A two-layer collaborative caching method for remote sensing satellite networks

Abstract

This invention discloses a two-layer collaborative caching method for remote sensing satellite networks, applied to a two-layer network composed of high-orbit and low-orbit satellite nodes. High-orbit satellite nodes serve as the global cache and backhaul backbone, while low-orbit satellite nodes act as edge cache access points directly serving ground users. When a low-orbit satellite node receives a user content request and its own cache misses, it executes the following core steps: First, based on the temporal and spatial attributes of the cached content, it calculates the content duplication rate with neighboring nodes and selects the node with the lowest duplication rate as the target query node. Second, it sends a query request to the target node to obtain a copy of the content. Finally, it distributes the obtained copy to the user. This invention significantly improves cache hit rate and content distribution efficiency through an intelligent collaborative query mechanism, effectively reducing transmission latency, and is particularly suitable for remote sensing satellite network environments with wide coverage and dynamic topology.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing satellite network communication technology, specifically relating to cache architecture design and content cache optimization technology in remote sensing satellite networks, and in particular a two-layer collaborative caching method for remote sensing satellite networks. Background Technology

[0002] Remote sensing satellite networks are composite systems based on remote sensing technology and satellite communication. They are widely used and have significant academic and practical value in fields such as Earth observation, environmental monitoring, resource management, and scientific research. With the rapid development of remote sensing technology and satellite communication, the massive amounts of data generated by remote sensing satellites need to be transmitted efficiently to ground users. However, traditional data transmission methods face many challenges, including the diversity of data stream characteristics, limited transmission time windows, and constraints on network transmission capacity. Remote sensing data is ultimately transmitted to users through multiple relays via multiple satellites and ground stations, consuming limited bandwidth resources and hindering the implementation of latency-sensitive services.

[0003] Recently, caching technology has proven to be an effective solution to this problem. Information-Centric Networking (ICN) has been proposed as a promising new network architecture. Thanks to the continuous improvement of onboard hardware and the decreasing cost of onboard equipment, content caching and computation on satellites have become routine operations for satellite systems. ICN decouples content from address, deploying reusable content to satellite nodes closer to users, thereby shortening content transmission distances and reducing retransmissions of identical content. ICN employs a content naming mechanism to separate content from location, using content identifiers to uniquely distinguish data, effectively solving the problem of duplicate requests for the same content, reducing network redundancy, and helping to alleviate link load and improve overall transmission efficiency in remote sensing satellite networks.

[0004] However, existing remote sensing satellite network caching technologies have several shortcomings, including network architecture that is not fully adapted to the collaborative working characteristics of high-orbit and low-orbit satellites, difficulty in accurately capturing the time-varying topology characteristics of satellite networks, content naming that does not match the specific attributes of remote sensing data, and caching strategies that lack dynamic perception of user interests. These shortcomings result in low utilization of cache resources, high data transmission latency, and poor efficiency in remote sensing data distribution. Summary of the Invention

[0005] This invention aims to address the shortcomings of existing remote sensing satellite network caching technologies, such as network architecture not being fully adapted to the collaborative working characteristics of high-orbit and low-orbit satellites, difficulty in accurately capturing the time-varying topology characteristics of satellite networks, content naming not conforming to the specific attributes of remote sensing data, and caching strategies lacking dynamic perception capabilities of user interests. These shortcomings result in low utilization of cache resources, high data transmission latency, and poor remote sensing data distribution efficiency.

[0006] A two-layer collaborative caching method for remote sensing satellite networks is applied to a two-layer network composed of high-orbit satellite nodes and low-orbit satellite nodes. The high-orbit satellite nodes serve as global caches and backhaul backbones, while the low-orbit satellite nodes serve as edge cache access points and directly serve ground users. The method includes:

[0007] When a low-Earth orbit satellite node serving a ground user receives a content request and its own cache is not hit, it performs the following steps:

[0008] Step 1. Based on the cached content duplication rate, determine the target query node from the neighboring satellite nodes of the low-orbit satellite node; wherein, the cached content duplication rate is used to characterize the similarity of cached content between two satellite nodes, and is calculated by comparing the temporal and spatial attributes of the cached content;

[0009] Step 2. Send a query request to the target query node to obtain a copy of the content;

[0010] Step 3. Distribute the obtained copy of the content to the requesting user.

