF nodes, containing the cache depth.

[0061] Operation monitoring and dynamic maintenance:

[0062] After PREOF is enabled, the following monitoring and maintenance operations should be performed continuously:

[0063] Redundant path monitoring: The packet loss rate and latency of redundant paths are monitored through the OAM mechanism; when abnormal packet loss occurs on a redundant path, a new redundant path is searched again, and faulty links or nodes are excluded during the search.

[0064] Delay difference threshold Dynamically updated: Updates dynamically according to changes in network status. Value; when When a change in value causes the current redundant path to no longer meet the requirements, a new redundant path is searched.

[0065] PREOF Exit Decision: When all anomalies in the target subnets have been repaired, and the prediction is in If the target is not exceeded again within the specified time, PREOF protection will be terminated. The termination process is as follows: first, cancel the replication action of the PRF node; then wait for the remaining replicas to be cleared from the network, with the waiting time not less than the maximum delay of the redundant path; finally, cancel the relevant configurations of the PEF node and POF node.

[0066] Furthermore, we deploy reinforcement learning agents based on deep Q-networks for optimization: , , , , , , , The value of ;

[0067] in:

[0068] The state space includes: the current load rate of each link, the latency of each hop, the packet loss rate of each hop, and the available cache of each node;

[0069] The reward function is designed as follows:

[0070]

[0071] in This is an indicator function for SLA violations, set to 1 if a violation occurs, and 0 otherwise; the resource utilization increment is the percentage of additional resource consumption after enabling PREOF compared to when it is not enabled. , , For weights.

[0072] The beneficial effects of this invention are as follows:

[0073] (1) This invention provides a deterministic network redundancy protection method, which can dynamically decide whether to enable PREOF, where to enable it, and how to select redundant paths based on the real-time network status, thereby meeting the SLA requirements of extremely low packet loss rate, low latency, and low jitter for deterministic transmission; wherein, through a three-level status assessment mechanism, PREOF is only enabled when the network transmission quality is abnormal or has a deterioration trend, thus avoiding the waste of network resources caused by always enabling redundancy protection.

[0074] (2) The present invention first sets PRF and PEF nodes based on minimum coverage. When no qualified redundant path can be found within the minimum coverage range, the present invention further provides an ordered search of PRF and PEF nodes in ascending order of extension distance, which avoids the high overhead of end-to-end exhaustive search. Moreover, it achieves a balance between latency difference, resource overhead and risk through a comprehensive cost function, which is more conducive to practical application.

[0075] (3) The present invention improves the findability of redundant paths while ensuring the protection effect of redundant paths through a three-level degradation mechanism of non-intersecting nodes, non-intersecting links, and minimum intersection.

[0076] (4) This invention incorporates end-to-end latency and POF cache depth into the latency difference threshold. The calculation ensures that the redundant path selection results meet the service quality requirements in all dimensions.

[0077] (5) Reinforcement learning is introduced to assist optimization on the basis of the structured decision framework, which takes into account the interpretability, safety and adaptability of the decision. Attached Figure Description

[0078] Figure 1 This is a flowchart illustrating the overall process of the deterministic network redundancy protection method of the present invention.

[0079] Figure 2 This is a flowchart of step S4.2, ordered extension search, in this invention;

[0080] Figure 3 This diagram illustrates the PRF node, PEF node, and the redundant paths between them.

[0081] Figure 4 This is the DetNet network domain topology of Example 1;

[0082] Figure 5 This is the DetNet network domain topology for Example 2. Detailed Implementation

[0083] Embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein similar or identical reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0084] I. Technical Solution

[0085] like Figure 1 As shown, this invention provides a method for providing redundancy protection for deterministic network traffic based on dynamic PREOF decision-making, comprising the following steps:

[0086] S1, Status Assessment

[0087] Through the OAM (Operations, Administration and Maintenance) mechanism, the DetNet controller continuously monitors the transmission quality metrics of the DetNet stream along the path, including: hop-by-hop packet loss rate, hop-by-hop latency, end-to-end packet loss rate, end-to-end maximum latency, and jitter; simultaneously, it predicts future time windows based on historical monitoring data. Predict the transmission quality within the system.

[0088] For the DetNet stream of interest (hereinafter referred to as the "target stream"), its status is divided into three levels—normal, warning, and protection—based on the current and predicted end-to-end packet loss rate. The specific classification rules are as follows:

[0089] 1) Normal state: The current end-to-end packet loss rate meets the SLA (Service Level Agreement) requirements, and it is predicted that it will decrease in the future. Continue to comply with the SLA within the specified time; when the target flow is in a normal state, the PREOF mechanism is not enabled.

[0090] 2) Warning state: The current end-to-end packet loss rate meets the SLA, but it is predicted that it will increase in the future. There is a tendency to exceed the SLA within the duration; when the target flow is in the warning state, the redundant path calculation and policy pre-deployment of subsequent steps S2 to S5 are performed, but the packet copying action is not performed on the PRF node for the time being.

[0091] 3) Protected state: When the end-to-end packet loss rate has exceeded or approached the SLA boundary, the target flow is in the protected state. The PREOF mechanism is enabled for the target flow. According to the pre-issued policy, packet replication is performed at the PRF node, and the packets are transmitted on the original path and the redundant path. Duplicate packets are eliminated at the PEF node, and packets are reordered at the POF node.

