Semantic residual routing method with quality of service protection

Abstract

The application discloses a semantic residual routing method with quality of service protection and belongs to the technical field of space-ground integration network. The application comprises the following steps: acquiring a task request, candidate semantic states and network states; determining residual data amounts corresponding to each candidate state; generating candidate paths; combining the states and the paths, calculating path performance by using the residual data amounts and judging whether the hard quality of service constraint is satisfied; establishing reference path performance and determining a relative protection boundary; adding combinations which simultaneously satisfy the hard constraint and the relative boundary to a feasible set, calculating joint scores and selecting an optimal one for transmission; and executing a safe rollback when no feasible combination exists. The method establishes quantitative coupling of semantic multiplexing, load and routing, sets double-layer quality of service boundaries which cannot be covered, takes state-path combinations as decision units, and solves the problem that load and performance calculation are disconnected.

Description

Technical Field

[0001] This invention relates to a semantic residual routing method with quality of service protection, belonging to the field of integrated air-ground-space network technology. Background Technology

[0002] The integrated space-ground-air network consists of low-Earth orbit satellites, airborne platforms, ground gateways, mobile terminals, edge nodes, and cloud nodes. Different links vary significantly in terms of propagation distance, visibility time, capacity, queue pressure, packet loss risk, and availability. The network topology continuously changes with satellite movement, airborne platform movement, and service load.

[0003] Traditional shortest path, capacity-first, reliability-first, or queue-aware routing typically assume that the amount of data to be transmitted is determined before route selection. In semantic state reuse scenarios, the actual transmission load is determined by the candidate semantic states used: the more suitable the candidate state is for the current task, the less corrective residuals need to be transmitted; when state aging, semantic drift, version mismatch, or high uncertainty occur, the residuals increase and may degenerate into complete transmission. Therefore, candidate state selection, residual load determination, and path selection are inextricably coupled.

[0004] Existing cached routing often requires traffic to pass through cache nodes or retrieve complete objects from cache nodes, which may result in additional detours. In semantic residual transmission, the cached state serves as conditional knowledge for the receiver to recover the target state, and the cache node does not need to be a physical forwarding node; however, if the most similar state is selected first and then the path is selected, it may result in a combination of excessively long propagation distances, insufficient reliability, or excessively low bottleneck capacity.

[0005] Simply incorporating latency, success rate, capacity, and semantic gains into a single weighted scoring function cannot establish an insurmountable quality of service boundary. Excessive weighting of similarity or reuse gains may offset significant network risks, causing high-scoring combinations to violate time-bound or reliable delivery requirements.

[0006] Furthermore, different candidate semantic states generate different residual data volumes, and the same path has different transmission service times and feasibility under different states. If path performance is calculated only based on the full data volume or a fixed data volume, it may incorrectly delete originally feasible small residual paths, or incorrectly accept state-path combinations with large actual residuals. Summary of the Invention

[0007] To address the problem of the disconnect between transmission load changes caused by candidate semantic states and path performance calculation, this invention provides a semantic residual routing method with quality of service protection.

[0008] The present invention provides a semantic residual routing method with quality of service protection, comprising:

[0009] Obtain task requests, candidate semantic state sets, and dynamic network states;

[0010] For each candidate semantic state, the corresponding target residual ratio is determined based on the semantic difference between the candidate semantic state and the task request and the state validity risk. Then, combined with the complete data volume, the corrective semantic residual data volume is obtained.

[0011] Based on the dynamic network state, a candidate path set is generated according to at least two path evaluation views;

[0012] Based on the candidate path set, candidate semantic states are combined with candidate paths. The performance index of the corresponding candidate path is calculated using the corrective semantic residual data volume corresponding to each candidate semantic state in the combination. Based on the performance index of the candidate path, it is determined whether the corresponding combination meets the hard service quality constraints of the task request itself.

[0013] Based on the performance metrics of the candidate paths, a reference path performance for the task request is established, and based on the reference path performance and the service quality requirements in the task request, a relative reference path protection boundary is determined.

[0014] The combination that simultaneously satisfies the hard quality of service constraint and the relative reference path protection boundary is added to the feasible combination set. The joint score of each combination in the feasible combination set is calculated. The combination with the highest joint score is selected from the feasible combination set as the target state-path combination. If the feasible combination set is empty, a safe rollback is performed. If no feasible combination can be generated by the safe rollback, the empty candidate semantic state and the safe path are used as the target state-path combination.

[0015] The corrective semantic residual data volume is transmitted according to the target state-path combination.

[0016] Preferably, the semantic difference includes the semantic similarity between the candidate semantic state and the task request, and the state validity risk includes the erroneous semantic state age and semantic version mismatch of the candidate semantic state.