[0011] Further, in step 1, determining the target query node based on the cache content duplication rate specifically involves: calculating the cache content duplication rate between the low-orbit satellite node and each adjacent satellite node, and selecting the adjacent satellite node with the lowest cache content duplication rate as the target query node.

[0012] Furthermore, the duplicate rate of the cached content is calculated using the following formula: Duplicate rate = |Intersection of cached content| / |Union of cached content|, where the calculation of intersection and union is based on the temporal and spatial attributes of the cached content.

[0013] Furthermore, prior to step 1, the method further includes: constructing a time-slot model to characterize the time-varying topology of the two-layer network; the construction of the time-slot model includes: discretizing the network operation cycle into multiple snapshots; merging adjacent snapshots with similar topologies to form a time-slot sequence with a stable topology; wherein, the similarity between snapshots is comprehensively evaluated by three indicators: Dice coefficient, spectral similarity, and degree-degree correlation.

[0014] Furthermore, the content request and content copy are identified using a hierarchical naming structure; the hierarchical naming structure includes at least: a spatial location field, a timestamp field, and a data source identifier field.

[0015] Furthermore, the decision to replace the cached content of the satellite node is made through a reinforcement learning model; the state space of the reinforcement learning model is defined by the current cached content set and the interest request set, and its reward function is dynamically calculated according to the change in content popularity.

[0016] Furthermore, the reward function is calculated as the difference between the popularity of the new cached content and the popularity of the replaced content when a cache replacement occurs; when only forwarding uncached content, the reward is the popularity of the new cached content; and when a cache hit occurs, the reward is zero.

[0017] A two-layer collaborative caching system for remote sensing satellite networks includes multiple cache nodes deployed on high-orbit and low-orbit satellites, the cache nodes being configured to perform the two-layer collaborative caching method for remote sensing satellite networks as described above.

[0018] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the two-layer collaborative caching method for remote sensing satellite networks as described above.

[0019] The beneficial effects of this invention are: (1) Achieving global coverage without blind spots and low-latency service. By integrating high-orbit and low-orbit satellites, transmission latency is significantly reduced while ensuring wide-area coverage. High-orbit satellites provide efficient data backhaul for low-orbit satellites, and low-orbit satellites directly provide low-latency, high-bandwidth services to users, thereby effectively solving the communication coverage problem in remote and extreme environments.

[0020] (2) Significantly improve network resource utilization and transmission efficiency. By transforming dynamic topology into stable time slot sequences through a time slot model, and combining it with a reinforcement learning-based caching strategy, network state awareness and content pre-caching are achieved, which effectively improves the cache hit rate. At the same time, through inter-satellite cooperation and broadcast distribution mechanisms, duplicate transmissions can be reduced, significantly improving spectrum utilization.

[0021] (3) Supports fine-grained and efficient remote sensing data management. The proposed hierarchical content naming system, combined with spatiotemporal feature identifiers, can significantly improve the retrieval efficiency of remote sensing data, effectively reduce data query and scheduling latency, and enhance the overall system response capability.

[0022] (4) It has adaptive and intelligent operation and maintenance capabilities. The interest-aware caching scheme based on reinforcement learning can dynamically predict the popularity trend of content and adaptively adjust the caching strategy, so that the system can maintain high caching performance in dynamic network environment, reduce the cost of manual intervention, and have long-term evolution capabilities.