[0092] The above prediction time window Calculate as follows:

[0093]

[0094] In the formula, This refers to the convergence time required for the strategy to be calculated and deployed to the data plane devices to take effect. The handover time required from the detection of an anomaly to the complete takeover by the redundant path. This is an additional safety margin pre-set by humans. Based on the above calculations... The value is set to ensure that the entire process of switching from the warning state to the protection state can be completed before the SLA is actually violated.

[0095] In summary:

[0096] Scenario 1: When the target flow is detected to change from normal state to warning state, the following steps are executed: S2~S6 to issue the PREOF policy, but the PREOF mechanism is not actually activated.

[0097] Scenario 2: When the target flow is detected to change from the warning state to the protection state, the PREOF mechanism is activated based on the issued PREOF policy, and step S7 is executed to implement operation monitoring and dynamic maintenance;

[0098] Scenario 3: When the target flow is detected to change directly from the normal state to the protection state, the following steps S2 to S7 are executed in full.

[0099] S2, Identify the target subnet

[0100] In the architecture required by the RFC 8655 standard, DetNet flows use explicit routing to specify the packet transmission path from the source node to the destination node. Specifically, in this invention, the path nodes listed in the explicit route must include at least every relay node along the path. This path is also called the original path, to distinguish it from the redundant path generated by PREOF.

[0101] When the target flow enters the warning state or protection state, the controller identifies links or continuous link segments with abnormal packet loss rates on the original path based on hop-by-hop OAM data, and defines the network area covered by the above links or continuous link segments as the target subnet.

[0102] Optionally, when there are several discontinuous target subnets on the original path, if the number of interval nodes between the target subnets is less than the preset merging threshold, they can be merged together with the interval nodes into a large target subnet; otherwise, subsequent steps S3 to S5 are executed independently for each target subnet.

[0103] S3, Determine the time delay difference threshold

[0104] Before performing a redundant path search, the controller first determines the delay difference threshold. . This is the upper limit of the allowable difference between the end-to-end delay of the redundant path and the original path between the PRF node and the PEF node. Its value is also subject to the following constraints:

[0105] 1) SLA constraints on end-to-end delay:

[0106]

[0107] In the formula, To address the end-to-end delay upper bound of the target flow based on the SLA agreement, This is the upper bound of the actual delay of the target stream being transmitted along the original path.

[0108] 2) POF cache depth constraint:

[0109]

[0110] In the formula, The maximum time window that a POF node can cache for a target flow; for periodic traffic, there is... ,in The maximum buffer depth that the device provides for the target stream packets. The message sending period; For safety margin, this embodiment takes... .

[0111] final Take the minimum value of the two constraints mentioned above, that is:

[0112]

[0113] The data is dynamically updated according to the network status, and the update cycle is consistent with the OAM statistical cycle. The measured values ​​are smoothed by an exponentially weighted moving average and then entered into the calculation.

[0114] S4, Determine the PRF and PEF node locations and redundant paths.

[0115] PRF nodes replicate DetNet packets and then forward them to both the original path and the redundant path, ensuring reliable packet delivery even if a path fails. PEF nodes record the sequence numbers of received DetNet packets and discard duplicate sequence number copies, refraining from forwarding them to downstream nodes. The determination of the location of PRF and PEF nodes and the redundant path involves two phases: minimum coverage attempt and ordered extension search.

[0116] S4.1, Minimal Coverage Attempt

[0117] like Figure 3 As shown on the left, the first relay node encountered when backtracking from the target subnet towards the source node (upstream) (i.e., the nearest upstream relay node in the target subnet) is selected as the PRF node, and the first relay node encountered when backtracking from the target subnet towards the destination node (downstream) (i.e., the nearest downstream relay node in the target subnet) is selected as the PEF node. A redundant path is searched between the PRF node and the PEF node that satisfies the following conditions: ① isolated from the original path, ② the time delay difference from the original path (denoted as ) (less than) ③ It does not form a loop with the upstream or downstream segments of the original path.

[0118] The latency mentioned here includes the latency of the message at the nodes along the path and the link latency, specifically including the processing latency, shaping latency, queuing latency, and transmission latency within the node as described in RFC9320, the propagation latency on the link, and the frame preemption latency.

[0119] The phase isolation mentioned here includes node disjointness and link disjointness, specifically using a hierarchical degradation mechanism, trying the following steps in order of priority:

[0120] Level 1: Nodes do not intersect, meaning that the redundant path and the original path have no common intermediate nodes;

[0121] Level 2: Links do not intersect, meaning that the redundant path has no common link with the original path, but is allowed to pass through common nodes (except for nodes within the target subnet).

[0122] Level 3: Minimal intersection, meaning the redundant path has the fewest common links and nodes with the original path (excluding nodes and links within the target subnet).

[0123] If multiple redundant paths meet the conditions under the current isolation level, select the path with the best latency difference. The minimum path. If no redundant path that meets the conditions is found at the first level, the search degenerates to the second level; if no path is found at the second level, the search degenerates to the third level, selecting the path that meets conditions ② and ③ and has the fewest common links and nodes with the original path.

[0124] If a redundant path that meets the conditions is found within the minimum coverage area, the PRF and PEF locations are confirmed, and the process proceeds to step S5; otherwise, step S4.2 is executed.

