Task-Driven Semantic Sharing Method for Integrated Space-Ground Networks
By constructing a dynamic heterogeneous network graph and a hierarchical semantic state cache set, and utilizing the instant task utility and synchronous net benefit model, the problem of the lack of task semantic sharing mechanism in the existing caching system is solved, and the efficient utilization of cache resources and closed-loop control of task reuse are realized.
Patent Information
- Application Number
- CN202611325301.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-17
AI Technical Summary
Existing caching systems lack a task-oriented semantic sharing mechanism, resulting in low task reuse rate, poor cache resource utilization, lack of quantitative basis for synchronization decisions, and disconnect between dynamic network modeling and semantic sharing control, making it difficult to operate in a closed loop under the same topology snapshot and the same time slot.
By constructing a dynamic heterogeneous network graph and a hierarchical semantic state cache set, and by comparing the utility of immediate tasks with the baseline of complete delivery without caching, we define effective hits, calculate the value of task-aware caching, set a synchronization net benefit model, implement risk gating and time slot synchronization budget, construct a data structure containing semantic vectors and historical task utility, and perform hierarchical semantic retrieval and proactive synchronization.
Effectively identify and reuse data with mission reuse value from different times and modes, reduce the load on inter-satellite links and satellite-to-ground links, optimize cache resource allocation, avoid resource waste and state obsolescence, and achieve closed-loop operation of each link under the same topology snapshot and the same time slot.
Smart Images

Figure CN123122632A_ABST
Abstract
Description
Technical Field
[0001] This application relates to a task-driven semantic sharing method for integrated space-ground-air-ground networks, belonging to the field of integrated space-ground-air-ground network technology. Background Technology
[0002] The integrated space-ground-ground network consists of low-Earth orbit satellites, near-space platforms, drones, ground gateways, edge servers, cloud nodes, and mobile users. It can provide wide-area sensing and communication services in areas that are difficult to cover with traditional ground infrastructure. The topology of this type of network changes continuously with orbital motion, platform maneuvering, and user movement. Inter-satellite links, satellite-to-ground links, air-to-ground links, and ground backhaul links have significant differences in capacity, propagation latency, reliability, and energy consumption.
[0003] Remote sensing monitoring, disaster response, traffic monitoring, marine observation, and low-altitude patrol missions often repeatedly request the status of the same area, target, or event at different times and in different modalities. If each request involves transmitting complete images, text, or sensor data from the data source, it would result in a large amount of redundant communication and place a high load on inter-satellite links and satellite-to-ground links.
[0004] Traditional caching systems primarily manage file copies, content names, or access frequency, often employing least recently used or least frequently used strategies for replacement. Task-oriented semantic sharing, however, does not require returning identical file copies, but rather a valid state sufficient to support the current task. Observations with different file identifiers, at different times, and in different modalities may have task reuse value; the same file identifier may also become invalid due to changes in context, semantic version upgrades, or increased state uncertainty.
[0005] Existing multi-level caching systems primarily aim to improve hit rates or reduce back-to-origin traffic. High-frequency accesses with low task utility or high risk may occupy scarce caches for extended periods, while low-frequency accesses that correspond to disaster emergencies and critical tasks may be prematurely eliminated.
[0006] Semantic state can be updated passively through request delivery or proactively through synchronization. Unconditional periodic synchronization consumes link and control resources; updating only when requested may render future popular states outdated for a long time.
[0007] Furthermore, dynamic network modeling, task semantic generation, hierarchical caching, state sharing, cache maintenance, synchronization decision-making, and performance evaluation are usually implemented by independent tools, making it difficult to ensure that each link operates in a closed loop under the same topology snapshot and the same time slot. It is also difficult to directly access the real satellite orbits, visibility, and time-varying link scenarios generated by STK. Summary of the Invention
[0008] To address the problems of existing caching systems that use file copies as shared objects and lack a task-oriented semantic closed-loop sharing mechanism, resulting in low task reuse rate, poor cache resource utilization, and lack of quantitative basis for synchronization decisions, this application provides a task-driven semantic sharing method for integrated space-ground-air networks.
[0009] This application discloses a task-driven semantic sharing method for integrated space-ground-air networks, comprising:
[0010] Step 1: Construct a dynamic heterogeneous network graph and a hierarchical semantic state cache set for the current time slot;
[0011] Step 2: Based on the dynamic heterogeneous network diagram of the current time slot, historical observation window data, and the control action set of the previous time slot, obtain the predicted belief state of multiple future time slots, and generate short-term predictive control quantities based on the predicted belief state.
[0012] Step 3: Obtain the current task request for the multimodal task;
[0013] Step 4: Retrieve semantic states in the hierarchical semantic state cache set in the order of near-end cache, residual semantic cache, and global semantic cache. Based on the short-term prediction control quantity and validity gating conditions, form a hierarchical candidate state set for the current task request.
