Edge caching and prefetching method based on cross-terminal behavior prediction

Abstract

The present application relates to the technical fields of satellite communication, edge computing, artificial intelligence prediction cache, and discloses an edge cache and prefetch method based on cross-terminal behavior prediction, comprising: collecting historical behavior data of a first terminal, inputting an AI prediction model, and obtaining a predicted request content ID of a second terminal within a preset time period; determining the predicted request content and the time limit label according to the predicted request content ID, and constructing cache data; sending the cache data meeting the preset cache condition to a target edge cache node for caching; when the second terminal initiates a request, calculating the distance between the original position information and the current position information of the second terminal; judging whether the distance belongs to a preset migration range, determining a target edge cache node, and matching the actual request content from the cache data of the target edge cache node. The present application solves the problem that single-user prediction in the prior art cannot adapt to satellite link dynamic idle and cross-user scenarios.

Description

Technical Field

[0001] This invention relates to the fields of satellite communication, edge computing, and artificial intelligence prediction caching technology, and in particular to an edge caching and prefetching method based on cross-terminal behavior prediction. Background Technology

[0002] In satellite communication's two-hop transmission mode, signals must be relayed via satellite to achieve bidirectional data transmission between the ground station and the terminal. Limited by the inherent physical transmission distance of the communication link, the round-trip latency is typically 400ms to 800ms. This directly causes a series of problems such as control command interaction delays and content request response stuttering, severely disrupting the continuity of service transmission, significantly reducing user experience, and becoming a key bottleneck restricting the high-quality development of satellite communication. To mitigate the negative impact of this latency, pre-caching technology has become one of the core solutions. By predicting and caching the content requested by users in advance, it effectively reduces the amount and frequency of real-time data transmission, thereby reducing the impact of link latency on services and improving the response efficiency of the satellite communication system. Currently, mainstream predictive caching technologies mainly include single-user historical behavior prediction and global popularity prediction.

[0003] However, existing caching technologies have significant technical shortcomings and are difficult to fully adapt to the special needs of satellite communication scenarios. First, predictive caching has limitations, failing to fully consider the dynamic changes in satellite link idle time slots, the strong correlation of cross-terminal user behavior, and the time sensitivity of content transmission, resulting in a lack of targeted caching deployment. Second, the prediction accuracy is insufficient, and the caching strategy is poorly adapted to the dynamic characteristics of satellite links, failing to fundamentally and effectively alleviate the content request stuttering problem caused by the high latency of satellite communication, and thus failing to meet the application requirements of high-quality satellite communication services. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an edge caching and prefetching method based on cross-terminal behavior prediction. This method solves problems such as the lack of consideration for dynamic link time slots, cross-terminal correlation, and timeliness in existing technologies; lack of targeted caching deployment; low prediction accuracy; poor adaptability; and inability to alleviate stuttering or adapt to satellite scenarios. This invention can adapt to special satellite communication scenarios in dual-hop satellite communication mode, meeting the application requirements of high-quality satellite communication services. The technical solution provided by this invention includes: The edge caching and prefetching method based on cross-terminal behavior prediction includes the following steps: S1. Collect historical behavior data of the first terminal, input it into the trained AI prediction model, and obtain the prediction request content ID of the second terminal within a preset time period; the AI ​​prediction model is trained based on the correlation features between the first terminal and the second terminal; S2. Determine the predicted request content and timeliness tag based on the predicted request content ID, and construct cache data; S3. Send the cached data that meets the preset caching conditions to the target edge cache node for caching; S4. When the second terminal initiates a request, calculate the distance between the original location information and the current location information of the second terminal; S5. Determine whether the distance falls within a preset migration range, identify the target edge cache node, and match the actual request content from the cached data of the target edge cache node.

[0005] Preferably, the training method for the AI ​​prediction model includes: The heterogeneous user behavior graphs of the first terminal and the second terminal are continuously updated. The graph attention network is used to learn the association features between the first terminal and the second terminal as a training set to train the AI ​​prediction model built by the Transformer encoder.

[0006] Preferably, the historical behavior data includes: request content ID, timestamp, geographical location, and first terminal state vector.

[0007] Preferably, the target edge cache node includes satellite-borne cache and ground edge station.

[0008] Preferably, the preset caching conditions include: The predicted value of the predicted request content ID is greater than the first preset threshold, and the confidence level is greater than the second preset threshold. The downlink bandwidth of the link from the satellite to the second terminal is in an idle time slot.