[0125] S4.2, Ordered Extended Search

[0126] like Figure 3 As shown on the right, a 1-hop relay is defined as the distance from one relay node to the next. The relay node closest to the target subnet in the upstream segment is called the "nearest upstream node," and the relay node closest to the target subnet in the downstream segment is called the "nearest downstream node." This represents the number of relay hops from the PRF node to the nearest upstream node. This represents the number of relay hops from the nearest downstream node to the PEF node.

[0127] like Figure 2 As shown, based on the preset extension distance The range of values ​​is arranged in ascending order, gradually targeting... For each value, perform the following steps to search for PRF and PEF nodes:

[0128] P1, for any extension distance Enumerate all that satisfy The node combinations (PRF, PEF) constitute a set. ;

[0129] P2, for sets Each combination in the search employs the hierarchical degradation mechanism in S4.1 to search for combinations that simultaneously satisfy "phase isolation" and "latency difference". < Redundant paths that "do not form loops":

[0130] 1) If no redundant path that meets the conditions is found, the combination is eliminated;

[0131] 2) If a redundant path that meets the conditions is found, then the combination and the corresponding redundant path are added to the candidates;

[0132] 3) If multiple redundant paths that meet the conditions are found, the combination is included as a candidate, and the path with the largest delay difference is selected. The path with the shortest length is selected as the candidate path for this combination;

[0133] P3, regarding the currently set extension distance :

[0134] 1) If no candidate combination exists, then... Increment by 1, and then search for PRF and PEF nodes again under the next value.

[0135] 2) If there is only one candidate combination, then the nodes of that combination and its candidate paths are selected as the final PRF nodes, PEF nodes and redundant paths.

[0136] 3) If there are multiple (i.e., more than 1) candidate combinations, calculate the comprehensive cost of each candidate combination (and its candidate paths) as follows, and select the combination with the smallest comprehensive cost for use as the final selected PRF node, PEF node, and redundant path:

[0137] 3.1) Calculate the cost of delay difference Resource costs Risk and cost :

[0138]

[0139]

[0140]

[0141] In the formula, for any candidate combination (PRF, PEF): delay difference cost Based on the time delay difference between its candidate path and the original path Divide by the delay difference threshold Acquisition; Resource Cost Based on the number of hops of its candidate path nodes Divide by the number of hops in the original path Gain; Risk and Cost The intersection degree between the candidate path and the original path is calculated, where nodes are non-intersecting. If only the links are disjoint or have minimal intersection, then , and This represents the number of nodes and links on the original path from PRF to PEF. and To eliminate the number of common nodes and common links between the candidate paths and the original paths after removing the target subnet, and The weighting coefficients are 0.3 and 0.7 in this embodiment.

[0142] 3.2) Calculate the dynamic weights among the latency cost, resource cost, and risk cost:

[0143]

[0144]

[0145] In the formula, For the Sigmoid function, The time delay difference sensitivity coefficient is used to control... The steepness of the change in time delay difference, with a default value of 6; The latency difference threshold is set to 0.5 by default. When the threshold is exceeded It began to increase significantly; This is a resource sensitivity coefficient used for control. The steepness of the curve as network resources change, with a default value of 5; This is the resource attention threshold, with a default value of 0.5. When it exceeds this value It began to increase significantly; The value is directly assigned as a constant; in this example, it is 0.2. Finally, , , The weights are obtained by normalizing the three factors. , , .

[0146] 3.3) Obtain the comprehensive cost through weighted summation:

[0147]

[0148] P4, based on the above extension distance If, during the progressive search for PRF and PEF nodes, no valid PRF and PEF nodes are found until the following termination conditions are met, an alarm will be reported on the controller, and the network operator will decide whether to accept service degradation:

[0149] Termination condition: The PRF node has been extended to the first relay node of the target flow, and the PEF node has been extended to the last relay node of the target flow; or the extension distance... The limit has been reached. .

[0150] S5, POF node settings

[0151] Based on the time delay difference between the redundant path and the original path determined in step S4 Set the POF nodes according to the following quantitative criteria:

[0152] Scenario 1: If ( If the target flow requires ordered packet transmission (to meet the maximum jitter promised by the SLA), a POF node must be configured. The POF node can be located either within the PEF node itself (i.e., the relay node has both PEF and POF functionality enabled) or in the first downstream relay node of the PEF node. The buffer size reserved for this target flow should be at least [size missing]. ×MTU bytes, where MTU is the maximum number of bytes in a single message that can be sent in the network. The buffer depth reserved by the device for the target stream packets. This indicates rounding up to the nearest integer.

[0153] Scenario 2: If If the target flow can accept out-of-order packets, then a POF node is not required.

[0154] S6, Strategy Compilation and Distribution

[0155] The controller compiles the PREOF policy generated in steps S4 and S5 into data plane forwarding entries and distributes them to the corresponding network devices through the southbound interface. The forwarding entries include: a destination flow identifier; replication entries for PRF nodes, containing an outgoing interface list and explicit routing information for redundant paths; elimination entries for PEF nodes; and sorting entries for POF nodes, containing cache depth.

