Task-driven efficient semantic state provisioning method

Abstract

The task-driven effective semantic state supply method solves the problem that the gating threshold is fixed and is cut off from the transmission amount control in the prior art, and the causal linkage of adaptive multiplexing and continuous supply driven by environment perception cannot be realized, and belongs to the technical field of network information processing.The present application comprises: obtaining a multi-modal task request, generating a task semantic vector and a task image;obtaining current network state information and cache state information, generating a short-time prediction control quantity;retrieving a candidate state from a hierarchical semantic state cache according to the task semantic vector and the short-time prediction control quantity, and performing qualification gating on the candidate state, wherein the determination threshold is dynamically adjusted by the short-time prediction control quantity;for the candidate state that passes the gating, determining a target corrective residual ratio according to the failure risk thereof relative to the task request;generating a corrective residual according to the complete data quantity and the target residual ratio, and the receiving end restores the target semantic state according to the candidate state and the corrective residual and writes back to the cache.

Description

Technical Field

[0001] This invention relates to a task-driven method for providing effective semantic state, belonging to the field of network information processing technology. Background Technology

[0002] Integrated space-ground-air networks, through the collaboration of low-Earth orbit satellites, near-space platforms, drones, ground gateways, edge nodes, and mobile terminals, provide wide-area information services for remote sensing monitoring, disaster response, marine observation, traffic monitoring, and low-altitude patrol inspection. The topology and link status of such networks change rapidly over time, and node storage, computing, and communication resources are limited. If each task transmits complete images, text, or sensor data from the source, it will generate a large amount of redundant communication and increase end-to-end latency, network congestion, and energy consumption.

[0003] Traditional content caching determines cache hits based on file identifiers, content names, or access frequency, and delivers a file copy upon a hit. Task-oriented semantic services do not require recovering all original data bit by bit, but rather obtaining a target state sufficient to support task judgment, decision-making, or control. Data from different times, nodes, and file identifiers may describe similar states of the same region or event, thus possessing cross-object identifier reuse value; objects with the same name may also lose their current task value due to disasters, sudden traffic changes, semantic model upgrades, or changes in observation uncertainty.

[0004] Information age alone cannot accurately characterize the aforementioned effectiveness. Earlier states in stable regions may still support the current task, while recent states in sudden events may quickly become invalid. Whether semantic states can be reused also depends on semantic drift between cached states and the current task, semantic version consistency, uncertainty on the requesting and caching sides, task priority, and task freshness requirements.

[0005] Existing semantic communication typically involves semantically encoding the current source data and optimizing its transmission. However, there is limited research on whether existing historical semantic states in the network can serve as conditional knowledge for the receiver, how much corrective information still needs to be supplemented, and how to safely degrade to complete transmission when reuse fails. Aggressive reuse based solely on similarity can easily provide outdated or erroneous states to high-priority tasks; while transmitting complete states in all cases fails to utilize distributed historical states to reduce communication overhead.

[0006] Furthermore, a fixed similarity threshold is difficult to adapt to dynamic networks and dynamic tasks. When network congestion is high, recent semantic demands are concentrated, and cache health is good, the scope of low-risk reuse can be appropriately expanded; when high-priority tasks increase, cache state risk rises, or semantic versions change, the gating threshold should be increased and more corrective residuals should be retained. Summary of the Invention

[0007] To address the problems of existing technologies where the gating threshold is fixed and disconnected from transmission volume control, making it impossible to achieve adaptive reuse driven by environment perception and causal linkage of continuous supply, this invention provides a task-driven effective semantic state supply method.

[0008] The present invention provides a task-driven method for efficient semantic state supply, comprising:

[0009] Obtain multimodal task requests, perform semantic extraction on the task requests, and generate task semantic vectors and task profiles;

[0010] Obtain current network status information and cache status information, and generate short-term predictive control variables based on the current network status information, the cache status information, and historical observations;

[0011] Based on the task semantic vector and the short-term prediction control quantity, candidate states are retrieved from the hierarchical semantic state cache, and eligibility gating is performed on the candidate states; wherein, the eligibility gating threshold is dynamically adjusted by the short-term prediction control quantity.

[0012] For the candidate states that have passed the qualification gating, a target corrective residual ratio is determined based on the failure risk of the candidate state relative to the task request; wherein, the target corrective residual ratio decreases as the prediction congestion index in the short-term predictive control increases, and increases as the task weight in the task profile increases.

[0013] A corrective residual is generated based on the total amount of data in the task request and the target corrective residual ratio, and the corrective residual is sent to the receiving end.

[0014] At the receiving end, the target semantic state is recovered based on the candidate state and the corrective residual, and the target semantic state is written back to the hierarchical semantic state cache.