[0023] In summary, this invention, through a combination of architectural innovation, model optimization, and intelligent decision-making, brings significant improvements in coverage, transmission efficiency, data management, and system autonomy, and possesses important technical value and application prospects. Attached Figure Description

[0024] Figure 1 This is a flowchart of the present invention;

[0025] Figure 2 This is a two-layer remote sensing satellite network architecture;

[0026] Figure 3 Flowchart of an interest-aware caching scheme based on reinforcement learning. Detailed Implementation

[0027] The present invention will now be described in detail with reference to the accompanying drawings. Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The directional terms such as left, center, right, top, and bottom in the embodiments of the present invention are only relative concepts or referenced to the normal use state of the product, and should not be considered restrictive.

[0028] This invention provides a two-layer collaborative caching method for remote sensing satellite networks, aiming to solve the problems of low caching efficiency and prolonged content distribution in existing satellite network technologies. The invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of this invention, providing detailed implementation methods and specific operation processes; however, the scope of protection of this invention is not limited to the following embodiments.

[0029] I. System Architecture and Scenario Overview

[0030] As attached Figure 2 As shown, this invention constructs a two-layer network caching architecture that integrates high-orbit satellites (GEO) and low-orbit satellites (LEO) to achieve global coverage without blind spots and low-latency service;

[0031] High-orbit satellite nodes: As upper-level nodes and global cache backbone of the network, they have wide-area continuous coverage capabilities. They are mainly responsible for obtaining remote sensing data from data sources (such as ground stations and other satellites) and providing content update and backhaul services for low-orbit satellites in large areas.

[0032] Low Earth Orbit (LEO) satellite nodes: As lower-layer network nodes and edge cache access points, they directly serve ground users within their coverage area; a single LEO satellite can provide large-area coverage within a diameter of hundreds of kilometers and support high-bandwidth, low-latency broadband Internet access; its core functions include: receiving and processing user requests, performing local cache queries, coordinating queries with neighboring satellites (including other LEO and GEO satellites), and distributing content to users;

[0033] Ground users: Located within the coverage area of ​​low-Earth orbit satellites, they have the ability to communicate directly with low-Earth orbit satellite nodes and initiate content requests to the low-Earth orbit satellites that serve them;

[0034] The application scenarios of this invention are mainly aimed at areas where ground base stations have not achieved full coverage, such as remote mountainous areas, oceans, deserts and disaster areas; all satellite nodes (GEO and LEO) have limited cache space and can use broadcasting to efficiently utilize spectrum resources to distribute content;

[0035] II. Detailed Steps of the Method and Flow

[0036] The main entity executing the method of this invention is the low-Earth orbit satellite node that receives the user request (hereinafter referred to as the "current node"); such as Figure 1 As shown, when the node's own cache is not hit, the following collaborative caching and query process will be triggered:

[0037] Step 0: Construct a time-slot model to characterize the time-varying topology of the network

[0038] Because the network topology of a satellite constellation changes drastically during its operational lifecycle due to the dynamic movement of satellite nodes and the frequent connection and disconnection of inter-satellite links, a time-slot model needs to be constructed to provide a stable network state foundation for subsequent caching decisions and routing queries. This model specifically includes:

[0039] Snapshot discretization: Discretize the operating cycle of the satellite network (such as one orbital cycle) into a series of very short static network snapshots. Each snapshot records the position of the satellite node and the inter-satellite link connection status at that moment.

[0040] Similarity assessment and snapshot merging: Directly using all snapshots can lead to data redundancy; therefore, this invention proposes to extract the essential topological pattern by merging adjacent snapshots with similar topology.

[0041] The following three indicators are used to comprehensively evaluate adjacent snapshots:

[0042] Dice coefficient: Measures the similarity between two layers based on the ratio of overlapping area to average size. The formula is:

[0043] (1)

[0044] in, Presentation layer The total number of links in the network, Presentation layer The total number of links in the network, Presentation layer With layers The total number of shared links, multiplied by 2 to emphasize the importance of the overlapping portion; Dice coefficient. The value range is [0,1];

[0045] Spectral similarity: This measure of similarity between two layers is achieved by analyzing the eigenvalues ​​of the network's matrix representation to capture its macroscopic topological structure and dynamic behavioral characteristics. The formula is as follows:

[0046] (2)

[0047] (3)