[0014] Step 5: For each semantic state in the hierarchical candidate state set, calculate the immediate task utility, select a delivery method to deliver the task response to the user based on the immediate task utility, determine the delivery result, and update the historical task utility and effective hit count stored in the semantic state based on the immediate task utility after delivery.
[0015] Step 6: Calculate the task-aware cache value based on the historical task utility and the number of valid hits stored in the semantic state. Write the semantic state into the corresponding target cache level according to the admission criteria. When the capacity of the target cache level is insufficient, evict the semantic state with the lowest task-aware cache value in the target cache level to obtain the updated hierarchical semantic state cache set.
[0016] Step 7: Each node obtains the pre-synchronization aggregated state risk value based on the predicted belief state, calculates the synchronization net benefit, and performs an active synchronization operation based on the synchronization net benefit to synchronize the semantic state to the target node.
[0017] As a preferred method, the specific method for calculating the concurrent net income in step seven includes:
[0018] Based on the predicted belief state, obtain the aggregate state risk value before synchronization, the expected residual load value before synchronization, and the expected task utility value before synchronization;
[0019] Calculate the aggregate state risk value, the expected residual load value, and the expected task utility value after synchronization;
[0020] The difference between the aggregation state risk value before and after synchronization, the difference between the expected residual load value before and after synchronization, and the difference between the expected task utility value after and before synchronization are assigned corresponding weight coefficients, summed, and then the synchronization cost is subtracted to obtain the net synchronization benefit.
[0021] Preferably, the method for performing active synchronization based on the aforementioned net synchronization benefit in step seven includes:
[0022] When the risk value of the aggregated state before synchronization is greater than or equal to the risk threshold, and the net benefit of synchronization is greater than or equal to the net benefit threshold, an active synchronization operation is performed.
[0023] The sum of the synchronization costs of all active synchronization operations is less than or equal to the upper limit of the synchronization budget for the current time slot.
[0024] As a preferred method, in step six, the method for calculating the value of the task-aware cache is as follows:
[0025]
[0026] Let m be the average age of the erroneous semantic states in the current time slot t. Let m be the number of valid hits for semantic state m. For the historical task utility of semantic state m, The storage occupancy of semantic state m. The weighting coefficient for the freshness of state term. The weighting coefficients for effectively hitting value items, The weighting coefficients for the historical task utility term. The weighting coefficient for the average state risk penalty term. The weighting coefficient for the relative storage usage penalty term. Let be the time constant for the decay of state freshness. semantic state Current cache layer The maximum cache capacity.
[0027] Preferably, step six, the method for writing the semantic state to the corresponding target cache level according to the admission criteria, includes:
[0028] When the immediate task utility of the semantic state is greater than the task utility of the uncached full delivery baseline, and the task-aware cache value is greater than or equal to the admission threshold of the target cache level, the semantic state is allowed to be written to the target cache level.
[0029] When the remaining capacity of the target cache level is insufficient to accommodate the semantic state to be written, the eviction operation is repeated in the target cache level until the capacity meets the write requirement:
[0030] The semantic state with the lowest task-aware cache value is selected for eviction, and the evictioned semantic state is removed from the hierarchical semantic state cache set.
[0031] Preferably, the method for generating the short-term predictive control quantity in step two includes:
[0032] The short-term predictive control variable includes a task-adaptive shared threshold.
[0033] Extract the queue pressure and link capacity of future time slots from the predicted belief state, and calculate the predicted congestion index. ;
[0034] According to the predicted congestion index Current task weight, average error semantic state age, cache age and cache health metrics Calculate the adaptive shared threshold for the task:
[0035]
[0036] in, For the current task request at the cache level In the current time slot Task adaptive sharing threshold, For cache level The basic sharing threshold, Weight of the current task The adjustment function, For cache level In the current time slot Average error semantic state age cache age The adjustment function, To cache health metrics The adjustment function, To predict congestion indicators The adjustment function, To adapt the lower limit of the shared threshold, An adaptive upper limit for the shared threshold;
[0037] In step four, a hierarchical candidate state set for the current task request is formed based on the task adaptive sharing threshold and validity gating conditions.
[0038] As a preferred option, the validity gating conditions include:
[0039] The semantic similarity between the semantic state and the current task request is greater than or equal to the task adaptive sharing threshold in the short-term predictive control variable.
[0040] The age of the erroneous semantic state is less than or equal to the maximum age of the erroneous semantic state allowed by the current task request;
[0041] The semantic version of the semantic state is consistent with the semantic version of the current task request.
[0042] Preferably, the present invention further includes step eight:
[0043] The delivery results, synchronized semantic state, synchronization cost, updated hierarchical semantic state cache set, and dynamic heterogeneous network graph of the current time slot are written back to the historical observation window data, and the control action set of the current time slot is formed according to the control action executed in this time slot.