[0015] Preferably, the step of acquiring the multimodal task request, performing semantic extraction on the task request, and generating a task semantic vector and a task profile includes:

[0016] The task request's multimodal data is acquired, including at least one of image modal data, text modal data, sensor modal data, and event modal data.

[0017] Each modal data is encoded separately to obtain the feature vector of each modality;

[0018] Normalize and dimensionality align the feature vectors of each modality to obtain the aligned features of each modality.

[0019] Based on the task relevance weight and modality reliability weight of each modality, the alignment features of each modality are weighted and fused to generate the task semantic vector;

[0020] Based on the task tags in the task request, the task tags are mapped to the task profile through a task mapping function; the task profile includes task type, task weight, maximum allowed completion time, freshness requirement, and semantic version.

[0021] Preferably, current network state information and cache state information are obtained, and short-term predictive control variables are generated based on the current network state information, the cache state information, and historical observations, including:

[0022] The system acquires dynamic network topology, link capacity, queue pressure, packet loss risk, recent semantic requirements, cache value, cache risk, and task pressure to construct a predictive belief state. Current network state information includes dynamic network topology, link capacity, queue pressure, and packet loss risk. The cache state information includes cache value and cache risk.

[0023] Based on historical observations and actions in the previous time slot, a short-term predicted state is generated using a prediction function.

[0024] Based on the short-term predicted state, the predicted congestion index and cache health index are calculated as the short-term predicted control variables.

[0025] Preferably, based on the task semantic vector and the short-term prediction control variable, candidate states are retrieved from the hierarchical semantic state cache, and eligibility gating is performed on the candidate states, including:

[0026] Retrieve cache states from the near-end semantic state cache, the regional semantic state cache, and the global semantic state cache respectively, and calculate the similarity between the semantic vector of the cache state and the semantic vector of the task.

[0027] Based on the predicted congestion indicators and cache health indicators, and combined with the task weights in the task profile, the judgment thresholds for each cache level are determined.

[0028] The cache states whose similarity is greater than or equal to the judgment threshold of the corresponding cache level and have the same semantic version and comparable objects as the task request are determined as candidate states that pass the qualification gate.

[0029] As a preferred option, the determination threshold is:

[0030] ;

[0031] in, Based on the basic gating threshold, As a task weight adjustment factor, To cache the age adjustment factor, This is the cache health adjustment factor obtained from the cache health metrics. This is the network congestion adjustment factor obtained from the predicted congestion index. To determine the lower limit of the threshold, To determine the upper limit of the threshold.

[0032] Preferably, for candidate states that have passed the qualification gating, the target corrective residual ratio is determined based on the failure risk of the candidate state relative to the task request, including:

[0033] Calculate the cache state age of the candidate states;

[0034] The cache state age is the difference between the current time and the update time of the candidate state;

[0035] The age of the request conditional error semantic state is calculated based on the cache state age, the semantic drift between the task semantic vector and the candidate state semantic vector, whether the semantic version of the candidate state is consistent with the semantic version of the task request, the request-side uncertainty of the task request, and the cache-side uncertainty of the candidate state, combined with the task weight in the task profile.

[0036] Based on the margin between the similarity and the judgment threshold, the reusable ratio is calculated, and the basic residual ratio is determined based on the reusable ratio.

[0037] Based on whether the age of the requested conditional error semantic state, the task weight, and the semantic version are consistent, the basic residual ratio is corrected to obtain a bounded target corrective residual ratio.

[0038] As a preferred option, the age of the requested conditional error semantic state is:

[0039] ;

[0040] in, Weights for the age-drift interaction term. For pure semantic drift weights, For version mismatch weights, As the uncertainty weight, This is a semantic version mismatch indicator function. A value of 1 indicates that the semantic version requested by the task is inconsistent with the semantic version of the candidate state; a value of 0 indicates that they are consistent. For the uncertainty of the requesting end, For cache-side uncertainty;

[0041] For cached state age, The semantic shift between the task semantic vector and the candidate state semantic vector. The task weights in the task profile are... This is the task weight index.

[0042] Preferably, generating corrective residuals based on the total amount of data in the task request and the target corrective residual ratio includes:

[0043] The actual amount of data transmitted is calculated as the product of the total amount of data and the proportion of the target corrective residual.

[0044] When the target corrective residual ratio is less than 1, a corrective residual with a data volume equal to the actual transmitted data volume is generated based on the difference between the candidate state and the target state corresponding to the task request.

[0045] When there are no candidate states that pass the qualification gate, set the target corrective residual ratio to 1 and perform complete data transmission.

[0046] Preferably, at the receiving end, recovering the target semantic state based on the candidate state and the corrective residual includes:

[0047] When the corrective residual is received, the candidate state is used as prior knowledge. Based on the task type and semantic version in the task profile, the candidate state and the corrective residual are fused through the conditional decoder to recover the target semantic state.

[0048] When complete data is received, the complete data is directly decoded to recover the target semantic state.