[0009] Preferably, the method for obtaining the idle time slots includes: obtaining idle time slot information of all satellite links through real-time telemetry data of the satellite network.

[0010] Preferably, before the second terminal initiates a request, the process includes: sending the cached data to a first edge cache node determined based on the original location information via a cloud platform for caching.

[0011] Preferably, when the second terminal makes a request, it includes: Calculate the distance between the original location information and the current location information of the second terminal; if the distance is within a first preset migration range, determine that the second terminal has not migrated, and use the first edge cache node as the target edge cache node to obtain the cached data corresponding to the actual requested content.

[0012] Preferably, before the second terminal makes a request, the method further includes: distributing the constructed cached data to multiple candidate edge cache nodes within a preset range for caching.

[0013] Preferably, when the second terminal makes a request, it further includes: Calculate the distance between the original location information and the current location information of the second terminal; if the distance is within a second preset migration range, determine that the second terminal has migrated within the preset range, determine the target edge cache node from multiple candidate edge cache nodes based on the current location information of the second terminal, and obtain the cache data corresponding to the actual requested content; If the distance is outside the second preset migration range, it is determined that the second terminal has migrated outside the preset range. The cached data that meets the preset caching conditions is sent to the target edge cache node determined according to the current location information through the cloud platform for caching, and the cached data corresponding to the actual request content is obtained.

[0014] Beneficial effects This invention addresses the shortcomings of traditional predictive caching techniques that fail to consider dynamic time slots in satellite links, cross-terminal correlations, and content timeliness. It improves prediction accuracy and satellite scene adaptability through cross-terminal feature learning, enhances resource utilization by leveraging timeliness tags, and reduces latency and congestion through near-terminal edge deployment, adapting to satellite communication characteristics and multi-terminal collaborative scenarios. Experiments show that this invention increases cache hit rate from 82% to 91%, reduces effective latency to <150ms, and saves 42% of bandwidth. It also overcomes the limitations of existing technologies where single-user prediction cannot adapt to dynamic idle satellite links and cross-user scenarios. Attached Figure Description

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

[0016] Figure 1 This is a schematic diagram of the edge caching and prefetching method based on cross-terminal behavior prediction provided in a preferred embodiment of the present invention. Detailed Implementation

[0017] 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.

[0018] Example 1

[0019] like Figure 1 As shown, this invention provides an edge caching and prefetching method based on cross-terminal behavior prediction.

[0020] Existing predictive caching technologies mainly include single-user historical behavior prediction and global popularity prediction; however, these methods do not consider dynamic time slots in the link, cross-terminal correlation, and timeliness, and their caching deployment lacks specificity; they suffer from low prediction accuracy, poor adaptability, and inability to alleviate lag or adapt to satellite scenarios. To solve these problems, the specific steps of the technical solution of this invention are as follows: S1. Collect historical behavior data of the first terminal, input it into the trained AI prediction model, and obtain the prediction request content ID of the second terminal within a preset time period; the AI ​​prediction model is trained based on the correlation features between the first terminal and the second terminal.

[0021] In this invention, the first terminal and the second terminal can be any two terminals participating in business interaction, such as handheld devices, ships, airplanes, etc. During data interaction, they can switch roles as data sender and data receiver according to the actual business scenario. Role division is not fixed, avoiding functional rigidity due to terminal numbering. Historical behavior data includes: request content ID, timestamp, geographical location, and the first terminal's state vector. In some preferred embodiments, the invention first deploys an AI prediction model trained on the associated features of the first and second terminals on a cloud platform for subsequent prediction of terminal user behavior data. The invention uses the historical behavior data of the first terminal as input to the cloud platform's AI prediction model, and through model feature mapping and prediction calculation, obtains the predicted request content corresponding to the predicted request content ID of the second terminal within a preset time period. In other preferred embodiments, the invention also has a built-in database on the cloud platform storing request content IDs, corresponding request content, and request time windows, allowing subsequent retrieval of the corresponding request content and time window via the request content ID.

[0022] S2. Determine the predicted request content and timeliness tag based on the predicted request content ID, and construct cache data.