[0048] in, and A β Representing layers and layers The adjacency matrix corresponding to the network, and Representing layers and layers The degree diagonal matrix corresponding to the network, and L (β) Representing layers and layers The Laplacian matrix corresponding to the network; , , …, Presentation layer The feature value sequence corresponding to the network; , , …, Presentation layer The feature value sequence corresponding to the network;

[0049] Degree-degree correlation: This method determines the similarity between two layers by capturing the similarity of the degrees of network nodes in different layers. The formula is as follows:

[0050] (4)

[0051] in, Indicates in snapshot and The existence values ​​are respectively and The probability of a random node. A positive value indicates a positive correlation between the two layers; conversely, a negative value indicates a negative correlation between the two layers.

[0052] After calculating the above indicators, the comprehensive inter-layer correlation is obtained by weighted summation, as shown in the formula:

[0053] (5)

[0054] in, , and These are the weighting functions corresponding to the three indicators, with values ​​ranging from [0,1]. ;

[0055] Determine the merging threshold and the time slot: Set a similarity threshold, which can be preset to a fixed value based on experience (e.g., 0.8), or dynamically determined by analyzing the statistical characteristics of the similarity distribution of historical snapshots (e.g., mean, inflection point); compare the calculated inter-layer correlation with the threshold, and merge adjacent snapshots with a correlation higher than the threshold;

[0056] Forming a time slot sequence: After merging, a smaller set of time slot sequences is obtained, in which the topology remains stable in each time period. Each time slot represents a relatively unchanging network topology state over a period of time. Subsequent caching strategies and routing decisions can be based on the current time slot.

[0057] Step 1: Determine the target query node based on the duplicate rate of cached content

[0058] After the current node fails to find the local cache, it needs to query other satellite nodes (adjacent LEOs or GEOs). To reduce duplicate queries and improve the hit probability, this invention does not select the next hop randomly or simply by the number of hops, but intelligently selects the next hop based on the duplicate rate of cached content.

[0059] Calculate the cache content duplication rate: For each adjacent satellite node of the current node, calculate the cache content duplication rate between them; this duplication rate represents the similarity between the cached contents of the two nodes. The lower the duplication rate, the higher the probability that the other party has the content that the current node lacks.

[0060] Calculation basis: Classification and matching are performed by comparing the time attributes (such as data collection time) and spatial attributes (such as the geographical range covered by the data) of the cached content;

[0061] Calculation formula: Duplication rate = |Intersection of cached content| / |Union of cached content|, where the intersection refers to the number of content items that are jointly cached by the two nodes and match in terms of spatiotemporal attributes; the union refers to the sum of the number of all cached content items of the two nodes.

[0062] Right now ,in, For repetition rate, Indicates satellite node A collection of cached content with spatiotemporal attributes. Adjacent satellite nodes A collection of cached content with spatiotemporal attributes;

[0063] Selecting the target node: After calculating the duplicate rate of cached content between the current node and all its neighboring nodes, the neighboring satellite node with the lowest duplicate rate is selected as the target query node. This selection strategy aims to maximize the "exploratory" nature of the query by prioritizing requests to the node with the greatest difference in cached content, thereby increasing the probability of finding the target content in a single query.

[0064] Step 2: Send a request to the target query node and retrieve a copy of the content.

[0065] The current node sends the user's content request (encapsulated as an interest packet) to the target query node determined in step 1. After receiving the request, the target query node queries its local cache. If a match is found, it prepares to return the data. If a match is still not found, the target node can continue to forward the request to its neighboring nodes or upper-layer GEO nodes according to its own strategy (e.g., based on content duplication rate or by querying the FIB table) until a content copy is found in the satellite network or back to the data source. Finally, the content copy will be transmitted back to the current node along the request path.

[0066] Step 3: Distribute copies of the content to requesting users

[0067] After the current node obtains a copy of the content, it leverages the advantages of satellite broadcast communication to simultaneously distribute the content to all ground users who requested it, thus efficiently completing the content distribution.