[0049] Preferably, the cache status includes the following fields: target object identifier, semantic vector, semantic version, update time, uncertainty, historical utility, hit count, cache level, cache storage usage, and source context.

[0050] The beneficial effects of this invention are as follows: This application dynamically predicts the network-semantic state, adaptively adjusts the gating threshold for cache retrieval, and continuously maps the changes in the gating threshold to the target residual ratio. This creates a causal closed loop between the reuse decision of candidate states and the control of the amount of transmitted data, solving the technical problems of fixed thresholds being unable to adapt to dynamic environments and the separation between reuse judgment and supply. This invention elevates the cache object from a file copy to a task-callable semantic state that simultaneously possesses content, time, version, uncertainty, source, level, and historical utility, enabling the discovery of task reuse relationships between different file identifiers and different observation times. This invention forms a unified task semantic vector through multimodal coding, normalization, dimension alignment, task relevance weighting, and modal reliability weighting, allowing observation content, task intent, time limit, freshness, and priority to jointly participate in state supply. This invention utilizes the network-semantic world model to output predicted congestion, semantic demand entropy, task pressure, and cache health indicators, enabling the candidate gating threshold to adjust with changes in the network and task environment. This invention unifies age-drift interaction, instantaneous semantic drift, version mismatch, joint uncertainty, and task weight into request conditional risks by using erroneous semantic state age, avoiding reuse based solely on timestamps or cosine similarity. This invention directly maps state validity risks to the proportion of bounded target residuals and the actual amount of transmitted data, and ensures supply security through minimum residual and complete transmission rollback mechanisms. The receiving end of this invention treats candidate states as conditional knowledge, transmitting only corrective residuals to restore the target semantic state; the restored state re-enters the hierarchical semantic cache, forming a closed loop. This invention's method employs online, causal processing and is applicable to integrated space-ground networks, digital twins, the Industrial Internet, vehicle-to-everything (V2X) networks, and edge sensing systems. Attached Figure Description

[0051] Figure 1 This is a flowchart of the method described in this application;

[0052] Figure 2 This is a schematic diagram illustrating the construction of multimodal task semantic vectors and the generation of task profiles in this application;

[0053] Figure 3 This is a closed-loop diagram showing the relationship between the AoISS risk of this application and the supply and status update of corrective residuals;

[0054] Figure 4 The following are simulation results of the reliability, latency, effective cache hit, semantic reuse, task utility and bandwidth utilization of this application under different request numbers. (a) is the packet delivery rate, (b) is the latency, (c) is the cache hit rate, (d) is the reuse gain, (f) is the bandwidth utilization and (e) is the semantic utility. Detailed Implementation

[0055] 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.

[0056] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0057] 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.

[0058] This task-driven effective semantic state supply method can be applied to integrated space-ground networks, satellite internet, low-altitude networks, edge sensing networks, digital twin networks, and task-oriented distributed information service systems. The method addresses technical problems such as cache hits failing to reflect task effectiveness, simple time freshness failing to characterize semantic state failure risk, fixed gating thresholds failing to adapt to dynamic task environments, and redundant communication caused by complete data transmission. It represents reusable information as a task-callable semantic state containing semantic content, observation time, semantic version, uncertainty, source, cache level, and historical task utility. The current request is represented as a task semantic state simultaneously containing multimodal content semantics and a task profile. A network-semantic world model is used to generate short-term predictive control variables and adjust the hierarchical candidate retrieval threshold accordingly. The age of erroneous semantic states quantifies the failure risk of candidate states relative to the current task, mapping the failure risk to a bounded corrective residual ratio. At the receiving end, the candidate state and corrective residual are fused to obtain the target semantic state, and the recovered target semantic state is written back to the hierarchical semantic state cache, forming a closed-loop supply. Figure 1 As shown, it includes the following steps.

[0059] Step 1: Constructing multimodal task semantic vectors and generating task profiles, refer to... Figure 2 ;

[0060] Get task request:

[0061]

[0062] in, and These are the source node and the destination node, respectively. For the target object, area, or event, For the complete data volume, For task semantic vectors, As task weight, To the maximum allowable completion delay, For the sake of freshness, For semantic version, Tag for task or scene For auxiliary metadata.

[0063] The task request includes multimodal data, which includes at least one of image modal data, text modal data, sensor modal data, and event modal data. The available modal set is as follows:

[0064] }

[0065] Among them, image modal data Used to represent observed content, including texture, spatial structure, and target distribution; text modal data Used to express textual intent, including categories, region descriptions, and task instructions; sensor modal data. Used to express environmental context, including observation time, platform information, environmental parameters, and confidence information; event modal data. Used to express service constraints, including task type, priority, deadline, and urgency.

[0066] Each modal data is encoded separately. The feature vector of any modality... Represented as:

[0067]

[0068] This corresponds to the modal encoder.