[0023] Timeliness tags are a core prerequisite for ensuring the accuracy of content services and improving resource utilization. They prevent content from being pushed out too early, wasting resources, or being pushed out too late, failing to meet immediate needs. They are also suitable for scenarios with limited satellite link resources, enabling precise resource scheduling. This embodiment uses a request time window to preset the timeliness tag. For example, if the request time window is "daily 8:10-8:30," the effective time can be set to 8:00 (5 minutes before the start of the time window to ensure content is prepared in advance), and the effective duration can be set to 30 minutes to cover the time window and ensure that the second terminal can obtain the content immediately upon request. Alternatively, a reasonable preset range can be determined by those skilled in the art based on actual conditions or on-site needs. In some preferred embodiments, the present invention constructs cached data based on the timeliness tag and the predicted request content ID corresponding to the predicted request content.

[0024] S3. The cached data that meets the preset caching conditions is sent to the target edge cache node for caching. The target edge cache node refers to a caching device deployed near the terminal access side (such as near a satellite communication base station or a concentrated terminal area). Its core function is to store satellite-related content data locally, shortening the terminal request transmission path, reducing satellite link transmission pressure, and improving response speed. Existing technologies mostly adopt a fixed caching mode, where edge cache nodes pre-store popular and common content without combining terminal-predicted request content and timeliness tags, caching only based on historical request frequency. This results in poor caching targeting, resource waste, and the failure to combine timeliness tags, leading to content expiration or missing peak request times. Inappropriate caching timing can cause link congestion or idleness, and a large deviation from the actual terminal requests.

[0025] To address the aforementioned technical issues, in some preferred embodiments, this invention uses the target content corresponding to the timeliness tag and behavior prediction data as cached data. It first determines whether current conditions, such as satellite link idle status, meet the cache transmission requirements. If the requirements are met, the cached data is sent to the target edge cache node for caching. The target edge cache node of this invention includes: satellite-borne cache and ground edge station.

[0026] S4. When the second terminal initiates a request, calculate the distance between the original location information and the current location information of the second terminal.

[0027] In some preferred embodiments, when the second terminal initiates a request, the present invention calculates the distance between the current location information of the second terminal and the original location information of the second terminal in order to determine whether the second terminal has undergone location migration.

[0028] S5. Determine whether the distance falls within a preset migration range, identify the target edge cache node, and match the actual request content from the cached data of the target edge cache node.

[0029] Existing edge computing-based caching solutions still have the following problems: Once the cache node is predetermined, when the terminal initiates a request, it does not consider whether the actual location of the terminal has shifted and still forces access to the original cache node, which can easily lead to excessive access distance, high transmission latency, and degraded service quality; if the strategy of "always switching to the nearest node" is adopted, the cache node will be frequently changed due to slight location jitter of the terminal, base station switching, etc., resulting in cache node jitter, frequent cache migration, and large computational overhead; there is a lack of a reasonable judgment mechanism for the degree of terminal location migration, making it impossible to achieve a balance between cache stability and access proximity, and it is difficult to take both cache stability and access proximity into account.

[0030] To address the aforementioned technical issues, in some preferred embodiments, this invention matches the distance with a preset migration range, dynamically determines the target edge cache node corresponding to the current request based on whether the distance falls within the preset migration range, and then matches the actual request content of the second terminal from the data already cached by the target edge cache node. This achieves the goal of adjusting cache nodes only when a valid location migration occurs at the terminal, ensuring access from the nearest location while avoiding frequent node switching, thus improving overall caching efficiency and system stability.

[0031] This embodiment effectively addresses the shortcomings of traditional predictive caching methods that fail to consider dynamic time slots in the link, cross-terminal correlation, and content timeliness. By learning cross-terminal associated features, it improves the accuracy of user behavior prediction and enhances the model's adaptability to satellite communication scenarios. Secondly, it achieves precise scheduling of cached content based on timeliness tags, avoiding resource waste from premature caching or failure to meet immediate needs due to late caching, significantly improving the resource utilization of satellite links and edge cache nodes. By pre-deploying target content to edge cache nodes close to the terminal, it shortens the data transmission path, avoids inherent transmission delays in satellite links, effectively alleviates terminal access lag issues, and reduces the probability of link congestion. Simultaneously, it dynamically selects target cache nodes based on location migration range, ensuring access from the nearest node while avoiding frequent switching of cache nodes due to slight terminal location fluctuations, balancing access efficiency and system stability. This embodiment is adaptable to the characteristics of limited satellite communication resources and dynamic link fluctuations, and is particularly suitable for satellite edge communication scenarios involving multiple types of terminals such as vehicle-mounted and portable satellite terminals, achieving low-latency and highly reliable content delivery in applications such as emergency communication and drone control.