[0017] Preferably, the target residual ratio is:

[0018] ;

[0019] in, Based on residuals, Let m be the age of the erroneous semantic state of task request i in time slot t. Let be the age normalization constant. It is a very small constant. As task weight, The weighting coefficient for the age of the erroneous semantic state. The weighting coefficients are the weights of the tasks. This is the weighting coefficient for semantic version mismatch. For the semantic version of task request i, For the semantic version of the candidate semantic state m; This is a semantic version mismatch indicator function; This represents the truncation function.

[0020] As preferred, performance metrics include end-to-end latency, delivery success rate, bottleneck capacity, average queue pressure, and hop count.

[0021] As a preferred option, hard service quality constraints include:

[0022] ;

[0023] For path In the time slot Capacity End-to-end delay For path In the time slot Delivery success rate For path In the time slot The bottleneck capacity The maximum allowed delay for task request i. Let i be the minimum delivery success rate for task request i. Let i be the minimum bottleneck capacity for task request i.

[0024] Preferably, the reference path performance includes reference latency. Reference loss and reference bottleneck capacity ;

[0025] The reference delay The fast path is the smaller of the end-to-end delay of the fast path and the end-to-end delay of the semantic candidate path. The fast path is the path with the smallest end-to-end delay in the candidate path set, and the semantic candidate path is the path induced by any candidate semantic state.

[0026] The reference loss The safe path is the path with the smallest packet loss rate among the candidate paths, which is the smaller value between the packet loss rate of the safe path and the packet loss rate of the semantic candidate path.

[0027] The reference bottleneck capacity The larger of the bottleneck capacity of the fast path and the bottleneck capacity of the semantic candidate path is given.

[0028] As a preferred option, the relative reference path protection boundary is:

[0029] ;

[0030] ;

[0031] ;

[0032] in, For path In the time slot Capacity End-to-end delay For path In the time slot Delivery success rate For path In the time slot The bottleneck capacity For time delay protection tolerance parameters; For packet loss protection tolerance parameters; This is the absolute protection tolerance parameter for packet loss; For capacity protection tolerance parameters; The reference delay of task request i in time slot t.

[0033] As the preferred option, the combined score is:

[0034]

[0035] in, For semantic reuse gain, The value of the candidate state. To predict congestion. For similarity, For path In the time slot Capacity End-to-end delay For path In the time slot Delivery success rate For path In the time slot The bottleneck capacity To average queue pressure for the path, For the number of jumps, The maximum allowed delay for task request i. , , , , , , , , , These are the weighting coefficients for the corresponding items.

[0036] Preferably, at least two path evaluation views include at least two of the following: latency view, capacity view, queue view, packet loss view, and restricted detour view;

[0037] Candidate paths are generated using the K-short-circuit algorithm on each view. The generated candidate paths are then deduplicated and loops are removed to obtain the candidate path set.

[0038] As a preferred option, a safe rollback should be performed in the following order, step by step:

[0039] Increase the target residual ratio;

[0040] Cancel the current candidate semantic state and select another candidate semantic state;

[0041] Switch to a safe path in the candidate path set; set the target residual ratio to 1 and perform full data transmission.

[0042] The beneficial effects of this invention are as follows: This application quantifies information such as the similarity of candidate semantic states, the age of erroneous semantic states, semantic versions, and task weights into a target residual ratio. This is then combined with the complete data volume to obtain a unique corrective semantic residual data volume for each candidate state. This dynamic residual bit volume is used as a direct input for calculating link service time and path end-to-end delay. This allows path performance evaluation to no longer be based on approximate estimations using a fixed data volume, but rather to precisely correspond to the actual transmission load. Thus, a quantitative coupling relationship is established between semantic multiplexing, transmission load, and dynamic routing. Path delay dynamically changes with the candidate semantic state, and the feasibility of the same path under different states is independently and accurately evaluated. Small residual paths that were mistakenly deleted under a fixed data volume evaluation are retained, and dangerous combinations with large actual residuals are accurately identified and correctly eliminated in subsequent screening. This fundamentally solves the problem of path misjudgment caused by incorrect data volume assumptions. Attached Figure Description

[0043] Figure 1 This is a flowchart of the semantic residual routing method with quality of service protection in this application;

[0044] Figure 2 This is a schematic diagram of the joint scoring and service quality screening of candidate semantic states and candidate paths in this application;

[0045] Figure 3 This is a diagram showing the relationship between the hard service quality constraints, relative reference path protection boundaries, and safe fallback in this application;

[0046] Figure 4The figures show the simulation results of average end-to-end latency and average bandwidth utilization under different request arrival rates in this application. (a) represents latency, and (b) represents bandwidth utilization. Detailed Implementation

[0047] 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.