[0156] S7, Operation Monitoring and Dynamic Maintenance

[0157] After PREOF is enabled, the controller continuously performs the following monitoring and maintenance operations:

[0158] S7.1, Redundant Path Monitoring: The packet loss rate and latency of redundant paths are monitored through the OAM mechanism. When abnormal packet loss occurs on a redundant path, steps S4 to S6 are re-executed to search for a new redundant path, excluding faulty links or nodes during the search.

[0159] S7.2, Delay Difference Threshold Dynamically updated: Updates dynamically according to changes in network status. Value. When If the value change causes the current redundant path to no longer meet the constraints, steps S4 to S6 are re-executed.

[0160] S7.3, PREOF Exit Decision: When all anomalies in the target subnets have been repaired, and the prediction is in If there is no further deterioration trend within the specified time, PREOF protection will be discontinued. The discontinuation process is as follows: first, cancel the replication action of the PRF node; then, wait for the remaining replicas to be cleared from the network, with a waiting time not less than the maximum delay of the redundant path; finally, cancel the relevant configurations of the PEF and POF nodes.

[0161] To address the aforementioned redundancy protection methods, reinforcement learning techniques can be used to assist in the optimization of S1, S3, and S4.

[0162] 1) Triggering decision optimization: By using a reinforcement learning agent, the evolution of network states is learned based on historical OAM data, which helps the controller to make more accurate predictions of SLA non-compliance and optimize the timing of switching between the three states.

[0163] 2) Dynamic adjustment and optimization: The reinforcement learning agent dynamically adjusts its cache based on the actual situation of the relay node's cache resources. When resources are plentiful, expand appropriately. thus relax This relaxes the search for redundant paths.

[0164] 3) Adaptive optimization of cost calculation parameters: The reinforcement learning agent learns latency sensitivity based on feedback from the effects of historical decisions. and attention threshold The optimal value; learning resource sensitivity and attention threshold The optimal value; learning risk cost calculation coefficient. , Optimal values ​​and weights The optimal value.

[0165] The reinforcement learning agent employs a two-stage training deployment approach: simulation pre-training and online fine-tuning. The first stage involves offline pre-training in a network simulation environment to obtain an initial policy. The second stage involves online fine-tuning in an actual network controller, continuously optimizing the policy based on actual decision-making performance. When the reinforcement learning agent's decisions lead to a violation of the Service Level Agreement (SLA), it automatically reverts to the rule-based baseline policy.

[0166] II. Example

[0167] Example 1, Minimum Coverage Search

[0168] like Figure 4 As shown, a DetNet network domain includes source terminal SRC, destination terminal DST, relay nodes N1-N7 (operating in the DetNet service and forwarding sublayers), and forwarding nodes T1-T33 (operating only in the DetNet forwarding sublayer). The link delays (uniform in both directions) between each node are labeled in the figure, in milliseconds (ms). The original path of DetNet flow F1 is SRC→N1→T1→N2→T8→T9→N3→T13→T16→N4→T20→T21→N5→T26→N6→T30→N7→T33→DST, with the node delay at each relay or forwarding node not exceeding 20μs (0.02ms). Therefore, the upper bound of the end-to-end delay of the original path of F1 is... =25 + 17 × 0.02 = 25.34 ms.

[0169] Stream F1 carries high-speed visual inspection services from SRC to DST at a rate of 60Mbps, with an SLA requirement of an upper limit for end-to-end latency. 40ms maximum jitter The response time is 2ms, and the packet loss rate is no more than 10%. -6 .

[0170] In this example, the IP network MTU is 1500 bytes, and the interval between adjacent packets in flow F1 is 200μs. The Alternate-Marking Method described in RFC9341 is used to perform hop-by-hop performance testing on flow F1. The marked packets are counted at the ingress and egress interfaces of the nodes to calculate the packet loss rate. The nodes complete the local statistics of packet loss, latency, and jitter with a sliding window of 10ms, and asynchronously report the statistical results to the control plane at a period of 20ms. This monitors real-time performance while avoiding excessive control plane signaling overhead caused by high-frequency reporting.

[0171] S1, State Assessment and Trigger Determination

[0172] In this example, the convergence time required for the policy calculation and deployment to the data plane devices to take effect is approximately 25ms. The switchover time required from detecting the anomaly to the redundant path fully taking over (including the time for the PRF node to activate replication and for S-BFD to confirm the connectivity of the redundant path) is approximately 50ms, with an additional safety margin of 10ms. Therefore, the prediction time window... =60 ms.

[0173] The controller continuously monitors the transmission quality metrics of flow F1 along its original path using the OAM mechanism. At a certain point, it detects that the end-to-end packet loss rate has increased to 6.5 × 10⁻⁶. -7 And predict the future The end-to-end packet loss rate will exceed the SLA threshold within the specified time. According to the three-level state division rule, the state of flow F1 is determined to be in the warning state. In the warning state, redundant path calculation and policy pre-deployment are performed in steps S2 to S5, but packet replication is not performed at the PRF node. The connectivity of the redundant path is only verified by the S-BFD method described in RFC7880 / 7881 / 7882.

[0174] S2, Identify the target subnet

[0175] Based on hop-by-hop OAM data, the controller identifies the link T13→T16 on the original path as exhibiting an abnormal packet loss rate. Since this is the only anomaly on the original path, the target subnet is T13→T16, with the nearest upstream relay node being N3 and the nearest downstream relay node being N4.