[0068] III. Specific Implementation of Key Supporting Technologies

[0069] (a) Content naming method that conforms to the characteristics of remote sensing data

[0070] This invention is designed based on an Information-Centric Networking (ICN) architecture. ICN is a content-centric network paradigm that decouples content from location, routing and retrieving content by name instead of traditional host addresses. In the remote sensing satellite network scenario, ground users request content by sending Interest Packets. Upon receiving an Interest Packet, a satellite node first searches its Content Store (CS) for matching content. If a match is found, a Data Packet is returned directly. If no match is found, the entry point of the request is recorded in the Pending Interest Table (PIT), and the Interest Packet is forwarded to other nodes for further searching based on the Forwarding Information Base (FIB) until the desired content is obtained. To achieve efficient retrieval and routing under this ICN architecture, this invention combines the unique spatiotemporal attributes of remote sensing data with user task attributes to describe the content, thereby designing a dedicated hierarchical content naming structure.

[0071] Interest package naming format:

[0072] (6)

[0073] in, This is a name prefix in interest packets, which can be used for routing and forwarding in ICNs; The longitude and latitude coordinates of the four corners of the remote sensing image are identified to reflect the spatial characteristics of the remote sensing data, i.e., the spatial location field. The timestamp field indicates the time when the remote sensing image was acquired. This requires identifying information such as satellite number, product number, and band number, i.e., the data source identification field; The system allocates fees to users and defines user priorities based on these fees. After receiving an interest packet, the network node can parse out its spatiotemporal attributes and then search for relevant entries in CS, PIT, and FIB. Once it finds the source data that meets the user's needs, it returns the data packet to the user.

[0074] Data packet naming format:

[0075] (7)

[0076] in, This is a name prefix for packets without routing control characters, only identifying the satellite node to which the content belongs. This naming structure supports name aggregation and its high scalability helps meet users' diverse needs for remote sensing content. Searching for specific content through a hierarchical naming structure has high retrieval efficiency and good compatibility with existing terrestrial network architecture.

[0077] This naming system decouples content from location and enables efficient aggregation and retrieval through spatiotemporal attributes, significantly improving the query efficiency of remote sensing data in satellite networks;

[0078] (II) Interest-aware caching scheme based on reinforcement learning

[0079] As attached Figure 3 As shown, this invention designs a caching decision scheme based on reinforcement learning; it aims to optimize the caching decision strategy for popular content in a two-layer remote sensing satellite network; the scheme utilizes content request data in the early period and the real-time storage status of cache nodes, and adopts a model-free reinforcement learning algorithm to dynamically adjust caching decisions in order to maximize content popularity matching and dynamically optimize the cached content of each satellite node.

[0080] 1. Problem Modeling:

[0081] To address the challenges of cache optimization in dynamic environments of remote sensing satellite networks, this invention models the cache decision problem of a single satellite node as a Markov Decision Process (MDP) and employs a model-free reinforcement learning method to achieve adaptive policy optimization. Within this framework, in each decision slot, when a cache node receives an interest packet request, its built-in agent observes the current network state. And select a caching action according to the established strategy. The action is executed; the environment then provides an immediate reward to assess its effectiveness, and the network enters a new state. The agent learns to maximize long-term cumulative rewards by repeatedly performing this "state-action-reward-new state" interaction process and continuously updating its decision-making model using reward signals. This achieves optimal long-term cache hit rate. The core innovation of this scheme lies in the targeted design of the state space, action space, and reward function to precisely adapt to the characteristics of remote sensing data and the dynamic nature of satellite networks. The specific definitions are as follows:

[0082] state space Defined as ,in, Represents an ordered set of cached content. This represents the time slot. The set of request interest packets arriving within the cache; the rationale for this state representation method is that the joint state space of cached content distribution and request patterns directly determines the cache efficiency value obtained by the system; based on the descending order of content popularity, the cache set can be represented as ;also, This indicates the state of the subsequent cache node after the caching strategy is executed. These are the interest requests to be processed in the next time slot;