[0044] As a preferred method, step five involves updating the number of valid hits, including:
[0045] The effective hit count is incremented by one only if the immediate task utility of the adopted semantic state is greater than the task utility of the uncached full delivery baseline.
[0046] Preferably, the real-time task utility in step five is calculated based on task weight, semantic quality, semantic sharing gain, delivery latency, energy consumption, actual data volume transmitted, and synchronization cost.
[0047] The beneficial effects of this application are as follows: This application performs hierarchical retrieval according to the order of near-end semantic cache, residual semantic cache, and global semantic cache. Semantic states are filtered based on task adaptive sharing thresholds and validity gating conditions, enabling the effective identification and reuse of data with task reuse value from different times and modalities. This avoids redundant transmission caused by task requests repeatedly retrieving complete data from the data source, reducing the load on inter-satellite and satellite-to-ground links. This invention selects the delivery method based on immediate task utility and calculates the task-aware cache value based on the historical task utility and effective hit count stored in the semantic state. The task-aware cache value is used as the admission and elimination criteria. States that are accessed frequently but have low task utility or high risk are eliminated due to their low cache value, while states that are infrequent but correspond to disaster emergencies and critical tasks are retained due to their high historical task utility. Cache resources are tilted towards states that truly support task decisions, preventing cache space from being crowded out by inefficient high-frequency data. This invention performs proactive synchronization operations based on the net synchronization benefit, synchronizing the semantic state to the target node. The net benefit of synchronization quantifies the improvement in risk, load, and utility before and after synchronization, transforming proactive synchronization from unconditional periodic execution or passive waiting for requests into on-demand push with quantifiable basis. Synchronization resources are prioritized for high-risk, high-value states, avoiding the waste of resources from unconditional periodic synchronization and the outdated states caused by passive updates. Attached Figure Description
[0048] Figure 1 This is a flowchart illustrating the method described in this application;
[0049] Figure 2 This is the overall architecture diagram of hierarchical semantic caching and semantic sharing for dynamic integrated space-ground networks as described in this application;
[0050] Figure 3 This is a closed-loop diagram showing the synchronization of cache admission, retention, eviction, and budget constraints in this application.
[0051] Figure 4 The following are simulation results of semantic cache hit rate, semantic reuse gain, task utility and bandwidth utilization under different content popularity skew. Among them, (a) is cache hit rate, (b) is reuse gain, (c) is task utility and (d) is bandwidth utilization.
[0052] Figure 5 The following are simulation results of the task utility, average latency, energy consumption per request, and average CPU utilization of this application under different network scales. Among them, (a) is the task utility, (b) is the average latency, (c) is the energy consumption per request, and (d) is the average CPU utilization.
[0053] Figure 6This is a trend of network CPU utilization and distribution of node CPU utilization in the 230-node and 1840-node scenarios of this application, where (a) is the 230-node scenario and (b) is the 1840-node scenario. Detailed Implementation
[0054] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0055] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.
[0056] The present application 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 application.
[0057] The technical problem this application aims to solve is that existing caching systems manage files / content in a file- / content-centric manner, lacking a task-semantic-oriented sharing mechanism. Specifically, this manifests as: a disconnect between cache lifecycle management and task value; high-frequency, inefficient states crowding out scarce cache space, while low-frequency, critical states are easily prematurely eliminated; proactive synchronization lacks risk-benefit assessment and budget constraints, unconditional periodic synchronization wastes link resources, and passive on-demand updates lead to outdated states; cached objects lack task semantics, and data with task reuse value across different times and modalities cannot be effectively identified and reused; dynamic network simulation and semantic sharing control are fragmented, with each stage implemented by independent tools, making closed-loop operation under the same topology snapshot difficult.
[0058] This application addresses the aforementioned issues through the following technical solutions: Valid hits are defined by comparing immediate task utility with a complete, cache-free delivery baseline; task-aware cache value is used as the basis for admission and elimination, directly linking cache lifecycle management to task value; a synchronization net benefit model quantifies risk improvement, load improvement, and utility enhancement before and after synchronization, setting three conditions: risk gating, net benefit gating, and time-slot synchronization budget, prioritizing synchronization resources for high-risk, high-value states; a semantic state data structure is constructed, including semantic vectors, semantic versions, state uncertainty, historical task utility, and valid hit counts; during retrieval, the similarity between task semantic vectors and cache state semantic vectors is calculated, enabling data with task reuse value to be identified and reused; dynamic heterogeneous network graph construction, predictive belief state updates, hierarchical semantic retrieval, delivery decisions, cache lifecycle management, budget-constrained synchronization, and closed-loop write-back are unified within the same control loop in each time slot; step eight writes back the execution results of this time slot to form the input for the next time slot, ensuring that each stage operates in a closed loop under the same topology snapshot and time-slot state.
[0059] The method in this embodiment is executed cyclically over discrete time slots t, referring to... Figure 1 The method of this implementation includes:
[0060] Step 1: Construct a dynamic heterogeneous network graph and a hierarchical semantic state cache set for the current time slot.