[0048] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0049] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0050] This embodiment of the semantic residual routing method with quality of service protection is used to select the optimal combination of candidate semantic states and candidate paths for task requests in an integrated air-space-ground network. The core concept of this method is to quantify the semantic differences and state validity risks between candidate semantic states and task requests into a target residual ratio, thereby obtaining corrective semantic residual data. This corrective semantic residual data is dynamic and is used as input for path performance evaluation, achieving quantitative coupling between semantic decision-making and network routing. Simultaneously, a two-layer insurmountable screening mechanism is constructed through task hard quality of service constraints and relative reference path protection boundaries to ensure that all selected combinations meet quality of service requirements in both absolute task requirements and relative network baseline dimensions. Finally, a joint score is calculated for the selected feasible combinations, and the optimal one is selected for transmission. When no feasible combination is found, a step-by-step safety fallback is performed. Specifically, this includes:

[0051] S1. Obtain the task request, candidate semantic state set, and dynamic network state;

[0052] In the current time slot At the beginning, obtain pending task requests. Obtain the candidate semantic state set and obtain the dynamic network state.

[0053] Task Request Includes the following parameters: Total data volume Task weight , is a real number greater than or equal to 1; maximum allowable delay Units are milliseconds; minimum delivery success rate , is a real number between 0 and 1; minimum bottleneck capacity .

[0054] The set of candidate semantic states is denoted as The set contains There are 10 candidate semantic states. Each candidate semantic state... It has the following corresponding attributes: similarity , is a real number between 0 and 1; Error semantic state age Semantic version State value Basic residuals , for the middle Real numbers between 1 and 0.

[0055] Dynamic network state in the current time slot The integrated space-air-ground network is represented as .

[0056] Node set Including satellite, aviation, gateway, edge, cloud, and user terminals:

[0057]

[0058] For a set of satellite nodes, For a collection of aviation platform nodes, For the set of gateway nodes, For the set of edge nodes, For cloud node collection, A collection of user terminals;

[0059] link The state vector is:

[0060]

[0061] in, For link nominal capacity, For link The propagation and fixed processing delay, For link queue pressure, For link The probability of packet loss, For link Link utilization For link Link type, For link The availability indicator is set to 1, which means the link is available, and 0 means the link is unavailable.

[0062] link The propagation and fixed processing delays are determined by the node distance and propagation speed, and the link The nominal capacity is determined by the link budget or equivalent capacity model. Packet loss and queue pressure are obtained from current network measurements or short-term predictions.

[0063] Calculate the remaining effective capacity of each link based on the link status:

[0064]

[0065] in, For link In the time slot The remaining effective capacity. If satellite movement causes link issues, From 1 to 0, when hour, The link and candidate paths containing it are deleted; when hour, If the link utilization rate approaches 1, the transmission service time will increase significantly and may trigger QoS filtering.

[0066] S2. Determine the target residual ratio and the amount of corrective semantic residual data;

[0067] The set of candidate semantic states obtained in S1 Based on this, for each candidate semantic state Based on the semantic differences between the candidate semantic state and the task request, as well as the risk of state validity, the corresponding target residual ratio is determined, and then combined with the complete data volume, the corrective semantic residual data volume is obtained.

[0068] Semantic differences include the semantic similarity between candidate semantic states and task requests. The lower the semantic similarity, the higher the proportion of the target residual.

[0069] State validity risks include incorrect semantic state age of candidate semantic states. Semantic version mismatch. Incorrect semantic state age. The larger the value, the higher the proportion of the target residual. Semantic version mismatch refers to the semantic version of task request i. Semantic version of candidate semantic state m Inconsistent.

[0070] Target residual ratio Determined in the following manner:

[0071]

[0072] in, For the task request corresponding to the candidate semantic state m In the time slot The target residual ratio; Based on residuals, For the task request corresponding to the candidate semantic state m In the time slot The age of the erroneous semantic state of t. Let be the age normalization constant. It is a very small constant. As task weight, The weighting coefficient for the age of the erroneous semantic state. The weighting coefficients are the weights of the tasks. This is the weighting coefficient for semantic version mismatch. For task request semantic version, For the semantic version of the candidate semantic state m; This is a semantic version mismatch indicator function; This represents the truncation function;

[0073] The similarity between a candidate semantic state and a task request determines the degree to which the state can be reused: the higher the similarity, the less corrective information needs to be transmitted. The age of an erroneous semantic state reflects the validity risk of the state: the older the state, the higher its uncertainty, requiring more corrective information to compensate for the biases caused by aging. Semantic version mismatch implies differences in encoding and decoding methods; when versions differ, additional residuals are needed to ensure correct decoding. Task weights reflect priority; high-weight tasks require more conservative residual strategies to ensure recovery quality.