[0176] S3, Determine the time delay difference threshold

[0177] Before performing a redundant path search, the controller first determines the delay difference threshold. It is subject to the following two constraints:

[0178] (1) SLA constraints on end-to-end delay: =40-25.34=14.66ms

[0179] (2) POF cache depth constraint:

[0180] In this example, the maximum buffer depth that the device hardware can provide for stream F1. 64 messages; message sending period = 0.2ms, the maximum time window that a POF node can buffer for stream F1. =12.8ms,

[0181] Safety margin =1.28ms, therefore =11.52ms

[0182] Take the minimum value of the two constraints: = 11.52ms

[0183] S4, Determine the PRF and PEF node locations and redundant paths.

[0184] S4.1, Minimal Coverage Attempt

[0185] Select the nearest upstream relay node N3 of the target subnet as the PRF node, and select the nearest downstream relay node N4 of the target subnet as the PEF node.

[0186] Search between N3 and N4 for segments that are isolated from the original path (N3→T13→T16→N4) and have a time delay difference from the original path. Redundant paths:

[0187] The original path delay is: 1+3+1+2×0.02=5.04ms (the delay of the three links N3→T13, T13→T16, and T16→N4, and the delay of the two nodes T13 and T16).

[0188] Search for redundant paths using a graded degradation mechanism:

[0189] Level 1: Nodes do not intersect (redundant paths do not share intermediate nodes with the original paths, i.e., they do not pass through T13 and T16).

[0190] exist Figure 4 In the topology, the possible paths from N3 to N4 (excluding T13 and T16) include:

[0191] Path A: N3→T12→T15→N4, link delay 3+3+3=9ms, node delay 2×0.02=0.04ms, total delay 9.04ms; =9.04-5.04=4ms

[0192] Path B: N3→T11→T14→T17→N4, link delay 2+2+2+3=9ms, node delay 3×0.02=0.06ms, total delay 9.06ms; =9.06-5.04=4.02ms

[0193] Both paths satisfy the node non-intersection constraint, and Select The shortest path, path A: N3→T12→T15→N4, is the redundant path. At this point, the positions of the PRF and PEF and the redundant path are determined, and there is no need to perform the ordered extension search in S4.2.

[0194] S5, POF node settings

[0195] Whether to set a POF node is determined based on quantification criteria. Because =4ms> Furthermore, F1 requires ordered message transmission, so a POF node must be configured. In this example, the POF location is set at N4 (i.e., N4 simultaneously enables both PEF and POF functions), representing the actual POF buffer depth reserved for flow F1. =20 messages, the corresponding buffer size is 20 × 1500 = 30000 bytes.

[0196] S6, Strategy Compilation and Distribution

[0197] The controller compiles the aforementioned PREOF policy into data plane forwarding table entries and distributes them to the corresponding network devices via the southbound interface (using the NETCONF protocol in this example). The forwarding table entries include:

[0198] (1) Target stream identifier: DetNet stream identifier of stream F1;

[0199] (2) Replicated entries of PRF node N3:

[0200] Outgoing interface 1: points to T13 (original path direction), explicit route is N3→T13→T16→N4;

[0201] Outgoing interface 2: points to T12 (redundant path direction), explicit route is N3→T12→T15→N4;

[0202] (3) Elimination entries of PEF node N4: For F1 messages from T16 and T15, identify and eliminate duplicate copies based on the sequence number;

[0203] (4) Sorting entries of POF node N4: The buffer depth is 20 packets (30KB), and the packets are reordered according to the sequence number before being forwarded downstream.

[0204] Since flow F1 is currently in an alert state, after the policy is issued, PRF node N3 will not perform replication on the data packets of flow F1 for the time being. It will only perform S-BFD detection on the redundant path N3→T12→T15→N4 to verify the connectivity of the redundant path.

[0205] S7, Operation Monitoring and Dynamic Maintenance

[0206] After the policy pre-deployment is completed, the controller continues to monitor the transmission quality of flow F1. After approximately 3 seconds, OAM data shows that the packet loss rate of the T13→T16 link has increased to 8×10⁻⁻⁻⁶. 7 The end-to-end packet loss rate exceeds 9×10⁻ 7 It is approaching the SLA boundary.

[0207] The controller upgrades the state of flow F1 from alert to protection. At this time, the controller sends a command to PRF node N3 to initiate the replication of data packets for flow F1. Starting from N3, each packet of flow F1 is forwarded simultaneously along both the original path (N3→T13→T16→N4) and the redundant path (N3→T12→T15→N4). PEF / POF node N4 performs elimination and reordering operations on packets from both paths to ensure that only one ordered packet is forwarded downstream.

[0208] After PREOF is enabled, the controller continuously performs the following monitoring and maintenance operations:

[0209] S7.1 Redundant Path Monitoring: The controller continuously monitors the packet loss rate and latency of redundant paths through an OAM mechanism (such as the aforementioned alternating marking method). If an abnormal packet loss rate occurs on a redundant path, the controller will re-execute steps S4 to S6 to search for a new redundant path (such as path B: N3→T11→T14→T17→N4), excluding abnormal links or nodes during the search.

[0210] S7.2, Dynamic update: The controller recalculates the measured values ​​using an exponentially weighted moving average smoothing process at 20ms intervals. Value. In this example, the delay difference between the current redundant path and the original path. The time remains at 4ms, which meets the constraints, so there is no need to search for redundant paths again.