[0061] At the start of time slot t, the status information of all nodes in the integrated space-ground-air network is acquired, including the location information of low-Earth orbit satellite nodes, near-space platforms and UAVs, ground nodes such as ground gateways and edge servers, and mobile user nodes, as well as visibility information, communication distance, link capacity, propagation delay, packet loss probability, queue pressure, and computing resources between nodes. If STK is already connected, the above information is directly read from the orbital parameters and visibility matrix exported by STK.
[0062] For any link e, determine availability:
[0063]
[0064] For link In the current time slot The availability indicator value, where 1 indicates availability and 0 indicates unavailability.
[0065] For link Both ends of the node in the current time slot The visibility flag, where 1 indicates visibility and 0 indicates invisibility;
[0066] For link Both ends of the node in the current time slot Communication distance, For link Maximum supported communication distance;
[0067] This is an indicator function that takes the value 1 when all conditions within the square brackets are true, and 0 otherwise.
[0068] Current time slot The set of available links is as follows:
[0069]
[0070] all The links constitute the set of available links for the current time slot. Nodes are divided into satellite, aviation, ground, and user layers, resulting in a dynamic heterogeneous network diagram:
[0071]
[0072] in, It is a collection of nodes in an integrated space-air-ground network, including satellite nodes, aviation nodes, ground nodes, edge nodes, cloud nodes, and user nodes. For the current time slot The set of available links, For the current time slot Link capacity vector for each available link;
[0073] For the current time slot Propagation delay vectors for each available link;
[0074] For the current time slot Packet loss probability vectors for each available link;
[0075] For the current time slot The unit data energy consumption attribute vector for each available link.
[0076] Each node maintains a semantic state set at each cache level. :
[0077]
[0078] These correspond to the near-end semantic cache, residual semantic cache, and global semantic cache, respectively. Each semantic state m:
[0079]
[0080] Each field describes the semantic state. Target object identifier semantic vectors Semantic version Update time Uncertainty Historical utility Valid hit count Current cache level Storage usage and source context .
[0081] The cache capacity is sufficient:
[0082]
[0083] For nodes At the cache level The maximum cache capacity.
[0084] Step 2: Based on the dynamic heterogeneous network diagram of the current time slot, historical observation window data, and the control action set of the previous time slot, obtain the predicted belief state of multiple future time slots, and generate short-term predictive control variables based on the predicted belief state.
[0085] Network-Semantic World Model Reading Historical observation window data and the set of control actions in the previous time slot If this is the first time the system is started, Includes only data from the current time slot. It is an empty set or a preset initial value.
[0086] Call the prediction function To obtain the future Predicted belief states for each time slot:
[0087]
[0088] from Extracting queue pressure for future time slot t+1 and link capacity Calculate and predict congestion indicators:
[0089]
[0090] in, This is a short-term prediction of the average queue pressure across the entire network links. This is the average link capacity of the entire network as predicted in a short time. To prevent positive decimals with a denominator of zero.
[0091] Calculate cache health metrics:
[0092]
[0093] in, For the current time slot A collection of semantic state caches across the entire network. A collection of semantic state caches for the entire network The total number of semantic states, The age decay coefficient for erroneous semantic states. semantic state In the current time slot The average age of the semantic error state;
[0094] Generate task adaptive sharing threshold:
[0095]
[0096] in, For the current task request at the cache level In the current time slot Task adaptive sharing threshold, For cache level The basic sharing threshold, Weight of the current task The adjustment function, For cache level In the current time slot Average error semantic state age cache age The adjustment function, To cache health metrics The adjustment function, To predict congestion indicators The adjustment function, To adapt the lower limit of the shared threshold, An adaptive upper limit for the shared threshold;
[0097] This threshold is used for hierarchical retrieval. The adjustment function's value rules are as follows: higher task weights, higher cache risk, or poor cache health result in a higher threshold; when predicted congestion is high and cache health is good, the threshold can be lowered within a safe boundary to expand the semantic sharing scope.
[0098] Step 3: Obtain the current task request from multiple modalities and generate a task profile;
[0099] The task semantic construction module obtains current task requests from user terminals or monitoring devices, including one or more of the following: images, text, sensor sequences, and event alarm information. It then performs unified encoding on the multimodal raw data, extracts semantic features, and generates the current task request. semantic vectors Parse task weights from the request. Maximum delay Maximum allowed error semantic state age and semantic version , , , , Create a task profile. Assign higher task weights to critical tasks such as disaster emergency response, and lower task weights to non-urgent tasks such as routine monitoring.