[0069] The feature vectors of each modality are normalized and their dimensions aligned. Specifically, each feature vector is normalized as follows:

[0070]

[0071] in It is a preset, extremely small positive number used to prevent the denominator from being zero.

[0072] Then, the normalized feature vectors Input modality alignment mappings are mapped to a unified semantic feature space to obtain modality alignment features with consistent dimensions:

[0073]

[0074] This is for modal alignment mapping.

[0075] Based on the task relevance weights and modality reliability weights of each modality, the alignment features of each modality are weighted and fused to generate a task semantic vector. Specifically, the task semantic vector is generated by fusing task relevance and modality reliability. :

[0076]

[0077] in, Modal indicators are available. For task relevance weights, This represents the modal reliability weight.

[0078] Norm is an L2 normalization operation that ensures the generated task semantic vector is normalized. It is a unit vector.

[0079] Based on the task tag in the task request Through the task mapping function Map task tags to task profiles:

[0080]

[0081] in, For task type, For task weights, To the maximum allowable completion delay, For the sake of freshness, This refers to semantic versioning. The same content can receive different priorities, freshness requirements, versioning, and residual control requirements under different tasks. Task profiling is not simply equivalent to content semantics. For the same forest image, routine ecological monitoring tasks and wildfire risk identification tasks can have different task weights, freshness requirements, and semantic versions.

[0082] This step unifies and quantizes multimodal heterogeneous inputs into standardized task semantic vectors and task profiles, providing a unified mathematical representation basis for subsequent candidate state retrieval and failure risk assessment.

[0083] Step 2: Construction of the network-semantic world model and generation of short-term predictive control variables;

[0084] Each cache state is represented as follows:

[0085]

[0086] Each field describes the target object identifier. semantic vectors Semantic version Update time Uncertainty Historical utility Number of hits Cache level Cache storage usage and source context .

[0087] Obtain current network status information and cache status information. The current network status information includes dynamic network topology, link capacity, queue pressure, and packet loss risk. The cache status information includes cache value and cache risk.

[0088] The predictive belief states maintained by the network-semantic world model are:

[0089]

[0090] in, For dynamic network graphs, For link capacity, For queue pressure, To mitigate the risk of packet loss, For recent semantic needs, For cache value, To mitigate caching risks, Due to task pressure.

[0091] Read historical observations Actions in the previous time slot Through prediction function Generate short-term predicted states. This includes historical observations. This is a sequence of historical belief states recorded over the past H time slots, including the action in the previous time slot. This records the supply decisions made at the previous moment (including the selected cache level, the proportion of residual data transferred, etc.). Historical observations... Actions in the previous time slot Input prediction function :

[0092]

[0093] Get the Future A sequence of predicted belief states for each time slot. Prediction function. For a pre-trained learnable time series prediction model, its parameters It has been obtained through offline training or online adaptive updates and is used to infer future belief states based on historical evolution patterns and the perturbation effect of the previous action.

[0094] Based on the short-term prediction status, predicted congestion metrics and cache health metrics are calculated as short-term prediction control variables. Specifically, from the short-term prediction status... Extract the predicted queue pressure and link capacity information, and calculate the predicted congestion index. :

[0095]

[0096] in, The predicted average queue pressure is extracted from the short-term predicted state. The predicted average link capacity is extracted from the short-term forecast state. To truncate the calculation results to Range. This metric combines average queue pressure and average link capacity; a higher value indicates more severe network congestion.

[0097] It also calculates cache health metrics. It retrieves a set of all current cache states. For each cached state in the set Obtain its normalized cache age Calculate the cache health metrics using the following formula:

[0098]

[0099] in, This is the age decay coefficient.

[0100] Cache health metrics The higher the value, the lower the overall risk of cached state validity, and the more inclined to reuse cached state.

[0101] In this step, the recent semantic demand entropy and task pressure index are also calculated as auxiliary situational observation values. The recent semantic demand entropy is calculated using the following formula:

[0102]

[0103] in, A collection of all task tag categories. For a moment Previous task categories in the window The frequency of occurrence of a task stress index reflects the concentration of recent task categories. The task stress index is calculated using the following formula:

[0104]

[0105] in, Indicates time window This is a collection of all task requests arriving within the specified timeframe. This task pressure metric reflects the average importance of recent requests and serves as a reference for upper-layer situational awareness and long-term strategy adjustments; it does not directly participate in the calculation of real-time gating thresholds.

[0106] This step compresses current network awareness data and cache status data into standardized predictive belief states, and generates predictions about future trends in network congestion and cache health through time-series forecasting.

[0107] Step 3: Hierarchical candidate status retrieval and eligibility gating;

[0108] Based on the task semantic vector output in step one and the short-time predictive control output from step two , Candidate states are retrieved from the hierarchical semantic state cache, and eligibility gating is performed on the candidate states. The eligibility gating threshold is dynamically adjusted by a short-time prediction control variable.