[0211] S7.3, PREOF Exit Detection: After a period of time, the controller's OAM data confirms that the performance of the T13→T16 link has returned to normal, and predicts that... If there is no further deterioration trend, then the PREOF exit process will be initiated:

[0212] (1) First, cancel the replication action of PRF node N3 to flow F1, and N3 resumes forwarding only along the original path;

[0213] (2) Then, wait for the remaining copies to be cleared in the network. The waiting time should not be less than the maximum delay of the redundant path. In this example, the waiting time is 12ms to leave a margin.

[0214] (3) Finally, cancel the elimination and sorting configuration of upstream F1 on PEF / POF node N4 and release the relevant cache resources.

[0215] At this point, the state of flow F1 returns to normal, and the PREOF protection exit is complete.

[0216] Example 2, Ordered Extended Search

[0217] In this embodiment, the deterministic flow F1 from the source terminal SRC to the destination terminal DST runs as follows: Figure 5 The DetNet network domain shown is... Figure 4 compared to, Figure 5 The only difference is that the link latency of T23-N5 is 13ms, while that of T23-N5 is 15ms. The parameters, controller mechanism, and settings of flow F1 are the same as in Example 1.

[0218] S1, State Assessment and Trigger Determination

[0219] The controller continuously monitors the transmission quality metrics of flow F1 along its original path using the OAM mechanism. At a certain point, the controller detects that the packet loss rate has increased to 6 × 10⁻⁶. -7 And predict the future The end-to-end packet loss rate within a given timeframe may exceed the 10% promised by the SLA. -6 The state of flow F1 is determined to be a warning state.

[0220] S2, Identify the target subnet

[0221] The controller detects from hop-by-hop OAM data that the target subnet causing the abnormal packet loss is segment T20→T21, with upstream relay node N4 and downstream relay node N5.

[0222] S3, Determine the time delay difference threshold

[0223] Constrained by SLA end-to-end delay: =40-25.34=14.66ms;

[0224] Constrained by POF cache depth: =12.8-1.28=11.52ms.

[0225] but =11.52ms.

[0226] S4, Determine the PRF and PEF node locations and redundant paths.

[0227] S4.1, Minimal Coverage Attempt

[0228] In the minimum coverage attempt, the PRF node is N4 and the PEF node is N5. The sum of the link delays of the original path N4→T20→T21→N5 is 1+1+1=3ms, the upper limit of node delay is 2×0.02=0.04ms, and the upper limit of total delay is 3.04ms.

[0229] Searching for redundant paths from N4 to N5 yields:

[0230] Path A: N4→T18→T23→N5, total delay is 3+2+13+2×0.02=18.04ms. =18.04-3.04=15ms>11.52ms, which does not meet the delay difference constraint and is therefore rejected.

[0231] Path B: N4→T19→T22→T24→N5, total delay is 2+1+1+15+3×0.02=19.06ms. =19.06-3.04=16.02ms>11.52ms, which does not meet the delay difference constraint and is therefore eliminated.

[0232] Exhaustive analysis revealed that within the minimum coverage area (PRF=N4, PEF=N5), there were no redundant paths that satisfied all constraints, necessitating an ordered extension search.

[0233] S4.2, Ordered Extended Search

[0234] From the extended distance Starting with =1, the ordered extension search steps described in P1~P4 are executed cyclically as follows:

[0235] P1, enumeration satisfies + Combinations with a value of 1;

[0236] P2, for each combination, determine the PRF and PEF node positions, and search for all phase isolation constraints and time delay differences that satisfy the hierarchical degradation mechanism. For paths that do not form a loop with the original path, the path with the smallest delay difference is selected as the candidate redundant path, and then its comprehensive cost is calculated:

[0237] Combination 1: PRF is N3, PEF is N5. The original path is N3→T13→T16→N4→T20→T21→N5, with a total upper limit of delay = 1+3+1+1+1+1+5×0.02=8.1ms. A candidate path 1 that meets the conditions is obtained: N3→T11→T14→T17→N5, satisfying the condition that the nodes are non-intersecting, with an upper limit of delay = 2+2+2+5+3×0.02=11.06ms. =11.06-8.1=2.96ms<11.52ms;

[0238] The overall cost is calculated below according to steps Q1~Q4:

[0239] Q1, Calculation of the three costs:

[0240] Delay difference cost = 2.96 / 11.52 = 0.257

[0241] Resource costs = 4 / 6 = 0.667

[0242] Risk Cost =0 (nodes do not intersect);

[0243] Q2, Cost Weight Calculation:

[0244] Delay difference sensitivity =6, latency difference attention threshold =0.5, then the delay difference cost weight based on SLA margin is = 0.189; Take resource sensitivity =5, resource attention threshold =0.5, the current network resource utilization rate of the redundant path. =0.35, resource utilization alarm threshold =0.8, then the resource cost weight based on network resource status is = 0.423; direct assignment = 0.2.

[0245] Q3, weight normalization, yields =0.233, =0.521, =0.243.

[0246] Q4, Weighted Overall Cost = 0.233×0.257+0.521×0.667+0= 0.407.