[0083] Action Space (A): When an interest packet arrives at a cache node, the node first parses its header information and queries its local content store (CS) to determine if the requested content exists. If the content is already cached, the node will directly return the corresponding data packet along the transmission path. Otherwise, the system will update the request frequency statistics for that interest packet and predict the probability of future requests based on this, thereby dynamically adjusting and processing the cache. If the node adopts a greedy strategy, that is, unconditionally caching all newly arrived content and evicting the least popular items, it may lead to over-replacement problems. Considering that the popularity of content is time-varying and random, caching decisions need to explore the possibility of all caching actions to accurately track the dynamic shift of popularity. Considering that the cache capacity in the router node is limited, let's assume... For each entry, the action space of a node can be defined as... ;in, to These represent replacing the first item in the cache. The most popular content (sorted in descending order of popularity) is used to store the most recently requested content; This indicates that content is not cached, and interest packets are forwarded directly to the next hop based on the Forwarding Information Table (FIB). This indicates that the requested content is already cached and no modification to the cache state is needed; specifically, if the requested content is not cached, the node needs to... to Choose an action from the options; otherwise, execute the action. ; When an agent performs an action Afterwards, the system records the cache reward and reflects the impact of the caching strategy on the system by updating the network environment status;

[0084] reward function Its core innovation lies in its design, which closely follows the real-time changes in content popularity.

[0085] (8)

[0086] in, This indicates the popularity of the content of the most recently responded request. This represents the popularity of the replaced cached content; when cache replacement occurs, the system calculates... The difference is used to reflect the dynamic changes in content popularity; generally, a positive difference indicates that the new cached content is more popular; conversely, a negative difference indicates that the new cached content is less popular than the original cached content; when an action is performed... When the system directly uses the popularity of the requested file as the reward, when an action is executed... At that time, since the cache state has not changed and the request is satisfied locally, the reward value is set to zero;

[0087] The reward function design of this invention differs significantly from existing solutions: we innovatively link the reward directly to the dynamic popularity difference, and achieve a more refined evaluation of the effectiveness of actions by comparing the popularity difference between the new cached content and the replaced content;

[0088] State transition: The transition probability is related to the popularity distribution of the content requested in the next time slot. Specific probability models are established for the three modes: replacement, forwarding, and hit. Specifically:

[0089] Define from the current state Transition to the next state The probability is The transition probability is entirely determined by the content requested in the next time slot. Popularity distribution The system decides to implement different state transition modes based on cache hits;

[0090] Firstly, cache miss and content replacement mode: when requesting content... When, perform the replacement action. The state transition probability under this mode is shown in formula (9):

[0091] (9)

[0092] Secondly, cache miss and request forwarding mode: when the request content At that time, the forwarding action is executed. The data packets received by the node are not cached, and the state transition probability in this mode is shown in formula (10).

[0093] (10)

[0094] Third, cache hit mode: when the requested content At that time, perform the retrieval action in the node's local cache. The state transition probability under this mode is shown in formula (11);

[0095] (11)

[0096] Compared with existing solutions, the core difference of this invention lies in binding the state transition probability completely with the dynamically changing content popularity distribution, and establishing accurate state transition models based on three specific operation scenarios such as replacement, forwarding, and hit. This not only reflects the spatiotemporal dynamics of network requests, but also enhances the adaptability and decision accuracy of the learning algorithm in the actual satellite caching environment through fine-grained probabilistic modeling.