[0109] The cached state is retrieved from the near-end semantic state cache (L1), the regional semantic state cache (L2), and the global semantic state cache (L3), respectively. The near-end semantic state cache (L1) is deployed on the node closest to the requesting end (e.g., terminal, base station, edge server), emphasizing low latency and high match confidence; the regional semantic state cache (L2) is deployed on regional aggregation nodes (e.g., regional gateway, regional data center), supporting cross-gateway and cross-regional reuse; and the global semantic state cache (L3) is deployed on the core network or cloud, storing states with broader coverage and longer update cycles. Retrieval is performed sequentially according to the priority order of L1, L2, and L3. For each level... Retrieve all cache states stored in this level of cache. And calculate the semantic vector of the cache state. With task semantic vector Similarity between them:

[0110]

[0111] because and Both are unit-oriented, and their inner product equals the cosine similarity, with a value ranging from 1 to 2. A larger value indicates closer semantic similarity. Semantic drift is also calculated. This is for use in step four.

[0112] According to the predicted congestion index and cache health metrics Combined with task weights in the task profile This involves determining the threshold values ​​for each cache level. Specifically, for each cache level... Get the average normalized cache age for this level. Obtain the basic gating threshold for this level. The judgment threshold is calculated using the following formula:

[0113]

[0114] in, As a task weight adjustment factor, To cache the age adjustment factor, To cache health adjustment factors, This is a network congestion adjustment factor. To determine the lower limit of the threshold, To determine the upper limit of the threshold; high task weight, high cache risk or poor cache health make the gating more stringent; when the cache is healthy and congestion is high, low-risk reuse can be expanded to a limited extent.

[0115] Cache states with a similarity greater than or equal to the corresponding cache level's threshold, and possessing the same semantic version and comparable objects as the task request, are identified as candidate states for passing the eligibility gating. Specifically, for each retrieved cache state... Perform a triple judgment:

[0116]

[0117] The first item determines similarity. Is it greater than or equal to this cache level? Judgment threshold The second item determines the semantic version of the cache state. Semantic version of the task request Whether they are consistent; the third item determines the target object identifier of the cache state. The target object identifier of the task request Comparability, i.e., whether they describe the same region or the same event, means that objects identified by different files but spatially or semantically overlapping are considered comparable. When all three criteria are met, the comparison is successful. The cached state is identified as a candidate state that passes the eligibility gating and is added to the candidate state set. :

[0118]

[0119] The cached state is excluded when any decision fails.

[0120] From the candidate state set The candidate state with the highest similarity is selected as the optimal matching state. If the candidate state set is empty, it is determined that there is no candidate state that passes the eligibility gating.

[0121] This step uses the predictive control output of the network-semantic world model to dynamically adjust the admission thresholds at each level, so that the retrieval criteria for candidate states can adapt to changes in network congestion and cache health status. At the same time, it ensures through triple hard gating (similarity, version, and object comparability) that only semantically valid and safe candidate states can enter the subsequent supply process.

[0122] Step 4: Failure risk assessment and determination of the target corrective residual ratio, referring to... Figure 3 ;

[0123] For candidate states that have passed the qualification gating, the target corrective residual ratio is determined based on the failure risk of the candidate state relative to the task request. Specifically, the target corrective residual ratio is correlated with the predicted congestion index in the short-time predictive control quantity. Increases and decreases, depending on the task weight in the task profile. It rises and rises.

[0124] Calculate the cache state age of the candidate states. The cache state age is the current time step. Update time of candidate states Difference:

[0125]

[0126] This value reflects the length of time that has elapsed since the candidate state was last updated.

[0127] Based on cache age Task semantic vector semantic vectors of candidate states semantic drift between Semantic versions of candidate states Semantic version of the task request Whether they are consistent, and the uncertainty of the requesting end of the task request. And the cache uncertainty of candidate states Combined with task weights in the task profile Calculate the age of the request conditional error semantic state:

[0128]

[0129] in, Weights for the age-drift interaction term. For pure semantic drift weights, For version mismatch weights, As the uncertainty weight, This is a semantic version mismatch indicator function. A value of 1 indicates that the semantic version requested by the task is inconsistent with the semantic version of the candidate state; a value of 0 indicates that they are consistent. For the uncertainty of the requesting end, For cache-side uncertainty; The task weight index is used; age and drift constitute an interaction term; the version term reflects inconsistency in the model or labeling system; the uncertainty term reflects the joint confidence level; and the task weight enables high-priority tasks to adopt more conservative strategies.