[0032] Example 2

[0033] Based on Example 1, this example provides a superior training method for cloud platform AI prediction models.

[0034] The heterogeneous user behavior graphs of the first terminal and the second terminal are continuously updated. The graph attention network is used to learn the association features between the first terminal and the second terminal as a training set to train the AI ​​prediction model built by the Transformer encoder.

[0035] Heterogeneous graphs are graph structures containing various types of nodes and edges. Compared to homogeneous graphs, they can more comprehensively represent the diversity and complex relationships in complex systems. In heterogeneous graphs, nodes can represent different types of entities, such as a first terminal (vehicle-mounted satellite terminal), a second terminal (portable satellite terminal), and content nodes (navigation maps, emergency communication data), while edges represent relationships between different types of entities, such as the vehicle-mounted satellite terminal's request for a navigation map, the portable satellite terminal's download of emergency communication data, and the association between two types of terminals sharing idle time slots on satellite links. In some preferred embodiments, the first and second terminals described in this invention assume different functional roles, such as a data sender and a data requester, or an operation control terminal and an instruction execution terminal, respectively. Due to differences in terminal division of labor and business positioning, the second terminal needs to rely on the first terminal to generate predictive data that can adapt to business interaction requirements. In other preferred embodiments, when the execution tasks, device functions, and business roles of the first and second terminals are completely consistent, mutual prediction between terminals or self-prediction by terminals can be achieved, and the results obtained from different prediction methods can remain consistent. The heterogeneous graph can clearly depict the behavioral relationships between users across terminals and the complex interactions between various entities, providing accurate data support for subsequent feature mining.

[0036] Graph Attention Network (GAT) is a network model that can focus on node relationships and efficiently learn node relationship features. In some preferred embodiments, this invention continuously updates the heterogeneous graph of user behavior between the first terminal and the second terminal, synchronously collects terminal behavior, content interaction, and link state data, and dynamically maintains the nodes and edges of the heterogeneous graph; then, it constructs a graph attention network adapted to the scenario, trains the model based on the updated heterogeneous graph, mines and outputs the relationship features between the two terminals, and provides support for subsequent prediction work.

[0037] The cloud platform AI prediction model refers to an artificial intelligence model deployed on a cloud platform for predictive analysis based on input features. Existing prediction methods often employ traditional machine learning models (such as logistic regression and random forests) or simple neural networks, directly inputting single behavioral features of the terminal for model training, failing to fully utilize the complex relationships between terminals. Some models use Transformer encoders, but these are not adapted to the characteristics of satellite terminal correlation features, simply applying a general architecture. This ignores the correlation patterns between terminals, leading to low prediction accuracy. Secondly, the general architecture cannot adapt to the temporal and correlational characteristics of satellite terminal correlation features, resulting in low training efficiency. Furthermore, the model's generalization ability is weak, and its prediction stability is insufficient when faced with fluctuations in terminal behavior and changes in link status.

[0038] To address the aforementioned technical challenges, in some preferred embodiments, this invention builds a cloud platform AI prediction model based on a Transformer encoder. The Transformer encoder used in this invention employs a standard layered stacked structure, consisting of several identical basic coding layers stacked sequentially. Each basic coding layer includes a multi-head self-attention unit, a layer normalization unit, a feedforward network unit, and residual connection branches. Utilizing the Transformer's self-attention mechanism, it accurately captures temporal and cross-correlation features in terminal-related characteristics, adapting to the features of satellite terminal scenarios. These related features are used as a training set, input into the constructed Transformer encoder model, and reasonable training parameters (learning rate, number of iterations, etc.) are set. The model parameters are adjusted in real-time using a validation set to avoid overfitting. After training, the model's prediction accuracy is verified using a validation set. Once the accuracy is achieved, the trained cloud platform AI prediction model is obtained for subsequent prediction tasks.

[0039] This embodiment can adapt to the changing patterns of terminal data over time and the correlation and coupling relationships between multiple entities, improving model training efficiency, prediction accuracy, and generalization stability, and providing reliable training model support for subsequent cross-terminal behavior prediction and satellite edge cache prefetching.

[0040] Example 3

[0041] Based on Embodiment 1, this embodiment provides a better preset caching condition.