[0097] 2. Decision-making and learning process:

[0098] To quantify the benefits of an agent's caching decisions under specific network states, we define a behavior value function; let... Indicates the moment of decision-making ,state Select action The strategy, then in the strategy The state-action value function under guidance can be expressed as formula (12);

[0099] (12)

[0100] Among them, discount factor It is used to adjust the relative importance of immediate rewards and long-term cumulative rewards, and to weigh short-term benefits against long-term gains; Indicates the state Next action The instant reward received; And respectively represent in the state space The current state and the next state in the context; And respectively represent in the action space The current action and the next action in the context; Indicates the state Execute action Transition to state The state transition probability; Representation Strategy Next, in the next state Use action State-action value function;

[0101] The caching decision process is divided into Value calculation and content storage; in the first stage, when an interest packet arrives at a node, the system calculates its... Values ​​are used to predict the probability of future requests; higher values... The value indicates that the content is more likely to be requested frequently and should be cached first; if cached... An increase in the value indicates rising content popularity; conversely, a decrease indicates that the cache value is lower than expected. The second phase involves storage decisions; when cache space is insufficient, content is prioritized for eviction. Content with lower values;

[0102] To address the uncertainty of transition probabilities caused by the popularity of dynamic content, this invention employs a model-free temporal difference learning algorithm; this algorithm is applied to state-action pairs. After the action is performed, combine the immediate reward and the next state. of The value is updated as shown in formula (13); this algorithm only depends on the previous value. Value, reward, and subsequent state-action pairs require no dynamic environment model and are suitable for dynamic network scenarios. After the value is updated, the agent transitions to the new state and executes the selected action, achieving adaptive caching optimization;

[0103] (13)

[0104] in, Indicates a time step; The learning rate is used to control the impact of newly acquired information on learning rates. The degree of impact of value updates; Discount factor; The first The state and actions performed in a step; These are the next state after the update and the next action selected. The immediate reward obtained after performing the current action; State-action value function;

[0105] Through continuous interaction and learning with the environment, the intelligent agent can eventually learn to predict content popularity trends, dynamically adjust caching strategies, and maximize long-term cache hit rate.

[0106] To optimize the agent's action selection mechanism in the state space, this invention employs the Greedy-Softmax strategy; this strategy balances exploration and exploitation based on probability and the known average reward of actions; each time an action is selected, the Greedy-Softmax strategy is used to optimize the action selection mechanism in the state space. The probability is determined using a Softmax action selection strategy. The probability adopts the Greedy action selection strategy; this strategy combines the Softmax strategy with the Greedy strategy, which balances the problem of incomplete exploration and slow algorithm convergence speed, and also takes into account the structure of Markov decision process in reinforcement learning task; the Greedy-Softmax action selection strategy can be described by formula (14).

[0107] (14)

[0108] in, Indicates a time step; This is the current state; For action space Optional actions; For the first Step strategy Next-state-action value function; A uniform random number is generated for each time slot. The value is a fixed value; the Softmax policy associates the action selection probability with its corresponding expected reward. The action with the larger the average reward has a higher probability of being selected, and the action probability distribution is based on the Boltzmann distribution, as shown in formula (15).

[0109] (15)

[0110] Among them, in state Let the size of the set of selectable actions be... The intelligent agent retrieves the corresponding action by looking up a table. value; It is an adjustable parameter used to adjust the degree of exploration; when When taking a smaller value, the strategy degenerates into a pure greedy strategy, that is, only selecting the optimal action; when When the value approaches 0, the Softmax strategy is equivalent to only utilizing the proposed solution; when... As the value approaches infinity, the Softmax strategy is equivalent to only exploring solutions.

[0111] This method can not only accurately reflect the popularity gain or loss brought about by each cache replacement, but also assign differentiated reward values ​​when performing different actions such as replacement, forwarding, and hit. This provides more accurate and directional learning feedback in the dynamic environment of satellite networks, effectively driving the agent to optimize towards improving the freshness of cached content and user satisfaction.

[0112] IV. Brief Description of Package Processing Flow

[0113] The proposed caching scheme achieves intelligent caching decisions by dynamically maintaining the ACTION table. The ACTION table adopts a five-tuple structure, including interest package name, cache tag, counter Count, Q value, and action tag. Among them, the counter Count is used to count the frequency of interest package requests, and the Q value is dynamically updated through reinforcement learning algorithm.