[0130] The age-drift interaction weight allows older observations in stable regions to be reused, while in regions of sudden change, even short-term semantic drift can rapidly increase risk. For example, in a forest area, the cached state may be older, but the age-drift term is suppressed when the current image, text labels, and sensor environment are highly consistent with the cached state. However, when a fire, smoke, or temperature anomaly occurs in the same area, the age of the requested conditionalized erroneous semantic state increases rapidly due to instantaneous drift and joint uncertainty. Task weights exponentially amplify risk. Lower task weights can be used for routine periodic monitoring, while higher task weights are used for disaster relief and critical infrastructure monitoring, allowing more correction information to be retained for the same candidate state.

[0131] When no candidate state passes the qualification gating, calculate the baseline failure risk:

[0132]

[0133] in, Represents the empty set. This serves as the benchmark risk coefficient.

[0134] Based on similarity With the judgment threshold The system first calculates the margin for similarity exceeding a judgment threshold, and then maps this margin to... Normalized relative margin of the interval:

[0135]

[0136] in, Indicates the maximum reusable ratio; This indicates the margin for determining the threshold; the reusable ratio. This reflects the proportion of information in candidate states that can be directly reused based on semantic similarity.

[0137] Based on the reusability ratio Determine the basic residual ratio:

[0138]

[0139] in, The minimum residual ratio is used to ensure that at least a minimum proportion of correction information is transmitted as a safety noise floor, even if the candidate states are highly similar.

[0140] Age based on request conditional error semantic state Task weight And whether the semantic versions are consistent, and the ratio of the basic residuals. After correction, the bounded target corrective residual ratio is obtained:

[0141]

[0142] This is a correction factor for the failure risk. This is a task weight adjustment factor. This is a version mismatch correction factor. It is the age half-life constant;

[0143] The final target corrective residual ratio satisfy When there are no candidate states that pass the qualification gating, the system sets the target corrective residual ratio to 1.

[0144] The corrective residual ratio of this application targets decreases as the predicted congestion index in the short-term predictive control quantity increases, and increases as the task weight in the task profile increases; the predicted congestion index... Although it did not directly appear in the aforementioned target corrective residual ratio In the calculation formula, but By adjusting the judgment threshold in step three Change similarity margin This changes the reusable ratio. and basic residual ratio Ultimately indirectly affected Therefore, when network congestion increases, the decision threshold decreases, the reusable ratio increases, the basic residual ratio decreases, and the target corrective residual ratio decreases accordingly; conversely, when network congestion decreases, the decision threshold increases, the reusable ratio decreases, the basic residual ratio increases, and the target corrective residual ratio increases accordingly. Similarly, task weights... That is, through step three Raising the judgment threshold indirectly increases the target residual ratio, which is further achieved through the formula mentioned above. The item directly increases the target residual ratio, so the higher the task weight, the higher the target corrective residual ratio.

[0145] This step quantifies the failure risk of candidate states into a computable request conditional error semantic state age, and maps it to a continuously adjustable target corrective residual ratio, so that the residual ratio can adaptively change with factors such as network congestion, task priority, semantic drift, version mismatch and uncertainty.

[0146] Step 5: Corrective residual generation and transfer, refer to Figure 3 ;

[0147] Based on the complete data volume in the task request and the proportion of target corrective residuals The corrective residual is generated and sent to the receiving end.

[0148] The system calculates the actual amount of data transmitted. For the complete data volume Proportion of target corrective residuals The product of:

[0149]

[0150] System computation semantic multiplexing gain This value reflects the proportion of the total data volume saved by reusing candidate states.

[0151] When the target corrective residual ratio If the value is less than 1, a safe candidate state is determined to exist. Based on the candidate state... The difference between the target state corresponding to the task request and the generated data volume is the actual amount of data transmitted. Corrective residuals Specifically, the complete semantic encoding of the candidate state is obtained, the target semantic encoding corresponding to the current task request is obtained, the difference information between the two is calculated, and a value of [size missing] is extracted from the difference information. Bit corrective residual Extraction methods can include selective truncation based on semantic importance or compressed encoding based on feature differences. The corrective residuals will then be used. Send from the source to the receiver via the currently available network link.

[0152] When no candidate state passes the eligibility gate, the target corrective residual ratio is set to 1, and full data transmission is performed. Specifically, this involves encoding the complete semantics of the current task request into the total data volume. Send to the receiving end. At this time... , .

[0153] This step transforms the failure risk assessment results into the actual amount of data transmitted. Through a binary branching mechanism that transmits residuals when the value is less than 1 and complete data when the value is equal to 1, it fully utilizes the buffer state to reduce transmission overhead while ensuring automatic and safe rollback to complete transmission when there is no reusable state, thus avoiding task failure due to excessive reuse.

[0154] Step Six: Receiver target semantic state recovery and cache update;

[0155] At the receiving end, the system determines the candidate status. and corrective residuals Restore the target semantic state and write it back to the hierarchical semantic state cache.