[0100] Step 4: Following the order of near-end semantic cache, residual semantic cache, and global semantic cache, retrieve semantic states from the hierarchical semantic state cache set. Based on short-term prediction control variables and validity gating conditions, form a hierarchical candidate state set for the current task request. Specifically:
[0101] Search level by level in the order of L1, L2, L3. For each cache level... ,from Extract each semantic state m and check whether it meets the validity gating condition one by one:
[0102] Similarity criteria:
[0103] Error semantic state age condition: ;
[0104] Semantic version consistency condition: ;
[0105] in, For the current task request Semantic vectors for medium tasks With semantic state semantic vectors The similarity between them.
[0106] For the current task request The age of the erroneous semantic state m in the current time slot t.
[0107] For the current task request Maximum allowed age of erroneous semantic states.
[0108] For the current task request The required semantic version. Semantic states that simultaneously meet all three conditions are included in the hierarchical candidate state set:
[0109]
[0110] Once a valid state is found at the current level, the process can be terminated early to reduce retrieval latency. If the candidate state set for a layer is empty, step five directly selects the full delivery method.
[0111] Step 5: For each semantic state in the hierarchical candidate state set, calculate the immediate task utility, select the delivery method to deliver the task response to the user based on the immediate task utility, determine the delivery result, and update the historical task utility and effective hit count stored in the semantic state based on the immediate task utility after delivery.
[0112] For each candidate semantic state m, compute the immediate task utility:
[0113]
[0114] in, For the current task request The immediate task utility; These are the weighting coefficients for the task's semantic quality.
[0115] For the current task request The semantic quality obtained; These are the weighting coefficients for the semantic shared gain term;
[0116] For the current task request Semantic shared gain; The weighting coefficient for the delivery delay penalty;
[0117] For the current task request The delivery delay is based on the propagation delay vector. calculate; For the current task request Maximum allowable delay;
[0118] This represents the weighting coefficient for the energy consumption penalty term; For the current task request Energy consumption;
[0119] The weighting coefficient for the penalty term based on the actual amount of data transmitted; For the current task request The actual amount of data transmitted;
[0120] The weighting coefficient for the synchronous cost penalty item;
[0121] For the current task request The synchronization costs incurred.
[0122] Calculate the task utility of a cache-free full delivery baseline .Compare and :
[0123] If there exists m such that The state with the highest immediate task utility is selected as the adopted state. 0;
[0124] Otherwise, no candidate state is used, and the full delivery is executed.
[0125] Delivery methods include three types: direct reuse, residual compensation, and complete delivery.
[0126] Direct reuse: When a valid candidate state exists and can be directly used for the current task, the semantic information is encapsulated into a task response and sent to the user.
[0127] Residual supplementation: When candidate states can serve as conditional knowledge but current change information is still required, combine semantic states with residual data for delivery.
[0128] Complete delivery: When no valid candidate states exist, complete multimodal data is delivered from the original data source.
[0129] Record the delivery results after delivery is completed, including Delivery method , , , .
[0130] The utility of using exponential smoothing to update historical tasks:
[0131]
[0132] in, Exponential smoothing coefficient for the utility of historical tasks
[0133] Update valid hit count:
[0134]
[0135] In other words, the effective hit count is incremented by one only when the immediate task utility of the adopted semantic state exceeds the baseline for complete delivery without caching. A traditional hit simply indicates that the file was found; an effective hit requires that the task utility from cache reuse exceeds the utility of retrieving the complete data from the origin server. This step, through the definition of an effective hit, directly links cache hit quality to the actual benefit of the task, solving the problem that managing hit rate or origin server traffic as a management objective fails to reflect the value of the task.
[0136] Step Six: Calculate the task-aware cache value based on the historical task utility and effective hit count stored in the semantic state. Write the semantic state to the corresponding target cache level according to the admission criteria. When the target cache level is insufficient, evict the semantic state with the lowest task-aware cache value in the target cache level, obtaining the updated hierarchical semantic state cache set. (Refer to...) Figure 3 , specifically:
[0137] For semantic states that need to be written to the cache, the computation task is aware of the cache value:
[0138]
[0139] in, Let m be the average age of the erroneous semantic states in the current time slot t. The number of valid hits for semantic state m. For the historical task utility of semantic state m, The storage occupancy of semantic state m. The weighting coefficient for the freshness of state term. The weighting coefficients for effectively hitting value items, The weighting coefficients for the historical task utility term. The weighting coefficient for the average state risk penalty term. The weighting coefficient for the relative storage usage penalty term. Let be the time constant for the decay of state freshness. semantic state Current cache layer The maximum cache capacity.
[0140] Admission requirements:
[0141]
[0142] semantic state In the current time slot The cached admission indicator; a value of 1 indicates that admission is allowed, and a value of 0 indicates that admission is not allowed. For target cache level In the current time slot The admission threshold.
[0143] Both sub-conditions must be met simultaneously: the immediate task utility is greater than the no-cached baseline, and the task-aware cache value is greater than or equal to the target cache level admission threshold.