[0074] Determining the target residual ratio Then, the amount of corrective semantic residual data corresponding to the candidate semantic state m is calculated as follows:

[0075]

[0076] in, For the task request corresponding to the candidate semantic state m In the time slot The amount of corrective semantic residual data; This represents the complete data volume.

[0077] For the candidate semantic state set The above calculation is performed on each candidate semantic state m to obtain a set of residual data indexed by the candidate semantic states. Therefore, a task request no longer corresponds to a single fixed load, but rather to a set of residual loads indexed by candidate states.

[0078] S2 quantifies semantic-level differences and risks into a network-level processable corrective semantic residual data volume. This data volume is no longer a fixed value but dynamically changes with different candidate semantic states m: the higher the semantic similarity, the younger the age of the erroneous semantic state, and the more consistent the semantic version with the task request, the smaller the residual data volume; conversely, the larger it is. This dynamic residual data volume establishes a quantitative coupling between semantic reuse, transmission load, and dynamic routing, providing dynamic input for the accurate calculation of path performance metrics in S4. The execution of S2 enables the same path to have different transmission service times and feasibility under different candidate semantic states, no longer relying on approximate estimations based on the complete data volume.

[0079] S3. Generate a set of candidate paths;

[0080] Based on the dynamic network state obtained from S1, evaluate the view according to at least two paths in the current network graph. A candidate path set is generated. At least two path evaluation views are required, including at least two of the following: latency view, capacity view, queue view, packet loss view, and restricted detour view. Candidate path set. :

[0081]

[0082] The set of paths generated for the low-latency view. The set of paths generated for a high-capacity view. The set of paths generated for low-queue views. The set of paths generated for a low packet loss view. A set of paths generated for a restricted detour view.

[0083] The latency view is used to retain paths with low propagation and processing overhead, the capacity view is used to retain paths with high service rates, the queue view is used to avoid congested nodes, the packet loss view is used to improve delivery success rate, and the restricted detour view is used to discover paths that are not the shortest but can meet QoS with small residuals.

[0084] Each view has a corresponding link cost:

[0085] The link cost in the latency view is:

[0086]

[0087] in This is the queue pressure conversion factor.

[0088] The link cost of the capacity view is:

[0089]

[0090] The link cost of the queue view is:

[0091]

[0092] The link cost for packet loss view is:

[0093]

[0094] The K-short-circuit algorithm is applied to each view to generate K shortest paths for each view, where K is a pre-configured positive integer. In addition to the K-short-circuit algorithm, this implementation can also use constrained shortest path, A* search, or Dijkstra's algorithm after link cost recalibration.

[0095] The paths generated by each view are combined to obtain an initial candidate path set. Then, deduplication and loop deletion are performed on the initial candidate path set. Paths containing loops, paths exceeding the maximum number of hops, and paths completely dominated by other paths are deleted.

[0096] After performing deduplication and loop deletion, a hop count limit is also applied: only items that meet the criteria are retained. The path, where Let P be the number of hops. This is the maximum number of additional hops allowed beyond the fast path hop count.

[0097] Also perform dominant deletion on candidate paths: if path With path Belonging to the same view, and It is comparable to [the previous one] in end-to-end latency, delivery success rate, and bottleneck capacity. If at least one is strictly better, then delete it. .

[0098] The set of paths after the above processing is the candidate path set, which includes fast paths. Safe path Semantic candidate paths And other candidate paths. Fast path The path with the minimum end-to-end delay from the candidate path set; safe path. The path with the minimum grouping loss rate in the candidate path set; semantic candidate path The path induced by any candidate semantic state;

[0099] In this embodiment, S2 and S3 belong to two processing stages: S2 determines the amount of corrective semantic residual data corresponding to each candidate semantic state and completes the semantic candidate compression process; S3 generates a candidate path set based on the dynamic network state. This two-stage structure is used to reduce the online search space, but it does not mean that the selection of candidate semantic states and candidate path selection are independent of each other. The final decision object is still a state-path combination consisting of a candidate semantic state and a candidate path, and S7 will perform joint scoring and selection in this combination space.

[0100] S3 generates a candidate path set covering multiple optimization directions, avoiding the omission of feasible paths suitable for small residual transmission by relying solely on the shortest path metric. Specifically, the latency view retains paths with low propagation and processing overhead; the capacity view retains paths with high service rates, enabling small residual data volumes to leverage the low transmission latency advantage of large-capacity links; the queue view avoids congested nodes; the packet loss view improves delivery success rate; and the restricted detour view discovers paths that are not the shortest but can meet service quality requirements with small residuals. Deduplication, loop removal, hop count limiting, and dominance removal constitute a pruning mechanism. Among them, dominance removal utilizes Pareto dominance relationships to remove inferior paths completely covered by other paths in terms of multi-dimensional performance, compressing the candidate set size without losing the optimal solution. The combined effect of multi-view path generation, hop count limiting, and dominance removal keeps the online search space of state-path combinations within an operable range, avoiding the exponential complexity problem caused by enumerating the complete Cartesian product of node-state-path, enabling this method to be executed online in large-scale dynamic integrated space-ground networks.