[0156] When receiving corrective residuals At that time, the system is in a candidate state. As prior knowledge, based on the task type in the task profile and semantic version Through condition decoder By fusing the candidate state with the corrective residual, the target semantic state is recovered.

[0157]

[0158] The fusion method can be differential decoding, conditional generation, or feature increment fusion. After the candidate state is fused with the corrective residual, the recovered target semantic state contains the recovered semantic vector. And complete semantic information.

[0159] Upon receiving complete data, the system directly decodes the complete data to recover the target semantic state. This involves inputting the received complete semantic encoding into the full decoder, which outputs complete semantic information. The recovered target semantic state is then written back to the hierarchical semantic state cache. Specifically, the system constructs a new cache state based on the recovered target semantic state.

[0160]

[0161] in, The recovered semantic vector, For the updated uncertainty, For the updated historical utility, This is the updated hit count. For cache level, Storage usage after recovery This is the updated source context.

[0162] Based on the cache value of the new cache state, the new cache state is written to the corresponding level in the near-end semantic state cache (L1), regional semantic state cache (L2), or global semantic state cache (L3). Specifically, the cache value of the new state is evaluated, and the state is written to the matching cache level according to a preset cache value threshold and level capacity. If the object identifier and semantic version of the state are the same as an existing state in the cache, the system overwrites the old state with the new state; if it is a new object, the system allocates storage space in the target level according to the cache replacement strategy and then writes it.

[0163] When the current task request is a complete transmission (without candidate state reuse), the system will also construct the restored target semantic state as a new cached state and write it to the hierarchical cache, following the same writing rules as described above. The system can also compare the supply utility of candidate reuse and complete transmission. Only when the candidate state task is valid and the communication benefit is positive will it enter the final supply set. That is, the system determines:

[0164]

[0165] in, To improve the semantic quality of reusing candidate states. The multiplexing provisioning method is adopted only when the multiplexing quality is higher than the full transmission reference quality; otherwise, it degenerates into full transmission.

[0166] This step utilizes candidate states as prior knowledge at the receiving end, and the complete target semantic state can be recovered solely through the transmitted corrective residuals, significantly reducing communication overhead. Simultaneously, the recovered state is written back to the cache, making it a reusable candidate for subsequent tasks, thus forming a complete causal closed loop.

[0167] In the deployment of an integrated space-ground network, dynamic topology can be generated by STK (Site-Time Killing) to determine satellite orbital positions and visibility, and then combined with aviation, maritime, and ground nodes to form a heterogeneous network. This implementation experiment uses 648 low-Earth orbit satellites and other nodes to form a dynamic scenario with 1840 nodes. For example... Figure 4 As shown, Figure 4 The SAGE-HC in this application is the solution. As the number of requests increases from 50 to 500, this application maintains controlled delivery reliability and average latency, and demonstrates advantages in terms of effective cache hit rate, semantic reuse ratio, task semantic utility and bandwidth utilization.

[0168] Experimental results show that, compared to the semantic transmission baseline, the overall framework improves the effective cache hit rate, semantic multiplexing gain, and task utility by 87.96%, 172.51%, and 34.08%, respectively, while reducing the number of transmitted bits by 7.76%. These results are used to illustrate technical relationships and do not limit specific parameters.

[0169] 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 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 embodiments.

Claims

1. A task-driven method for providing effective semantic state, characterized in that, include: Obtain multimodal task requests, perform semantic extraction on the task requests, and generate task semantic vectors and task profiles; Obtain current network status information and cache status information, and generate short-term predictive control variables based on the current network status information, the cache status information, and historical observations; Based on the task semantic vector and the short-term prediction control quantity, candidate states are retrieved from the hierarchical semantic state cache, and eligibility gating is performed on the candidate states; wherein, the eligibility gating threshold is dynamically adjusted by the short-term prediction control quantity. For the candidate states that have passed the qualification gating, a target corrective residual ratio is determined based on the failure risk of the candidate state relative to the task request; wherein, the target corrective residual ratio decreases as the prediction congestion index in the short-term predictive control increases, and increases as the task weight in the task profile increases. A corrective residual is generated based on the total amount of data in the task request and the target corrective residual ratio, and the corrective residual is sent to the receiving end. At the receiving end, the target semantic state is recovered based on the candidate state and the corrective residual, and the target semantic state is written back to the hierarchical semantic state cache.

2. The task-driven effective semantic state supply method according to claim 1, characterized in that, The process of acquiring multimodal task requests, performing semantic extraction on the task requests, and generating task semantic vectors and task profiles includes: The task request's multimodal data is acquired, including at least one of image modal data, text modal data, sensor modal data, and event modal data. Each modal data is encoded separately to obtain the feature vector of each modality; Normalize and dimensionality align the feature vectors of each modality to obtain the aligned features of each modality. Based on the task relevance weight and modality reliability weight of each modality, the alignment features of each modality are weighted and fused to generate the task semantic vector; Based on the task tags in the task request, the task tags are mapped to the task profile through a task mapping function; the task profile includes task type, task weight, maximum allowed completion time, freshness requirement, and semantic version.