[0042] In some preferred embodiments, the preset caching conditions of the present invention include: the predicted value of the predicted request content ID is greater than a first preset threshold and the confidence level is greater than a second preset threshold; the downlink bandwidth of the link from the satellite to the second terminal is in an idle time slot. Since the storage space and bandwidth of the terminal are limited, blind full caching is not allowed. The present invention first judges the predicted value of the predicted request content ID corresponding to the cached data. If the predicted value is greater than the first preset threshold, the condition is met. The prediction confidence level output by the AI ​​prediction model is obtained through the cloud platform, and the confidence level is compared with the second preset threshold to determine whether the confidence level is greater than the second preset threshold. The prediction confidence level is used to characterize the reliability of the model's prediction of the prediction behavior data of the second terminal. The higher the confidence level, the greater the reference value of the prediction result. At the same time, the idle time slot information of all links from the satellite to the second terminal is collected through real-time telemetry data of the satellite network. The idle time slot status corresponding to the downlink bandwidth of the satellite link to the second terminal is extracted from it, so that the cloud can perceive the load of the satellite downlink in real time. Once the downlink has "remaining bandwidth" or "idle time slot", cached data is immediately inserted.

[0043] In satellite communication systems, the uplink and downlink are typically independent in terms of physical layer and resource allocation. Whether in FDD (Frequency Division Duplex) or TDD (Time Division Duplex) mode, the system manages uplink / downlink frequency or time resources separately. The "prefetching and caching" behavior of this invention mainly involves the downlink process of data flowing from the cloud / satellite to edge nodes. Even if the uplink (terminal sending back commands) is busy, prefetching can be performed as long as there are spare resources in the downlink. In this invention, caching data when "downlink bandwidth has idle time slots" is primarily to maximize resource utilization. Satellite link bandwidth is expensive, and waiting for the "entire link" (both uplink and downlink simultaneously) to be idle is extremely rare in actual business operations, leading to very low prefetching efficiency. Secondly, this invention avoids peak business periods. Prefetching is essentially a form of "background traffic," and this invention utilizes only fragmented downlink time (i.e., idle time slots) to transmit predicted data to avoid interfering with the user's ongoing real-time downlink tasks. When all three conditions are met simultaneously—the probability value of the predicted request content ID corresponding to the cached data is greater than the first preset threshold, the prediction confidence is greater than the second preset threshold, and the downlink bandwidth from the satellite link to the second terminal is in an idle time slot—it is determined that the current conditions meet the requirements for sending cached data. The cloud platform then sends the cached data (including the target content with the time expiration tag) to the edge cache node corresponding to the second terminal for caching. If any condition is not met, the cached data sending operation will not be performed temporarily, and the system will wait for the next condition judgment cycle to re-evaluate.

[0044] This embodiment filters out unreliable prediction results by limiting the prediction of the requested content ID to three conditions: the predicted value is greater than a first preset threshold, the prediction confidence is greater than a second preset threshold, and the downlink bandwidth from the satellite to the second terminal is in an idle time slot. This avoids wasting resources on invalid caching and avoids link conflicts between cache transmission and normal terminal communication. By utilizing idle link resources, it ensures the accuracy and timeliness of cache prefetching, further optimizes the utilization rate of satellite link and edge cache resources, and improves the stability and rationality of satellite edge cache prefetching.

[0045] Example 4

[0046] Based on Embodiment 1, this embodiment provides a better method for prefetching cached data.

[0047] To address the high latency issue of satellite link transmission and improve the request response speed of the second terminal, in some preferred embodiments, this invention sends the cached data to a first edge cache node determined based on the original location information via a cloud platform before the second terminal initiates a request. When the second terminal makes a request, the distance between the original location information and the current location information of the second terminal is calculated. If the distance is within a first preset migration range, it is determined that the second terminal has not migrated, and the first edge cache node is selected as the target edge cache node to obtain the cached data corresponding to the actual requested content. This embodiment eliminates the need for forwarding through the cloud platform and obtaining data from a remote cloud server; it can directly read the cached corresponding data from the first edge cache node. This invention can minimize the data transmission path, avoid the impact of satellite link transmission latency, significantly improve the content request response speed of the second terminal, and further reduce the transmission pressure on the satellite link, avoiding link congestion.

[0048] This embodiment achieves content deployment based on the original location through a pre-caching mechanism, reducing satellite link transmission latency; by setting a first preset migration range to filter terminal micro-movements, it avoids frequent switching of cache nodes due to location jitter, reduces cache migration and data synchronization overhead, and improves the stability of the caching system while ensuring access proximity, further improving terminal request response speed and service reliability.