[0114] Interest packet processing flow: When a cache node in the cache network receives an interest packet, it first queries the Content Storage (CS). If the cache is hit, the content data packet is returned directly, and the hit count and the Count value in the ACTION table are updated. Otherwise, it queries the PIT table of pending interests. If a matching entry exists, new interface information is added; otherwise, it queries the Forwarding Information Base (FIB). Finally, it queries the FIB. When the FIB hits, the request frequency of the interest packet is embedded in the packet header and forwarded to the upstream interface; otherwise, the interest packet is discarded and the process ends.

[0115] Packet processing flow: When a data packet is received by the cache node from the content source, it first queries the PIT table. If no match is found, the data packet is discarded. If the PIT table is verified, the ACTION table is queried. If an entry is found in the ACTION table, the CS state is updated. Then, the data packet is forwarded to the requesting node along the PIT record path, and the process ends.

[0116] 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 the present invention is defined by the appended claims and their equivalents.

Claims

1. A two-layer collaborative caching method for remote sensing satellite networks, characterized in that, The method is applied to a two-layer network consisting of high-orbit satellite nodes and low-orbit satellite nodes, wherein the high-orbit satellite nodes serve as a global cache and backhaul backbone, and the low-orbit satellite nodes serve as edge cache access points and directly serve ground users; the method includes: When a low-Earth orbit satellite node serving a ground user receives a content request and its own cache is not hit, it performs the following steps: Step 1. Based on the cached content duplication rate, determine the target query node from the neighboring satellite nodes of the low-orbit satellite node; wherein, the cached content duplication rate is used to characterize the similarity of cached content between two satellite nodes, and is calculated by comparing the temporal and spatial attributes of the cached content; Step 2. Send a query request to the target query node to obtain a copy of the content; Step 3. Distribute the obtained copy of the content to the requesting user.

2. The two-layer collaborative caching method for remote sensing satellite networks according to claim 1, characterized in that, In step 1, determining the target query node based on the cache content duplication rate specifically involves: calculating the cache content duplication rate between the low-orbit satellite node and each adjacent satellite node, and selecting the adjacent satellite node with the lowest cache content duplication rate as the target query node.

3. The two-layer collaborative caching method for remote sensing satellite networks according to claim 2, characterized in that, The duplicate rate of cached content is calculated using the following formula: Duplicate rate = |Intersection of cached content| / |Union of cached content|, where the calculation of intersection and union is based on the temporal and spatial attributes of the cached content.

4. The two-layer collaborative caching method for remote sensing satellite networks according to claim 1, characterized in that, Before step 1, the method further includes: constructing a time-slot model to characterize the time-varying topology of the two-layer network; the construction of the time-slot model includes: discretizing the network operation cycle into multiple snapshots; merging adjacent snapshots with similar topologies to form a time-slot sequence with a stable topology; wherein, the similarity between snapshots is comprehensively evaluated by three indicators: Dice coefficient, spectral similarity, and degree-degree correlation.

5. The two-layer collaborative caching method for remote sensing satellite networks according to claim 1, characterized in that, The content requests and content copies are identified using a hierarchical naming structure; The hierarchical naming structure includes at least: a spatial location field, a timestamp field, and a data source identifier field.

6. The two-layer collaborative caching method for remote sensing satellite networks according to claim 1, characterized in that, The decision to replace the cached content of the satellite node is made through a reinforcement learning model; the state space of the reinforcement learning model is defined by the current cached content set and the interest request set, and its reward function is dynamically calculated according to the change of content popularity.

7. The two-layer collaborative caching method for remote sensing satellite networks according to claim 6, characterized in that, When a cache replacement occurs, the reward function is calculated as the difference between the popularity of the new cached content and the popularity of the replaced content; when only forwarding uncached content, the reward is the popularity of the new cached content; and when a cache hit occurs, the reward is zero.

8. A two-layer collaborative caching system for remote sensing satellite networks, characterized in that, It includes multiple cache nodes deployed on high-orbit and low-orbit satellites, the cache nodes being configured to perform a two-layer collaborative caching method for remote sensing satellite networks as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the two-layer collaborative caching method for remote sensing satellite networks as described in any one of claims 1 to 7.

Images