3. The task-driven effective semantic state supply method according to claim 1, characterized in that, The step of obtaining current network state information and cache state information, and generating short-term predictive control variables based on the current network state information, the cache state information, and historical observations, includes: The system acquires dynamic network topology, link capacity, queue pressure, packet loss risk, recent semantic requirements, cache value, cache risk, and task pressure to construct a predictive belief state. Current network state information includes dynamic network topology, link capacity, queue pressure, and packet loss risk. The cache state information includes cache value and cache risk. Based on historical observations and actions in the previous time slot, a short-term predicted state is generated using a prediction function. Based on the short-term predicted state, the predicted congestion index and cache health index are calculated as the short-term predicted control variables.

4. The task-driven effective semantic state supply method according to claim 3, characterized in that, The step of retrieving candidate states from the hierarchical semantic state cache based on the task semantic vector and the short-term prediction control variable, and performing eligibility gating on the candidate states, includes: Retrieve cache states from the near-end semantic state cache, the regional semantic state cache, and the global semantic state cache respectively, and calculate the similarity between the semantic vector of the cache state and the semantic vector of the task. Based on the predicted congestion indicators and cache health indicators, and combined with the task weights in the task profile, the judgment thresholds for each cache level are determined. The cache states whose similarity is greater than or equal to the judgment threshold of the corresponding cache level and have the same semantic version and comparable objects as the task request are determined as candidate states that pass the qualification gate.

5. The task-driven effective semantic state supply method according to claim 4, characterized in that, The determination threshold is: ； in, Based on the basic gating threshold, As a task weight adjustment factor, To cache the age adjustment factor, This is the cache health adjustment factor obtained from the cache health metrics. This is the network congestion adjustment factor obtained from the predicted congestion index. To determine the lower limit of the threshold, To determine the upper limit of the threshold.

6. The task-driven effective semantic state supply method according to claim 1, characterized in that, The step of determining the target corrective residual ratio based on the failure risk of the candidate state relative to the task request for the candidate state that has passed the qualification gating includes: Calculate the cache state age of the candidate states; The cache state age is the difference between the current time and the update time of the candidate state; The age of the request conditional error semantic state is calculated based on the cache state age, the semantic drift between the task semantic vector and the candidate state semantic vector, whether the semantic version of the candidate state is consistent with the semantic version of the task request, the request-side uncertainty of the task request, and the cache-side uncertainty of the candidate state, combined with the task weight in the task profile. Based on the margin between the similarity and the judgment threshold, the reusable ratio is calculated, and the basic residual ratio is determined based on the reusable ratio. Based on whether the age of the requested conditional error semantic state, the task weight, and the semantic version are consistent, the basic residual ratio is corrected to obtain a bounded target corrective residual ratio.

7. The task-driven effective semantic state supply method according to claim 6, characterized in that, The age of the requested conditional error semantic state is: ； in, Weights for the age-drift interaction term. For pure semantic drift weights, For version mismatch weights, As the uncertainty weight, This is a semantic version mismatch indicator function. A value of 1 indicates that the semantic version requested by the task is inconsistent with the semantic version of the candidate state; a value of 0 indicates that they are consistent. For the uncertainty of the requesting end, For cache-side uncertainty; For cached state age, The semantic shift between the task semantic vector and the candidate state semantic vector. The task weights in the task profile are... This is the task weight index.

8. The task-driven effective semantic state supply method according to claim 1, characterized in that, The step of generating corrective residuals based on the total amount of data in the task request and the target corrective residual ratio includes: The actual amount of data transmitted is calculated as the product of the total amount of data and the proportion of the target corrective residual. When the target corrective residual ratio is less than 1, a corrective residual with a data volume equal to the actual transmitted data volume is generated based on the difference between the candidate state and the target state corresponding to the task request. When there are no candidate states that pass the qualification gate, set the target corrective residual ratio to 1 and perform complete data transmission.

9. The task-driven effective semantic state supply method according to claim 1, characterized in that, The step of recovering the target semantic state at the receiving end based on the candidate state and the corrective residual includes: When the corrective residual is received, the candidate state is used as prior knowledge. Based on the task type and semantic version in the task profile, the candidate state and the corrective residual are fused through the conditional decoder to recover the target semantic state. When complete data is received, the complete data is directly decoded to recover the target semantic state.

10. The task-driven effective semantic state supply method according to claim 1, characterized in that, The cache status includes the following fields: target object identifier, semantic vector, semantic version, update time, uncertainty, historical utility, hit count, cache level, cache storage usage, and source context.

Images