[0049] Example 5

[0050] Based on Example 1, this example provides another superior method for prefetching cached data.

[0051] To address the issues of long request back-to-origin distances, high satellite link loads, insufficient cache coverage, and service interruptions or increased latency after the terminal moves beyond the pre-cached area when the first edge cache node cannot continue providing nearby services after the terminal moves over a large area, the cache data prefetching method of the present invention includes, in some preferred embodiments, distributing the constructed cache data to multiple candidate edge cache nodes within a preset range for caching before the second terminal makes a request. When the second terminal makes a request, the distance between the original location information and the current location information of the second terminal is calculated; if the distance is within a second preset migration range, it is determined that the second terminal has migrated within the preset range, and the target edge cache node is determined from the multiple candidate edge cache nodes based on the current location information of the second terminal, and the cache data corresponding to the actual request content is obtained; if the distance is outside the second preset migration range, it is determined that the second terminal has migrated outside the preset range, and the cache data that meets the preset caching conditions is sent to the target edge cache node determined based on the current location information through the cloud platform for caching, and the cache data corresponding to the actual request content is obtained. This embodiment employs a multi-point pre-caching and hierarchical scheduling strategy. Before the second terminal initiates a request, the constructed cached data is pre-distributed to multiple candidate edge cache nodes within a preset range to expand the cache coverage. When the second terminal initiates a request, the distance between its original location and its current location is calculated. If the distance is within a second preset migration range, it is determined that the terminal is migrating within the preset area. Based on the current location information, the target node is determined from the nearest of the pre-cached candidate edge cache nodes, and the corresponding cached data is directly obtained. If the distance exceeds the second preset migration range, it is determined that the terminal has moved out of the preset cache area. The cloud platform then sends the cached data that meets the conditions to the newly determined target edge cache node based on the current location in real time, and then provides content services to the terminal.

[0052] This embodiment expands the cache coverage by using multi-point pre-caching, improving the cache hit rate when the terminal moves within the area and avoiding frequent source calls. Through a hierarchical location judgment mechanism, the cache is hit as close as possible within the movement range, and the cache is pushed back in real time when it is outside the range, taking into account both mobility adaptability and service continuity. This reduces the pressure of repeated transmission on the satellite link and ensures that the terminal can still obtain low-latency and highly reliable edge caching services in large-scale movement scenarios, which is especially suitable for highly mobile satellite terminal scenarios.

[0053] Example 6

[0054] Based on any one of the embodiments from one to five, this embodiment provides a preferred method for replacing cached data in a target edge cache node.

[0055] Existing satellite edge cache nodes mostly use simple strategies such as first-in-first-out (FIFO) and least recently used (LRU) for data replacement, or judge based on the single dimension of the historical popularity of content. Although some solutions combine popularity, they do not associate them with terminal behavior prediction results, nor do they take into account the timeliness of data and the characteristics of satellite scenarios.

[0056] The replacement strategy is simplistic and does not incorporate terminal behavior prediction, which can easily lead to the accidental removal of high-value data and the retention of invalid content, resulting in a low cache hit rate. It is not adapted to the characteristics of satellite scenarios, ignores the dynamic time slots of the link and the timeliness of content, and makes critical data easy to be replaced, increasing the pressure of repeated transmission on the link. The replacement logic is fixed and cannot dynamically adapt to changes in terminal behavior and link status, making cache replacement uncontrollable and affecting service continuity.

[0057] To address the aforementioned technical problems, in some preferred embodiments, the method for replacing cached data according to the present invention includes: when the cache capacity of the edge cache node is lower than a third preset threshold, calculating the replacement priority based on the most recent access time and predicted popularity weight of each cached data, and replacing the cached data in the edge cache node according to the replacement priority. First, the remaining cache capacity of the edge cache node is monitored in real time. When the cache capacity is detected to be lower than the third preset threshold, the data replacement process is triggered. For each cached data in the cache node, two core indicators are collected synchronously: one is the most recent access time of the data, representing the actual usage frequency of the terminal; the other is the predicted popularity weight of each data based on the behavior prediction results of Embodiment 1, which reflects the probability and importance of the data being requested by the terminal in the future. Second, combining the most recent access time and the predicted popularity weight, the replacement priority of each cached data is calculated uniformly. In this embodiment, the lower the priority, the more likely it is to be replaced. Finally, according to the calculated replacement priority, the cached data in the edge cache node is replaced in an orderly manner, removing the data with the lowest priority, and simultaneously storing new target cached data into the node, completing the update of the cached data.