[0144] After admission, check the remaining capacity of the target cache level. If the capacity is insufficient, evict the state with the lowest value:
[0145]
[0146] Repeat this process until the capacity meets the write requirements, resulting in an updated hierarchical semantic state cache set. This step uses task-aware cache value as the sole criterion for admission and eviction, prioritizing cache resources for states with high effective hit rate, high historical utility, low risk, and low storage consumption, thus solving the problem of the disconnect between cache lifecycle management and task value.
[0147] Step 7: Each node obtains the pre-synchronization aggregated state risk value based on the predicted belief state, calculates the synchronization net benefit, and performs an active synchronization operation based on the synchronization net benefit to synchronize the semantic state to the target node.
[0148] The budget-constrained synchronization module performs synchronization decisions for each node n. It obtains the pre-synchronization aggregate state risk value from the predicted belief state. Expected residual load value before synchronization Expected task utility value before synchronization ,from Retrieve all semantic states maintained by the current node. Assuming a synchronization operation is performed, deduce the state after synchronization:
[0149] Aggregate state risk value after synchronization ), Expected residual load value after synchronization Expected task utility value after synchronization .
[0150] Calculate concurrent net income:
[0151]
[0152] in, For node n in the current time slot Synchronization cost. Triggering conditions:
[0153]
[0154] Node n in the current time slot The active synchronization indicator value is set to 1 to indicate that active synchronization is performed, and 0 to indicate that active synchronization is not performed. This represents the risk threshold for the aggregation state. This is the threshold for synchronous net income;
[0155] Verify synchronization budget constraints:
[0156]
[0157] If the total synchronization cost of eligible nodes exceeds the budget limit, nodes will be selected for synchronization in descending order of net synchronization benefit. Nodes not selected will not be synchronized in this time slot.
[0158] For the selected nodes to perform active synchronization, send the semantic state to the target node. Correction is performed after synchronization is complete.
[0159]
[0160] in, , These are the semantic state vectors of semantic state m before and after active synchronization, respectively. To proactively synchronize and update operators; This is the correction amount of the semantic vector after active synchronization of semantic state m; This is the correction amount for the semantic version after the semantic state m is actively synchronized; This is the correction amount for the uncertainty of the semantic state m after active synchronization;
[0161] Update semantic vector, semantic version, state uncertainty, most recent synchronization time, and source metadata.
[0162] Step 8: Write back the delivery results, synchronized semantic state, synchronization cost, updated hierarchical semantic state cache set, and dynamic heterogeneous network graph of the current time slot to the historical observation window data, and form the control action set for the next time slot based on the control actions executed in this time slot.
[0163] The closed-loop evaluation and state write-back module writes the synchronized semantic state to the corresponding nodes and cache levels. The virtual queue is updated.
[0164] The cache capacity of node n in the virtual queue can be updated as follows:
[0165]
[0166] in, For node n in the current time slot The load of writing to the cache, For node n in the current time slot Load eviction from cache This refers to the net write volume or capacity service volume allowed for node n in each time slot.
[0167] The semantic risk virtual queue for node n is:
[0168]
[0169] in, The semantic risk virtual queue backlog value for node n in the current time slot t;
[0170] The aggregation state risk value generated by node n in the current time slot t;
[0171] The residual service queue for link e is:
[0172]
[0173] in, This represents the residual service queue backlog value of link e in the current time slot t. Let e be the service rate of link e in the current time slot t. This represents the new arrival traffic on link e in the current time slot t.
[0174] Add the current time slot data using a sliding window mechanism. ,include , The system includes delivery results, synchronous action records, virtual queue status, current task request records, and comprehensive task utility statistics. Control actions executed within this time slot form... It includes shared state selection, admission and delisting actions, delivery path selection, and synchronization actions.
[0175] For long-term online operation, Lyapunov functions can be further constructed:
[0176]
[0177] in, The Lyapunov function value for the current time slot t. The cache capacity and virtual queue backlog value for the current time slot t;
[0178] In each time slot, the following drift-penalized approximation problem can be solved within the instantaneously feasible set:
[0179]
[0180] This is the set of control actions for the current time slot t, including shared state selection, cache admission and eviction, residual delivery, path and synchronization actions; The adjustment weighting coefficients between long-term task utility and link queue stability, cache queue stability, and semantic risk queue stability; In order to control the action The current task request is below. Long-term task utility in the current time slot t; In order to control the action The new arrival traffic on link e in the current time slot t; In order to control the action Below, the load of node n writing to the cache in the current time slot t; In order to control the action Below, the aggregate state risk value generated by node n in the current time slot t; This is the set of current task requests arriving in the current time slot t; Let be the set of instantaneous actionable actions for the current time slot t;
[0181] As an optional enhancement step, it does not affect the basic closed-loop process.
[0182] Effect verification
[0183] To verify the effectiveness of the technical solution of this application, multiple sets of experiments were conducted in the STK simulation environment. Figures 4 to 6 Key simulation results are presented.