[0101] S4. Calculate the performance indicators of state-path combinations and determine hard service quality constraints.

[0102] After obtaining the corrective semantic residual data volume corresponding to each candidate semantic state in S2 and the candidate path set in S3, the candidate semantic states and candidate paths are combined to form a state-path combination. For each combination, the corrective semantic residual data corresponding to the candidate semantic state m in that combination is used. Calculate the performance metrics of the candidate path P, and determine whether the combination meets the hard service quality constraints of the task request itself based on the performance metrics.

[0103] The performance metrics for path P include end-to-end latency. Delivery success rate Bottleneck capacity Average queue pressure and number of jumps .

[0104] End-to-end delay Calculated as follows: For links ,link In terms of data volume The time delay is:

[0105]

[0106] The end-to-end delay of path P is the sum of the delays of all links on the path:

[0107]

[0108] In the above calculations, It is the amount of corrective semantic residual data obtained by S2 calculation, and its value varies depending on the candidate semantic state m.

[0109] Delivery success rate Calculate as follows:

[0110]

[0111] in, Let be the packet loss probability for link e. The packet loss rate for the path is . .

[0112] Bottleneck capacity Calculate as follows:

[0113]

[0114] Average queue pressure Calculate as follows:

[0115]

[0116] Number of jumps This represents the number of links contained in path P.

[0117] After calculating the above performance metrics, it is determined whether the combination meets the hard quality of service constraints of the task request itself. The hard quality of service constraints are:

[0118]

[0119] For task request Maximum allowable delay For task request Minimum delivery success rate, For task request The minimum bottleneck capacity;

[0120] All three conditions must be met simultaneously. If any one condition is not met, the combination is determined not to meet the hard quality of service constraints. Perform the above judgment on each combination separately and record the result of whether each combination meets the hard quality of service constraints.

[0121] like Figure 2 As shown, after calculating the above performance metrics, the performance metrics for each state-path combination and the judgment results regarding whether the combination meets the hard service quality constraints are collected and used as input data for subsequent steps. Specifically, for each state-path combination... The collected data includes: similarity obtained by S1 Error semantic state age State value The residual ratio calculated by S2 and corrective semantic residual data volume The end-to-end path delay calculated by S4 Delivery success rate Bottleneck capacity Average queue pressure , number of jumps And whether the combination meets hard service quality constraints; The judgment result.

[0122] S4 achieves quantitative coupling between semantic decision-making and network routing. The link delay formula directly uses the dynamic residual data volume calculated by S2 as the input of the transmitted data volume, and the end-to-end path delay dynamically changes with different candidate semantic states. For the same path P, it carries candidate semantic states. residual data volume With carrying candidate semantic state residual data volume At the same time, the end-to-end latency is different; for the same candidate semantic state m, the latency varies in different paths. and During uplink transmission, end-to-end latency varies. Using state-path combinations as the smallest decision unit, the performance metrics of each combination are independently recalculated, avoiding local optima caused by selecting states before or after paths. Hard service quality constraints mark combinations that violate absolute task requirements as infeasible, providing the first layer of judgment results for S6's two-layer screening.

[0123] S5. Establish reference path performance and determine the relative reference path protection boundary;

[0124] After obtaining the performance metrics of each path in the candidate path set in S4, a reference path performance for the task request is established based on the performance metrics of the candidate paths. Based on the reference path performance and the service quality requirements in the task request, the relative reference path protection boundary is determined.

[0125] Reference path performance includes reference latency. Reference loss and reference bottleneck capacity .

[0126] Task Request Reference delay in time slot t End-to-end delay for the fast path End-to-end latency of semantic candidate paths The smaller value in:

[0127]

[0128] Task Request Reference loss in time slot t Packet loss rate for the safe path Packet loss rate with the semantic candidate path The smaller value in:

[0129]

[0130] Task Request Reference bottleneck capacity in time slot t The bottleneck capacity of the fast path and the bottleneck capacity of the semantic candidate path. The larger value in:

[0131]

[0132] After establishing the reference path performance, the relative reference path protection boundary is determined based on the reference path performance and the service quality requirements in the task requests. The relative reference path protection boundary includes:

[0133] Delay protection boundary:

[0134]

[0135] in, This is the time delay protection tolerance parameter.