[0247] Combination 2: PRF is N4, PEF is N6. The original path is N4→T20→T21→N5→T26→N6, with a total upper limit of delay = 1+1+1+2+2+4×0.02=7.08ms. A candidate path 2 that meets the conditions is obtained: N4→T19→T22→T24→T27→T29→N6, satisfying the condition that the nodes are non-intersecting, with an upper limit of delay = 2+1+1+2+2+2+5×0.02=10.1ms. =10.1-7.08=3.02ms<11.52ms;

[0248] The corresponding comprehensive cost is calculated as follows: =0.400

[0249] P3, based on comprehensive cost, the final solution has PRF of N4, PEF of N6, and redundant path N4→T19→T22→T24→T27→T29→N6. =3.02ms.

[0250] P4. When d=1, a redundant path satisfying the constraints has been found, and the extended search process ends. The positions of PRF and PEF, as well as the redundant path, are then finally determined.

[0251] S5, POF node settings

[0252] In this example, the POF node is set up on N6 (meaning that N6 enables both PEF and POF functions). The actual POF buffer depth reserved for flow F1 is 16 packets, and the buffer size is 16 × 1500 = 24000 bytes.

[0253] S6, Strategy Compilation and Distribution

[0254] The controller compiles the aforementioned PREOF policy into data plane forwarding table entries and distributes them to the corresponding network devices via the NETCONF protocol. The forwarding table entries include:

[0255] (1) Target stream identifier: DetNet stream identifier of stream F1;

[0256] (2) Replicated entries of PRF node N4:

[0257] Outgoing interface 1: points to T20 (original path), explicit route is N4→T20→T21→N5→T26→N6;

[0258] Outgoing interface 2: points to T19 (redundant path), explicit route is N4→T19→T22→T24→T27→T29→N6;

[0259] (3) Elimination entries of PEF node N6: For F1 messages from T26 and T29, identify and eliminate duplicate copies based on the sequence number;

[0260] (4) Sorting entries of POF node N6: The buffer depth is 16 packets (24KB). The packets are reordered according to the sequence number and then forwarded downstream.

[0261] Since flow F1 is currently in an alert state, after the policy is issued, PRF node N4 will not perform replication on the data packets of flow F1 for the time being, but will only detect the connectivity of redundant paths through S-BFD.

[0262] S7. Operation monitoring and dynamic maintenance, the process is similar to that in Example 1, and will not be repeated here.

[0263] Example 3, Reinforcement Learning Assistance

[0264] This embodiment, based on Embodiment 1, further illustrates the implementation method of reinforcement learning-assisted optimization:

[0265] The controller deploys a reinforcement learning agent based on a deep Q-network. The state space includes, but is not limited to: the current load rate of each link, the latency of each hop, the packet loss rate of each hop, and the available buffer space of each node. The action space can include: three-level state transition decisions and the maximum buffer depth provided for the F1 flow. Delay difference sensitivity and attention threshold Resource sensitivity and attention threshold Risk cost coefficient , and weight The value of the reward function is designed as follows:

[0266]

[0267] in This is the indicator function for SLA violations, set to 1 if a violation occurs and 0 otherwise; the resource utilization increment is the percentage of additional resource consumption after enabling PREOF compared to when it is not enabled; the decision switching frequency penalizes policy changes that are too frequent.

[0268] First, an initial policy was obtained through 100,000 offline pre-training cycles in a DetNet network simulation environment. Then, it was loaded into the reinforcement learning agent and fine-tuned online with a small learning rate. A safety monitoring mechanism was implemented in the controller: if an SLA violation occurred within three consecutive monitoring cycles, the reinforcement learning agent's decisions were automatically reverted to the rule-based baseline policy, and the abnormal scenario was recorded for subsequent offline analysis and model improvement.

[0269] This invention is not limited to the above-described embodiments. Any obvious improvements, substitutions, or modifications that can be made by those skilled in the art without departing from the essence of this invention are within the scope of protection of this invention.

Claims

1. A deterministic network redundancy protection method, characterized in that: S1.1, Identify the target flow status as follows: Normal state: along the original path, both currently and in the future. Throughout the period, the packet loss rate met the target requirements; Warning status: Along the original path, the current packet loss rate meets the target requirements, but it is predicted that the loss rate will increase in the future. The target will be exceeded within the specified timeframe; Protected state: Along the original path, the current packet loss rate does not meet the target requirements; S1.2, Execute protection actions based on the target flow status: When the target flow changes from a normal state to an alert state, the PREOF strategy is issued; When the target flow changes from the warning state to the protection state, the PREOF mechanism is executed based on the issued policy.

2. The deterministic network redundancy protection method according to claim 1, characterized in that: In the PREOF strategy, the PRF and PEF node locations and redundant paths are determined as follows: Phase 1, Minimum Coverage Attempt: The nearest upstream node of the target subnet is designated as the PRF node and the nearest downstream node as the PEF node. If there is a redundant path between the two nodes that meets the requirements, the PRF and PEF node positions and the redundant path are determined. If there is no redundant path between the two nodes that meets the requirements, then proceed to stage 2. Phase 2, Ordered Extended Search: By extension distance The search proceeds step by step in ascending order, for any of the extended distances... : P1, enumerate all satisfying... The node combinations (PRF, PEF), where This represents the number of relay hops from the PRF node to the nearest upstream node. This represents the number of relay hops from the nearest downstream node to the PEF node. P2, for each combination of nodes, search for redundant paths that meet the requirements; P3 selects the redundant path with the lowest overall cost and its corresponding combination as the final PRF, PEF nodes and redundant paths.