[0184] Figure 4 The figures show the simulation results of semantic cache hit rate, semantic reuse gain, task utility, and bandwidth utilization under different content popularity skew levels. SAGE-HC represents the method of this application. The simulation uses a Zipf distribution to describe the request probability of different semantic objects, and gradually transitions from a relatively uniform request distribution to a request distribution dominated by a few popular objects by changing the skew parameter. Figure 4 As shown, with changes in Zipf skewness, the system in this application can maintain a high effective cache hit rate and semantic reuse by utilizing near-end caching, residual semantic caching, and global semantic caching, while maintaining a stable balance between task utility and bandwidth utilization. This result demonstrates that the dynamic sharing threshold and task-aware cache value can adapt to changes in the concentration of content demand.
[0185] In content traffic generation, the Zipf distribution is used to describe the request probability of different semantic objects:
[0186]
[0187] in, For semantic objects The probability of a request. The sequence number of the semantic objects, sorted from highest to lowest popularity. The total number of semantic objects. This is the skew parameter; a larger one... This indicates that requests are concentrated on a small number of popular objects, and is used to verify the adaptability of tiered caching to different content distributions.
[0188] Figure 5The simulation results for task utility, average latency, energy consumption per request, and average CPU utilization under different network scales are shown in the graphs. SAGE-HC represents the method of this application. Energy consumption per request statistics include transmission, processing, and synchronization energy consumption. Figure 5 As shown, as the network scale gradually expands, the system can still maintain stable task utility and controlled average latency, energy consumption per request, and average CPU utilization.
[0189]
[0190] in, The transmission energy consumption coefficient per unit data volume transmitted via link e. The energy consumption coefficient for performing a unit computational operation on node n. For the current task request The number of computational operations required during the node processing phase. For the current task request Synchronous energy consumption For the current task request The set of processing nodes that have been processed. For the current task request The set of links on the transmission path used. Figure 6 This diagram illustrates the network CPU utilization trends and node CPU utilization distribution in scenarios with 230 and 1840 nodes, with SAGE-HC representing the method used in this application. CPU utilization is normalized to the computational load of multimodal coding, hierarchical retrieval, effectiveness evaluation, and control decisions, based on the available computing power of each node. For example... Figure 6 As shown, in the 230-node and 1840-node scenarios, the network CPU utilization trend and node-level distribution remain within a manageable range.
[0191] Node CPU utilization can be expressed as:
[0192]
[0193] in, The computational cost of multimodal coding for node n in the current time slot t; The computational cost for hierarchical retrieval of node n in the current time slot t; The computational cost for node n in the current time slot t for effectiveness evaluation; The computational cost of node n for control decision-making in the current time slot t; The available computing power of node n in the current time slot t;
[0194] In a STK simulation implementation configuration, 18 orbital planes with 36 LEO satellites per plane, totaling 648 LEO satellites, are used, combined with airborne, maritime, ground, and mobile nodes to form a dynamic space-ground-air scenario with up to 1840 nodes. The simulation time step is 10 seconds, generating 361 topology snapshots. The above configuration is used for verification and does not limit the actual constellation size. The paper's experiments, under the same topology snapshots, request sequences, link implementations, and random seeds, compare the overall system of this application with queue-aware, capacity-aware, latency-aware, content-based LRU, utility-aware, and centralized ideal schemes. The results show that this application achieves communication benefits through task-conditional sharing, hierarchical caching, and budget synchronization without sacrificing runaway latency, energy consumption, or computational overhead.
[0195] While this application 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 this application. 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 this application 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 task-driven semantic sharing method for integrated space-ground-air-ground networks, characterized in that: include: Step 1: Construct a dynamic heterogeneous network graph and a hierarchical semantic state cache set for the current time slot; Step 2: Based on the dynamic heterogeneous network diagram of the current time slot, historical observation window data, and the control action set of the previous time slot, obtain the predicted belief state of multiple future time slots, and generate short-term predictive control quantities based on the predicted belief state. Step 3: Obtain the current task request for the multimodal task; Step 4: Retrieve semantic states in the hierarchical semantic state cache set in the order of near-end cache, residual semantic cache, and global semantic cache. Based on the short-term prediction control quantity and validity gating conditions, form a hierarchical candidate state set for the current task request. Step 5: For each semantic state in the hierarchical candidate state set, calculate the immediate task utility, select a delivery method to deliver the task response to the user based on the immediate task utility, determine the delivery result, and update the historical task utility and effective hit count stored in the semantic state based on the immediate task utility after delivery. Step 6: Calculate the task-aware cache value based on the historical task utility and the number of valid hits stored in the semantic state. Write the semantic state into the corresponding target cache level according to the admission criteria. When the capacity of the target cache level is insufficient, evict the semantic state with the lowest task-aware cache value in the target cache level to obtain the updated hierarchical semantic state cache set. Step 7: Each node obtains the pre-synchronization aggregated state risk value based on the predicted belief state, calculates the synchronization net benefit, and performs an active synchronization operation based on the synchronization net benefit to synchronize the semantic state to the target node.