[0136] Packet loss protection boundary:

[0137]

[0138] in, For packet loss protection tolerance parameters; This is the absolute protection tolerance parameter for packet loss;

[0139] Capacity protection boundary:

[0140]

[0141] in, This refers to the capacity protection tolerance parameter.

[0142] For high-priority disaster response tasks, a smaller [specification threshold] can be set. Larger Larger and more stringent , , For routine periodic monitoring tasks, a relatively lenient protection tolerance can be adopted. Therefore, the protection boundary is configured according to the task level, rather than using a uniform fixed routing standard.

[0143] S5 establishes a dynamic protection baseline based on the performance of the best path in the current network, rather than using a fixed threshold. Fast paths provide optimal references for latency and capacity, while safe paths provide optimal references for reliability. The relative protection boundary allows for some performance degradation of semantic paths relative to the current network baseline, but does not allow for unlimited expansion. This protection boundary is independent of the scoring function and does not participate in the weighting of coefficients; no scoring gain can compensate for deleted combinations. This completely solves the problem in traditional weighted scoring methods where high semantic gains are offset by high latency or high packet loss penalties through weighting coefficients, preventing semantic gains from being obtained through excessive detours, reliability degradation, or bottleneck capacity sacrifices.

[0144] S6. Screen feasible combinations, calculate joint scores, and perform safe rollback;

[0145] Based on the judgment results of whether each combination meets the hard quality of service constraints recorded in S4, and the relative reference path protection boundary determined in S5, for each state-path combination... Perform the filtering. Simultaneously, record whether each combination meets the relative reference path protection boundary. , , The judgment results are then obtained. At this point, the complete data for each state-path combination includes: similarity, age of the erroneous semantic state, state value, residual ratio, actual number of bits, path latency, delivery success rate, bottleneck capacity, average queue pressure, hop count, feasibility of hard quality of service constraints, and feasibility of the relative reference path protection boundary. This complete dataset is used for subsequent calculation of the joint score.

[0146] Combinations that simultaneously satisfy hard quality of service constraints and relative reference path protection boundaries are added to the set of feasible combinations. .

[0147] For feasible combinations Make a judgment: If If empty, then execute a safe rollback; if If the set is not empty, calculate the joint score of each combination in the feasible combination set.

[0148] Combined score Calculate as follows:

[0149]

[0150] in, For semantic reuse gain, The value of the candidate state. To predict congestion. , , , , , , , , , These are the weighting coefficients for the corresponding items.

[0151] The scoring is used to rank combinations within the safe and feasible set, but not to determine infeasible combinations. In other words, even if a combination has high similarity, large reuse gain, or high state value, it cannot re-enter the candidate set by adjusting the weighting coefficients if it violates any of the protection conditions.

[0152] For feasible combinations Each combination in Calculate the joint score separately. If the set of feasible combinations is not empty, proceed to S7 after calculation. If the set of feasible combinations is empty, perform a safe rollback, using the empty candidate semantic state and the safe path as the target state-path combination, and proceed to S8.

[0153] The safety rollback is executed step by step in the following order until an executable solution is found:

[0154] Level 1 backoff: Keep the candidate semantic state unchanged and increase the target residual ratio. Increase the step size based on the current value, recalculate the corrective semantic residual data volume, and re-execute the filtering process from S4 to S6. If a feasible combination exists under the increased residual ratio, use that combination as the target state-path combination and proceed to S8.

[0155] Second-level backoff: If the first-level backoff still cannot produce a feasible combination, cancel the current candidate semantic state and select another candidate semantic state. From the set of candidate semantic states... Select the next candidate state ,by Repeat the filtering process from S2 to S6. If a feasible combination exists, use that combination as the target state-path combination and proceed to S8.

[0156] Third-level fallback: If the second-level fallback still cannot produce a feasible combination, switch to a safe path in the candidate path set. The fixed path is a safe path. Traverse the set of candidate semantic states Check if there exists a combination that simultaneously satisfies both hard quality of service constraints and relative reference path protection boundaries. If so, use this combination as the target state-path combination and proceed to S8.

[0157] This order can preserve some semantic benefits when conditions permit, while ensuring that an executable solution eventually exists.

[0158] Fourth-level rollback: If none of the above rollback levels can produce a feasible combination, set the target residual ratio to 1 and perform a full data transfer. Cancel semantic reuse, using empty candidate semantic states and safe paths. As the target state-path combination, proceed to S8.

[0159] Based on the complete dataset described above for each combination, the joint score for that combination is calculated using the joint score formula. Each input term in the joint score formula is derived from the aggregated data for that combination.