3. The deterministic network redundancy protection method according to claim 2, characterized in that: The search requirements for the redundant path include: ① isolation from the original path, and ② time delay difference from the original path. Less than the delay difference threshold ③ It does not form a loop with the upstream or downstream segments of the original path; Search according to the following graded degradation mechanism: Level 1: Redundant paths have no common intermediate nodes with the original paths; Level 2: Redundant paths have no common link with the original path, but are allowed to pass through common nodes; Level 3: The redundant path has the fewest common links and nodes with the original path; Starting from level one, if no redundant path satisfying conditions ② and ③ is found at the current level, the process degenerates to the next level; if multiple redundant paths satisfying the conditions exist at the current level, the delay difference is selected. The shortest path.

4. The deterministic network redundancy protection method according to claim 3, characterized in that: The comprehensive cost is calculated as follows: In the formula, For the overall cost, As a cost of time delay, For the cost of resources, For the cost of risk, , , These are the weighting coefficients; The costs are calculated as follows: In the formula, This represents the number of hops for redundant path nodes between the PRF and PEF. This represents the hop count for the corresponding original path nodes. and This represents the number of nodes and links on the original path from PRF to PEF. and To determine the number of common nodes and common links between redundant paths and the original paths after removing the target subnet, and These are the weighting coefficients; Weighting coefficient , , Calculate as follows: In the formula, For the Sigmoid function, The time delay difference sensitivity coefficient, The threshold for focusing on latency difference. For resource sensitivity coefficient, For resource attention threshold, Take a constant value.

5. The deterministic network redundancy protection method according to claim 4, characterized in that: The weighting coefficient and The sensitivity coefficients for time delay difference are 0.3 and 0.7 respectively. The threshold for latency difference is 6. The resource sensitivity coefficient is 0.

5. The resource attention threshold is 5. It is 0.

5. It is 0.

2.

6. The deterministic network redundancy protection method according to claim 4, characterized in that: The time delay difference threshold Determine as follows: In the formula, To address the end-to-end delay upper bound of the target flow based on the SLA agreement, This is the upper bound of the actual delay of the target stream being transmitted along the original path. This refers to the maximum time window that a POF node can cache for a target flow, where periodic traffic has... , The maximum buffer depth that the device provides for the target stream packets. The message sending period; For safety margin, .

7. The deterministic network redundancy protection method according to claim 1, characterized in that: Configure as follows : In the formula, This refers to the convergence time required for the strategy to be calculated and deployed to the data plane devices to take effect. The handover time required from the detection of an anomaly to the complete takeover by the redundant path. This is a preset additional safety margin.

8. The deterministic network redundancy protection method according to claim 1, characterized in that: In the PREOF strategy, for the determined PRF, PEF nodes and redundant paths, the POF nodes are configured as follows: like If the target flow requires ordered message transmission, then a POF node is set; like If the target flow can accept out-of-order packets, then it is permissible not to set a POF node; The POF node is located either within the PEF node itself or as the first relay node downstream of the PEF node; the buffer size reserved for the target flow is at least [size missing]. ×MTU bytes; MTU is the maximum number of bytes that can be sent in a single message in the network. The buffer depth reserved by the device for the target stream packets. Indicates rounding up. This represents the time delay difference between the redundant path and the original path. For message sending period, This represents the maximum jitter.

9. The deterministic network redundancy protection method according to claim 6, characterized in that: Strategy compilation and distribution: PREOF policies are compiled into data plane forwarding entries and distributed to the corresponding network devices through the southbound interface. The forwarding entries include: target flow identifier; replication entries of PRF nodes, containing the outgoing interface list and explicit routing information of redundant paths; elimination entries of PEF nodes; and sorting entries of POF nodes, containing the cache depth. Operation monitoring and dynamic maintenance: After PREOF is enabled, the following monitoring and maintenance operations should be performed continuously: Redundant path monitoring: The packet loss rate and latency of redundant paths are monitored through the OAM mechanism; when abnormal packet loss occurs on a redundant path, a new redundant path is searched again, and faulty links or nodes are excluded during the search. Delay difference threshold Dynamically updated: Updates dynamically according to changes in network status. Value; when When a change in value causes the current redundant path to no longer meet the requirements, a new redundant path is searched. PREOF Exit Decision: When all anomalies in the target subnets have been repaired, and the prediction is in If the target is not exceeded again within the specified time, PREOF protection will be terminated. The termination process is as follows: First, the replication action of the PRF node is cancelled; then, the remaining replicas are allowed to be cleared from the network, and the waiting time is not less than the maximum delay of the redundant path; finally, the relevant configurations of the PEF node and POF node are cancelled.

10. The deterministic network redundancy protection method according to claim 6, characterized in that: Deploying and optimizing reinforcement learning agents based on deep Q-networks: , , , , , , , The possible values ​​of ; in: The state space includes: the current load rate of each link, the latency of each hop, the packet loss rate of each hop, and the available cache of each node; The reward function is designed as follows: in This is an indicator function for SLA violations, set to 1 if a violation occurs, and 0 otherwise; the resource utilization increment is the percentage of additional resource consumption after enabling PREOF compared to when it is disabled. , , For weights.

Images