2. The task-driven semantic sharing method for integrated space-ground-air-ground networks according to claim 1, characterized in that, The specific methods for calculating the concurrent net income in step seven include: Based on the predicted belief state, obtain the aggregate state risk value before synchronization, the expected residual load value before synchronization, and the expected task utility value before synchronization; Calculate the aggregate state risk value, the expected residual load value, and the expected task utility value after synchronization; The difference between the aggregation state risk value before and after synchronization, the difference between the expected residual load value before and after synchronization, and the difference between the expected task utility value after and before synchronization are assigned corresponding weight coefficients, summed, and then the synchronization cost is subtracted to obtain the net synchronization benefit.
3. The task-driven semantic sharing method for integrated space-ground-air-ground networks according to claim 2, characterized in that, The method for performing active synchronization based on the synchronized net income in step seven includes: When the risk value of the aggregated state before synchronization is greater than or equal to the risk threshold, and the net benefit of synchronization is greater than or equal to the net benefit threshold, an active synchronization operation is performed. The sum of the synchronization costs of all active synchronization operations is less than or equal to the upper limit of the synchronization budget for the current time slot.
4. The task-driven semantic sharing method for integrated space-ground-air-ground networks according to claim 1, characterized in that, In step six, the method for calculating the value of the task-aware cache is as follows: Let m be the average age of the erroneous semantic states in the current time slot t. The number of valid hits for semantic state m. For the historical task utility of semantic state m, The storage occupancy of semantic state m. The weighting coefficient for the freshness of state term. The weighting coefficients for effectively hitting value items, The weighting coefficients for the historical task utility term. The weighting coefficient for the average state risk penalty term. The weighting coefficient for the relative storage usage penalty term. Let be the time constant for the decay of state freshness. semantic state Current cache layer The maximum cache capacity.
5. The task-driven semantic sharing method for integrated space-ground-air-ground networks according to claim 1, characterized in that, Step six, the method for writing the semantic state to the corresponding target cache level according to the admission criteria, includes: When the immediate task utility of the semantic state is greater than the task utility of the no-cache full delivery baseline, and the task-aware cache value is greater than or equal to the admission threshold of the target cache level, the semantic state is allowed to be written to the target cache level. When the remaining capacity of the target cache level is insufficient to accommodate the semantic state to be written, the eviction operation is repeated in the target cache level until the capacity meets the write requirement: The semantic state with the lowest task-aware cache value is selected for eviction, and the evictioned semantic state is removed from the hierarchical semantic state cache set.
6. The task-driven semantic sharing method for integrated space-ground-air-ground networks according to claim 1, characterized in that, In step two, the method for generating short-time predictive control variables includes: The short-term predictive control variable includes a task-adaptive shared threshold. Extract the queue pressure and link capacity of future time slots from the predicted belief state, and calculate the predicted congestion index. ; According to the predicted congestion index Current task weight, average error semantic state age, cache age and cache health metrics Calculate the adaptive shared threshold for the task: in, For the current task request at the cache level In the current time slot Task adaptive sharing threshold, For cache level The basic sharing threshold, Weight of the current task The adjustment function, For cache level In the current time slot Average error semantic state age cache age The adjustment function, To cache health metrics The adjustment function, To predict congestion indicators The adjustment function, To adapt the lower limit of the shared threshold, An adaptive upper limit for the shared threshold; In step four, a hierarchical candidate state set for the current task request is formed based on the task adaptive sharing threshold and validity gating conditions.
7. The task-driven semantic sharing method for integrated space-ground-air-ground networks according to claim 6, characterized in that, The validity gating conditions include: The semantic similarity between the semantic state and the current task request is greater than or equal to the task adaptive sharing threshold in the short-term predictive control variable. The age of the erroneous semantic state is less than or equal to the maximum age of the erroneous semantic state allowed by the current task request; The semantic version of the semantic state is consistent with the semantic version of the current task request.
8. The task-driven semantic sharing method for integrated space-ground-air-ground networks according to claim 6, characterized in that, It also includes step eight: The delivery results, synchronized semantic state, synchronization cost, updated hierarchical semantic state cache set, and dynamic heterogeneous network graph of the current time slot are written back to the historical observation window data, and the control action set of the current time slot is formed according to the control action executed in this time slot.
9. The task-driven semantic sharing method for integrated space-ground-air-ground networks according to claim 1, characterized in that, In step five, the method for updating the number of valid hits includes: The effective hit count is incremented by one only if the immediate task utility of the adopted semantic state is greater than the task utility of the uncached full delivery baseline.
10. The task-driven semantic sharing method for integrated space-ground-air-ground networks according to claim 1, characterized in that, The real-time task utility in step five is calculated based on task weight, semantic quality, semantic sharing gain, delivery latency, energy consumption, actual data transmission volume, and synchronization cost.