[0058] This embodiment prioritizes retaining high-value, high-demand data while replacing low-value, low-demand invalid data, significantly improving the cache hit rate of edge cache nodes and making full use of limited cache resources. It also considers the correlation between predicted popularity and timeliness to ensure that critical data such as emergency communication and core navigation are not easily replaced, thereby improving the stability of terminal services. At the same time, it dynamically triggers replacement based on capacity thresholds to avoid blind replacement, reduce repeated satellite link transmissions caused by invalid data replacement, further reduce satellite link resource consumption, and improve the efficiency and rationality of overall cache data prefetching.

[0059] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. An edge caching and prefetching method based on cross-terminal behavior prediction, characterized in that, Includes the following steps: S1. Collect historical behavior data of the first terminal, input it into the trained AI prediction model, and obtain the prediction request content ID of the second terminal within a preset time period; the AI ​​prediction model is trained based on the correlation features between the first terminal and the second terminal; S2. Determine the predicted request content and timeliness tag based on the predicted request content ID, and construct cache data; S3. Send the cached data that meets the preset caching conditions to the target edge cache node for caching; S4. When the second terminal initiates a request, calculate the distance between the original location information and the current location information of the second terminal; S5. Determine whether the distance falls within a preset migration range, identify the target edge cache node, and match the actual request content from the cached data of the target edge cache node.

2. The edge caching and prefetching method based on cross-terminal behavior prediction as described in claim 1, characterized in that, The training methods for the AI ​​prediction model include: The heterogeneous user behavior graphs of the first terminal and the second terminal are continuously updated. The graph attention network is used to learn the association features between the first terminal and the second terminal as a training set to train the AI ​​prediction model built by the Transformer encoder.

3. The edge caching and prefetching method based on cross-terminal behavior prediction as described in claim 1, characterized in that, The historical behavior data includes: request content ID, timestamp, geographical location, and first terminal state vector.

4. The edge caching and prefetching method based on cross-terminal behavior prediction as described in claim 1, characterized in that, The target edge cache node includes satellite-borne cache and ground edge station.

5. The edge caching and prefetching method based on cross-terminal behavior prediction as described in claim 1, characterized in that, The preset caching conditions include: The predicted value of the predicted request content ID is greater than the first preset threshold, and the confidence level is greater than the second preset threshold. The downlink bandwidth of the link from the satellite to the second terminal is in an idle time slot.

6. The edge caching and prefetching method based on cross-terminal behavior prediction as described in claim 5, characterized in that, The method for obtaining the idle time slots includes: obtaining idle time slot information for all satellite links through real-time telemetry data of the satellite network.

7. The edge caching and prefetching method based on cross-terminal behavior prediction as described in claim 1, characterized in that, Before the second terminal initiates a request, the process includes: sending the cached data to a first edge cache node determined based on the original location information via the cloud platform for caching.

8. The edge caching and prefetching method based on cross-terminal behavior prediction as described in claim 7, characterized in that, When the second terminal makes a request, it includes: Calculate the distance between the original location information and the current location information of the second terminal; if the distance is within a first preset migration range, determine that the second terminal has not migrated, and use the first edge cache node as the target edge cache node to obtain the cached data corresponding to the actual requested content.

9. The edge caching and prefetching method based on cross-terminal behavior prediction as described in claim 1, characterized in that, Before the second terminal makes a request, the process also includes: distributing the constructed cached data to multiple candidate edge cache nodes within a preset range for caching.

10. The edge caching and prefetching method based on cross-terminal behavior prediction as described in claim 9, characterized in that, When the second terminal makes a request, it also includes: Calculate the distance between the original location information and the current location information of the second terminal; if the distance is within a second preset migration range, determine that the second terminal has migrated within the preset range, determine the target edge cache node from multiple candidate edge cache nodes based on the current location information of the second terminal, and obtain the cache data corresponding to the actual requested content; If the distance is outside the second preset migration range, it is determined that the second terminal has migrated outside the preset range. The cached data that meets the preset caching conditions is sent to the target edge cache node determined according to the current location information through the cloud platform for caching, and the cached data corresponding to the actual request content is obtained.

Images