[0160] S6 employs a two-layer screening process to ensure that all combinations entering the scoring stage have passed dual verification against both absolute task requirements and relative network baselines—both conditions are indispensable. The separation of scoring and screening prevents weighting coefficients from eroding the security boundary—the score is only used to differentiate between combinations that have passed security review and does not assume any security determination function. The joint score comprehensively evaluates multiple dimensions such as semantic similarity, reuse gain, state value, delivery success rate, bottleneck capacity, normalized latency, queue pressure, semantic risk, hop count, and predicted congestion, providing a fine-grained ranking of combinations that pass the screening.

[0161] S7. Select the target state-path combination;

[0162] When feasible combinations When the set of feasible combinations is not empty, select the combination with the highest joint score from the set of feasible combinations as the target state-path combination:

[0163]

[0164] Among them, among them, The optimal candidate semantic state for task request i Find the optimal candidate path for task request i; traverse the set of feasible combinations. Compare the combined scores of all combinations. The combination with the highest score is selected as the final decision. If multiple combinations have the same highest score, they are selected according to the following priority order: priority is given to delivery success rate. The highest-performing combination; if still tied, prioritize end-to-end latency. The smallest combination; if still tied, prioritize the number of jumps. The smallest combination.

[0165] S7 selects the solution that best combines semantic benefits and network performance from feasible combinations. Since S6 has removed all insecure combinations, the selection scope of S7 is limited to a safe set where both absolute task requirements and relative network baselines are met. This optimal solution is a joint optimality of state and path; any pairing of candidate semantic states with any candidate path under non-sequential decision-making is independently evaluated, ensuring that optimal combinations are not overlooked. The parallel processing rule guarantees the determinism of route selection.

[0166] S8. Transmit corrective semantic residual data according to the target state-path combination;

[0167] The transmission operation is performed according to the target state-path combination. The target state-path combination comes from one of the following two approaches: when the set of feasible combinations is not empty, it comes from the one selected by S7. When the set of feasible combinations is empty and a safe rollback is performed, the target state-path combination generated from the S6 safe rollback.

[0168] The corrective semantic residual data corresponding to the target candidate semantic state is encapsulated into data packets and sent to the receiving end along the target path.

[0169] The candidate semantic state serves as prior knowledge for the receiver: the target candidate semantic state itself is not transmitted with the corrective semantic residual data packet. The target candidate semantic state can be provided to the receiver in advance through control plane identification, pre-deployment, existing receiver state, or previously synchronized state. After receiving the corrective semantic residual data packet, the receiver uses the locally acquired target candidate semantic state as prior knowledge to decode and recover the received corrective semantic residual, obtaining the target state requested by the task. Since the target candidate semantic state is not required to be located on the physical forwarding path for transmitting the corrective semantic residual, there is no need to change the residual physical path to access cache nodes.

[0170] When the safety rollback is executed to level four, along the safety path Transmit complete data volume It does not rely on any candidate semantic state for recovery, and the receiving end directly receives the complete data.

[0171] After S8 completes execution, the task request... In the current time slot The processing flow has ended.

[0172] For any path P, the marginal effect of residual load on path delay is:

[0173]

[0174] This equation shows that, with a fixed path, reducing residual load inevitably reduces transmission service time, but propagation delay and queue penalties still exist. Therefore, semantic multiplexing cannot replace quality of service protection. This marginal impact is implicitly reflected in the calculation of end-to-end delay in S4.

[0175] like Figure 4 As shown, SAGE-HC represents the method of this application. Under the same dynamic network topology, request sequence, and link conditions, as the request arrival rate increases from 50 to 300, the average end-to-end latency of the method of this application remains at approximately 15.9 ms, while the capacity-first and shortest path methods increase from approximately 47 ms to approximately 49.5 ms. At an arrival rate of 300, the average bandwidth utilization of this application is approximately 0.10, while that of the capacity-first and shortest path methods is approximately 0.92 and 0.94, respectively.

[0176] The above results demonstrate that the lower communication load is not achieved through delayed services or uncontrolled detours, but rather through a combination of residual load reduction and QoS-protected path selection. The overall experiments also show that, at comparable delivery reliability, the framework described in this application reduces the average latency by 9.14% and 8.75% compared to capacity-aware and deep reinforcement learning routing, respectively.

[0177] In the ablation results, after removing QoS protection, the hit rate and the number of transmitted bits changed little, but the average latency increased from 41.189ms to 42.630ms, and the task utility decreased by 13.75%. This indicates that QoS protection is not responsible for generating semantic hits, but rather for limiting the network implementation method of semantic benefits.

[0178] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A semantic residual routing method with quality of service protection, characterized in that, include: Obtain task requests, candidate semantic state sets, and dynamic network states; For each candidate semantic state, the corresponding target residual ratio is determined based on the semantic difference between the candidate semantic state and the task request and the state validity risk. Then, combined with the complete data volume, the corrective semantic residual data volume is obtained. Based on the dynamic network state, a candidate path set is generated according to at least two path evaluation views; Based on the candidate path set, candidate semantic states are combined with candidate paths. The performance index of the corresponding candidate path is calculated using the corrective semantic residual data volume corresponding to each candidate semantic state in the combination. Based on the performance index of the candidate path, it is determined whether the corresponding combination meets the hard service quality constraints of the task request itself. Based on the performance metrics of the candidate paths, a reference path performance for the task request is established, and based on the reference path performance and the service quality requirements in the task request, a relative reference path protection boundary is determined. The combination that simultaneously satisfies the hard quality of service constraint and the relative reference path protection boundary is added to the feasible combination set. The joint score of each combination in the feasible combination set is calculated. The combination with the highest joint score is selected from the feasible combination set as the target state-path combination. If the feasible combination set is empty, a safe rollback is performed. If no feasible combination can be generated by the safe rollback, the empty candidate semantic state and the safe path are used as the target state-path combination. The corrective semantic residual data volume is transmitted according to the target state-path combination.

2. The semantic residual routing method with quality of service protection according to claim 1, characterized in that, The semantic difference includes the semantic similarity between the candidate semantic state and the task request, and the state validity risk includes the erroneous semantic state age and semantic version mismatch of the candidate semantic state.

3. The semantic residual routing method with quality of service protection according to claim 2, characterized in that, The target residual ratio is: ； in, Based on residuals, Let m be the age of the erroneous semantic state of task request i in time slot t. Let be the age normalization constant. It is a very small constant. As task weight, The weighting coefficient for the age of the erroneous semantic state. The weighting coefficients are the weights of the tasks. This is the weighting coefficient for semantic version mismatch. For the semantic version of task request i, For the semantic version of the candidate semantic state m; This is a semantic version mismatch indicator function; This represents the truncation function.

4. The semantic residual routing method with quality of service protection according to claim 1, characterized in that, The performance metrics include end-to-end latency, delivery success rate, bottleneck capacity, average queue pressure, and hop count.

5. The semantic residual routing method with quality of service protection according to claim 1, characterized in that, The hard service quality constraints include: ； For path In the time slot Capacity End-to-end delay For path In the time slot Delivery success rate For path In the time slot The bottleneck capacity The maximum allowed delay for task request i. Let i be the minimum delivery success rate for task request i. Let i be the minimum bottleneck capacity for task request i.

6. The semantic residual routing method with quality of service protection according to claim 1, characterized in that, The reference path performance includes reference delay. Reference loss and reference bottleneck capacity ; The reference delay The fast path is the smaller of the end-to-end delay of the fast path and the end-to-end delay of the semantic candidate path. The fast path is the path with the smallest end-to-end delay in the candidate path set, and the semantic candidate path is the path induced by any candidate semantic state. The reference loss The safe path is the path with the smallest packet loss rate among the candidate paths, which is the smaller value between the packet loss rate of the safe path and the packet loss rate of the semantic candidate path. The reference bottleneck capacity The larger of the bottleneck capacity of the fast path and the bottleneck capacity of the semantic candidate path is given.

7. The semantic residual routing method with quality of service protection according to claim 6, characterized in that, The relative reference path protection boundary is: ； ； ； in, For path In the time slot Capacity End-to-end delay For path In the time slot Delivery success rate For path In the time slot The bottleneck capacity For time delay protection tolerance parameters; For packet loss protection tolerance parameters; This is the absolute protection tolerance parameter for packet loss; For capacity protection tolerance parameters; The reference delay of task request i in time slot t.

8. The semantic residual routing method with quality of service protection according to claim 3, characterized in that, The joint score is: ； in, For semantic reuse gain, The value of the candidate state. To predict congestion; For similarity, For path In the time slot Capacity End-to-end delay For path In the time slot Delivery success rate For path In the time slot The bottleneck capacity To average queue pressure for the path, For the number of jumps, The maximum allowed delay for task request i. , , , , , , , , , These are the weighting coefficients for the corresponding items.

9. The semantic residual routing method with quality of service protection according to claim 1, characterized in that, The at least two path evaluation views include at least two of the following: latency view, capacity view, queue view, packet loss view, and restricted detour view; Candidate paths are generated using the K-short-circuit algorithm on each view. The generated candidate paths are then deduplicated and loops are removed to obtain the candidate path set.

10. The semantic residual routing method with quality of service protection according to claim 6, characterized in that, The safety rollback is executed in the following order, level by level: Increase the target residual ratio; Cancel the current candidate semantic state and select another candidate semantic state; Switch to a safe path in the candidate path set; set the target residual ratio to 1 and perform full data transmission.

Images